Sie sind auf Seite 1von 287

Weak KAM Theorem in

Lagrangian Dynamics
Preliminary Version
Number 10
Albert FATHI
Lyon, Version 15 June 2008
ii
Contents
Preface vii
Introduction ix
0.1 The Hamilton-Jacobi Method . . . . . . . . . . . . xi
1 Convex Functions: Legendre and Fenchel 1
1.1 Convex Functions: General Facts . . . . . . . . . . 1
1.2 Linear Supporting Form and Derivative . . . . . . 7
1.3 The Fenchel Transform . . . . . . . . . . . . . . . . 11
1.4 Dierentiable Convex Functions and Legendre Trans-
form . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Quasi-convex functions . . . . . . . . . . . . . . . . 28
1.6 Exposed Points of a Convex Set . . . . . . . . . . . 30
2 Calculus of Variations 37
2.1 Lagrangian, Action, Minimizers, and Extremal Curves 37
2.2 Lagrangians on Open Subsets of R
n
. . . . . . . . 40
2.3 Lagrangians on Manifolds . . . . . . . . . . . . . . 48
2.4 The Euler-Lagrange Equation and its Flow . . . . 51
2.5 Symplectic Aspects . . . . . . . . . . . . . . . . . . 54
2.6 Lagrangian and Hamiltonians . . . . . . . . . . . . 58
2.7 Existence of Local Extremal Curves . . . . . . . . 63
2.8 The Hamilton-Jacobi method . . . . . . . . . . . . 71
3 Calculus of Variations for a Lagrangian Convex in
the Fibers: Tonellis Theory 81
3.1 Absolutely Continuous Curves. . . . . . . . . . . . 81
3.2 Lagrangian Convex in the Fibers . . . . . . . . . . 89
3.3 Tonellis Theorem . . . . . . . . . . . . . . . . . . 95
iii
iv
3.4 Tonelli Lagrangians . . . . . . . . . . . . . . . . . . 98
3.5 Hamilton-Jacobi and Minimizers . . . . . . . . . . 101
3.6 Small Extremal Curves Are Minimizers . . . . . . 103
3.7 Regularity of Minimizers . . . . . . . . . . . . . . . 106
4 The Weak KAM Theorem 109
4.1 The Hamilton-Jacobi Equation Revisited . . . . . . 109
4.2 Dominated Functions and the Ma ne Critical Value 118
4.3 Defect and Calibration of Curves . . . . . . . . . . 126
4.4 Minimal Action for a Given Time . . . . . . . . . . 135
4.5 The Lax-Oleinik Semi-group. . . . . . . . . . . . . 144
4.6 The Lax-Oleinik semi-group . . . . . . . . . . . . . 144
4.7 Existence of Negative Weak KAM Solutions . . . . 153
4.8 Invariant Measures and Ma nes Critical Value . . 158
4.9 The Symmetrical Lagrangian . . . . . . . . . . . . 161
4.10 The Mather Function on Cohomology. . . . . . . . 163
4.11 Dierentiability of Dominated Functions . . . . . . 167
4.12 Mathers Set. . . . . . . . . . . . . . . . . . . . . . 175
4.13 Complements . . . . . . . . . . . . . . . . . . . . . 178
4.14 Examples . . . . . . . . . . . . . . . . . . . . . . . 181
5 Conjugate Weak KAM Solutions 183
5.1 Conjugate Weak KAM Solutions . . . . . . . . . . 183
5.2 Aubry Set and Ma ne Set. . . . . . . . . . . . . . . 186
5.3 The Peierls barrier. . . . . . . . . . . . . . . . . . . 190
5.4 Chain Transitivity . . . . . . . . . . . . . . . . . . 197
6 A Closer Look at the Lax-Oleinik semi-group 201
6.1 Semi-convex Functions . . . . . . . . . . . . . . . . 201
6.1.1 The Case of Open subsets of R
n
. . . . . . 201
6.2 The Lax-Oleinik Semi-group and Semi-convex Func-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.3 Convergence of the Lax-Oleinik Semi-group . . . . 207
6.4 Invariant Lagrangian Graphs . . . . . . . . . . . . 210
7 Viscosity Solutions 213
7.1 The dierent forms of Hamilton-Jacobi Equation . 213
7.2 Viscosity Solutions . . . . . . . . . . . . . . . . . . 214
7.3 Lower and upper dierentials . . . . . . . . . . . . 223
v
7.4 Criteria for viscosity solutions . . . . . . . . . . . . 230
7.5 Coercive Hamiltonians . . . . . . . . . . . . . . . . 232
7.6 Viscosity and weak KAM . . . . . . . . . . . . . . 233
8 More on Viscosity Solutions 235
8.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . 235
8.2 Construction of viscosity solutions . . . . . . . . . 236
8.3 Quasi-convexity and viscosity subsolutions . . . . . 242
8.4 The viscosity semi-distance . . . . . . . . . . . . . 249
8.5 The projected Aubry set . . . . . . . . . . . . . . . 252
8.6 The representation formula . . . . . . . . . . . . . 255
9 Ma nes Point of View 259
9.1 Ma nes potential . . . . . . . . . . . . . . . . . . . 259
9.2 Semi-static and static curves . . . . . . . . . . . . 262
vi
Preface
The project of this book started from my work published in the
Comptes Rendus de lAcademie des Sciences, see [Fat97b, Fat97a,
Fat98a, Fat98b].
I gave several courses and lectures on the material presented
there.
The project went through several versions. The rst version
was in French. It was produced for the Graduate course Syst`emes
lagrangiens et theorie dAubry-Mather, that I gave at the Ecole
Normale Superieure de Lyon during Spring Semester 1998. The
French set of notes has circulated widely. Daniel Massart and
Ezequiel Maderna caught up a large amount of mistakes in the
French version. The rst set of notes in english were a translated
and improved version of lectures notes in French, and consited of
versions of chapter 1 to 5. It was done while I was on sabbatical
during Spring Semester 2000 at the University of Geneva. I wish to
thank the Swiss National Scientic foundation for support during
that time. This rst version was distributed and used at the Ecole
dete en geometrie held at Universite de Savoie June, 15-22,
2000. A certain number of typing mistakes were found by people
attending the Ecole dete en geometrie
After adding chapter 6, we incorporated some of the improve-
ments suggested by Alain Chenciner and Richard Montgomery.
The subsequent versions, besides improvements, contained a
couple of chapters on viscosity solutions of the Hamilton-Jacobi
equation, especially the connection with the weak KAM theorem,
and a last brief one making the connection with Ma nes point of
view. The opportunity to teach a course of DEA in Lyon in 2001-
2002 and 2002-2003 was instrumental in the expansions in this set
of notes.
vii
viii
The seventh version was done in Pisa. I had the privilige of
giving a seires of Lectures in Winter 2005 in the Centro di Giorgi
at the Scuola Normale Superiore in Pisa. This seventh version is
a major revision of the sixth.
In this version 8, we have incorporated several typing mistakes
picked up by Maxime Zavidovique.
The present tenth version is prepared for a course given at the
Summer School Dynamical Systems: Theoretical and Applied
Hamiltonian Dynamics held at t Instituto Superior Tecnico in
Lisbon 16-20 June 2008. It has gone through a major revision of
chapter 4. I have incorporated a new proof found in June 2008
of the Weak KAM Theorem that is more elementary than the
previous ones in that it only uses the order properties for the Lax-
Oleinik semi-group and some compacrness arguments. It avoids
any appeal to a xed point theorem. We hope that the simple
arguments may be used in other contexts. We kept as a second
proof the one using a xed point theorem, since we consider it as
much more natural and almost forced on us by the compactness
obtained from Flemings Lemma.
A lot of people have helped me for a better understanding
of the subject, it is impossible to mention them all, among the
ones that I can remember vividly in chronological order: John
Mather, Michel Herman, Nicole Desolneux, Daniel Massart, Denis
Serre (without whom, I would have never realized that there was
a deep connection with viscosity solutions), Jean-Christophe Yoc-
coz, Francis Clarke, Gabriel & Miguel Paternain, Gonzalo Con-
treras, Renato Itturiaga, Guy Barles, Jean-Michel Roquejore,
Ezequiel Maderna, Patrick Bernard, Italo Capuzzo-Dolcetta, Pier-
marco Cannarsa, Craig Evans. Special thanks to Alain Chenciner
for his drive to understand and improve this subject. Last but not
least Antonio Siconol, we have been enjoying now a long a solid
collaboration, a large number of the improvements in these set of
notes is due to the numerous conversation that we have specialy
on the viscosity theory aspects.
Starting with the French notes, Claire Desecures helped a lot
in the typing.
Lyon, 14 June 2008
Introduction
The object of this course is the study of the Dynamical System
dened by a convex Lagrangian. Let M be a compact C

manifold
without boundary. We denote by TM the tangent bundle and by
: TM M the canonical projection. A point of TM will be
denoted by (x, v) with x M and v T
x
M =
1
(x). In the
same way, a point of the cotangent bundle T

M will be denoted
by (x, p) with x M and p T

x
M a linear form on the vector
space T
x
M.
We consider a function L : TM R of class at least C
3
.
We will call L the Lagrangian. As a typical case of L, we can
think of L(x, v) =
1
2
g
x
(v, v) where g is a Riemannian metric on
M. There is also the case of more general mechanical systems
L(x, v) =
1
2
g
x
(v, v) V (x), with g a Riemannian metric on M and
V : M R a function.
The action functional L is dened on the set of continuous
piecewise C
1
curves : [a, b] M, a b by
L() =
_
b
a
L((s), (s))ds
We look for C
1
(or even continuous piecewise C
1
) curves :
[a, b] M which minimize the action L() =
_
b
a
L((s), (s))ds
among the C
1
curves (or continuous piecewise C
1
) : [a, b] M
with the ends (a) and (b) xed. We will also look for curves
which minimize the action among the curves homotopic to with
xed endpoints or even for curves which achieve a local minimum
of the action among all curves homotopic with same endpoints.
The problem is tackled using dierential calculus on a func-
tional space. We rst look for the critical points of the action
ix
x
L() on the space of curves
(
1
x,y
([a, b], M) = : [a, b] M [ of class C
1
and (a) = x, (b) = y.
Such a curve which is a critical point is called an extremal curve
for the Lagrangian L. If an extremal curve is C
2
, it is possible to
show that the curve satises the Euler-Lagrange equation which,
in a system of coordinates, is written as
L
x
((t), (t))
d
dt
(
L
v
((t), (t)) = 0.
If the second partial vertical derivative

2
L
v
2
(x, v) is non-degenerate
at each point of TM we then see that we can solve for (t). It
results that there is a vector eld
(x, v) X
L
(x, v)
on TM such that the speed curves t ((t), (t)) of extremal
curves for the Lagrangian are precisely the solutions of this vec-
tor eld X
L
. The (local) ow
s
: TM TM of this vector eld
X
L
is called the Euler-Lagrange ow of the Lagrangian L. By
denition, a curve : [a, b] M is an extremal curve if and only
if ((s), (s)) =
sa
((a), (a)), for all s [a, b].
As TM is not compact, it may happen that
s
is not dened
for all s R, which would prevent us from making dynamics. It
will be supposed that L veries the two following conditions
(1) with x xed v L(x, v) is C
2
-strictly convex, i.e. the sec-
ond partial vertical derivative

2
L
v
2
(x, v) is dened strictly positive,
as a quadratic form;
(2) L(x, v) is superlinear in v, i.e.
lim
v
L(x, v)
|v|
+,
where || is a norm coming from a Riemannian metric on M.
Since all the Riemannian metrics are equivalent on a compact
manifold, this condition (2) does not depend on the choice of the
Riemannian metric.
Condition (2) implies that the continuous function L : TM
R is proper, i.e. inverse images under L of compact sets are com-
pact.
xi
Conditions (1) and (2) are of course satised for the examples
given above.
The function H(x, v) =
L
v
(x, v)vL(x, v) is called the Hamil-
tonian of the system. It is invariant by
s
. Under the assumptions
(1) and (2), this function H : TM R is also proper (in fact su-
perlinear). The levels H
1
(c), c R are thus compact subsets of
TM. As each trajectory of
s
remains in such compact set, we con-
clude from it that
s
is dened for all s R, as soon as L satises
conditions (1) and (2). We can, then, study the Euler-Lagrange
ow using the theory of Dynamical Systems.
0.1 The Hamilton-Jacobi Method
A natural problem in dynamics is the search for subsets invari-
ant by the ow
s
. Within the framework which concerns us the
Hamilton-Jacobi method makes it possible to nd such invariant
subsets.
To explain this method, it is better to think of the Hamiltonian
H as a function on cotangent bundle T

M. Indeed, under the


assumptions (1) and (2) above, we see that the Legendre transform
L : TM T

M, dened by
L(x, v) = (x,
L
v
(x, v)),
is a dieomorphism of TM onto T

M. We can then regard H as


a function on T

M dened by
H(x, p) = p(v) L(x, v), where p =
L
v
(x, v).
As the Legendre transform L is a dieomorphism, we can use it
to transport the ow
t
: TM TM to a ow

t
: T

M T

M
dened by

t
= L
t
L
1
.
Theorem 0.1.1 (Hamilton-Jacobi). Let be a closed 1-form on
M. If H is constant on the graph Graph() = (x,
n
) [ x M,
then this graph is invariant by

t
.
We can then ask the following question:
xii
Given a xed closed 1-form
0
, does there exist another
closed 1-form cohomologous with
0
such that H is constant on
the graph of ?
The answer is in general negative if we require to be contin-
uous. It is sometimes positive, this can be a way to formulate the
Kolmogorov-Arnold-Moser theorem, see [Bos86].
However, there are always solutions in a generalized sense. In
order to explain this phenomenon, we will rst show how to reduce
the problem to the 0 cohomology class. If
0
is a xed closed 1-
form, let us consider the Lagrangian L

0
= L
0
, dened by
L

0
(x, v) = L(x, v)
0,x
(v).
Since
0
is closed, if we consider only curves with the same xed
endpoints, the map
_


0
is locally constant. It follows that
L

0
and L have the same extremal curves. Thus they have also
the same Euler-Lagrange ow. The Hamiltonian H

0
associated
with L

0
veries
H

0
(x, p) = H(x,
0,x
+p).
By changing the Lagrangian in this way we see that we have only
to consider the case
0
= 0.
We can then try to solve the following problem:
Does there exist a constant c R and a dierentiable function
u : M R such that H(x, d
x
u) = c, for all x M?
There is an integrated version of this question using the semi-
group T

t
: (
0
(M, R) (
0
(M, R), dened for t 0 by
T

t
u(x) = infL() +u((0)) [ : [0, t] M, (t) = x.
It can be checked that T

t+t

= T

t
T

, and thus T

t
is a (non-
linear) semigroup on (
0
(M, R).
A C
1
function u : M R, and a constant c R satisfy
H(x, d
x
u) = c, for all x M, if and only if T

t
u = uct, for each
t 0.
Theorem 0.1.2 (Weak KAM). We can always nd a Lipschitz
function u : M R and a constant c R such that T

t
u = uct,
for all t 0.
xiii
The case M = T
n
, in a slightly dierent form (viscosity solu-
tions) is due to P.L. Lions, G. Papanicolaou and S.R.S. Varadha-
ran 87, see [LPV87, Theorem 1, page 6]. This general version was
obtained by the author in 96, see [Fat97b, Theor`eme1, page 1044].
Carlsson, Haurie and Leizarowitz also obtained a version of this
theorem in 1992, see [CHL91, Theorem 5.9, page 115].
As u is a Lipschitz function, it is dierentiable almost every-
where by Rademachers Theorem. It can be shown that H(x, d
x
u) =
c at each point where u is dierentiable. Moreover, for such a func-
tion u we can nd, for each x M, a C
1
curve
x
:] , 0] M,
with
x
(0) = x, which is a solution of the multivalued vector eld
grad
L
u(x) dened on M by
grad
L
u(y) = L
1
(y, d
y
u).
These trajectories of grad
L
u are minimizing extremal curves.
The accumulation points of their speed curves in TM for t
dene a compact subset of TM invariant under the Euler-Lagrange
ow
t
. This is an instance of the so-called Aubry and Mather
sets found for twist maps independently by Aubry and Mather in
1982 and in this full generality by Mather in 1988.
We can of course vary the cohomology class replacing L by
L

and thus obtain other extremal curves whose speed curves


dene compact sets in TM invariant under
t
. The study of these
extremal curves is important for the understanding of this type of
Lagrangian Dynamical Systems.
xiv
Chapter 1
Convex Functions:
Legendre and Fenchel
Besides some generalities in the rst two sections, our main goal
in this chapter is to look at the Legendre and Fenchel transforms.
This is now standard material in Convex Analysis, Optimization,
and Calculus of Variations. We have departed from the usual
viewpoint in Convex Analysis by not allowing our convex functions
to take the value +. We think that this helps to grasp things
on a rst introduction; moreover, in our applications all functions
have nite values. In the same way, we have not considered lower
semi-continuous functions, since we will be mainly working with
convex functions on nite dimensional spaces.
We will suppose known the theory of convex functions of one
real variable, see for example [RV73, Chapter 1]or [Bou76, Chapitre
1].
1.1 Convex Functions: General Facts
Denition 1.1.1 (Convex Function). Let U be a convex set in
the vector space E. A function f : U R is said to be convex if
it satises the following condition
x, y U, t [0, 1], f(tx + (1 t)y) tf(x) + (1 t)f(y).
1
2
The function f is said to be strictly convex if it satises the fol-
lowing stronger condition
x ,= y U, t ]0, 1[, f(tx + (1 t)y) < tf(x) + (1 t)f(y).
It results from the denition that f : U R is convex if and
only if for every line D E the restriction of f on D U is a
convex function (of one real variable).
Proposition 1.1.2. (i) An ane function is convex (an ane
function is a function which is the sum of a linear function and a
constant).
(ii) If (f
i
)
i
I is a family of convex functions : U R, and
sup
iI
f
i
(x) < + for each x U then sup
iI
f
i
is convex.
(iii) Let U be an open convex subset of the normed space E.
If f : U R is convex and twice dierentiable at x U, then
D
2
f(x) is non-negative denite as a quadratic form. Conversely,
if g : U R admits a second derivative D
2
g(x) at every point
x U, with D
2
g(x) non-negative (resp. positive) denite as a
quadratic form, then g is (resp. strictly) convex.
Properties (i) and (ii) are immediate from the denitions. The
property (iii) results from the case of the functions of a real vari-
able by looking at the restrictions to each line of E.
Denition 1.1.3 (C
2
-Strictly Convex Function). Let be a in the
vector space E. A function f : U R, dened on the convex
subset U of the normed vector space E, is said to be C
2
-strictly
convex if it is C
2
, and its the second derivative D
2
f(x) is positive
denite as a quadratic form, for each x U.
Exercise 1.1.4. Let U be an open convex subset of the normed
space E, and let f : U R be a convex function.
a) Show that f is not strictly convex if and only if there exists a
pair of distinct points x, y U such that f is ane on the segment
[x, y] = tx + (1 t)y [ t [0, 1].
b) If f is twice dierentiable at every x U, show that it
is strictly convex if and only if for every unit vector v E the
set x U [ D
2
f(x)(v, v) = 0 does not contain a non trivial
segment parallel to v. In particular, if D
2
f(x) is non-negative
denite as a quadratic form at every point x U, and the set of
3
points x where D
2
f(x) is not of maximal rank does not contain a
non-trivial segment then f is strictly convex.
Theorem 1.1.5. Suppose that U is an open convex subset of
the topological vector space E. Let f : U R be a convex
function. If there exists an open non-empty subset V U with
sup
xV
f(x) < +, then f is continuous on U.
Proof. Let us rst show that for all x U, there exists an open
neighborhood V
x
of x, with V
x
U and sup
yVx
f(y) < +. In-
deed, if x / V , we choose z
0
V . The intersection of the open set
U and the line containing x and z
0
is an open segment contain-
ing the compact segment [x, z
0
]. We choose in this intersection, a
point y near to x and such that y / [x, z
0
], thus x ]y, z
0
[, see g-
ure 1.1. It follows that there exists t with 0 < t
0
< 1 and such that
x = t
0
y +(1t
0
)z
0
. The map H : E E, z x = t
0
y +(1t
0
)z
sends z
0
to x, is a homeomorphism of E, and, by convexity of U, it
maps U into itself. The image of V by H is an open neighborhood
V
x
of x contained in U. Observe now that any point x

of V
x
can
be written as the form x

= t
0
y + (1 t
0
)z with z V for the
same t
0
]0, 1[ as above, thus
f(x

) = f(t
0
y + (1 t
0
)z)
t
0
f(y) + (1 t
0
)f(z)
t
0
f(y) + (1 t
0
) sup
zV
f(z) < +.
This proves that f is bounded above on V
x
.
Let us now show that f is continuous at x U. We can
suppose by translation that x = 0. Let V
0
be an open subset of U
containing 0 and such that sup
yV
0
f(y) = M < +. Since E is
a topological vector space, we can nd an open set

V
0
containing
0, and such that t

V
0
V
0
, for all t R with [t[ 1. Let us
suppose that y

V
0
(

V
0
), with 1. We can write y = z
+
and y = z

, with z
+
, z



V
0
(of course z

= z
+
, but this
is irrelevant in our argument). As y = (1 )0 + z
+
, we obtain
f(y) (1 )f(0) +f(z
+
), hence
y

V
0
(

V
0
), f(y) f(0) (M f(0)).
4
U
V
V
x
z
0
x
y
Figure 1.1:
We can also write 0 =
1
1+
y +

1+
z

, hence f(0)
1
1+
f(y) +

1+
f(z

) which gives (1 + )f(0) f(y) + f(z

) f(y) + M.
Consequently
y

V
0
(

V
0
), f(y) f(0) M +f(0).
Gathering the two inequalities we obtain
y

V
0
(

V
0
), [f(y) f(0)[ (M f(0)).
Corollary 1.1.6. A convex function f : U R dened on an
open convex subset U of R
n
is continuous.
Proof. Let us consider n+1 anely independent points x
0
, , x
n

U. The convex hull of x
0
, , x
n
has a non-empty interior. By
convexity, the map f is bounded by max
n
i=0
f(x
i
) on .
Most books treating convex functions from the point of view of
Convex Analysis do emphasize the role of lower semi-continuous
convex functions. When dealing with nite valued functions, the
following exercise shows that this is not really necessary.
5
Exercise 1.1.7. Let U be an open subset of the Banach space E.
If f : U R is convex, and lower semi-continuous show that it is
in fact continuous. [Indication: Consider the sequence of subsets
C
n
= x U [ f(x) n, n N. Show that one of these subsets
has non-empty interior.]
We recall that a function f : X Y , between the metric
spaces X, Y , is said to be locally Lipschitz if, for each x X,
there exists a neighborhood V
x
of x in X on which the restriction
f
|Vx
is Lipschitz.
Theorem 1.1.8. Let E be a normed space and U E an open
convex subset. Any convex continuous function f : U R is a
locally Lipschitz function.
Proof. In fact, this follows from the end of the proof of Theorem
1.1.5. We now give a direct slightly modied proof.
We x x U. Since f is continuous, there exists r ]0, +[
and M < + such that
sup
y

B(x,r)
[f(y)[ M.
We have used the usual notation

B(x, r) to mean the closed ball
of center x and radius r.
Let us x y, y


B(x, r/2). We call z the intersection point
of the boundary B(x, r) = x

E [ |x

x| = r of the closed
ball

B(x, r) with the line connecting y and y

such that y is in the


segment [z, y

], see gure 1.2. We of course have |zy

| r/2. We
write y = tz + (1 t)y

, with t [0, 1[, from which it follows that


yy

= t(zy

). By taking the norms and by using |zy

| r/2,
we see that
t |y

y|
2
r
.
The convexity of f gives us f(y) tf(z)+(1t)f(y

), from which
we obtain the inequality f(y) f(y

) t(f(z) f(y

)). It results
that
f(y) f(y

) 2tM
4M
r
|y y

|,
and by symmetry
y, y


B(x, r/2), [f(y) f(y

)[
4M
r
|y y

|.
6
z
y
y

Figure 1.2:
Corollary 1.1.9. If f : U R is convex with U R
n
open and
convex, then f is a locally Lipschitz function.
We recall Rademachers Theorem, see [EG92, Theorem 2, page
81]or [Smi83, Theorem 5.1, page 388].
Theorem 1.1.10 (Rademacher). A locally Lipschitz function de-
ned on open subset of R
n
and with values in R
m
is Lebesgue
almost everywhere dierentiable.
Corollary 1.1.11. A convex function f : U R, where U is
open convex of R
n
, is Lebesgue almost everywhere dierentiable.
It is possible to give a proof of this Corollary not relying on
Rademachers Theorem, see [RV73, Theorem D, page 116]. We
conclude this section with a very useful lemma.
Lemma 1.1.12. Let f : V R be a convex function dened on
an open subset V of a topological vector space.
(a) A local minimum for f is a global minimum.
7
(b) If f is strictly convex, then f admits at most one minimum.
Proof. (a) Let x
0
be a local minimum. For y V and t [0, 1]
and close to 1 we have
f(x
0
) f(tx
0
+ (1 t)y) tf(x
0
) + (1 t)f(y),
thus (1 t)f(x
0
) (1 t)f(y) for t close to 1. It follows that
f(y) f(x
0
).
(b) It results from the convexity of f that the subset x [
f(x) is convex. If = inf f, we have x [ f(x) = inf f =
x [ f(x) inf f. If f is strictly convex this convex set cannot
contain more than one point.
1.2 Linear Supporting Form and Derivative
As is usual, if E as a vector space (over R) we will denote by
E

= Hom(E, R) its algebraic dual space. We will indierently


use both notations p(v) or p, v) to denote the value of v E
under the linear form p E

.
Denition 1.2.1 (Supporting Linear Form). We say that the
linear form p E

is a supporting linear form at x


0
U for the
function f : U R, dened on U E, if we have
x U, f(x) f(x
0
) p(x x
0
) = p, x x
0
).
We will denote by SLF
x
(f) the set of supporting linear form at x
for f, and by SLF(f) the graph
SLF(f) =
xU
x SLF
x
(f) U E

.
In the literature, the linear form p is also called subderiva-
tive of f at x
0
or even sometimes subgradient. We prefer to call
it supporting linear form to avoid confusion with the notion of
subdierential that we will introduce in another chapter.
Example 1.2.2. a) If f : R R, t [t[ then SLF
0
(f) = [1, 1],
for t > 0, SLF
t
(f) = 1, and for t < 0, SLF
t
(f) = 1.
b) If g : R R, t t
3
then SLF
t
(g) = , for every t R.
The following Proposition is obvious.
8
Proposition 1.2.3. The set SLF
x
(f) is a convex subset of E

.
Moreover, if we endow E

with the topology of simple convergence


on E (weak topology) then SLF
x
(f) is also closed.
Here is the relation between supporting linear form and deriva-
tive.
Proposition 1.2.4. Let f : U R be a function dened on an
open subset U of the normed space E.
a) If f is dierentiable at some given x U then SLF
x
(f)
Df(x), i.e. it is either empty or equal to the singleton Df(x).
b) If E = R
n
, and all partial derivatives f/x
i
(x), i = 1, . . . , n,
exist at some given x U, then SLF
x
(f) is either empty or re-
duced to the single linear form (a
1
, . . . , a
n
)

n
i=1
a
i
f/x
i
(x).
Proof. a) If SLF
x
(f) ,= , let p be a supporting linear form of f
at x. If v E is xed, for all > 0 small we have x +v U and
thus f(x +v) f(x) p(v). Dividing by and taking the limit
as goes to 0 in this last inequality, we nd Df(x)(v) p(v). For
linear forms this implies equality, because a linear form which is
0 everywhere has to be 0.
b) We denote by (e
1
, . . . , e
n
) the canonical base in R
n
. Let us
consider a point x = (x
1
, . . . , x
n
) R
n
where all partial deriva-
tives exist. This implies that the function of one variable h
f(x
1
, . . . , x
i1
, h, x
i+1
, . . . , x
n
) is dierentiable at x
i
, hence by part
a), if p SLF
x
(f), we have p(e
i
) = f/x
i
(x). Since this is true
for every i = 1, . . . , n, therefore the map p must be (a
1
, . . . , a
n
)

n
i=1
a
i
f/x
i
(x).
We have not imposed any continuity in the denition of a sup-
porting linear form for a function f. This is indeed the case under
very mild conditions on f, as we will see presently.
Proposition 1.2.5. Let U be an open subset of the topological
vector space E, and let f : U R be a function. Suppose that
f is bounded from above on a neighborhood of x
0
U, then any
supporting linear form of f at x
0
is continuous.
Proof. Let V be a neighborhood of 0 such that V = V , and f is
dened and bounded from above by K < + on x
0
+V . Since V
9
is symmetrical around 0, for each v V , we have
p(v) f(x
0
+v) f(x
0
) 2K
p(v) = p(v) f(x
0
v) f(x
0
) 2K,
hence the linear form p is thus bounded on a nonempty open sub-
set, it is therefore continuous.
As is customary, if E is a topological vector space, we will
denote by E

the topological dual space of E, namely E

is the subset formed by the continuous linear forms. Of course


E

= E

if E is nite-dimensional. If E is a normed space, with


norm ||, then E

is also a normed space for the usual norm


|p| = supp(v) [ v E, |v| 1.
In the case of continuous map, we can improve Proposition 1.2.3.
Proposition 1.2.6. Suppose that f : U R is a continuous
function dened on the topological vector space E. If we endow E

with the topology of simple convergence on E (weak topology),


then the graph SLF(f) is a closed subset of U E

.
The proof of this Proposition is obvious.
Exercise 1.2.7. Let f : U R be a locally bounded function
dened on the open subset U of the normed space E. (Recall that
locally bounded means that each point in U has a neighborhood on
which the absolute value of f is bounded)
a) Show that for every x U, we can nd a constant K, and a
neighborhood V such that for every y V and every p SLF
y
(f)
we have |p| K. [Indication: see the proof of Theorem 1.4.1]
b) If E is nite dimensional, and f is continuous, show the
following continuity property: for every x U, and every neigh-
borhood W of SLF
x
(f) in E

= E

, we can nd a neighborhood V
of x such that for every y V we have SLF
y
(f) W.
As we will see now the notion of linear supporting form is
tailored for convex functions.
Proposition 1.2.8. If the function f : U R, dened on the
convex subset U of the vector space E, admits a supporting linear
form at every x U, then f is convex.
10
Proof. Let us suppose that x
0
= y + (1 t)z with y, z U and
t [0, 1]. If p is a supporting linear form at x
0
, we have
f(y) f(x
0
) p(y x
0
) and f(z) f(x
0
) p(z x
0
),
hence
tf(y) + (1 t)f(z) f(x
0
) p(t(y x
0
) + (1 t)(z x
0
))
= p(ty + (1 t)z x
0
) = 0.
The following theorem is essentially equivalent to the Hahn-
Banach Theorem.
Theorem 1.2.9. Let U be a convex open subset of the locally
convex topological vector space E. If f : U R is continuous and
convex, then we can nd a supporting linear form for f at each
x U.
Proof. As f is continuous and convex, the set
O = (x, t) [ x U, f(x) < t
is open, non-empty, and convex in E R. Since (x
0
, f(x
0
)) is not
in O, by the Hahn-Banach Theorem, see [RV73, Theorem C, page
84] or [Rud91, Theorem, 3.4, page 59], there exists a continuous
and non identically zero linear form : E R R and such that
(x, t) O, (x, t) > (x
0
, f(x
0
)).
We can write (x, t) = p
0
(x) +k
0
t, with p
0
: E R a continuous
linear form and k
0
R. Since (x
0
, t) > (x
0
, f(x
0
)) for all
t > f(x
0
), we see that k
0
> 0. If we dene p
0
= k
1
0
p
0
, we get
p
0
(x)+t p
0
(x
0
)+f(x
0
), for all t > f(x), therefore f(x)f(x
0
)
( p
0
)(x x
0
). The linear form p
0
is the supporting linear form
we are looking for.
The following Proposition is a straightforward consequence of
Theorem 1.2.9 and Proposition 1.2.4
11
Proposition 1.2.10. Let f : U R be a continuous convex
function dened on an open convex subset U of the normed space
E. If f is dierentiable at x
0
then the derivative Df(x
0
) is the
only supporting linear form of f at x
0
. In particular, we have
x U, f(x) f(x
0
) Df(x
0
)(x x
0
).
Corollary 1.2.11. Let f : U R be a continuous convex func-
tion dened on an open convex subset U of a normed space. If f
is dierentiable at x
0
, then x
0
is a global minimum if and only if
Df(x
0
) = 0.
Proof. Of course, if the derivative exists at a minimum it must be
0, this is true even if f is not convex. The converse, which uses
convexity, follows from the inequality
f(y) f(x
0
) Df(x
0
)(y x
0
) = 0
given by Proposition 1.2.10 above.
Corollary 1.2.12. If U R
n
is open and convex and f : U R
is a convex function, then, for almost all x, the function f admits
a unique supporting linear form at x.
Proof. This is a consequence of Proposition 1.2.10 above and Rade-
machers Theorem 1.1.10.
Exercise 1.2.13. Let U be an open and convex subset of R
n
.
Suppose that f : U R is convex and continuous. Show that if
f admits a unique supporting linear form p
0
at x
0
then Df(x
0
)
exists, and is equal to p
0
. [Indication: For each x U 0, choose
p
x
SLF
x
(f), and prove that
p
0
(x x
0
) f(x) f(x
0
) p
x
(x x
0
).
Conclude using exercise 1.2.7.
1.3 The Fenchel Transform
Recall that for a topological vector E, we denote its topological
dual by E

.
12
Denition 1.3.1 (Fenchel Transform). If L : E R is function,
the Fenchel transform of L, denoted by H (or L

if we want to
refer explicitly to L), is the function H : E

] , +] dened
by
H(p) = sup
vE
p, v) L(v).
We will call Fenchels formula the relation between H and L.
The everywhere satised inequality
p, v) L(v) +H(p),
is called the Fenchel inequality.
It is easily seen that H(0) = inf
vE
L(v) and that H(p)
L(0), for all p E

.
We have not dened H on E

because it is identically + on
E

under a very mild hypothesis on L.


Exercise 1.3.2. If L : E R is bounded on some non-empty
open subset of the normed space E, show that if we extend the
Fenchel H to E

, using the same denition, then H is identically


+ on E

.
Usually H assumes the value + even on E

. To give a case
where H is nite everywhere, we must introduce the following
denition.
Denition 1.3.3 (Superlinear). Let E be a normed space. A map
f : E ] , +] is said to be superlinear, if for all K < +,
there exists C(K) > such that f(x) K|x| + C(K), for all
x E.
When E is nite-dimensional, all norms are equivalent hence
the notion of superlinearity does not depend on the choice of a
norm.
Exercise 1.3.4. 1) Show that f : E R, dened on the normed
space E, is superlinear if and only if lim
x
f(x)
x
= + and f
is bounded below.
2) If f : E R is continuous on the nite dimensional vector
space E, show that it is superlinear if and only if
lim
x
f(x)
|x|
= +.
13
Proposition 1.3.5. Let L : E R be a function, dened on the
normed space E, and let H be its Fenchel transform.
(1) If L is superlinear, then H is nite everywhere. It is even
bounded on bounded subsets of E.
(2) If H is nite everywhere, it is convex.
(3) If L is bounded on bounded subsets of E, then H is su-
perlinear. In particular, if L is continuous, and E is nite-
dimensional, then H is superlinear.
Proof. Let us show (1). We know that H is bounded below by
L(0). It remains to show it is nite an bounded from above on
each subset p E

[ |p| K, for each K < +. By the


superlinearity of L, there exists C(K) > such that L(v)
K|v| + C(K), for all v E, and thus for p E

such that
|p| K, we have
p, v) L(v) |p| |x| K|x| C(|p|) C(|p|) < +.
From which follows sup
pK
H(p) C(|p|) < +.
Property (2) results from the fact that H is an upper bound
of a family of functions ane in p.
Let us show (3). We have
H(p) sup
v=K
p, v) sup
v=K
L(v).
But sup
v=K
p, v) = K|p|, and sup
v=K
L(v) < + by the
hypothesis, since the sphere v E [ |v| = K is bounded. If E
is nite dimensional, bounded sets are compact, and therefore, if
L is continuous, it is bounded on bounded subsets of E.
Theorem 1.3.6 (Fenchel). Let us suppose that L : E R is
superlinear on the normed space E.
(i) The equality p
0
, v
0
) = H(p
0
) +L(v
0
) holds if and only if p
0
is a supporting linear form for L at v
0
.
(ii) If L is convex and dierentiable everywhere then p, v) =
H(p) +L(v) if and only if p = DL(v). Moreover
v E, H DL(v) = DL(v)(v) L(v).
14
(iii) If we have L(v) = sup
pE
p, v) H(p), for each v E, then
L is convex. Conversely, if L is convex and continuous then
L(v) = sup
pE
p, v) H(p), for each v E.
Proof. Let us show (i). If L(v) L(v
0
) p
0
, v v
0
), we nd
p
0
, v
0
) L(v
0
) p
0
, v) L(v), for all v E, and thus H(p
0
) =
p
0
, v
0
) L(v
0
). Conversely, by Fenchels inequality p
0
, v)
H(p
0
) + L(v), for all v E. If we subtract from this inequal-
ity the equality p
0
, v
0
) = H(p
0
) +L(v
0
), we obtain p
0
, v v
0
)
L(v) L(v
0
).
Part (ii) follows from part (i) since for a dierentiable func-
tion the only possible supporting linear form is the derivative, see
Proposition 1.2.4.
Let us show (iii). If L(v) = sup
pE
p, v) H(p), then, the
function L is convex as a supremum of ane functions. Conversely,
by (i) we always have L(v) p, v) H(p). Therefore L(v)
sup
pE
p, v) H(p). If L is convex, let p
0
be a linear supporting
form for L at v, by (ii), we obtain L(v) = p
0
, v) H(p
0
) and thus
L(v) = sup
pE
p, v) H(p).
Exercise 1.3.7. Let L : E R be superlinear on the normed
space E, and let H be its Fenchel transform. Denote by A
L
the
set of ane continuous functions v p(v) +c, p E

, c R, such
that L(v) p(v) +c, for each v E. If L

: E R is dened by
L

(v) = sup
fA
L
f(v), show that
L

(v) = sup
pE

p, v) H(p).
[Indication: An ane function f = p + c, p E

, c R, is in A
L
if and only if c H(p).]
Proposition 1.3.8. Suppose that L : E R is continuous and
superlinear on the nite-dimensional linear space E, and H : E

R is its Fenchel transform.


(i) H is everywhere continuous, and superlinear.
(ii) For every p E

, there exists v E such that p, v) =


H(p) +L(v).
(iii) If L is convex, for every v E, there exists p E

such that
p, v) = H(p) +L(v).
15
Proof. We are assuming that E is nite-dimensional, and that L
is continuous. Therefore, in part (i), the continuity follows from
the convexity of H, see 1.1.6, and the superlinearity follows from
part (iii) of Theorem 1.3.6.
We now prove part (ii). Since lim
v+
L(x, v)/|v| = +,
and [p(v)[ |p||v|, we see that
lim
v+
[p(v) L(x, v)]
|v|
= .
Hence the supremum H(x, p) of the continuous function p()
L(x, ) is the same as the supremum of its restriction to big enough
bounded sets. Since bounded sets in E are compact, the supremum
H(x, p) is achieved.
For part (iii), we remark that E = E

, and that L is the


Fenchel transform of H, by part (ii) of Fenchels Theorem 1.3.6,
therefore we can apply part (ii) of the present Proposition.
Corollary 1.3.9. If E is nite-dimensional and L : E R is
everywhere dierentiable and superlinear, then DL : E E

is
surjective.
Proof. This follows from part (ii) of Fenchels Theorem 1.3.6 to-
gether with part (ii) of Proposition 1.3.8 (note that L is continuous
since it is dierentiable everywhere).
We will need some bered version of the results in this section.
We will have to consider locally trivial nite-dimensional vector
bundle : E X, where X is a Hausdor topological space. We
will use the notation (x, v) for a point in E to mean x X and
v E
x
=
1
(x), with this notation : E X is the projection
on the rst coordinate (x, v) x.
We denote, as is customary by

: E

X the dual vector


bundle.
We recall that a continuous norm on : E X is a continuous
function (x, v) |v|
x
such that v |v|
x
is a norm on the
ber E
x
, for each x X. Such a norm induces a dual norm on

: E

X dened, for p E

x
, in the usual way by
|p[
x
= supp(v) [ v E
x
, |v[
x
1.
The following result is classical.
16
Proposition 1.3.10. Let : E X be a locally trivial vector
bundle with nite-dimensional bers over the Hausdor topologi-
cal space X, then all continuous norms on this bundle are equiva-
lent above compact subsets of X. This means that for each com-
pact subset C X, and each pair ||, ||

of continuous norms,
there exists constants , , with > 0, and such that
(x, v) E, x C
1
|v|
x
|v|

x
|v|
x
.
Proof. We do it rst for the case of the trivial bundle xR
n
X,
with X compact. It is not dicult to see that it suces to do it
with ||
x
a xed norm independent of x, for example the Euclidean
norm on R
n
, which we simply denote by ||. The set S = Xv
R
n
[ |v| = 1 is compact and disjoint from 0, therefore by
continuity the two bounds = inf
(x,v)S
|v|
x
, = sup
(x,v)S
|v|

x
are attained, hence they are nite and ,= 0. It is not dicult to
see by homogeneity that
(x, v) X R
n
, |v| |v|

x
|v|.
For the case of a general bundle, if C X is compact, we can
nd a nite number U
1
, . . . , U
n
of open subsets of X such that the
bundle is trivial over each U
i
, and C U
1
U
n
. Since X is
Hausdor, we can write C = C
1
C
n
, with C
i
compact, and
included in U
i
. From the rst part of the proof two norms on the
bundle are equivalent above each C
i
, hence this is also the case of
their (nite) union C.
Denition 1.3.11 (Superlinear Above Compact subsets). Sup-
pose that : E X is a nite-dimensional locally trivial vec-
tor bundle over the topological space X. We say that a function
L : E X is superlinear above compact subsets if for every com-
pact subset C X, and every K 0, we can nd a constant
A(C, K) > such that
(x, v) E, x C L(x, v) K|v|
x
+A(C, K),
where ||
x
is a xed continuous norm on the vector bundle E.
When X is compact we will say that L is superlinear instead
of superlinear above compact subsets. Of course in that case, it
suces to verify the condition of superlinearity with K = X.
17
Of course, the condition above is independent of the choice
of the continuous norm on the vector bundle, since all norms are
equivalent by Proposition 1.3.10. We have not dened the concept
of uniform superlinearity for a general X because it depends on
the choice of the norm on the bundle, since if X is not compact
not all norms are equivalent.
Theorem 1.3.12. Suppose L : E R is a continuous function
on the total space of the nite-dimensional locally trivial vector
bundle : E X. We consider

: E

X, the dual vector


bundle and dene H : E

R by
H(x, p) = sup
vEx
p(v) L(x, v).
If L is superlinear above compact subsets of X, and X is a Haus-
dor locally compact, topological space, then H is continuous and
superlinear above compact subsets of X.
Proof. Since continuity is a local property, and X is Hausdor lo-
cally compact, without loss of generality, we can assume X com-
pact, and : E X trivial, therefore E = X R
n
. We choose a
norm || on R
n
.
Fix K 0, we can pick C > such that
(x, v) X R
n
, L(x, v) (K + 1)|v| +C.
If we choose R > 0 such that R + C > sup
xX
L(x, 0) (this is
possible since the right hand side is nite by the compactness of
X), we see that for each x X, v R
n
,and each p R
n
satisfying
|p| K, |v| R, we have
p(v) L(x, v) |p|v|| (K + 1)|v| C
RC < sup
xX
L(x, 0)
L(x, 0) = p(0) L(x, 0).
Therefore for |p| K, we have H(x, p) = sup
vR
p(v)L(x, v).
Since v R
n
[ |v| R is compact, we see that H is continuous
on the set X p R
n
[ |p| K. But K 0 is arbitrary,
therefore the function H is continuous everywhere.
18
We prove superlinearity above compact subsets of X. Us-
ing the same argument as in nal part the proof of Proposition
1.3.10 above, we can without loss of generality suppose that X
is compact, and that the bundle is the trivial bundle X R
n

X. For a xed K, remark that by compactness, and continuity


A = supL(x, v) [ x X, v R
n
, |v| K is nite. Therefore
H(x, p) p(v) A, for each v R
n
, satisfying |v| K. If we
take the supremum over all such vs, since K|p| = supp(v) [ v
R
n
, |v| K, we get H(x, p) K|p| A.
Denition 1.3.13 (Convex in the Fibers). Let L : E R be
a continuous function on the total space of the nite-dimensional
locally trivial vector bundle : E X, where X is a Hausdor
space. We will say that a Lagrangian L on the manifold M is
convex in the bers, if the restriction L
|Ex
is convex for each x X.
In fact, for convex functions superlinearity above compact sets
is not so dicult to have, because of the following not so well
known theorem.
Theorem 1.3.14. Suppose L : E R is a continuous function
on the total space of the nite-dimensional locally trivial vector
bundle : E X, where X is a Hausdor space. If L is convex
in the bers, then L is superlinear above each compact subsets of
X if and only if L
|Ex
is superlinear, for each x X.
Proof. Of course, it is obvious that if L is superlinear above each
compact subset, then each restriction L
|Ex
is superlinear.
Suppose now that L
|Ex
is convex and superlinear for each x
X, to prove that L is linear above compact subsets of X, again by
the same argument as in nal part the proof of Proposition 1.3.10
above, we can without loss of generality that X is compact, and
that the bundle is the trivial bundle X R
n
X.
We choose a xed norm || on R
n
. For given x
0
X, and
K 0, we will show that there exists a neighborhood V
x
0
of x
0
and C(x
0
, K) > such that
x V
x
0
, v R
n
, L(x, v) K|v| +C(x
0
, K). (*)
A compactness argument nishes the proof.
19
We now prove (*). We choose C
1
> such that
v R
n
, L(x
0
, v) (K + 1)|v| +C
1
.
We then pick R > 0 such that R +C
1
L(x
0
, 0) + 1. Now if p
R
n
, and v R
n
satisfy respectively |p|
x
0
K, and |v|
x
0
= R,
we see that
L(x
0
, v) p(v) (K + 1)|v| +C
1
K|v|
R +C
1
L(x
0
, 0) + 2.
Since the set (v, p) R
n
R
n
[ |v|
x
0
= R, |p|
x
0
K is
compact, and L is continuous, we can nd a neighborhood V
x
0
of
x
0
in X such that for each x V
x
0
, v R
n
, and each p R
n
, we
have
|v| = R, |p| K L(x, v) p(v) > L(x, 0).
This implies that for xed x V
x
0
, and p R
n
satisfying |p|
K, the convex function L(x, ) p() achieves its minimum on
the compact set v R
n
[ |v| R in the interior of that set.
Therefore, the convex function L(x, ) p() has a local minimum
attained in v R
n
[ |v| < R. By convexity this local minimum
must be global, see 1.1.12. Therefore, dening C = infL(x, v)
p(v) [ x X, |v|
x
R, |p|
x
K, we observe that C is nite by
compactness, and we have
(x, v, p) V
x
0
R
n
R
n
, |p| K L(x, v) p(v) C.
Taking the inmum of the right hand side over all |p|
x
K, we
get
(x, v) V
x
0
R
n
, L(x, v) K|v| C.
1.4 Dierentiable Convex Functions and Le-
gendre Transform
Theorem 1.4.1. Let U be an open convex subset of R
n
. If f :
U R is convex and dierentiable at each point of U, then f is
C
1
.
20
Proof. We x x U. Let r ]0, [ be such that the closed ball

B(x, r) is contained in U. Let us set M = sup


y

B(x,r)
[f(y)[ <
+. For h, k

B(0,
r
2
), we have
f(x +h +k) f(x +k) Df(x +k)(h), (*)
taking the supremum over all h such that |h| = r/2, we obtain
|Df(x + k)| 4M/r. Since the ball p R
n
[ |p| 4M/r is
compact, it is enough to see that if k
n
0 and Df(x + k
n
) p,
then p = Df(x). But taking the limit in the inequality (), we get
k

B(0, r/2), f(x +h) f(x) p, h).
It results that Df(x) = p, since we have already seen that at a
point where a function is dierentiable only its derivative can be
a supporting linear form, see Proposition 1.2.4.
Exercise 1.4.2. Let K be a compact topological space and U an
open convex subset of R
n
. If L : K U R is continuous and
such that for each k K, the map U R : v L(k, v) is convex
and everywhere dierentiable, then
L
v
: K U (R
n
)

, (k, v)
L
v
(k, v) is continuous. [Indication: Adapt the proof of Theorem
1.4.1.]
Denition 1.4.3 (Legendre Transform). Let L : U R be a C
1
function, with U R
n
open. The Legendre transform associated
with L is the map L : U R
n
, v DL(v).
We can rephrase part (ii) of Fenchels Theorem 1.3.6 and Corol-
lary 1.3.9 in the following way:
Proposition 1.4.4. Let L : R
n
R be C
1
, convex and super-
linear, then its Legendre transform L : R
n
R
n
is surjective.
Moreover, if we denote by H : R
n
R its Fenchel transform
then p, v) = H(p) +L(v) if and only if p = DL(v), and we have
v R
n
, H L(v) = DL(v)(v) L(x, v).
In particular, the surjectivity of L is a consequence of super-
linearity of L.
We are interested in nding out, for a C
1
convex function L :
R
n
R, when its Legendre transform L : R
n
R
n
is bijective.
It is easy to understand when L is injective.
21
Theorem 1.4.5. Suppose L : U R is a C
1
convex function,
dened on an open subset U of R
n
. Its associated Legendre trans-
form L is injective if and only if L is strictly convex.
Proof. Let p R
n
. We have p = DL(x) if and only if DL
p
(x) = 0
where L
p
(x) = L(x) p(x). Hence x is a point where the function
L
p
reaches its minimum, see 1.2.11. If L is strictly convex so is
L
p
. However a strictly convex function can achieve its minimum
at most at one point.
Conversely, if L is injective, the convex function L(x)DL(x
0
)(x)
has only x
0
as a critical point and hence, and again by Corollary
1.2.11, it reaches its minimum only at x
0
. If x
0
= tx + (1 t)y
with t ]0, 1[, x ,= x
0
and y ,= x
0
, we therefore have
L(x) DL(x
0
)(x) > L(x
0
) DL(x
0
)(x
0
)
L(y) DL(x
0
)(y) > L(x
0
) DL(x
0
)(x
0
).
Since t > 0 and (1 t) > 0, we obtain
tL(x)+(1t)L(y)DL(x
0
)(tx + (1 t)y)
. .
x
0
) > L(x
0
)DL(x
0
)(x
0
),
hence tL(x) + (1 t)L(y) > L(x
0
).
We would like now to prove the following theorem.
Theorem 1.4.6. Let L : R
n
R be C
1
, and convex. If L is its
Legendre transform, then the following statements are equivalent:
(1) The function L is strictly convex, and superlinear.
(2) Its Legendre transform L : R
n
R
n
is a homeomorphism.
(3) Its Legendre transform L : R
n
R
n
is bijective.
Proof. We rst show that (1) implies (3). If (1) is true then from
Proposition 1.4.4, we know that L is surjective, and from Theorem
1.4.5 it is injective.
The fact that (3) implies (2) follows from Brouwers Theo-
rem on the invariance of the domain see [Dug66, Theorem 3.1,
page 358]. (Note that one can obtain a proof independent from
Brouwers Theorem by using Theorem 1.4.13 below.)
22
We now prove that (2) implies (1). Another application of
Theorem 1.4.5 shows that L is strictly convex. It remains to show
the superlinearity. Since L is a homeomorphism, the set A
K
= x [
|DL(x)| = K is compact, and L(A
K
) = p R
n
[ |p| = K,
thus
v R
n
, K|v| = sup
xA
K
DL(x)(v).
As L(v) DL(x)(v) +L(x) DL(x)(x) we see that
L(v) K|v| + inf
xA
K
[L(x) DL(x)(x)],
but inf
xA
K
[L(x)DL(x)(x)] > , because A
K
is compact and
L is of class C
1
.
When it comes to Lagrangians, Analysts like to assume that
they are superlinear, and Geometers prefer to assume that its as-
sociated Legendre transform is bijective. The following Corollary
shows that for C
2
-strictly convex Lagrangians, these hypothesis
are equivalent.
Corollary 1.4.7. Let L : R
n
R be a C
2
convex function. Its
associated Legendre transform L is a C
1
dieomorphism from R
n
onto its dual space R
n
if and only if L is superlinear, and C
2
-
strictly convex.
Proof. Suppose that L = DL is a C
1
dieomorphism. By the
previous Theorem 1.4.6, the map L is superlinear. Moreover, the
derivative DL(v) = D
2
L(v) is an isomorphism, for each v R
n
.
Therefore D
2
L(v) is non degenerate as a bilinear form, for each
v R
n
. Since, by the convexity of L, the second derivative D
2
L(v)
is non negative denite as a quadratic form, it follows that D
2
L(v)
is positive denite as a quadratic form, for each v R
n
.
Conversely, suppose L superlinear, and C
2
-strictly convex .
Then DL : R
n
R
n
is a homeomorphism by Theorem 1.4.6.
Moreover, since DL(v) = D
2
L(v), the derivative DL(v) is thus
an isomorphism at every point v R
n
. By the Local Inversion
Theorem, the inverse map L
1
is also C
1
.
In the sequel of this section, we will discuss some aspects of
the Legendre transform that will not be used in this book. They
nonetheless deserve to be better known.
23
We start with the notion of proper map.
Denition 1.4.8 (Proper Map). A map f : X Y , between
the topological spaces X and Y , is said to be proper if for ev-
ery compact subset K of the target space Y , the inverse image
f
1
(K) X is also compact.
The main properties of proper maps are recalled in the follow-
ing exercise.
Exercise 1.4.9. Let f : X Y be a proper continuous map
between metric spaces.
1) Show that for each closed subset F X, the image f(F) is
closed in Y . [Indication: Use the fact that if a sequence converges,
then the subset formed by this sequence together with its limit is
compact.
2) Conclude that f is a homeomorphism as soon as it is bijec-
tive.
3) Show that a continuous map f : R
n
R
m
is proper if and
only if
lim
x+
|f(x)| = +.
Theorem 1.4.10. Let L : U R be a C
1
convex function, where
U is an open convex subset of R
n
. If its associated Legendre
transform L : U R
n
is proper, then L is surjective.
We need some preliminaries in order to prove the theorem.
Lemma 1.4.11 (Of the Minimum). Let f :

B(x, r) R be a
function which has a derivative at each point of

B(x, r). If f
achieves its minimum at x
0


B(x, r), a closed ball in a normed
space, then Df(x
0
)(x
0
x) = |Df(x
0
)|r = |Df(x
0
)||x
0
x|.
Proof. Without loss of generality, we can suppose x = 0. For all
y

B(0, r) and for all t [0, 1], we have
f(ty + (1 t)x
0
) f(x
0
),
thus, the function
y
: [0, 1] R, t f(ty + (1 t)x
0
) has a
minimum at t = 0, its derivative at 0, namely Df(x
0
)(y x
0
), is
thus 0. Hence Df(x
0
)(y x
0
) 0, for each y

B(0, r), and
24
consequently Df(x
0
)(x
0
) Df(x
0
)(y), for each y

B(0, r). It
follows that
Df(x
0
)(x
0
) = inf
y

B(0,r)
Df(x
0
)(y) = |Df(x
0
)|r.
If Df(x
0
) = 0, we also have the second part of the required equal-
ities. If Df(x
0
) ,= 0, then x must be on the boundary B(0, r) of

B(0, r) and we again have the second part of the equalities.


Corollary 1.4.12. Let f : U R be a C
1
convex function dened
on the open convex subset U of R
n
. If the derivative Df(x) is never
the 0 linear form, for x U, then for each compact subset K U,
we have
inf
xU
|Df(x)| = inf
xU\K
|Df(x)|.
Proof. The inequality inf
xU
|Df(x)| inf
xU\K
|Df(x)| is ob-
vious. If we do not have equality for some compact subset K,
then
inf
xU
|Df(x)| < inf
xU\K
|Df(x)|,
and therefore
inf
xU
|Df(x)| = inf
xK
|Df(x)|,
If we set K
0
= x U [ |Df(x)| = inf
zU
|Df(z)|, it follows
that K
0
is closed, non-empty, and contained in K, therefore K
0
is
compact. Moreover
x U K
0
, |Df(x)| > inf
zU
|Df(z)|.
Since K
0
is compact there exists r > 0 such that the closed set

V
r
(K
0
) = x [ d(x, K
0
) r is contained in the open set U. As
this set

V
r
(K
0
) is also compact, there exists x
0


V
r
(K
0
) such
that f(x
0
) = inf
x

Vr(K
0
)
f(x). Necessarily x is on the boundary of

V
r
(K
0
), because otherwise x
0
would be a local minimum of f and
therefore Df(x
0
) = 0, which is excluded. Hence d(x
0
, K
0
) = r.
By compactness of K
0
, we can nd x K
0
such that d(x
0
, x) = r.
Since

B(x, r)

V
r
(K
0
) and x
0


B(x, r), we also have f(x
0
) =
25
inf
y

B(x,r)
f(y). By the previous Lemma 1.4.11, we must have
Df(x
0
)(x
0
x) = |Df(x
0
)|r. The convexity of f gives
f(x) f(x
0
) Df(x
0
)(x x
0
)
f(x
0
) f(x) Df(x)(x
0
x),
hence Df(x)(x x
0
) Df(x
0
)(x x
0
) = |Df(x
0
)|r. As |x
x
0
| = r, we get
Df(x)(x x
0
) |Df(x)||x x
0
| = |Df(x)|r.
This implies |Df(x)| |Df(x
0
)|, which is absurd. In fact, we
have |Df(x)| = inf
zU
|Df(z)|, since x K
0
, and |Df(x
0
)| >
inf
zU
|Df(z)|, because x
0
/ K
0
.
Proof of theorem 1.4.10. Fix p R
n
, the Legendre transform of
L
p
= Lp is Lp, it is thus also proper. By the previous Corollary,
it must vanish at some point in U, because inf
x/

B(0,r)
|DL
p
(x)|
, when r .
Theorem 1.4.13. Let L : R
n
R be a C
1
convex function. Its
associated Legendre transform L : R
n
R
n
is proper if and only
if L is superlinear.
Proof. Let us suppose L superlinear. By convexity we have L(0)
L(x) DL(x)(0x) and thus DL(x)(x) L(x)L(0) from which
we obtain
|DL(x)| DL(x)
_
x
|x|
_

L(x)
|x|

L(0)
|x|
,
by the superlinearity of L, we do have |DL(x)| , when |x|
.
The proof of the converse is very close to the end of the proof of
Theorem 1.4.6. If L is proper, the set A
K
= x [ |DL(x)| = K
is compact. Moreover, since L is necessarily surjective, see 1.4.10,
we have L(A
K
) = p R
n
[ |p| = K, and thus
v R
n
, K|v| = sup
xA
K
DL(x)(v).
As L(v) DL(x)(v) +L(x) DL(x)(x) we see that
L(v) K|v| + inf
xA
K
[L(x) DL(x)(x)],
26
but inf
xA
K
[L(x)DL(x)(x)] > , because A
K
is compact and
L is of class C
1
.
We would like to conclude this section with a very nice theorem
due to Minty see [Min64, Min61] (see also [Gro90, 1.2. Convexity
Theorem]). In order to give the best statement, we recall some
notions about convex subsets of a nite-dimensional vector space
E. If C F is a convex subset, we will denote by A(C) the
ane subspace generated by C, the relative interior relint(C) of
C is its interior as a subset of A(C).
Theorem 1.4.14 (Minty). If L : R
n
R is a C
1
convex function,
then the closure of the image L(R
n
) = DL(R
n
) of its associated
Legendre transform L is convex. Moreover L contains the relative
interior of its closure.
In order to prove this theorem, we will need the following
lemma.
Lemma 1.4.15. Let L : R
n
R be a C
1
convex function. If
p / L(R
n
), then there exists v R
n
0 such that L(x)(v) p(v)
for all x R
n
. If p is not in the closure of L(R
n
), then, moreover,
there exists > 0 such that L(x)(v) +p(v) for all x R
n
.
Proof of Theorem 1.4.14. To simplify notations, we call C the clo-
sure of L(R
n
). Observe that a linear form on R
n
is of the form
p p(v), where v R
n
. Therefore the Lemma above 1.4.15 shows
that a point in the complement of C, can be strictly separated by
a hyperplane from L(R
n
), and hence from its closure C. This
implies the convexity of C.
It remains to prove the second statement.
We rst assume that the ane subspace generated by L(R
n
)
is the whole of R
n
. We have to prove that the interior of C is
contained in L(R
n
). Suppose that p
0


C is not contained in
L(R
n
), by Lemma 1.4.15 above we can nd v R
n
0 with
L(z)(v) p
0
(v) for all z R
n
, therefore p(v) p
0
(v) for all p in
the closure C of L(R
n
). This is clearly impossible since p
0


C
and v ,= 0.
To do the general case, call E the ane subspace generated
by L(R
n
). Replacing L by LDL(0), we can assume that 0 E,
and therefore E is a vector subspace of R
n
. Changing bases, we
27
can assume that E = R
k
0 R
n
. Since DL(x) E =
R
k
0, we see that L/x
i
is identically 0 for i > 0, and
therefore L(x
1
, . . . , x
n
) depends only on the rst k variables, so we
can write L(x
1
, . . . , x
n
) =

L(x
1
, . . . , x
k
), with

L : R
k
R convex
and C
1
. It is obvious that DL and D

L have the same image in


R
k
= R
k
0 R
n
, therefore the image of D

L generates
anely R
k
. We can therefore apply the rst case treated above
to nish the proof.
Proof of Lemma 1.4.15. We x p
0
/ L(R
n
) = DL(R
n
). The func-
tion L
p
0
= Lp
0
is convex. As its derivative is never the 0 linear
map, it does not have a local minimum. To simplify the notations
let us set f = L p
0
. For each integer k 1, by Lemma 1.4.11
applied to f, and to the ball

B(0, k), we can nd x
k
with |x
k
| = k
such that
Df(x
k
)(x
k
) = |Df(x
k
)||x
k
|.
The convexity of f gives
y R
n
, Df(y)(y x
k
) f(y) f(x
k
) Df(x
k
)(y x
k
),
hence
y R
n
, Df(y)(y x
k
) Df(x
k
)(y x
k
). (*)
In particular, we have Df(0)(x
k
) Df(x
k
)(x
k
) = |Df(x
k
)||x
k
|,
and thus |Df(0)| |Df(x
k
)|. Taking a subsequence, we can
suppose that |x
k
| , that Df(x
k
) p

, and that x
k
/|x
k
|
v

, with v

of norm 1. Dividing both sides the inequality ()


above by |x
k
|, and using the equality
Df(x
k
)(x
k
) = |Df(x
k
)||x
k
|,
and taking limits we obtain
y R
n
, Df(y)(v

) |p

|,
we rewrite it as
y R
n
DL(y)(v

) p
0
(v

) +|p

|.
It then remains to observe that p

= 0 implies that Df(x


k
) =
DL(x
k
)p
0
tends to 0 and thus p
0
is in the closure of DL(R
n
).
28
Exercise 1.4.16. Suppose that L : R
n
R is C
1
, and strictly
convex. Show that the image L(R
n
) is a convex open subset of
R
n
.
Does this result remain true, for L : U R C
1
, and strictly
convex, on the open subset U R
n
?
1.5 Quasi-convex functions
At some point, we will need a class of functions more general than
the convex ones.
Denition 1.5.1 (Quasi-convex). Let C E be a convex subset
of the vector space E. A function f : C R is said to be quasi-
convex if for each t R, the subset f
1
(] , t]) is convex.
Proposition 1.5.2. Let f : C R be a function dened on the
convex subset C of the vector space E.
1 The function f is quasi-convex if and only
x, y C, [0, 1], f(x + (1 )y) max(f(x), f(y)).
2 If f is quasi-convex then for every x
1
, , x
n
C and every

1
, ,
n
[0, 1], with

n
i=1

i
= 1, we have
f(
n

i=1

i
x
i
) max
1in
f(x
i
).
Proof. We prove (2) rst. Suppose f is quasi convex. Since f
1
(]
, max
in
f(x
i
)]) is convex and contains x
1
, , x
n
necessarily

n
i=1

i
, x
i
f
1
(], max
in
f(x
i
)].
To nish proving (1), suppose conversely that
x, y C, [0, 1], f(x + (1 )y) max(f(x), f(y)).
If x, y are in f
1
(] , t]), then f(x) and f(y) are t. Therefore
any convex combination x+(1)y satises f(x+(1)y)
max(f(x), f(y)) t, and hence x+(1)y f
1
(] , t]).
29
Example 1.5.3. 1) Any convex function is quasi-convex.
2) Any monotonic function : I R, where I is an interval
in R, is quasi-convex.
We need a slight generalization of property (2) of Proposition
1.5.2. We start with a lemma. Although we do not need it in its
full generality, it is nice to have the general statement.
Lemma 1.5.4. Suppose that (X, /, ) is a probability measure,
and that : X C is a measurable function with value in
a convex subset C of a nite-dimensional normed space E. If
_
X
|(x)| d(x) < +, then
_
X
(x) d(x) is contained in C.
Proof. We will do the proof by induction on dimC. Recall that
the dimension of a convex set is the dimension of the smallest
ane subspace that contains it. If dimC = 0, then by convexity
C is reduced to one point and the result is trivial.
We assume now that the result is true for every n < dimC.
Replacing E by an ane subset we might assume that C has a
non empty interior. Therefore the convex set C is contained in
the closure of its interior

C, see Lemma 1.5.5 below. Let us dene
v
0
=
_
X
(x)d(x). If v
0
/

C, since

C is open and convex, by
Hahn-Banach Theorem there exists a linear form : E R such
that (v) > (v
0
), for each v

C. Since C is contained in the
closure of

C, we obtain
v C, (v) (v
0
).
Therefore, we have the inequality
x X, (x) (v
0
). ()
If we integrate this inequality we get
_
X
(x) d(x) (v
0
). By
linearity of , the integral
_
X
(x)d(x) is equal to (
_
X
(x) d(x)),
hence to (v
0
), by the denition of v
0
. This means that the in-
tegration of the inequality () leads to an equality, therefore we
have (x) = (v
0
), for -almost every x X. It follows that,
on a set of full -measure (x)
1
((v
0
)) C. But the subset

1
((v
0
)) C is convex and has a dimension < dimC = dimE,
because it is contained in the ane hyperplane
1
((v
0
)). By
induction
_
X
(x) d(x)
1
((v
0
)) C.
30
Lemma 1.5.5. If C is a convex subset of the topological vector
space E, then its interior

C is convex. Moreover if

C is non-empty,
then C is contained in the closure of

C.
Proof. Suppose x

C, and y C. For t > 0, the map H
t
: E
E, z tz + (1 t)y is a homeomorphism of E. Moreover, for
if 0 < t 1, by convexity of C, we have H
t
(C) C, therefore
tx + (1 t)y = H
t
(x) H
t
(

C) C. Since H
t
(

C) is open, we
obtain that tx + (1 t)y

C, for 0 < t 1. This implies the
convexity of

C. Now, if

C is non-empty, we can nd x
0


C, for
y C, and 0 < t 1, we know that tx
0
+ (1 t)y

C. Since
y = lim
t0
tx
0
+ (1 t)y. Therefore C is contained in the closure
of

C.
Proposition 1.5.6. Suppose that f : C R is a quasi-convex
function dened on the convex subset C of the nite dimensional
normed space E. If (X, /, ) is a probability space, and : X
C is a measurable function with
_
X
|(x)| d(x) < +, then
f(
_
X
(x) d(x)) sup
xX
f((x)).
Proof. The set D = c C [ f(c) sup
xX
f((x)) is convex,
and, by denition, contains (x) for every x X. Therefore by
Lemma 1.5.4, we obtain
_
X
(x) d(x) D.
1.6 Exposed Points of a Convex Set
Let us recall the denition of an extremal point.
Denition 1.6.1 (Extremal Point). A point p in a convex set C
is said to be extremal if each time we can write p = tx + (1 t)y,
with x, y C and t [0, 1], then p = x or p = y.
Theorem 1.6.2 (Krein-Milman). If K is a convex compact subset
of a normed space, then K is the closed convex envelope of its
extremal points.
The proof of the Krein-Milman Theorem can be found in most
books on Functional Analysis, see [Bou81, Theor`eme 1, page II.59],
[RV73, Theorem D, page 84] or [Rud91, Theorem 3.23, page 75]
31
C
B A
D
Figure 1.3: Stadium: The four points A, B, C, D are extremal but
not exposed.
Let us recall that an ane hyperplane in a R-vector space E,
determines two open (resp. closed) half-spaces. If H is the set of
points where the ane function a : E R is 0, then the two
open (resp. closed) half-spaces are given by a > 0 and a < 0 (resp.
a 0 and a 0). An hyperplane H is said to be an hyperplane of
support of a subset A E if AH ,= and A is entirely contained
in one of the two closed half-spaces determined by H.
We will need a concept a little ner than that of extremal
point, it is the concept of exposed point.
Denition 1.6.3 (Exposed Point). Let C be a convex subset of a
normed space. A point p of C is exposed, if there is a hyperplane
H of support of C with H C = p.
An exposed point is necessarily an extremal point (exercise).
The converse is not necessarily true, as it can be seen on the
example of a stadium, see gure 1.3.
Theorem 1.6.4 (Straszewicz). If C is a convex compact subset
of R
n
, then C is the closed convex envelope of its exposed points.
Proof. We will use the Euclidean norm on R
n
. Let us denote by
C
1
the closure of the convex envelope of the set of the exposed
points of C. Let us suppose that there exists x C C
1
. As
32
y
C
1 e x a
Figure 1.4: Proof of Straszewiczs Theorem.
closed subset of C, the set C
1
is also compact. By the Theorem
of Hahn Banach, we can nd a hyperplane H strictly separating
x from C
1
. We consider the line D orthogonal to H and passing
through x. We denote by a the intersection DH, see gure 1.4.
If c R
n
, we call c
D
the orthogonal projections of c on D. We
set d = sup
cC
1
|c c
D
|. By compactness of C
1
, and continuity
of c c
D
, this sup d is nite.
Let us x y a point in D on the same side of H as C
1
, we can
33
write
d(y, c) =
_
|y c
D
|
2
+|c c
D
|
2

_
|y a|
2
+d
2
|y a|

1 +
d
2
|y a|
2
|y a|
_
1 +
d
2
2|y a|
2
_
= |y a| +
d
2
2|y a|
.
Since x / H, we get x ,= a, and 0 < |a x|. Therefore, since d
is nite, for y far away on D so that d
2
/[2|y a|] < |a x|, we
obtain
d(y, c) < |y a| +|a x|.
But x, a, y are all three on the line D, and a is between x and y,
hence |y a| +|a x| = |y x|. It follows that for y far away
enough
c C
1
, d(y, c) < |y x|. (*)
Let us then set R = sup
cC
d(y, c). We have R |y x| because
x C. This supremum R is attained at a point e C. By ()
we must have e / C
1
. The hyperplane

H tangent, at the point
e, to the Euclidean sphere S(y, R) = x R
n
[ |x y| = R
is a hyperplane of support for C which cuts C only in e, since
C

B(y, R). Therefore e is an exposed point of C and e / C
1
.
This is a contradiction since C
1
contains all the exposed points of
C.
Theorem 1.6.5. Suppose L : R
n
R is convex and superlinear.
We consider the graph of L
Graph L = (x, L(x)) [ x R
n
R
n
R.
Any point of GraphL belongs to the closed convex envelope of the
exposed points of the convex set
Graph

L = (x, t)[t L(x),


34
formed by the points of R
n
R which are above GraphL.
In fact, for each x R
n
we can nd a compact subset C
R
n
R such that (x, L(x)) is in the closed convex envelope of the
exposed points of Graph

L which are in C.
Proof. Let p be a linear form of support for L at x
0
. Since L(x)
L(x
0
) p(x x
0
), the function

L dened by

L(v) = L(v + x
0
)
p(v)L(x
0
) is 0 everywhere and takes the value 0 at 0. It is also
superlinear, since L is so. Moreover Graph

L is obtained starting
from Graph L using the ane map (v, t) (vx
0
, tp(v)L(x
0
)),
therefore the exposed points of Graph

L are the images by this
ane map of the exposed points of GraphL. From what we have
just done without loss of generality, we can assume that x
0
=
0, L(0) = 0, L 0, and that we have to show that the point (0, 0)
is in the closure of the convex envelope of the exposed points of
Graph

L. For this, we consider the convex subset


C = (x, t) R
n
R [ L(x) t 1.
The exposed points of C are either of the form (x, 1) or of the
form (x, L(x)) with L(x) < 1. These last points are also exposed
points of Graph

L, see Lemma 1.6.6 below. By the superlinearity


of L, the subset C is compact. We apply Straszewicz Theorem to
conclude that (0, 0) is in the closure of the convex envelope of the
exposed points of C. We can then gather the exposed points of C
of the form (x, 1) and replace them by their convex combination.
This allows us to nd, for each n 1, exposed points of C of the
form (x
i,n
, L(x
i,n
)), 1 i
n
, with L(x
i,n
) < 1, a point y
n
with
(y
n
, 1) C, and positive numbers
1,n
, . . . ,
n,n
and
n
such that

n
+

n
i=1

i,n
= 1 and
(0, 0) = lim
n

n
(y
n
, 1) +
n

i=1

i,n
(x
i,n
, L(x
i,n
)).
As L(x
i,n
) 0 we see that
n
0 and

n
i=1

i,n
L(x
i,n
) 0. It
follows that
n
=

n
i=1

i,n
1, since
n
+
n
= 1. Moreover,
since C is compact, the y
n
are bounded in norm, therefore
n
y
n

0, because
n
0. It results from what we obtained above that
(0, 0) = lim
n
n

i=1

i,n

n
(x
i,n
, L(x
i,n
)).
35
This is the required conclusion, because

n
i=1

i,n
/
n
= 1 and the
(x
i,n
, L(x
i,n
)) are exposed points of Graph

L.
Lemma 1.6.6. Let C be a convex subset in the topological vector
space E. Suppose H is a hyperplane containing x
0
C. If there
exists a neighborhood V of x
0
such that H is a hyperplane of
support of C V , then H is a hyperplane of support of C.
Moreover, if H C V = x
0
then x
0
is an exposed point of
C.
Proof. This is almost obvious. Call H
+
is a closed half-space de-
termined by H and containing C V . If v E, then the open
ray D
+
v
= x
0
+ tv [ t > 0 is either entirely contained in H
+
or
disjoint from it. Now if x C, by convexity of C, for t 0 small
enough tx + (1 t)x
0
= x
0
+ t(x x
0
) C V H
+
, therefore
the open ray D
+
xx
0
H
+
. But x = x
0
+ 1(x x
0
) D
+
xx
0
.
Suppose H C V = x
0
. If y ,= x
0
and y C H then
the ray D
+
yx
0
is contained in H, therefore for every t [0, 1]
small enough ty +(1 t)x
0
H C V . This is impossible since
ty + (1 t)x
0
,= x
0
for t > 0.
36
Chapter 2
Calculus of Variations
Our treatment of the Calculus of Variations is essentially the clas-
sical treatment (up to Tonelli) of the one-dimensional setting in
a modern setting. We have mainly used [Cla90], [Mn] and the
appendix of [Mat91]. After most of it was typed we learned from
Bernard Dacorogna the existence of an excellent recent introduc-
tion to the subject [BGH98].
In this chapter, we treat general Lagrangians (i.e. not neces-
sarily convex in the bers). In the second chapter, we will treat the
Lagrangians convex in bers, therefore all properties concerning
existence of minimizing curves will be in next chapter.
2.1 Lagrangian, Action, Minimizers, and Ex-
tremal Curves
In this chapter (and the following ones) we will us the standard
notations that we have already seen in the introduction, namely:
If M is a manifold (always assumed C

, and without bound-


ary), we denote by TM its tangent bundle, and by : TM M
the canonical projection. A point of TM is denoted by (x, v),
where x M, and v T
x
M =
1
(x). With this notation, we
of course have (x, v) = x. The cotangent bundle is

: T

M
M. A point of T

M is denoted by (x, p), where x M, and


p T

x
M = L(T
x
M R).
37
38
Denition 2.1.1 (Lagrangian). A Lagrangian L on the manifold
M is a continuous function L : TM R.
Notice that although L is a function on TM, we will nonethe-
less say that L is a Lagrangian on M.
Denition 2.1.2 (Action of a Curve). If L is a Lagrangian on
the manifold M, and : [a, b] M is a continuous piecewise C
1
curve, with a b, the action L() of for L is
L() =
_
b
a
L((s), (s)) ds.
We are interested in curves that minimize the action.
Denition 2.1.3 (Minimizer). Suppose L is a Lagrangian on M.
If ( is some set of (parametrized) continuous curves in M, we will
say that : [a, b] M is a minimizer for the class ( if for every
curve : [a, b] M, with (a) = (a), (b) = (b), and (, we
have L() L().
If ( is the class of continuous piecewise C
1
curves, then mini-
mizers for this class are simply called minimizers.
It should be noticed that to check that : [a, b] M is a
minimizer for some class, we only use curves parametrized by the
same interval, and with the same endpoints.
In order to nd minimizers, we will use dierential calculus so
that minimizers are to be found among citical points of the action
functional L. In section 2.2 we will rst treat the linear case, i.e.
the case where M is an open subset of R
n
. In section 2.3 we will
treat the case of a general manifold.
We conclude this section with some denitions that will be
used in the following sections.
Denition 2.1.4 (Non-degenerate Lagrangian). If L is a C
2
Lag-
rangian on the manifold M, we say that L is non-degenerate if for
each (x, v) TM the second partial derivative
2
L/v
2
(x, v) is
non-degenerate as a quadratic form.
Notice that the second partial derivative
2
L/v
2
(x, v) makes
sense. In fact, this is the second derivative of the restriction of L to
39
the vector space T
x
M, and denes therefore a quadratic form on
T
x
M. In the same way, the rst derivative L/v(x, v) is a linear
form on T
x
M, and therefore L/v(x, v) (T
x
M)

= T

x
M.
Denition 2.1.5 (Global Legendre Transform). If L is a C
1
Lag-
rangian on the manifold M, we dene the global Legendre trans-
form

L : TM T

M associated to L by

L(x, v) = (x,
L
v
(x, v)).
Of course, if L is C
r
, then

L is C
r1
.
Proposition 2.1.6. If L is a C
r
Lagrangian, with r 2, on the
manifold M, then the following statements are equivalent
(1) the Lagrangian L is non-degenerate;
(2) the global Legendre transform

L : TM T

M is a C
r1
local dieomorphism;
(3) the global Legendre transform

L : TM T

M is a C
r1
local dieomorphism.
Moreover, the following statements are equivalent
(i) the Lagrangian L is non-degenerate, and

L is injective;
(ii) the global Legendre transform

L : TM T

M is a (global)
C
r1
dieomorphism onto its image;
(iii) the global Legendre transform

L : TM T

M is a (global)
C
r1
dieomorphism onto its image.
Proof. Statements (1), (2), and (3) above are local in nature, it
suces to prove them when M is an open subset of R
n
.We use the
canonical coordinates on M R
n
, TM = M R
n
, and T

M =
M R
n
. In these coordinates , at the point (x, v) TM, the
derivative D

L(x, v) : R
n
R
n
R
n
R
n
of the global Legendre
transform

L has the following matrix form
D

L(x, v) =
_
Id
R
n

2
L
xv
(x, v)
0

2
L
v
2
(x, v)
_
40
therefore D

L(x, v) is invertible as a linear function R
n
R
n

R
n
R
n
if and only if

2
L
v
2
(x, v) is non-degenerate as a quadratic
form. The equivalence of (1), (2), and (3) (resp. (i), (ii), and (iii))
is now a consequence of the inverse function theorem.
Finally a last denition for this section.
Denition 2.1.7 (C
r
Variation of a Curve). Let M be an ar-
bitrary dierentiable manifold. Let us consider a C
r
curve :
[a, b] M. A variation of class C
r
of is a map : [a, b]], [
M of class C
r
, where > 0, such that (t, 0) = (t), for all
t [a, b]. For such a variation, we will denote by
s
the curve
t (t, s) which is also of class C
r
.
2.2 Lagrangians on Open Subsets of R
n
We suppose that M is an open subset contained in R
n
. In that
case TM = M R
n
, and the canonical projection : TM M
is the projection on the rst factor.
We study the dierentiability properties of L, for this we have
to assume that L is C
1
.
Lemma 2.2.1. Suppose that L is a C
1
Lagrangian the open subset
M of R
n
. Let ,
1
: [a, b] R
n
be two continuous piecewise C
1
curves, with ([a, b]) M. The function L( + t
1
) is dened
for t small. It has a derivative at t = 0, which is given by
d
dt
L( +t
1
)
|t=0
=
_
b
a
DL[(s), (s)](
1
(s),
1
(s)) ds
=
_
b
a
_
L
x
[(s), (s)](
1
(s)) +
L
v
[(s), (s)](
1
(s))
_
ds.
Proof. Both , and
1
are continuous, hence the map : [a, b]
R R
n
(s, t) (s) + t
1
(s) is continuous, and therefore uni-
formly continuous on [a, b] [1, 1]. Since (s, 0) = (s) is in the
open subset M, for every s [a, b], we conclude that there exists
> 0 such that ([a, b] [, ]) M. Therefore the action of the
curve (, t) = +t
1
is dened for every t [, ].
41
Pick F a nite subset of [a, b] such that both , and
1
are
dierentiable at each point of [a, b] F. The function : ([a, b]
F) R dened by
(s, t) = L((s) +t
1
(s), (s) +t
1
(s)),
has a partial derivative with respect to t given by

t
(t, s) = DL[(s) +t
1
(s), (s) +t
1
(s)](
1
(s),
1
(s)).
Moreover, this partial derivative is uniformly bounded on ([a, b]
F) [1, 1], because DL is continuous, and the curves ,
1
are
continuous, and piecewise C
1
. Therefore we can dierentiate L(+
t
1
) =
_
b
a
L((s) (s)) dt under the integral sign to obtain the de-
sired result.
Exercise 2.2.2. Suppose that L is a C
1
Lagrangian on the open
subset M of R
n
. If : [a, b] M is a Lipschitz curve, then (s)
exists almost everywhere. Show that the almost everywhere dened
function s L((s), (s)) is integrable. If
1
: [a, b] M is also
Lipschitz, show that L( + t
1
) is well dened for t small, nite,
and dierentiable.
Denition 2.2.3 (Extremal Curve). An extremal curve for the
Lagrangian L is a continuous piecewise C
1
curve : [a, b] M
such that
d
dt
L( +t
1
)
t=0
= 0, for every C

curve
1
: [a, b] R
n
satisfying
1
= 0 in the neighborhood of a and b.
By lemma 2.2.1, it is equivalent to say that
_
b
a
_
L
x
((s), (s))(
1
(s)) +
L
v
((s), (s))(
1
(s))
_
ds = 0,
for each curve
1
: [a, b] M of class C

which satises
1
= 0
in the neighborhood of a and b.
Remark 2.2.4. If is an extremal curve, then for all a

, b

[a, b],
with a

< b

, the restriction [[a

, b

] is also an extremal curve.


The relationship between minimizers and extremal curves is
given by the following proposition.
42
Proposition 2.2.5. If L is a Lagrangian on M, and : [a, b] M
is a C
r
curve, with r 1 (resp. continuous piecewise C
1
) curve,
which minimizes the action on the set of C
r
(resp. continuous
piecewise C
1
), then is an extremal curve for L.
Proposition 2.2.6 (Euler-Lagrange). Let us assume that L is
a Lagrangian is of class C
2
on the open subset M of R
n
. If :
[a, b] M is a curve of class C
2
, then is an extremal curve if
and only if it satises the Euler-Lagrange equation
d
dt
L
v
((t), (t)) =
L
x
((t), (t)), (E-L)
for all t [a, b].
Proof. Since L and are both C
2
, if
1
: [a, b] M is C

and
vanishes in the neighborhood of a and b, then the map
t
L
v
[(t), (t)](
1
(t))
is C
1
and is 0 at a and b. It follows that
_
b
a
d
dt
_
L
v
[(t), (t)](
1
(t))
_
dt = 0,
which implies
_
b
a
L
v
[(t), (t)](
1
(t)) dt =
_
b
a
d
dt
_
L
v
((t), (t))
_
(
1
(t)) ds.
We thus obtain that is an extremal curve if and only if
_
b
a
_
L
x
((t), (t))
d
dt
_
L
v
((t), (t))
__
(
1
(t)) ds = 0,
for every C

curve
1
: [a, b] M satisfying
1
= 0 in the neigh-
borhood of a and b. It is then enough to apply the following
lemma:
Lemma 2.2.7 (Dubois-Raymond). Let A : [a, b] L(R
n
, R) =
R
n
be a continuous map such that
_
b
a
A(t)(
1
(t))dt = 0, for each
C

curve
1
: [a, b] R
n
which vanishes in the neighborhood of
a and b, then A(t) = 0, for all t [a, b].
43
Proof. Suppose that there exists t
0
]a, b[ and v
0
R
n
such that
A(t
0
)(v
0
) ,= 0. Replacing v
0
by v
0
if necessary, we can sup-
pose that A(t
0
)(v
0
) > 0. We x > 0 small enough so that
A(t)(v
0
) > 0, for all t [t
0
, t
0
+ ] ]a, b[. We then choose
C

curve : [a, b] [0, 1] with = 0 outside of the inter-


val [t
0
, t
0
+ ] and (t
0
) = 1. Of course
_
b
a
A(t)((t)v
0
)dt =
0, but
_
b
a
A(t)((t)v
0
)dt =
_
t
0
+
t
0

(t)A(t)(v
0
) dt and the function
(t)A(t)(v
0
) is continuous, non-negative on [t
0
, t
0
+ ] and
(t
0
)A(t
0
)(v
0
) > 0, since (t
0
) = 1, hence its integral cannot
vanish. This is a contradiction.
In the remainder of this section, we show that, under natural
assumptions on the Lagrangian L, the extremal curves which are
C
1
or even continuous piecewise C
1
are necessarily of class C
2
, and
must thus verify the Euler-Lagrange equation.
Lemma 2.2.8. Let L be a Lagrangian on the open subset M of
R
n
, and let : [a, b] M be an extremal curve of class C
1
for L,
then there exists p R
n
such that
t [a, b],
L
v
((t), (t)) = p +
_
t
a
L
x
((s), (s)) ds.
Proof. If
1
: [a, b] M is C

and vanishes in the neighborhood


of a and b, then the map
t
__
t
a
L
x
((s), (s)) ds
_
(
1
(t))
is C
1
and is 0 at a and b. It follows that
_
b
a
d
dt
___
t
a
L
x
((s), (s)) ds
_
(
1
(t))
_
dt = 0,
which implies
_
b
a
L
x
((t), (t))(
1
(t)) dt =
_
b
a
__
t
a
L
x
((s), (s)) ds
_
(
1
(t)) dt.
Thus the condition
_
b
a
_
L
x
((t), (t))(
1
(t)) +
L
v
((t), (t))(
1
(t))
_
dt = 0
44
is equivalent to
_
b
a
_
L
v
((t), (t)]
_
t
a
L
x
((s), (s)) ds
_
(
1
(t)) dt = 0.
It is then enough to apply the following lemma:
Lemma 2.2.9 (Erdmann). If A : [a, b] R
n
is a continuous
function such that
_
b
a
A(t)(
1
(t)) dt = 0, for every curve
1
:
[a, b] R
n
of C

class and vanishing in the neighborhood of


a and b, then, the function A(t) is constant.
Proof. Let us choose
0
: [a, b] R of class C

with
_
b
a

0
(t)dt =
1 and
0
= 0 in a neighborhood of a and b. Let
1
: [a, b] R
n
be a C

curve which is 0 in the neighborhood of a and b, then, if


we set C =
_
b
a

1
(s) ds, the curve
1
: [a, b] R
n
, dened by

1
(t) =
_
t
a

1
(s) C
0
(s) ds
is C

and is 0 in a neighborhood of a and b. Therefore, since



1
(t) =
1
(t) C
0
(t), we have
_
b
a
A(t)[
1
(t) C
0
(t)] dt = 0,
consequently
_
b
a
A(t)[
1
(t)]dt
_
b
a
[
0
(t)A(t)](C)dt = 0. ()
If we dene p =
_
b
a

0
(t)A(t)dt R
n
, then
_
b
a
[
0
(t)A(t))](C) dt is
nothing but p(C). On the other hand by the denition of C and
the linearity of p, we have that p(C) =
_
b
a
p(
1
(t))dt. We then can
rewrite the equation () as
_
b
a
[A(t) p](
1
(t)) dt = 0.
Since
1
: [a, b] R
n
is a map which is subject only to the two
conditions of being C

and equal to 0 in a neighborhood of a and


b, Dubois-Raymonds Lemma 2.2.7 shows that A(t) p = 0, for
all t [a, b].
45
Remark 2.2.10. Of course, the proofs of both the Dubois-Raymond
and the Erdmann lemmas are reminiscent of now classical proofs
of analogous statements in Laurent Schwartzs Theory of Distri-
butions, but these statements are much older.
Corollary 2.2.11. If L is a non-degenerate, C
r
Lagrangian, with
r 2, on the open subset M of R
n
, then every extremal C
1
curve
is C
r
.
Proof. Let : [a, b] M be a C
1
extremal curve. Let us x
t
0
and consider (x
0
, v
0
) = ((t
0
), (t
0
)) TM. From proposi-
tion 2.1.6, the Legendre transform

L : (x, v) (x,
L
v
(x, v)) is
a local dieomorphism Let us call / a local inverse of

L with
/(x
0
,
L
v
(x
0
, v
0
)) = (x
0
, v
0
). The map / is of class C
r1
. By
continuity of and , for t near to t
0
, we have
((t), (t)) = /
_
(t),
L
v
((t), (t))

. ()
But, by lemma 2.2.8, we have
L
v
((t), (t)) = p +
_
t
a
L
x
((s), (s)) ds.
It is clear that the right-hand side of this equality is of class C
1
.
Referring to () above, as / is C
r1
, we see that ((t), (t)) is also
of class C
1
, for t near to t
0
. We therefore conclude that is C
2
.
By induction, using again (), we see that is C
r
.
Corollary 2.2.12. Suppose that the Lagrangian L on M is of
class C
r
, r 2, and that its global Legendre transform

L : TM
T

M is a dieomorphism on its image



L(TM)), then every con-
tinuous piecewise C
1
extremal curve of L is in fact C
r
, and must
therefore satisfy the Euler-Lagrange equation.
Proof. The assumption that

L is a dieomorphism implies by
proposition 2.1.6 that L is non-degenerate. Therefore by corollary
2.2.11, we already know that the extremal C
1
curves are all C
r
.
Let : [a, b] M be a continuous piecewise C
1
extremal curve.
Let us consider a nite subdivision a
0
= a < a
1
< < a
n
= b
such that the restriction [[a
i
, a
i+1
] is C
1
. Since [[a
i
, a
i+1
] is also
an extremal curve, we already know that [[a
i
, a
i+1
] is C
r
. Using
46
that is an extremal curve, for every C

curve
1
: [a, b] M
which is equal to 0 in a neighborhood of a and b, we have
_
b
a
L
x
[(t), (t)](
1
(t)) +
L
v
[(t), (t)](
1
(t)) dt = 0. ()
Since [[a
i
, a
i+1
] is of class at least C
2
, we can integrate by parts
to obtain
_
a
i+1
a
i
L
v
[(t), (t)](
1
(t))dt =
L
v
[(a
i+1
),

(a
i+1
)](
1
(a
i+1
))

L
v
[(a
i
),
+
(a
i
)](
1
(a
i
))
_
a
i+1
a
i
_
d
dt
_
L
v
[(s), (s)]
__
(
1
(s)) ds,
where

(t) is the left derivative and


+
(t) is the right derivative
of at t [a, b]. Using this, we conclude that
_
a
i+1
a
i
L
x
[(t), (t)](
1
(t)) +
L
v
[(t), (t)](
1
(t)) dt =
_
a
i+1
a
i
L
x
[(t), (t)](
1
(t))
_
d
dt
_
L
v
[(s), (s)]
__
(
1
(s)) ds
+
L
v
[(a
i+1
),

(a
i+1
)](
1
(a
i
))
L
v
[(a
i
),
+
(a
i
)](
1
(a
i
))
=
L
v
[(a
i+1
),

(a
i+1
)](
1
(a
i
))
L
v
[(a
i
),
+
(a
i
)](
1
(a
i
))
where the last equality holds, because [[a
i
, a
i+1
] is a C
2
extremal
curve, and therefore must satisfy the Euler-Lagrange equation (E-
L) on the interval [[a
i
, a
i+1
]. Summing on i, and using (), we
get that
n1

i=1
_
L
v
[(a
i+1
),

(a
i+1
)](
1
(a
i
))
L
v
[(a
i
),
+
(a
i
)](
1
(a
i
))
_
= 0,
for every C

curve
1
: [a, b] M which vanishes in a neighbor-
hood of a and b. For 1 i n 1, we can choose the C

curve

1
, vanishing in a neighborhood of the union of the two intervals
[a, a
i1
] and [a
i+1
, b], and taking at a
i
an arbitrary value xed in
advance. This implies that
i = 1, . . . , N,
L
v
((a
i
),

(a
i
)) =
L
v
((a
i
),
+
(a
i
)).
47
The injectivity of the Legendre transform gives

(a
i
) =
+
(a
i
).
Hence, the curve is in fact of class C
1
on [a, b] and, consequently,
it is also of class C
r
.
The proof of following lemma is essentially the same as that of
lemma 2.2.1.
Lemma 2.2.13. If is a C
2
variation of the C
2
curve : [a, b]
M with values in the open subset M of R
n
, then the map s
L(
s
) is dierentiable and its derivative in 0 is
d
ds
L(
s
)
s=0
=
_
b
a
DL[((t), (t))]
_

s
(t, 0),

2

st
(t, 0)
_
dt.
We now obtain a characterization of extremal curves that does
not use the fact that M is contained in an open subset of an
Euclidean space.
Lemma 2.2.14. A C
2
curve : [a, b] M is an extremal curve
of the Lagrangian L if and only if
d
ds
L(
s
)
s=0
= 0, for any C
2
variation of such that (a, s) = (a), (b, s) = (b) for s in a
neighborhood of 0.
Proof. The variations of the type (t) +b
1
(t) with
1
of C

class
are particular variations of class C
2
.
It thus remains to be seen that, if is a C
2
extremal curve,
then we have
d
ds
L(
s
)
s=0
= 0 for variations of class C
2
of such
that (a, s) = (a) and (b, s) = (b). That results from the
following theorem.
Theorem 2.2.15 (First Variation Formula). If : [a, b] M is
a C
2
extremal curve, then for any variation of class C
2
of , we
have
d
ds
L(
s
)
s=0
=
L
v
[(b), (b)]
_

s
(b, 0)
_

L
v
[(a), (a)]
_

s
(a, 0)
_
.
Proof. We have
d
ds
L(
s
)
s=0
=
_
b
a
DL[(t), (t)]
_

s
(t, 0),

2

ts
(t, 0)
_
dt
=
_
b
a
_
L
x
[(t), (t)]
_

s
(t, 0)
_
+
L
v
[(t), (t)]
_

2

ts
(t, 0)
_
_
dt.
48
As is a C
2
extremal curve, it satises the Euler-Lagrange equa-
tion
L
x
[(t), (t)] =
d
dt
_
L
v
((t), (t))
_
,
plugging in, we nd
d
ds
L(
s
)
s=0
=
_
b
a
_
d
dt
_
L
v
_
((t), (t))
_
_

s
(t, 0)
_
+
L
v
[(t), (t)]
_

2

ts
(t, 0)
_
dt.
However the quantity under the last integral is nothing but the
derivative of the function t
L
v
[(t), (t)](

s
(t, 0)) which is of
class C
1
thus
d
ds
L(
s
)
s=0
=
L
v
[(b), (b)]
_

s
(b, 0)
_

L
v
[(a), (a)]
_

s
(a, 0)
_
.
2.3 Lagrangians on Manifolds
We consider an arbitrary C

manifold M endowed with a Lag-


rangian L of class C
r
, with r 2.
Lemma 2.3.1. Consider : [a, b] [c, d] M of class C
2
. Dene

s
: [a, b] M by
s
(t) = (t, s). The map s L(
s
) is C
1
.
Proof. To simplify notation we assume that 0 [c, d] and we show
that s L(
s
) is C
1
on some interval [, ], with > 0. We can
cover the compact set ([a, b]0) by a nite family of coordinate
charts. We then nd a subdivision a
0
= a < a
1
< < a
n
= b
such that ([a
i
, a
i+1
] 0)) is contained in a U
i
the domain of
denition of one of these charts. By compactness, for small
enough, we have ([a
i
, a
i+1
] [, ]) U
i
, for i = 0, . . . , n 1.
Transporting the situation via the chart to an open set in R
n
, we
can apply 2.2.13 to obtain that s
_
a
i+1
a
i
L[(t, s),

t
(t, s)] dt is
C
1
on some interval [, ]. It is now enough to notice that
L(
s
) =
n1

i=0
_
a
i+1
a
i
L
_
(s, t),

t
(s, t)

dt,
to be able to nish the proof.
49
We can then introduce the concept of an extremal curve for
the C
2
Lagrangian L in the case of C
2
curves : [a, b] M with
values in arbitrary manifold M.
Denition 2.3.2 (Extremal C
2
Curve). A C
2
curve : [a, b]
M is an extremal curve for the C
2
Lagrangian L, if for each C
2
variation : [a, b]] , [ M of , with (t, s) = (t) in the
neighborhood of (a, 0) and (b, 0), we have
d
ds
L(
s
)
s=0
= 0.
Remark 2.3.3. By lemma 2.2.14, if the curve and the Lagrangian
are of class C
2
, this denition coincides with the denition given
for the case where the manifold is an open subset of R
n
.
Lemma 2.3.4. If : [a, b] M is a C
2
extremal curve and
[a

, b

] [a, b] then, the restriction [[a

, b

] is also an extremal
curve
Proof. For any C
2
variation : [a

, b

]] , [ M of [[a

, b

],
with (t, s) = (t) in the neighborhood of (a

, 0) and (b

, 0), we
nd

with 0 <

, and > 0 such that with (t, s) = (t) for


every (t, s) ([a

, a

+] [b

, b

]) [

]. We can therefore
extend [[a

, b

] [

] to

[a, b] [

] by (t, s) = (t), for


t / [a

, b

] [

]. It is clear that

is a C
2
variation, with
(t, s) = (t) in the neighborhood of (a, 0) and (b, 0). Moreover,
for s [

], the dierence L(

s
)L(
s
) is equal to L([[a, a

])+
L([[b

, b]).
Theorem 2.3.5 (Euler-Lagrange). Suppose L is a C
2
Lagrangian
on the manifold M. Let : [a, b] M, be a C
2
curve. If
is extremal, then, for each subinterval [a

, b

] [a, b] such that


([a

, b

]) is contained in a domain U of a coordinate chart, the


restriction [[a

, b

] satises (in coordinates) the Euler-Lagrange


equation.
Conversely, if for every t
0
[a, b], we can nd an > 0 and a
domain U of a coordinate chart such that ([t
0
, t
0
+] [a, b])
U and [[t
0
, t
0
+][a, b] satises in the chart the Euler-Lagrange
equation, then the curve is an extremal curve.
Proof. If is an extremal curve, then [[a

, b

] is also an extremal
curve. Since ([a

, b

]) U, where U is a domain of a coordi-


nate chart, we can then transport, via the coordinate chart, the
50
situation to an open subset of R
n
hence [[a

, b

] must verify the


Euler-Lagrange equation.
To prove the second part, we remark that by compactness s,
we can nd a subdivision a
0
= a < a
1
< < a
n
= b, and a
sequence U
0
, . . . , U
n1
of domains of coordinate charts such that
([a
i
, a
i+1
]) U
i
. If is a variation of class C
2
of , we can nd
> 0 such that ([a
i
, a
i+1
] [, ]) U
i
, i = 1, , n). The
rst variation formula 2.2.15 shows that
d
ds
L(
s
[[a
i
, a
i+1
])
s=0
=
L
v
[(a
i+1
), (a
i+1
)]
_

s
(a
i+1
)
_

L
v
[(a
i
), (a
i
)]
_

s
(a
i
)
_
.
Adding these equalities, we nd
d
ds
L(
s
)
s=0
=
L
v
[(b), (b)]
_

s
(b, 0)
_

L
v
[(a), (a)]
_

s
(a, 0)
_
.
If (a, s) = (a) and (s, b) = (b) in a neighborhood of s = 0,
we get that both

s
(a, 0), and

s
(b, 0) are equal to 0,therefore the
second member of the equality above is 0.
The previous proof also shows that the rst variation formula
is valid in the case of arbitrary manifolds.
Theorem 2.3.6 (First Variation Formula). Let L be a C
2
Lag-
rangian on the manifold M. If : [a, b] M is a C
2
extremal
curve, for each C
2
variation : [a, b]] , [ M, (t, s) (t, s)
of , we have
d
ds
L(
s
)
s=0
=
L
v
[ ((b), (b)]
_

s
(b, 0)
_

L
v
[(a), (a)]
_

s
(a, 0)
_
.
By the same type chart by chart argument, using proposition
2.2.12, we can show the following proposition.
Proposition 2.3.7. Suppose the C
r
Lagrangian L, with r 2,
on the manifold M is such that its global Legendre transform

L :
TM T

M is a dieomorphism onto its image. If : [a, b] M


is a C
k
curve, with k 1, (resp. a continuous piecewise C
1
curve)
which is a minimizer for the class of C
k
curves (resp. of continuous
piecewise C
1
curves), then is an extremal of class at least C
r
.
51
2.4 The Euler-Lagrange Equation and its
Flow
We will rst consider an open subset M of R
n
and a non-degenerate
C
r
Lagrangian L : TM R, with r 2. We will assume that
for each (x, v) TM the bilinear form
2
L/v
2
(x, v) is non-
degenerate. It follows from 2.2.11 that every C
1
(locally) extremal
curve : [a, b] M is necessarily of class C
2
, and satises the
Euler-Lagrange equation
L
x
((t), (t)) =
d
dt
L
v
((t), (t)),
and hence by dierentiation we obtain

2
L
v
2
[(t), (t)]( (t), )=
L
x
[(t), (t)]()

2
L
xv
[(t), (t)]( (t), ),
where this is to be understood as an equality between elements
of R
n
. Since
2
L/v
2
(x, v) is non-degenerate, we can in fact
solve for (t), and therefore we see that satises a second order
dierential equation. This suggests to dene a vector eld X
L
on
TM = M R
n
by
X
L
(x, v) = (v,

X
L
(x, v)) T
(x,v)
(TM),
where, due to the non-degeneracy of
2
L/v
2
(x, v) , the function

X
L
is uniquely dened by

2
L
v
2
(x, v)[

X
L
(x, v), ] =
L
x
(x, v)()

2
L
xv
(x, v)(v, ).
This function

X
L
is C
r2
, if L is C
r
.
From our previous computation, if is a curve satisfying the
Euler-Lagrange equation, its speed curve t ((t), (t)) is an
integral curve of X
L
. Conversely, since the rst of the coordinates
of X
L
(x, v), (v,

X
L
(x, v)) is v, the solutions of this vector eld are
curves of the form t ((t), (t)) with : [a, b] M of class
C
2
, and (t) =

X
L
((t), (t)), and therefore satises the Euler-
Lagrange equation. Thus, these integral curves are the curves of
the form t ((t), (t)) with : [a, b] M of class C
2
which
satisfy the Euler-Lagrange equation, in other words with an
extremal curve.
52
Theorem 2.4.1. Let L be a C
r
Lagrangian on M, with r 2.
Assume that for each (x, v) TM the bilinear form
2
L/v
2
(x, v)
is non-degenerate. Then for every (x, v) TM, we can nd an
extremal : [, ] M with (0) = x, and (0) = v; moreover,
if L is at least C
3
, then any two such extremals coincide on their
common domain of denition.
Proof. Suppose (x, v) TM is given. Since X
L
is at least con-
tinuous, we can apply the Cauchy-Peano theorem, see [Bou76], to
nd an integral curve of X
L
dened on some interval [, ],
with > 0 and passing through (x, v) at time t = 0. But as we
have seen above such a solution if of the form (t) = ((t), (t)),
with an extremal. This is obviously the required extremal.
If L is C
3
, then X
L
is C
1
, and we therefore have uniqueness
of solution by the Cauchy-Lipschitz theorem. Therefore, if
1
:
[[

] M is another extremal with


1
(0) = x, and
1
(0) = v,
then
1
(t) = (
1
(t),
1
(t)) is another solution of X
L
with the same
initial condition as , therefore =

on the intersection of their


domain of denition.
We can, then, summarize what we obtained in the following
theorem.
Theorem 2.4.2 (Euler-Lagrange). Let M be an open subset of
R
n
. If L is a Lagrangian on M of class C
r
, with r 2, and for every
(x, v) TM the quadratic form

2
L
v
2
(x, v)) is non-degenerate, then
there exists one and only one vector eld X
L
on TM such that the
solutions of X
L
are precisely the curves the form t ((t), (t))
where : [a, b] M is an extremal curve of L. This vector eld
is of class C
r2
. The vector eld X
L
is called the Euler-Lagrange
vector eld of the Lagrangian L.
Everything in this theorem was proved above, but maybe we
should say a word about the uniqueness. In fact, we can obtain
X
L
(x, v) from the extremals of L. In fact, if we choose an extremal
: [, ] M with (0) = x, and (0) = v (this is possible by
theorem 2.4.1), then X
L
(x, v) is nothing but the speed at 0 of the
curve t ((t), (t)).
Let us extend this result to the case of an arbitrary manifold
M.
53
Theorem 2.4.3 (Euler-Lagrange). Let M be a dierentiable man-
ifold. If L : TM R is a C
r
Lagrangian, with r 2, and for
every (x, v)
inTM the quadratic form

2
L
v
2
(x, v)) is non-degenerate, then there
exists one and only one vector eld X
L
on TM such that the
solutions of X
L
are precisely of the form t ((t), (t)) where
: [a, b] M is an extremal curve of L. This vector eld is of
class C
r2
.
Proof. By theorem 2.4.2 above, for every open subset U M
which is contained in the domain of a coordinate chart, we can
nd such a vector eld X
U
L
on TU for which the solutions are
precisely of the form t ((t), (t)) where : [a, b] U with
an extremal curve having values in U. But if U and V are
two such open subsets, for both restrictions X
U
L
[U V , X
U
L
[U V ,
the solution curves are the curves t ((t), (t)) where is an
extremal curve of L whose image is contained in U V , so they
both coincide with X
UV
L
.
Denition 2.4.4 (The Euler-Lagrange Vector Field and its Flow).
If L : TM R is a C
r
Lagrangian, with r 2, and for every
(x, v) TM the quadratic form

2
L
v
2
(x, v)) is non-degenerate, the
vector eld X
L
dened by theorem 2.4.3 above is called the Euler-
Lagrange vector eld of the Lagrangian L. If L is C
r
with r 3,
then, by the Cauchy-Lipschitz theorem, the eld X
L
generates a
partial ow on TM of class C
r2
. We will denote this partial ow
by
L
t
and we will call it the Euler-Lagrange ow of the Lagrangian
L.
In fact, under the stronger hypothesis that the global Legendre
transform

L : TM T

M, dened by

L(x, v)(x, v) =
_
x,
L
v
(x, v)
_
,
is a C
1
dieomorphism onto its image, we will see in Theorem 2.6.5
that the partial ow is of class C
r1
and that, even if L is only of
class C
2
, the vector eld X
L
is uniquely integrable and generates
a partial ow
L
t
of class C
1
which is also called in that case the
Euler-Lagrange ow of the Lagrangian L.
54
2.5 Symplectic Aspects
Let M be dierentiable manifold of class C

, denote by

:
T

M M the canonical projection of the cotangent space T

M
onto M and denote by T

: TT

M TM the derivative of

.
On T

M, we can dene a canonical dierential 1-form called the


Liouville form. Thus the value this form at a given (x, p) T

M
is a linear map
(x,p)
: T
(x,p)
(T

M) R. To dene it, we just


remark that the two linear maps T
(x,p)

: T
(x,p)
(T

M) T
x
M,
and p : T
x
M R can be composed, and thus we can dene
(x,p)
by
W T
(x,p)
(T

M),
(x,p)
(W) = p[T
(x,p)

(W)].
To understand this dierential 1-form on the T

M manifold, we
take a chart : U (U) R
n
, we can consider the associated
chart T

: T

U T

((U)) = (U) R
n
. In these charts the
canonical projection

is nothing but the projection (U)R


n

(U) on the rst factor. This gives us coordinates (x


1
, . . . , x
n
) on
U, and therefore coordinates (x
1
, , x
n
, p
1
, , p
n
) on T

U such
that the projection

is nothing but (x
1
, , x
n
, p
1
, , p
n
)
(x
1
, , x
n
). A vector W T
(x,p)
(T

M) has therefore coordinates


(X
1
, , X
n
, P
1
, , P
n
), and the coordinates of T
(x,p)

(W)
T
x
U are (X
1
, , X
n
). It follows that
(x,p)
(W) =

n
i=1
p
i
X
i
.
Since X
i
is nothing but the dierential form dx
i
evaluated on W,
we get that
[T

U =
n

i=1
p
i
dx
i
.
We therefore conclude that is of class C

.
Let be a dierential 1-form dened on the open subset U of
M. This 1-form is a section : U T

U, x
x
. The graph of
is the set
Graph() = (x,
x
) [ x U T

M.
Lemma 2.5.1. Let be a dierential 1-form dened on the open
subset U of M. We have

= ,
where

is the pull-back of the Liouville form on T

M by the
map : U T

U
55
Proof. Using coordinates charts, it suces to verify in the case
where U is an open subset of R
n
. Using the canonical coordinates
on U R
n
, we can write
x
=

n
i=1

i
(x)dx
i
. As a map : U
T

U it is thus given in these coordinates by


(x
1
, . . . , x
n
) (x
1
, . . . , x
n
,
1
(x), . . . ,
n
(x)).
But in these coordinates, it is clear that the pull-back

is

n
i=1

i
(x)dx
i
= .
By taking, the exterior derivative of we can dene a sym-
plectic structure on T

M. To explain what that means, let us


recall that a symplectic form on a vector space E is an alternate
(or antisymmetric) bilinear form a : E E R which is non-
degenerate as a bilinear form, i.e. the map a

: E E

, x
a(x, ) is an isomorphism.
Lemma 2.5.2. If the nite dimensional vector space E admits a
symplectic form, then its dimension is even.
Proof. We choose a basis on E. If A is the matrix of a in this
base, its transpose
t
A is equal to A (this reects the antisymme-
try). Therefore taking determinants, we get det(A) = det(
t
A) =
det(A) = 1
dimE
det(A). The matrix of a

: E E

, using
on E

the dual basis, is also A. the non degeneracy of a

gives
det(A) ,= 0. It follows that 1
dimE
= 1, and therefore dimE is
even.
Denition 2.5.3 (Symplectic Structure). A symplectic structure
on a C

dierentiable manifold V is a C

closed dierential 2-
form on V such that, for each x V , the bilinear form
x
:
T
x
V T
x
V R is a symplectic form on the vector space T
x
V .
As an exterior derivative is closed, to check that = d is a
symplectic form on T

M, it is enough to check the non-degeneracy


condition. We have to do it only in an open subset U of R
n
, with
the notations introduced higher, we see that
= d =
n

i=1
dp
i
dx
i
=
n

i=1
dx
i
dp
i
,
56
it is, then, easy to check the non-degeneracy condition. In fact,
using coordinates we can write a W T
(x,p)
(T

M) as
W =
n

i=1
X
i

x
i
+
n

i=1
P
i

p
i
,
Therefore
(x,p)
(W, ) =

n
i=1
X
i
dp
i

n
i=1
P
i
dx
i
, and
(x,p)
(W, ) =
0 implies X
i
= P
i
= 0, for i = 1, . . . , n, by the independence of
the family (dx
1
, . . . , dx
n
, dp
1
, . . . , dp
n
).
In the following, we will suppose that V is a manifold provided
with a symplectic structure . If H is a C
r
function dened on
the open subset O of V , By the fact that

x
is an isomorphism,
we can associate to H a vector eld X
H
on O well dened by

x
(X
H
(x), ) = d
x
H().
Since is non-degenerate, the vector eld X
H
is as smooth as
the derivative of H, therefore it is C
r1
. In particular, if H is
C
2
, then the solutions of the vector eld X
H
dene a partial ow

H
t
: O O.
Denition 2.5.4 (Hamiltonian Flow). Suppose H : O R is
a C
1
function, dened on the open subset O of the symplectic
manifold V . The Hamiltonian vector of H is the vector eld X
H
on O, uniquely dened by

x
(X
H
(x), ) = d
x
H(),
where is the symplectic form on V . If, moreover, the function
H is C
r
, the vector X
H
eld is C
r1
. Therefore for r 2, the
partial ow
H
t
generated by H exists, it is called the Hamitonian
ow of H, and H is called the Hamiltonian of the ow
H
t
.
Lemma 2.5.5. H : O R is a C
2
function, dened on the open
subset O of the symplectic manifold V , then H is constant on the
orbits of its Hamiltonian ow
H
t
.
Proof. We must check that d
x
H(X
H
(x)) is 0, for all x O. But
d
x
H(X
H
(x)) is
x
(X
H
(x), X
H
(x)) vanishes because the bilinear
form
x
is alternate.
57
Denition 2.5.6 (Lagrangian Subspace). In a vector space E en-
dowed with a symplectic bilinear forma : EE R, a Lagrangian
subspace is a vector subspace F of E with dimE = 2 dimF and
a is identically 0 on F F.
Lemma 2.5.7. Let F be a subspace of the vector space E which
is Lagrangian for the symplectic form a on E. If x E is such
that a(x, y) = 0, for all y F, then, the vector x is itself in F.
Proof. Dene F

= x E [ y F, a(x, y) = 0. We have F

F, since a is 0 on F F. Since a

: E E

is an isomorphism, the
dimension of F

is the same as that of its image a

(F

) = p
E

[ p[F = 0. This last subspace can be identied with the dual


(E/F)

of the quotient of E by F. Therefore dimF

= dimE
dimF = 2 dimF dimF = dimF. Therefore F

= F.
Denition 2.5.8 (Lagrangian Submanifold). If V is a symplectic
manifold, a Lagrangian submanifold of V is a submanifold N of
class at least C
1
, and such that the subspace T
x
N of T
x
V is, for
each x N, a Lagrangian subspace for the symplectic bilinear
form
x
.
By the lemma 2.5.7 above, if x N, any vector v T
x
V such
that
x
(v, v

) = 0, for all v

T
x
N, is necessarily in T
x
N.
Lemma 2.5.9. If is a C
1
dierential 1-form on the manifold
M, then the graph Graph() of is a Lagrangian submanifold of
T

M if and only if is a closed form.


Proof. Indeed,by lemma 2.5.1, we have =

, and thus also


d =

d =

. However, the form regarded as map


of M T

M induces a dieomorphism of C
1
class of M on
Graph(), consequently d = 0 if and only if [ Graph() = 0.
Theorem 2.5.10 (Hamilton-Jacobi). Let H : O R be a C
2
function dened on the open subset O of the symplectic manifold
V . If N O is a connected C
1
Lagrangian submanifold of V , it is
locally invariant by the partial ow
H
t
if and only if H is constant
on N.
58
Proof. If H is constant on N, we have

x
N, d
x
H[T
x
N = 0,
and thus
x
(X
H
(x), v) = 0, for all v T
x
N, which implies that
X
H
(x) T
x
N. By the theorem of Cauchy-Peano [Bou76], if N is
of class C
1
(or Cauchy-Lipschitz, if N is of C
2
class), the restric-
tion X
H
[N has solutions with values in N. By uniqueness of the
solutions of X
H
in O (which holds because X
H
is C
1
on O), the
solutions with values in N must be orbits of
H
t
. We therefore
conclude that N is invariant by
H
t
as soon as H is constant on
N. Conversely, if N is invariant by
H
t
, the curves t
H
t
(x)
have a speed X
h
(x) for t = 0 which must be in T
x
N, therefore
X
h
(x) T
x
N and d
x
H[N = (X
H
(x), ) vanishes at every point
of N, since N is a Lagrangian submanifold. By connectedness of
N, the restriction H[N is constant.
2.6 Lagrangian and Hamiltonians
Denition 2.6.1 (Hamiltonian). If L is a C
1
Lagrangian on the
manifold M, its Hamiltonian

H : TM R is the function dened
by

H(x, v) =
L
v
(x, v)(v) L(x, v).
Obviously, if L is C
k
, with k 1, then, its associated Hamil-
tonian

H is of class C
k1
.
Proposition 2.6.2. Suppose that L is a C
2
Lagrangian on the
manifold M. If : [a, b] M is a C
2
extremal curve then the
Hamiltonian

H is constant on its speed curve t ((t), (t)).
In particular, the Hamiltonian

H is invariant under the Euler-
Lagrange ow
L
t
when it exists.
Proof. We have

H((t), (t)) =
L
v
((t), (t))( (t)) L((t), (t)),
We want to show that its derivative is zero. This is a local result,
we can suppose that takes its values in a chart on M, and there-
fore use coordinates. Performing the dierentiation with respect
59
to t in coordinates, after simplications, we get
d
dt

H((t), (t)) =
_
d
dt
_
L
v
((t), (t))
_
_
( (t))
L
x
((t), (t))( (t)).
This last quantity is zero since satises the Euler-Lagrange equa-
tion, see 2.2.6.
Suppose that the restriction of the global Legendre transform

L to some open subset O M is a dieomorphism onto its image

O =

L(O), we will dene the function H

O
:

O R by H

O
=

H
(

L[O)
1
, where (

L[O)
1
:

O O is the inverse of the restriction

L[O. This function H

O
is also called the (associated) Hamiltonian.
Proposition 2.6.3. Let L be a C
k
Lagrangian, with k 2, on
a manifold M. Suppose that the restriction

L[O of the global
Legendre transform to the open O M is a dieomorphism onto
its image

O. Then the Hamiltonian H

O
=

H (

L[O)
1
is also
of class C
k
on the open subset

O T

M. If M = U is an open
subset of R
n
, then, in the natural coordinates on TU and T

U,
we have
H

O
p
(x, p) =v
H

O
x
(x, p) =
L
x
(x, v),
with L(x, v) = p.
Proof. To simplify notations we will set H = H

O
. We know that

L
is C
k1
. It follows that the dieomorphism (

L[O)
1
is also C
k1
,
and hence H is C
k1
, since obviously

H is C
k1
. therefore H is
at least C
1
We then take coordinates to reduce the proof to the
case where M = U is an open subset of R
n
. Using the canonical
coordinates on R
n
, we write (x, v) = (x
1
, , x
n
, v
1
, , v
n
). By
denition of

H, and H, we have
H
_
x,
L
v
(x, v)
_
= L(x, v) +
n

j=1
L
v
j
(x, v)v
j
. (*)
60
If we dierentiate both sides with respect to the variable v
i
, we
nd
n

j=1
H
p
j
_
x,
L
v
(x, v)
_

2
L
v
i
v
j
(x, v) =
n

j=1
v
j

2
L
v
i
v
j
_
x,
L
x
(x, v)
_
,
for all i = 1, . . . , n. Since we have
D

L(x, v) =
_
Id
R
n

2
L
xv
(x, v)
0

2
L
v
2
(x, v)
_
the matrix
_

2
L
v
i
v
j
_
is invertible, for (x, v) O. Thereore we
obtain
H
p
j
_
x,
L
v
(x, v)
_
= v
j
.
If p =
L
x
(x, v) = L(x, v), we indeed found the rst equation. If
we dierentiate both sides of the equality () with respect to the
variable x
i
, using what we have just found, we obtain
H
x
i
_
x,
L
v
(x, v)
_
+
n

j=1
v
j

2
L
x
i
v
j
(x, v) =
n

j=1

2
L
x
i
v
j
(x, v)v
j

L
x
i
(x, v),
hence
H
x
i
[x,
L
v
(x, v)] =
L
x
i
.
As

L[O is a dieomorphism of class C
k1
and L is of class C
k
,
writing the formulas just obtained as
H
p
(x, p) =p
2

L
1
(x, p)
H
x
(x, p) =
L
x
[

L
1
(x, p)]
where p
2
is the projection of TU = U R
n
on the second factor,
we see that the derivative of H is C
k1
, and thus H is C
k
.
61
Let us recall that, for U an open subset of R
n
, the Euler-
Lagrange ow is the ow of the vector eld X
L
on TU = U R
n
dened by X
L
(x, v) = (x, v
1
, v,

X
L
(x, v)) with
L
x
(x, v) =

2
L
v
2
(x, v)(

X
L
(x, v), ) +

2
L
xv
(x, v)(v, ). (**)
Suppose now that

L[O is a diieomorphism onto its image for the
open subset O TM. Since the dieomorphism

L is C
k1
, with
k 2, we can transport by this dieomorphism the vector eld
X
L
[O to a vector eld dened on

O.
Theorem 2.6.4. Let L be a C
k
Lagrangian, with k 2, on
a manifold M. Suppose that the restriction

L[O of the global
Legendre transform to the open O M is a dieomorphism onto
its image

O. If we transport on

O =

L(O) the Euler-Lagrange
vector eld X
L
, using the dieomorphism

L[O, we nd on

O the
Hamiltonian vector eld X
H
O
associated to H
O
=

H (

L[O)
1
.
In particular, even if k = 2, the Euler-Lagrange vector eld X
L
is uniquely integrable on O, therefore the partial Euler-Lagrange
ow
L
t
is dened and C
1
on O. More generally, for every r 2,
the Euler-Lagrange ow
L
t
on O is of class C
r1
.
Proof. Let us x (x, v) O. We set p =
L
v
(x, v). As the Euler-
Lagrange vector eld is of the form X
L
(x, v) = (x, v, v,

X
L
(x, v)),
we have
T
(x,v)

L
_
X
L
(x, v)
_
=
_
x, p, v,
L
v
(x, v)(

X
L
(x, v)) +
L
x
(x, v)(v)
_
.
But
L
v
(x, v) =

2
L
v
2
et
L
x
=

2
L
xv
. Using equation (), we then
nd
T

L
_

X(x, v)
_
=
_
x, p, v,
L
x
(x, v)
_
=
_
x, p,
H
p
,
H
x
_
,
62
this is precisely X
H
O
, because

(x,p)
(
H
O
p

x

H
O
x

p
, )
= (
n

i=1
dx
i
dp
i
)(
H
O
p

x

H
O
x

p
, )
=
n

i=1
H
O
p
i
dp
i
+
H
x
i
dx
i
= dH
O
.
Since the Hamiltonian is C
k
, the vector eld X
H
O
is of class C
k1
,
its local ow
H
O
t
is, by the theorem of Cauchy-Lipschitz, well
dened and of class C
k1
. However, the dieomorphism

L[O is of
class C
k1
and sends the vector eld X
L
on X
H
O
, consequently
the local ow of X
L
[O is well dened and equal to (

L[O)
1

H
O
t

L
which is C
r1
.
The following theorem is clearly a consequence of 2.6.4, propo-
sitions 2.6.3 and 2.1.6.
Theorem 2.6.5. Suppose that L is a non-degenerate C
r
Lagran-
gian, with r 2, on the manifold M. Then for r = 2, the the
Euler-Lagrange vector eld X
L
is uniquely integrable and denes
a local ow
L
t
which is of class C
1
. More generally, for every
r 2, the Euler-Lagrange ow
L
t
is of class C
r1
.
Remark 2.6.6. When L is non degenerate of class C
r
, with r 2,
we can dene
L
=

L, where is the canonical symplectic form
on T

M. Since

L is a local dieomorphism, for each (x, v) TM,
the bilinear form
L
(x,v)
is non-degenerate. Obviously, the 2-form is
of class C
r2
. To be able to say that
L
is closed, we need to have
r 3. Under this condition
L
denes a symplectic structure
of class C
r2
on TM, and we can interpret what we have done
by saying that X
L
is the Hamiltonian ow associated to

H by

L
. Notice however that we cannot conclude from this that X
L
is
uniquely integrable because

H is only C
r1
, and we do not gain
one more degree of dierentiabilty, like in proposition 2.6.3, for
the Hamiltonian. Note that this also gives an explanation for the
invariance of

H by the Euler-Lagrange ow.
63
2.7 Existence of Local Extremal Curves
We will use the following form of the inverse function theorem.
Theorem 2.7.1 (Inverse Function). Let U be an open subset of
R
m
and K a compact space. We suppose that : K U R
m
is a continuous map such that
(1) for each k K, the map
k
: U R
m
, x (k, x) is C
1
,
(2) The map

x
: K U L(R
m
, R
m
), (k, x)

x
(k, x) is
continuous.
If C U is a compact subset such that
(i) for each k K, and each x C, the derivative

x
(k, x) is
an isomorphism,
(ii) for each k K, the map
k
is injective on C,
then there exists an open subset V such that
(a) we have the inclusions C V U,
(b) for each k K, the map
k
induces a C
1
dieomorphism of
V on an open subset of R
m
.
Proof. Let || denote a norm on R
m
, and let d denote its associ-
ated distance. Let us show that there is an integer n such that, if
we set
V
n
=
_
x [ d(x, C) <
1
n
_
,
then, the restriction
k
[V
n
is injective, for each k K. We argue
by contradiction. Suppose that for each positive integer n we can
nd v
n
, v

n
U and k
n
K with
v
n
,= v

n
, d(v
n
, C) <
1
n
, d(v

n
, C) <
1
n
and (k
n
, v
n
) = (k
n
, v

n
).
By the compactness of C and K, we can extract subsequences
v
n
i
, v

n
i
and k
n
i
which converge respectively to v

, v

C and
k

K. By continuity of , we see that (k

, v

) = (k

, v

).
From (ii), it results that v

= v

. Since v
n
,= v

n
, we can set u
n
=
64
vnv

n
vnv

. Extracting a subsequence if necessary, we can suppose


that u
n
i
u

. This limit u

is also of norm 1. As v
n
i
and v

n
i
both converge to v

= v

, for i large, the segment between v


n
i
and v

n
i
is contained in the open set U. Hence for i big enough we
can write
0 = (k
n
i
, v
n
i
) (k
n
i
, v

n
i
)
=
_
1
0

v
(k
n
i
, sv
n
i
+ (1 s)v

n
i
)(v
n
i
v

n
i
) ds,
dividing by |v

n
i
v
n
i
| and taking the limit as n
i
, we obtain
0 =
_
1
0

v
(k

, v

)(u

) ds
=

v
(k

, v

)(u

).
However

v
(k

, v

) is an isomorphism, since (k

, v

) K
C. But |u

| = 1, this is a contradiction. We thus showed the


existence of an integer n such that the restriction of
k
on V
n
is
injective, for each k K. The continuity of (k, v)

v
(k, v)
and the fact that

v
(k, v) is an isomorphism for each (k, v) in the
compact set KC, show that, taking n larger if necessary, we can
suppose that

v
(k, v) is an isomorphism for each (k, v) K V
n
.
The usual inverse function theorem then shows that
k
restricted
to V
n
is a local dieomorphism for each k K. Since we have
already shown that
k
is injective on V
n
, it is a dieomorphism of
V
n
on an open subset of R
m
.
The following lemma is a simple topological result that de-
serves to be better known because it simplies many arguments.
Lemma 2.7.2. Let X be a topological space, and let Y be a
locally compact locally connected Hausdor space. Suppose that
: X U Y is continuous, where U is an open subset of Y
, and that, for each x X, the map
x
: U Y, y (x, y)
is a homeomorphism onto an open subset of Y . Then, the map
: X U X Y, (x, y) (x, (x, y)) is an open map, i.e. it
maps open subsets of X U to open subsets of X Y . It is thus
a homeomorphism onto an open subset of X Y .
65
Proof. It is enough to show that if V is open and relatively com-
pact in Y , with

V U, and x
0
X, y
0
Y are such that
y
0

x
0
(V ), then, there exists a neighborhood W of x
0
in X
and a neighborhood N of y
0
in Y , such that
x
(V ) N, for each
x W. In fact, this will show the inclusion W N (W V ).
As
x
0
(V ) is an open set containing y
0
, there exists N, a compact
and connected neighborhood of y
0
in Y , such that N
x
0
(V ).
Since V =

V V is compact and N
x
0
(V ) = , by continuity
of , we can nd a neighborhood W of x
0
such that
x W,
x
(V ) N = . (*)
We now choose y
0
V , such that
x
0
( y
0
) = y
0
. Since N is
a neighborhood of y
0
and is continuous, taking W smaller if
necessary we can assume that
x W,
x
( y
0
) N. (**)
By condition (), for x W, we have
x
(V ) N =
x
(

V ) N,
therefore the intersection
x
(V ) N is both open and closed as a
subset of the connected space N. This intersection is not empty
because it contains
x
( y
0
) by condition (). By the connectedness
of N, this of course implies that
x
(V ) N = N.
Lemma 2.7.3 (Tilting). Let || be a norm on R
n
. We denote
by

B

(0, R) (resp.

B

(0, R)) the open (resp. closed) ball of R


n
of center 0 and radius R for this norm. We suppose that K is a
compact space and that , , C
1
and C
2
are xed > 0 numbers,
with C
1
> C
2
.
Let : K] , [

(0, C
1
+ ) R
n
be continuous map
such that
(1) for each xed k K, the map (t, v) (k, t, v) has every-
where a partial derivative

t
, and this partial derivative is
itself C
1
;
(2) the map and its partial derivatives

t
,

2

t
2
,

2

vt
=

v
_

are continuous on the product space K], [

(0, C
1
+
);
66
(3) for each k K and each v

B

(0, C
1
+), we have

t
(k, 0, v) = v;
(4) for each (k, v) K B(0, C
1
+), we have
(k, 0, v) = (k, 0, 0).
Then, there exists > 0 such that, for each t [, 0[]0, ]
and each k K, the map v (k, t, v) is a dieomorphism of

(0, C
1
+/2) onto an open subset of R
n
, and moreover
(k, t, v) [ v

B

(0, C
1
) x R
n
[ |x (k, 0, 0)| C
2
[t[.
Proof. Let us consider the map
(k, t, v) =
(k, t, v) (k, 0, v)
t
,
dened on K ([, 0[]0, ])

B

(0, C
1
+). We can extend it
by continuity at t = 0 because
(k, t, v) =
_
1
0

t
(k, st, v) ds ()
The right-hand side is obviously well-dened for t = 0, and equal
to /t(k, 0, v) = v. Moreover, upon inspection of the right-hand
side of (), the extension : K] , [

(0, C
1
+) R
n
is
such that for each xed k K, the map (t, v) (k, t, v) is C
1
,
with

t
(k, t, v) =
_
1
0

t
2
(k, st, v)s ds,

v
(k, t, v) =
_
1
0

vt
(k, st, v) ds.
Therefore both the partial derivatives /t, /v are continu-
ous on the product space K] , [

(0, C
1
+). Let us then
dene the map

: K] , [

(0, C
1
+) R R
n
by

(k, t, v) = (t, (k, t, v)).


67
To simplify, we will use the notation x = (t, v) to indicate the
point (t, v) RR
n
= R
n+1
. The map

is obviously continuous,
and the derivative

/x is also continuous on the product space


K] , [

(0, C
1
+). Since (k, 0, v) = v, for each (k, v)
K

B

(0, C
1
+), we nd that

x
(k, 0, v) =
_
1 0

t
(k, 0, v) Id
R
n
_
,
where we used a block matrix to describe a linear map from the
product RR
n
into itself. It follows that

x
(k, 0, v) is an isomor-
phism for each (k, v) K

B

(0, C
1
+).
Since K and

B

(0, C
1
+ /2) are compact, using the inverse
function theorem 2.7.1, we can nd
1
> 0 and

]/2, [ such
that, for each k K, the map (t, v) (k, t, v) is a C
1
dieomor-
phism from the open set ]
1
,
1
[

(0, C
1
+

) onto an open
set RR
n
. It follows that, for each (k, t) K]
1
,
1
[, the map
v

(k, t, v) is a C
1
dieomorphism

B

(0, C
1
+/2) onto some
open subset of R
n
. By lemma 2.7.2, we obtain that the image of
K]
1
,
1
[

(0, C
1
) by the map (k, s, v) (k,

(k, s, v)) is
an open subset of K R R
n
. This open subset contains the
compact subset K 0

B

(0, C
2
), since

(k, 0, v) = (0, v).
We conclude that there exists > 0 such that the image of K]

1
,
1
[

(0, C
1
) by map the (k, s, v) (k,

(k, s, v)) contains
K[, ]

B

(0, C
2
). Hence, for (k, t) K[, ], the image
of

B

(0, C
1
) by the map v (k, t, v) contains

B

(0, C
2
).
Since we have
(k, s, v) = s(k, s, v) +(k, 0, v)
(k, 0, v) = (k, 0, 0),
we can translate the results obtained for in terms of . This
gives that, for s ,= 0, and [s[ , the map v (k, s, v) is also
a dieomorphism of

B

(0, C
1
+ /2) on an open subset of R
n
and that the image of

B

(0, C
1
) by this map contains the ball

B((k, 0, 0), C
2
s).
Theorem 2.7.4 (Existence of local extremal curves). Let L :
TM M be a non-degenerate C
r
Lagrangian, with r 2. We
68
x a Riemannian metric g on M. For x M, we denote by ||
x
the norm induced on T
x
M by g. We call d the distance on M
associated with g.
If K M is compact and C [0, +[, then, there exists
> 0 such that for x K, and t [, 0[]0, ], the map
L
t
is
dened, and induces a dieomorphism from an open neighborhood
of v T
x
M [ |v|
x
C onto an open subset of M. Moreover,
we have

L
t
(v T
x
M [ |v|
x
C) y M [ d(y, x) C[t[/2.
To prove the theorem, it is enough to show that for each
x
0
M, there exists a compact neighborhood K, such that the
conclusion of the theorem is true for this compact neighborhood
K. For such a local result we can assume that M = U is an
open subset of R
n
, with x
0
U. In the sequel, we identify
the tangent space TU with U R
n
and for x U, we identify
T
x
U = x R
n
with R
n
. We provide U R
n
with the natu-
ral coordinates (x, v) = (x
1
, , x
n
, v
1
, , v
n
). We start with a
lemma which makes it possible to replace the norm obtained from
the Riemannian metric by a constant norm on R
n
.
Lemma 2.7.5 (Distance Estimates). For each > 0, there exists
an open neighborhood V of x
0
with

V compact U and such that
(1) for each v T
x
U

= R
n
and each x

V we have
(1 )|v|
x
0
|v|
x
(1 +)|v|
x
0
;
(2) for each (x, x

)

V we have
(1 )|x x

|
x
0
d(x, x

) (1 +)|x x

|
x
0
.
Proof. For (1), we observe that, for x x
0
, the norm |v|
x
tends
uniformly to 1 on the compact set v [ |v|
x
0
= 1, by continuity
of the Riemannian metric. Therefore for x near to x
0
, we have
v R
n
, (1 ) <
_
_
_
_
v
|v|
x
0
_
_
_
_
x
< (1 +).
For (2), we use the exponential map exp
x
: T
x
U U, induced
by the Riemannian metric. It is known that the map exp : TU =
69
UR
n
UU, (x, v) (x, exp
x
v) is a local dieomorphism on a
neighborhood of (x
0
, 0), that exp
x
(0) = x, and D[exp
x
](0) = Id
R
n.
Thus, there is a compact neighborhood

W of x
0
, such that any pair
(x, x

)

W

W is of the form (x, exp
x
[v(x, x

)]) with v(x, x

) 0
if x and x

both tend to x
0
. The map (x, v) exp
x
(v) is C
1
,
therefore, using again exp
x
(0) = x, D[exp
x
](0) = Id
R
n, we must
have
exp
x
v = x +v +|v|
x
0
k(x, v),
with lim
v0
k(x, v) = 0, uniformly in x

W. Since d(x, exp
x
v) =
|v|
x
, for v small, it follows that for x, x

close to x
0
|x x

|
x
d(x, x

)
=
|v(x, x

)+|v(x, x

)|
x
0
k(x, v(x, x

))|
x
|v(x, x

)|
x
.
We can therefore conclude that
xx

x
d(x,x

)
1, when x, x

x
0
.
But we also have
xx

x
xx

x
0
1 when x x
0
, we conclude that
d(x,x

)
xx

x
0
is close to 1, if x and x

are both in a small compact


neighborhood of x
0
.
Proof of the theorem 2.7.4. Let us give and two > 0 numbers,
with enough small to have
C
1
<
C
1 +
+

2
1
2(1 )
<
1
1 +
.
We set C
1
= C/(1 +). Let

W U be a compact neighborhood
of x
0
. Since

W

B
x
0
(0, C
1
+ ) is compact, there exists > 0
such that
L
t
is dened on

W

B
x
0
(0, C
1
+ ) for t ] , [.
We then set (x, t, v) =
L
t
(x, v). The map is well dened on

W[, ]

B
x
0
(0, C
1
+). Moreover, since t
L
t
(x, v) is the
speed curve of its projection t (t, x, v), we have

L
t
(x, v) = ((x, t, v),

t
(x, t, v)),
and
(x, 0, v) = x, and

t
(x, 0, v) = v.
70
Since the ow
L
t
is of class C
r1
, see theorem 2.6.5, both maps
and

t
are of class C
r1
, with respect to all variables. Since r
2, we can then apply the tilting lemma 2.7.3, with C
1
= C/(1+)
and C
2
= C/2(1 ), to nd > 0 such that, for each x

W,
and each t [, 0[]0, ], the map (x, v)
L
t
(x, v) induces
a C
1
dieomorphism from v [ |v|
x
0
< /2 +C/(1 +) onto an
open subset of R
n
with

L
t
(x, v) [ |v|
x
0

C
1 +
y R
n
[ |y x|
x
0

Ct
2(1 )
.
Since
L
0
(x, v) = x, taking

W and > 0 smaller if necessary,
we can assume that

W V and

L
t
(x, v) [ t [, ], x

W, v

B(0, C/(1 +) +) V,
where V is given by lemma 2.7.5. Since

W V , by what we
obtained in lemma 2.7.5, for x

W, for every R 0, we have
v T
x
U [ |v|
x
0

R
1 +
v T
x
U [ |v|
x
R
v T
x
U [ |v|
x
0

R
1
.
As

W is compact and contained in the open set V , for t > 0 small
and x

W, we have
y M [ d(y, x)
Ct
2
V,
hence again by lemma 2.7.5
y M [ d(y, x)
Ct
2
y V [ |y x|
x
0

Ct
2(1 )
.
Therefore by the choices made, taking > 0 smaller if necessary,
for t [, ], and x

W, the map
L
t
is a dieomorphism from
a neighborhood of v T
x
U [ |v|
x
C onto an open subset of
U such that

L
t
(v T
x
U [ |v|
x
C) y V [ d(y, x)
C[t[
2
.
71
2.8 The Hamilton-Jacobi method
We already met an aspect of the theory of Hamilton and Jacobi,
since we saw that a connected Lagrangian submanifold is invariant
by a Hamiltonian ow if and only if the Hamiltonian is constant
on this submanifold. We will need a little more general version for
the case when functions depend on time.
We start with some algebraic preliminaries.
Let a be an alternate 2-form on a vector space E. By denition,
the characteristic subspace of a is
ker a = E [ a(, ) = 0,
i.e. ker a is the kernel of the linear map a
#
: E E

, a(, ).
If E is of nite dimension, then Ker(a) = 0 if and only if a is a
symplectic form. Since a space carrying a symplectic form is of
even dimension, we obtain the following lemma.
Lemma 2.8.1. Let a be an alternate 2-form on the vector space
E is provided with the alternate bilinear 2-form. If the dimension
of E is nite and odd, then Ker(a) is not reduced to 0.
We will need the following complement.
Lemma 2.8.2. Let a be an alternate bilinear 2-form on the vec-
tor space E of nite odd dimension. We suppose that there is a
codimension one subspace E
0
E such that the restriction a[E
0
is a symplectic form, then, the dimension of the characteristic sub-
space Ker(a) is 1.
Proof. Indeed, we have E
0
Ker(a) = 0, since E
0
is symplectic.
As E
0
is a hyperplane, the dimension of Ker(a) is 1. But we
know by the previous lemma that Ker(a) ,= 0.
Denition 2.8.3 (Odd Lagrangian Subspace). Let E be a vector
space of dimension 2n + 1 provided with an alternate 2-form a,
such that the dimension of Ker(a) is 1. A vector subspace F of E
is said to be odd Lagrangian if dimF = n + 1 and the restriction
a[F is identically 0.
Lemma 2.8.4. Let E be a vector space of dimension 2n + 1
provided with an alternate 2-form a, such that the dimension of
Ker(a) is 1. If F is an odd Lagrangian subspace, then Ker(a) F.
72
Proof. We set F

= E [ f F, a(, f) = 0. We have
F

F + Ker(a), therefore dima


#
(F

) = dimF

1, since
Ker(a)
#
= Ker(a), where a
#
: E E

, a(, ). But a
#
(F

)
E

[ (F) = 0 which is of dimension n. It follows that


dimF

1 n and thus dimF

n+1. Since F

F +Ker(a)
and dimF = n + 1, we must have F

= F and Ker(a) F.
In the sequel of this section, we x a manifold M and O an open
subset of its cotangent space T

M. We denote by

: T

M M
the canonical projection.
We suppose that a C
2
Hamiltonian H : O R is given. We
denote by X
H
the Hamiltonian vector eld associated to H, and
by
H
t
the local ow of X
H
. We dene the dierential 1-form
H
on O R by

H
= Hdt,
where is the Liouville form on T

M. More precisely, we should


write

H
= p

1
(H p
1
) dt,
where p
1
: T

M R T

M is the projection on the rst factor


and dt is the dierential on T

MR of the projection T

MR
R on the second factor. The exterior derivative
H
= d
H
denes a dierential 2-form which is closed on O R. We have

H
= p

1
+ (p

1
dH) dt,
where = d is the canonical symplectic form on T

M. If
(x, p, t) (O R), then, the tangent space T
(x,p,t)
(O R) =
T
(x,p)
(T

M) R is of odd dimension. Since, moreover, the re-


striction (
H
)
(x,p,t)
[T
(x,p)
T

M is nothing but the symplectic form


= d, the lemmas above show that the characteristic space of
(
H
)
(x,p,t)
is of dimension 1, at each point (x, p, t) O R.
Lemma 2.8.5. At a point (x, p, t) in O R, the characteristic
subspace of
H
is generated by the vector X
H
+

t
, where X
H
is
the Hamiltonian vector eld on O associated with H.
Proof. The vector eld X
H
+

t
is never 0, because of the part

t
,
it is then enough to see that

H
(X
H
+

t
, ) = 0.
73
But, we have

H
_
X
H
+

t
,
_
= (X
H
, ) + (dH dt)
_
X
H
+

t
,
_
= dH +dH(X
H
)dt dH
= 0,
since dH(X
H
) = (X
H
, X
H
) = 0.
Denition 2.8.6 (Odd Lagrangian Submanifold). We say that a
C
1
submanifold V of OR is odd Lagrangian for
H
, if dimV =
dimM + 1 and the restriction
H
[V is identically 0. This last
condition is equivalent to the fact that the restriction
H
[V is
closed as dierential 1-form.
Lemma 2.8.7. If the C
1
submanifold V of OR is odd Lagran-
gian for
H
, then, the vector eld X
H
+

t
is tangent everywhere
to V .
It is not dicult to see that the local ow of X
H
+

t
on OR
is

H
s
(x, p, t) = (
H
s
(x, p), t +s).
Corollary 2.8.8. If the C
1
submanifold V of O R is odd Lag-
rangian for
H
, then, it is invariant by the local ow
H
s
.
Proof. Again since we are assuming that V is only C
1
, the re-
striction of X
H
+

t
to V which is tangent to V is only C
0
, as
a section V TV . We cannot apply the Cauchy-Lipschitz theo-
rem. Instead as in the proof of 2.5.10, if (x, p, t) V , we apply the
Cauchy-Peano [Bou76] to nd a a curve in V which is a solution
of the vector eld. Then we apply the uniqueness in O where the
vector eld to conclude that this curve in V is a part of an orbit
of the ow.
Lemma 2.8.9. the form
H
is preserved by the ow
H
s
.
Proof. By the Cartan formula, a closed dierential form is pre-
served by the ow of the vector eld X, if and only if the exterior
derivative d[(X, )] of the form (X, ) is identically 0. However
in our case

H
(X
H
+

t
, ) = 0.
74
Let us consider U, an open subset of M, and a local C
1
section
s : U]a, b[ T

MR of projection

Id
R
: T

MR MR
such that s(x, t) O, for each (x, t) U]a, b[. If we set s(x, t) =
(x, p(x, t), t), then the image of the section s is an odd Lagrangian
submanifold for
H
if and only if the form: s

[H(x, p(x, t)) dt]


is closed. If we choose coordinates x
1
, . . . . . . , x
n
in a neighborhood
of a point in U, we get
s

[ H(x, p(x, t)) dt] = H(x, p(x, t)) dt +


n

i=1
p
i
(x, t) dx
i
and thus the image of the section s is an odd Lagrangian sub-
manifold if and only if the dierential 1-form H(x, p(x, t)) dt +

n
i=1
p
i
(x, t) dx
i
is closed. If this is the case and U is simply
connected, this form is then exact and there is a C
2
function
S : U]a, b[R such that
dS = H(x, p(x, t)) dt +
n

i=1
p
i
(x, t) dx
i
,
which means that we have
S
x
= p(x, t) and
S
t
= H(x, p(x, t)).
This brings us to the Hamilton-Jacobi equation
S
t
+H(x,
S
x
) = 0. (H-J)
Conversely, any C
2
solution S : U]a, b[ R of this equation
gives us an invariant odd Lagrangian submanifold, namely the
image of the section s : U]a, b[ T

M R dened by s(x, t) =
_
x,
S
x
(x, t), t
_
. Indeed, with this choice and using coordinates, we
nd that
s

(
H
) =
n

i=1
S
x
i
(x, t) dx
i
H(x, p(x, t)) dt.
Since S satises the Hamilton-Jacobi equation, we have
s

(
H
) =
S
t
(x, t) dt +
n

i=1
S
x
i
(x, t) dx
i
,
75
which implies that s

(
H
) = dS is closed.
The following theorem partly summarizes what we obtained:
Theorem 2.8.10. Let H : O R be a C
2
Hamiltonian, dened
on the open subset O of the cotangent space T

M of the manifold
M. We denote by
H
t
the local ow of the Hamiltonian vector eld
X
H
associated with H. Let U be an open subset of M and a, b R,
with a < b. Suppose that the C
2
function S : U]a, b[R is such
that
(1) for each (x, t) U]a, b[, we have
_
x,
S
x
(x, t)
_
O;
(2) the function S satises the Hamilton-Jacobi equation
S
t
+H(x,
S
x
) = 0.
We x (x
0
, t
0
) a point in U]a, b[, and to simplify notations we
denote the point
H
t
_
x
0
,
S
x
(x
0
, t
0
)
_
by (x(t), p(t)). If ], []a, b[
is the maximum open interval such that
H
t
_
x
0
,
S
x
(x
0
, t
0
)
_
is de-
ned, and its projection x(t) is in U, for each t ], [, then, we
have
t ], [, p(t) =
S
x
(x(t), t +t
0
).
Moreover, for t tending to or , the projection x(t) leaves every
compact subset of U.
Proof. By what we have already seen, the image of the section
s : U]a, b[ T

M R, (x, t) dened by s(x, t) =


_
x,
S
x
(x, t), t
_
is odd Lagrangian for
H
, it is thus invariant by the local ow
H
t
of X
H
+

t
. If we denote by ]
0
,
0
[ the maximum open interval
such that
H
t
_
x
0
,
S
x
(x
0
, t
0
), t
0
_
is dened and in the image of the
section s, it results from the invariance that
t ]
0
,
0
[, p(t) =
S
x
(x(t), t +t
0
).
We of course have ]
0
,
0
[], [. If we suppose that
0
< ,
then x(
0
) U and, by continuity of the section s, we also have

0
_
x
0
,
S
x
(x
0
, t
0
), t
0
_
= s(x(
0
)). By the invariance of the image
of s by
H
t
, we then nd > 0 such that
H
t
_
x
0
,
S
x
(x
0
, t
0
), t
0
_
=
76

H
t
0
_
s(x(
0
))
_
is dened and in the image of the section s, for
each t [
0
,
0
+]. This contradicts the denition of
0
.
It remains to see that x(t) comes out of every compact sub-
set of U, for example, when t . Indeed if this would not be
the case, we could nd a sequence t
i
such that x(t
i
) would
converge to a point x

U, but, by continuity of s, the se-


quence
H
t
i
_
x
0
,
S
x
(x
0
, t
0
)
_
= s[x(t
i
)] would converge to a point in
T

U, which would make it possible to show that


H
t
_
x
0
,
S
x
(x
0
, t
0
)
_
would be dened and in T

U for t near to and t > . This con-


tradicts the denition of .
We know consider the problem of constructing (local) solu-
tions of the Hamilton-Jacobi equation. We will include in this
construction a parameter that will be useful in the sequel.
Theorem 2.8.11 (Method of Characteristics). Let H : O R
be a C
2
Hamiltonian, dened on the open subset O of cotangent
space T

M of the manifold M. We denote by


H
t
the local ow
of the Hamiltonian vector eld X
H
associated with H. Let U
be open in M and let K be a compact space. We suppose that
S
0
: K U R is a function such that
(1) for each k K, the map S
0,k
: U R is C
2
;
(2) the map (k, x) (x,
S
0
x
(k, x)) is continuous on the product
KU with values in T

M, and its image is contained in the


open subset O T

M;
(3) the derivative with respect to x of (k, x) (x,
S
0
x
(k, x)) is
also continuous on the product K U (this is equivalent to
the continuity of the (k, x)

2
S
0
x
2
(k, x) in charts contained
in U).
Then, for each open simply connected subset W, with

W compact
and included in U, there exists > 0 and a continuous map S :
K W] , [ satisfying
(i) for each (k, x) K W, we have S(k, x, 0) = S
0
(k, x);
(ii) for each k M, the map S
k
: W] , [ R, (x, t)
S(k, x, t) is C
2
, and satises the Hamilton-Jacobi equation
S
k
t
+H(x,
S
k
x
) = 0;
77
(iii) both maps (k, x, t) DS
k
(x, t), (k, x, t) D
2
S
k
(x, t) are
continuous on the product K W] , [.
Proof. We want to nd a function S(k, x, t) such that S(k, x, 0) =
S
0
(k, x), and S
k
(x, t) = S(k, x, t) is a solution of the Hamilton-
Jacobi equation, for each k K. As we already know the graph
(x,
S
k
x
, t) [ x U T

M R must be invariant by the local


ow
H
t
. This suggests to obtain this graph like a part of the
image of the map
k
dened by

k
(x, t) =
H
t
_
x,
S
0
x
(k, x), 0
_
= (
H
t
[x,
S
0
x
(k, x)], t).
The map (k, x, t) =
k
(x, t) is well dened and continuous on
an open neighborhood | of K U 0 in K O R. The
values of are in OR. Moreover, the map
k
is C
1
on the open
subset |
k
= (x, t) [ (k, x, t) |, and both maps (k, x, t)

k
t
(k, x, t), (k, x, t)

k
x
(k, x, t) are continuous on |.
Given its denition, it is not dicult to see that
k
is injec-
tive. Let us show that the image of
k
is a C
1
odd-Lagrangian
submanifold of O R. Since
k
(x, t) =
H
t
_
x,
S
0
x
(k, x), 0
_
and

H
t
is a local dieomorphism which preserves
H
, it is enough to
show that the image of the derivative of
k
at a point of the form
(x
0
, 0) is odd Lagrangian for
H
(and thus of dimension n+1 like
|
k
). Since we have

k
(x, 0) = (x,
S
0
x
(k, x), 0),

k
(x
0
, t) =
H
t
_

k
(x
0
, 0)
_
,
we see that the image of D
k
(x
0
, 0) is the sum of the subspace
E = T
(x
0
,
S
0
x
(k,x
0
))
Graph(d
x
S
0
) 0 T

k
(x
0
,0)
(T

x
M R)
and the subspace generated by X
H
(x
0
,
S
0
x
(k, x
0
)) +

t
. Note now
that the restriction
H
[T

x
M0 is , and that Graph(d
x
S
0
) is a
Lagrangian subspace for . Moreover the vector X
H
_
x
0
,
S
0
x
(k, x
0
)
_
+

t
is not in E and generates the characteristic subspace of
H
.
Therefore the image of D
k
(x
0
, 0) is indeed odd Lagrangian for

H
. This image is the tangent space to the image
k
at the point

k
(x
0
, 0). We thus have shown that the whole image of
k
is an odd
78
Lagrangian submanifold for
H
. In the remainder of this proof, we
denote by

the projection

Id
R
: T

MR MR. Let us
show that the derivative of

k
: |
k
M R is an isomorphism
at each point (x, 0) U0. Indeed, we have

k
(x, 0) = (x, 0),
for each x U, and

k
(x, t) = (

H
t
(x,
S
0
x
(k, x)), t), therefore,
by writing the derivative D(

k
)(x, 0) : T
x
M R T
x
M R in
matrix form, we nd a matrix of the type
D(

k
)(x, 0) =
_
Id
TxM

0 1
_
which is an isomorphism. As K

W0 is a compact subset of |
and the derivative of

k
is an isomorphism at each point (x, 0)
U 0, we can apply the inverse function theorem 2.7.1 to nd
an open neighborhood

W of

W in U and > 0 such that, for each
k K, the map

k
induces a C
1
dieomorphism of

W] , [
onto an open subset of M R. Since moreover

k
(x, 0) = x,
we can apply lemma 2.7.2 to obtain > 0 such that

k
(

W]
, [)

W [, ], for each k K. Let us then dene the
C
1
section
k
: W] , [ T

M R of the projection

:
T

M R M R by

k
(x, t) =
k
_
(

k
)
1
(x, t)].
Since
k
is a section of

, we have

k
(x, t) = (x, p
k
(x, t), t),
with p
k
(x, t) T

x
M, and p
k
(x, 0) =
S
0
x
(k, x). Moreover, the
image of
k
is an odd Lagrangian submanifold for
H
, since it is
contained in the image of
k
, consequently, the dierential 1-form

H
= p
k
(x, t) H(x, p
k
(x, t))dt
is closed on W] , [. However, the restriction of this form
to W 0 is
S
0
x
(k, x) = dS
0,k
which is exact, therefore, there
exists a function S
k
: W] , [ R such that S
k
(x, 0) = S
0,k
(x)
and dS
k
= p(x, t) H(x, p
k
(x, t))dt. We conclude that
S
k
x
=
p
k
(x, t) and
S
k
t
= H(x, p
k
(x, t)). Consequently, the function
S
k
is a solution of the Hamilton-Jacobi equation and S
k
(x, t) =
S
0
(k, x)
_
t
0
H(x, p
k
(x, s)) ds, which makes it possible to see that
(k, x, t) S
k
(x, t) is continuous. The property (iii) results from
the two equalities
S
k
x
= p
k
(x, t) and
S
k
t
= H(x, p
k
(x, t)).
79
Corollary 2.8.12. Let H : O R be a C
2
Hamiltonian, dened
on the open subset O of the cotangent space T

M of the manifold
M. Call
H
t
the local ow of the Hamiltonian vector eld X
H
associated to H. Denote by d a distance dening the topology of
M. The open ball of center x and radius r for this distance will
be denoted by

B
d
(x, r).
For every compact C O, we can nd , > 0 such that, for
each (x, p) C, there exists a C
2
function S
(x,p)
:

B
d
(x, )]
, [ R, with
S
(x,p)
x
(x, 0) = p, and satisfying the Hamilton-
Jacobi equation
S
(x,p)
t
+H
_
x,
S
(x,p)
x
_
= 0.
Proof. Since we do not ask that S
(x,p)
depends continuously on
(x, p), it is enough to show that, for each point (x
0
, p
0
) in O,
there is a compact neighborhood C of (x
0
, p
0
), contained in O,
and satisfying the corollary. We can of course assume that M
is an open subset of R
n
. In this case T

M = M R
n
. To
begin with, let us choose a compact neighborhood of (x
0
, p
0
) of
the form

U K O, with

U M and K R
n
. The function
S
0
: K U R dened by
S
0
(p, x) = p(x),
is C

and veries (x,


S
0
x
) = (x, p) U K O. Let us choose
a neighborhood compact

W of x
0
with

W U. By the previous
theorem, there exists > 0, and a function S : KW], [R
satisfying the following properties
S(p, x, 0) = S
0
(p, x) = p(x), for each (p, x) K W;
for each p K, the map S
p
: W] , [ R is C
2
, and is
a solution of the Hamilton-Jacobi equation.
We now choose

V a compact neighborhood of x
0
contained in W.
By compactness of

V , we can nd > 0 such that B
d
(x, ) W,
for each x

V . It then remains to take C =

V K and to dene,
for (x, p) C, the function S
(x,p)
: B
d
(x, )] , [ R by
S
(x,p)
(y, t) = S
p
(y, t).
It is not dicult to obtain the required properties of S
(x,p)
from
those of S
p
.
80
Chapter 3
Calculus of Variations for
a Lagrangian Convex in
the Fibers: Tonellis
Theory
The goal of this chapter is to prove Tonellis theorem which estab-
lishes the existence of minimizing extremal curves. This theory
requires the convexity of the Lagrangian in the bers, and the use
of absolutely continuous curves.
Another good reference for this chapter is [BGH98]. Again, we
have mainly used [Cla90], [Mn] and the appendix of [Mat91]. How-
ever, we have departed from the usual way of showing minimizing
properties using Mayer elds. Instead we have systematically used
(local solutions) of the the Hamilton-Jacobi equation, since this is
the main theme of this book.
3.1 Absolutely Continuous Curves.
Denition 3.1.1 (Absolutely Continuous Curve). A curve :
[a, b] R
n
is said to be absolutely continuous, if for each > 0,
there exists > 0 such that for each family ]a
i
, b
i
[
iN
of disjoint
intervals included in [a, b], and satisfying

iN
b
i
a
i
< , we
have

iN
|(b
i
) (a
i
)| < .
81
82
It is clear that such an absolutely continuous map is (uni-
formly) continuous.
Theorem 3.1.2. A curve : [a, b] R
n
is absolutely continuous
if and only if the following three conditions are satised
(1) the derivative (t) exists almost everywhere on [a, b];
(2) the derivative is integrable for the Lebesgue measure on
[a, b].
(3) For each t [a, b], we have (t) (a) =
_
t
a
(s) ds.
Proofs of this theorem can be found in books on Lebesgues
theory of integration, see for example [WZ77, Theorem 7.29, page
116]. A proof can also be found in [BGH98, Theorem 2.17]
Lemma 3.1.3. Suppose that : [a, b] R
k
is an absolutely
continuous curve. 1) If f : U R
m
is a locally Lipschitz map,
dened on the neighborhood U of the image ([a, b]), then f :
[a, b] R
m
is also absolutely continuous.
2) We have
t, t

[a, b], t t

, |(t

) (t)|
_
t

t
| (s)| ds.
Proof. To prove the rst statement, since ([a, b]) is compact, we
remark that, cutting down U if necessary, we can assume that f is
(globally) Lipschitz. If we call K a Lipschitz constant for f, then
for each family ]a
i
, b
i
[
iN
of disjoint intervals included in [a, b], we
have

iN
|f (b
i
) f (a
i
)| K

iN
|(b
i
) (a
i
)|.
Therefore the absolute continuity of f follows from that of .
To prove the second statement, we choose some p R
k
with
|p| = 1. The curve p : [a, b] R is absolutely continuous
with derivative equal almost everywhere to p( (t)). Therefore by
83
theorem 3.1.2 above
p((t

) (t)) = p (t

) p (t))
=
_
t

t
p( (s)) ds
|eq
_
t

t
|p|| (s)| ds
=
_
t

t
| (s)| ds.
It now suces to observe that
|(t

) (t)| = sup
p=1
p((t

) (t))
to nish the proof.
The following proposition will be useful in the sequel.
Proposition 3.1.4. Let
n
: [a, b] R
k
be a sequence of abso-
lutely continuous curves. We suppose that the sequence of deriva-
tives
n
: [a, b] R
k
(which exist a.e.) is uniformly integrable for
the Lebesgue measure m on [a, b], i.e. for each > 0, there exists
> 0, such that if A [a, b] is a Borel subset with its Lebesgue
measure m(A) < then
_
A
|
n
(s)| ds < , for each n N.
If for some t
0
[a, b] the sequence
n
(t
0
) is bounded in norm,
then there is a subsequence
n
j
: [a, b] R
k
which converges
uniformly to a curve : [a, b] R
k
. The map is absolutely
continuous, and the sequence of derivatives
n
j
converges to the
derivative in the weak topology (L
1
, L

), which means that


for each function : [a, b] R
k
measurable and bounded, we
have
_
b
a
(s)(
n
(s)) ds
n+
_
b
a
(s)( (s)) ds.
Proof. We rst show that the sequence
n
is equicontinuous. If
> 0 is xed, let > 0 be the corresponding given by the condition
that the sequence
n
is uniformly integrable. If t < t

are such that


t

t < then, for each n N


|
n
(t

)
n
(t)|
_
t
t

|
n
(s)| ds
< .
84
Since the
n
(t
0
) form a bounded sequence, and the
n
are equicon-
tinuous, by Ascolis theorem, we can nd a subsequence
n
j
such
that
n
j
converges uniformly to : [a, b] R
n
, which is therefore
continuous. Let us show that is absolutely continuous. We x
> 0, and pick the corresponding > 0 given by the condition
that the sequence
n
is uniformly integrable. If (]a
i
, b
i
[)
iN
is a se-
quence of disjoint open intervals, and such that

iN
(b
i
a
i
) < ,
we have
n N,

iN
|
n
(b
i
)
n
(a
i
)[|

iN
_
b
i
a
i
|
n
(s)| ds
=
_
S
iN
]a
i
,b
i
[
|
n
(s)|ds
< ,
Taking the limit we also get

iN
|(b
i
) (a
i
)| .
The curve is thus absolutely continuous. The derivative :
[a, b] R
n
therefore exists, for Lebesgue almost any point of
[a, b], and we have
t, [a, b], (t) (t

) =
_
t
t

(s) ds.
It remains to show that
n
j
tends to in the (L
1
, L

) topology.
The convergence of
n
j
to shows that for any interval [t, t

]
[a, b] we have
_
t

t

n
j
(s) ds =
n
j
(t

)
n
j
(t) (t

) (t) =
_
t

t
(s) ds.
If U is an open subset of [a, b] let us show that
_
U

n
j
(s) ds
_
U
(s) ds. We can write U =

iI
]a
i
, b
i
[ with the ]a
i
, b
i
[ disjoint,
where I is at most countable. If the set of the indices I is nite, we
apply what precedes each interval ]a
i
, b
i
[, and adding, we obtain
the convergence
_
U

n
j
(s) ds
_
U
(s) ds.
85
To deal with the case were I is innite and countable, we x > 0
and choose the corresponding > 0 given by the fact that the
sequence
n
is uniformly integrable. We can nd a nite subset
I
0
I such that

iI\I
0
(b
i
a
i
) < . We have
n N,

iI\I
0
|
n
(b
i
)
n
(a
i
)| < .
Taking limits, we then obtain

iI\I
0
|(b
i
) (a
i
)| .
We pose U
0
=

iI
0
]a
i
, b
i
[, so we can write
|
_
U\U
0

n
(s) ds| = |

iI\I
0

n
(b
i
)
n
(a
i
)|

iI\I
0
|
n
(b
i
)
n
(a
i
)|
< ,
and also
|
_
U\U
0
(s) ds| .
As I
0
is nite, we have
lim
j
_
U
0

n
j
(s) ds =
_
U
0
(s) ds.
We conclude that
limsup
j
|
_
U

n
j
(s) ds
_
U
(s) ds| 2.
Since is arbitrary, we see that
_
U

n
j
(s) ds
_
U
(s) ds.
If we take now an arbitrary measurable subset A of [a, b], we
can nd a decreasing sequence (U

)
N
of open subset with A
86

N
U

and the Lebesgue measure m(U

) decreasing to m(A),
where m is the Lebesgue measure. We x > 0, and we choose
the corresponding > 0 given by the fact that the sequence
n
is uniformly integrable. As the U

form a decreasing sequence of


sets of nite measure and m(U

) m(A), by Lebesgues theorem


of dominated convergence, we have
_
U

(s) ds
_
A
(s) ds.
We now x an integer big enough to satisfy
|
_
U

(s) ds
_
A
(s)ds| ,
and m(U

A) < . By the choice of , we have


n N, |
_
U

\A

n
(s) ds|
_
U

\A
|
n
(s)|ds .
Since U

is a xed open set, taking the limit for j , we obtain


_
U


n
j
(s) ds
_
U

(s) ds.
We conclude that
limsup
j
|
_
A

n
j
(s) ds
_
A
(s) ds| 2,
and thus, since > 0 is arbitrary
_
A

n
j
(s) ds
_
A
(s) ds.
We then consider the vector space L

([a, b], R
k
), formed by the
: [a, b] R
k
measurable bounded maps, and provided with
the standard norm ||

= sup
t[a,b]
|(t)|. The subset c formed
by the characteristic functions
A
, where A [a, b] is measurable,
generates a vector subspace

c which is dense in L

([a, b], R
k
), see
exercise 3.1.5 below. The maps
n
: L

([a, b], R
k
) R dened
by

n
() =
_
b
a
(s)(
n
(s)) ds
87
are linear and continuous on L

([a, b], R
k
, m), with
|
n
|
_
b
a
|
n
(s)|ds.
We also dene : L

([a, b], R
k
, m) R by
() =
_
b
a
(s)( (s)) ds,
which is also linear and continuous on L

([a, b], R
k
, m). We now
show that the sequence of norms |
n
| is bounded. Applying the
fact that the sequence of derivatives
n
is uniformly integrable
with = 1, we nd the corresponding
1
> 0. We can then cut the
interval [a, b] into [(b a)/
1
] + 1 intervals of length
1
, where
[x] indicates the integer part of the real number x. On each one
of these intervals, the integral of |
n
| is bounded by 1, hence
n N, |
n
|
_
b
a
|
n
(s)| ds [(b a)/
1
] + 1.
As
n
j
() (), for c, by linearity the same conver-
gence is true for

c. As

c is dense in L

([a, b], R
k
, m) and
sup
nN
|
n
| < +, a well-known argument of approximation (see
the exercise below 3.1.6) shows, then, that
L

([a, b], R
k
, m),
n
j
() ().
Exercise 3.1.5. Consider the vector space L

([a, b], R
k
), formed
by the measurable bounded maps : [a, b] R
k
, and provided
with the standard norm ||

= sup
t[a,b]
|(t)|. The subset
c formed by the characteristic functions
A
, where A [a, b]
is measurable, generates a vector subspace which we will call

c.
Show that

c which is dense in L

([a, b], R
k
). [Indication: Con-
sider rst : [a, b] ] K, K[, with K R
+
, and dene
n
=

n1
i=n
(iK/n)
A
i,n
, where A
i,n
is the set t [a, b] [ iK/n
(x) < (i + 1)K/n.]
Exercise 3.1.6. Let
n
: E F be a sequence of continuous
linear operators between normed spaces. Suppose that the sequence
88
of the norms |
n
| is bounded, and that there exists a continuous
linear operator : E F and a subset c E generating a dense
subspace of E such that
n
(x) (x) for every x c. Show then
that
n
(x) (x) for every x E
We can replace in the denition of an absolutely continuous
curve, the norm by any distance Lipschitz-equivalent to a norm.
This makes it possible to generalize the denition of absolutely
continuous curve to a manifold.
Denition 3.1.7 (Absolutely continuous Curve). Let M be a
manifold, we denote by d the metric obtained on M from some
xed Riemannian metric.
A curve : [a, b] M is said to be absolutely continuous if, for
each > 0, there exists > 0 such that for any countable family
of disjoint intervals (]a
i
, b
i
[)
iN
all included in [a, b] and satisfying

iN
(b
i
a
i
) < , we have

iN
d((b
i
), (a
i
)) < .
Denition 3.1.8. We denote by (
ac
([a, b], M) the space of ab-
solutely continuous curves dened on the compact interval [a, b]
with values in the manifold M. This space (
ac
([a, b], M) is pro-
vided with the topology of uniform convergence.
Lemma 3.1.9. If : [a, b] M is absolutely continuous, then,
the derivative (t) T
(t)
M exists for almost every t [a, b]. If
([a

, b

]) U, where U is the domain of denition of the coordi-


nate chart : U R
n
, then, we have
t [a

, b

], (t) (a

) =
_
t
a

d( )
dt
(s) ds.
Exercise 3.1.10. Suppose that the manifold M is provided with
a Riemannian metric. We denote by d the distance induced by
this Riemannian metric on M, and by ||
x
the norm induced on
the ber T
x
M, for x M. If : [a, b] M is an absolutely
continuous curve, show that
d((b), (a))
_
b
a
| (s)|
(s)
ds.
[Indication: Use lemmas 2.7.5, and 3.1.3.]
89
Denition 3.1.11 (Bounded Below). Let L : TM R be a
Lagrangian on the manifold M. We will say that L is bounded
below above every compact subset of M, if for every compact
K M, we can nd a constant C
K
R such that
(x, v)
1
(K), L(x, v) C
K
,
where : TM M is the canonical projection.
If the Lagrangian L : TM R is bounded below above every
compact subset of the manifold, we can dene the action of the
Lagrangian for an absolutely continuous curve admitting possibly
+ as a value. Indeed, if (
ac
([a, b], M), since ([a, b]) is
compact, there exists a constant C

R such that
t [a, b], v T
(t)
M, L((t), v) C

.
In particular, the function t L((t), (t))C

is well dened and


positive almost everywhere for the Lebesgue measure. Therefore
_
b
a
[L((t), (T)) C

] ds makes sense and belongs to [0, +]. We


can then set
L() =
_
b
a
L((t), (t))ds
=
_
b
a
[L((t), (t)) C

]ds +C

(b a).
This quantity is clearly in R+. It is not dicult to see that
the action L() is well dened (i.e. independent of the choice of
lower bound C

).
3.2 Lagrangian Convex in the Fibers
In this section, we consider a manifold M provided with a Rie-
mannian metric of reference. For x M, we denote by ||
x
the
norm induced on T
x
M by the Riemannian metric. We will denote
by d the distance induced on M by the Riemannian metric. The
canonical projection of TM on M is as usual : TM M.
We will consider a C
1
Lagrangian L on M which is convex in
the bers, and superlinear above every compact subset of M, see
90
denitions 1.3.11 and 1.3.13. Note that, for Lagrangians convex in
the bers, the superlinearity above compact subset is equivalent to
the restriction L[T
x
M : T
x
M R is superlinear, for each x M,
see 1.3.14. It is clear that a Lagrangian superlinear above every
compact subset of M is also bounded below above every compact
subset of M, therefore we can dene the action for absolutely
continuous curves.
Theorem 3.2.1. Suppose that L : TM R is a C
1
Lagrangian
convex in the bers, and superlinear above compact subsets of M.
If a sequence of curves
n
(
ac
([a, b], M) converges uniformly to
the curve : [a, b] M and
liminf
n
L(
n
) < +,
then the curve is also absolutely continuous and
liminf
n
L(
n
) L().
Proof. Let us start by showing how we can reduce the proof to the
case where M is an open subset of R
k
, where k = dimM.
Consider the set K = ([a, b])

n

n
([a, b]), it is compact
because
n
converges uniformly to . By superlinearity of L above
each compact subset of M, we can nd a constant C
0
such that
x K, v T
x
M, L(x, v) C
0
.
If [a

, b

] is a subinterval of [a, b], taking C


0
as a lower bound of
L(
n
(s),
n
(s)) on [a, b] [a

, b

], we see that
n, L(
n
[[a

, b

]) L(
n
) C
0
[(b a) (b

)].
It follows
[a

, b

] [a, b], liminf


n
L(
n
[[a

, b

]) < +. (*)
By continuity of : [a, b] M, we can nd a nite sequence a
0
=
a < a
1
< < a
p
= b and a sequence of domains of coordinate
charts U
1
, . . . , U
p
such that ([a
i1
, a
i
]) U
i
, for i = 1, . . . , p.
Since
n
converges uniformly to , forgetting the rst curves
n
if
necessary, we can assume that
n
([a
i1
, a
i
]) U
i
, for i = 1, . . . , p.
91
By condition (), we know that liminf
n
L(
n
[[a
i1
, a
i
]) < , it
is then enough to show that this condition implies that [[a
i1
, a
i
]
is absolutely continuous and that
L([[a
i1
, a
i
]) liminf
n
L(
n
[[a
i1
, a
i
]),
because we have liminf(
n
+
n
) liminf
n
+ liminf
n
, for
sequences of real numbers
n
and
n
. As the U
i
are domains of
denition of coordinates charts, we do indeed conclude that it is
enough to show the theorem in the case where M is an open subset
of R
k
.
In the sequel of the proof, we will thus suppose that M = U is
an open subset of R
k
and thus TU = U R
k
and that ([a, b]) and
all the
n
([a, b]) are contained in the same compact subset K
0
U.
Let us set = liminf
n
L(
n
). Extracting a subsequence such
that L(
n
) < + and forgetting some of the rst curves
n
,
we can suppose that
L(
n
) and n N, L(
n
) + 1 < +.
We denote by || a norm on R
k
.
Lemma 3.2.2. If C 0 is a constant, K U is compact, and
> 0, we can nd > 0 such that for each x, y U with x K
and |y x| < , and for each v, w R
k
with |v| C, we have
L(y, w) L(x, v) +
L
v
(x, v)(w v) .
Proof. Let us choose
0
> 0 such that the set

0
(K) = y R
k
[ x K, |y x|
0

is a compact subset of U.
We denote by A the nite constant
A = sup|
L
x
(x, v)| [ x K, |v| C.
Since L is uniformly superlinear above every compact subset of
M = U, we can nd a constant C
1
> such that
y

V

0
(K), w R
k
, L(y, w) (A + 1)|w| +C
1
.
92
We then set
C
2
= supL(x, v)
L
x
(x, v)(v) [ x K, |v| C.
By compactness the constant C
2
is nite. We remark that if |w|
C
2
C
1
, then for y

V

0
(K), x K and |v| C, we have
L(y, w) (A+ 1)|w| +C
1
A|w| + (C
2
C
1
) +C
1
= A|w| +C
2

L
v
(x, v)(w) +
_
L(x, v)
L
v
(x, v)(v)
_
= L(x, v) +
L
v
(x, v)(w v).
It then remains to nd
0
, so that we satisfy the sought
inequality when |w| C
2
C
1
. But the set
(x, v, w) [ x K, |v| C, |w| C
2
C
1

is compact and L(x, w) L(x, v) +


L
v
(x, v)(w v) by convexity
of L in the bers of the tangent bundle TU. It follows that for
> 0 xed, we can nd > 0 with
0
and such that, if
x K, |y x| , |v| C et |w| C
2
C
1
, we have
L(y, w) > L(x, v) +
L
x
(x, v)(w v) .
We return to the sequence of curves
n
: [a, b] U which
converges uniformly to : [a, b] U. We already have reduced
the proof to the case where ([a, b])

nN

n
([a, b]) is included
in a compact subset K
0
U with U an open subset of R
k
. We
now show that the derivatives
n
are uniformly integrable. Since
L is superlinear above each compact subset of U, we can nd a
constant C(0) > such that
x K
0
, v R
k
, L(x, v) C(0).
We recall that
n N, L(
n
) + 1 < +.
93
We then x > 0 and we take A > 0 such that
+ 1 C(0)(b a)
A
< /2.
Again by the superlinearity of L above compact subsets of the
base M, there exists a constant C(A) > such that
x K
0
, v R
k
, L(x, v) A|v| +C(A).
Let E [a, b] be a measurable subset, we have
C(A)m(E) +A
_
E
|
n
(s)|ds
_
E
L(
n
(s),
n
(s)) ds
and also
C(0)(b a m(E))
_
[a,b]\E
L(
n
(s),
n
(s)) ds.
Adding the inequalities and using L(
n
) + 1, we nd
[C(A) C(0)]m(E) +C(0)(b a) +A
_
E
|
n
(s)|ds + 1,
this in turn yields
_
E
|
n
(s)|ds
+ 1 C(0)(b a)
A
+
[C(0) C(A)]
A
m(E)
/2 +
[C(0) C(A)]
A
m(E).
If we choose > 0 such that
C(0)C(A)
A
< /2, we see that
m(E) < n N,
_
E
|
n
(s)| ds < .
This nishes to proves the uniform integrability of the sequence
n
.
We can then conclude by proposition 3.1.4 that
n
converges to
in the (L
1
, L

) topology. We must show that lim


n
L(
n
)
L().
Let C be a xed constant, we set
E
C
= t [a, b] [ | (t)| C.
94
We x > 0 and we apply lemma 3.2.2 with this , the constant C
xed above and the compact set K = K
0
([a, b])

nN

n
[a, b]
to nd > 0 as in the conclusion of the lemma 3.2.2. Since
n

uniformly, there exists an integer n
0
such that for each n n
0
,
we have |
n
(t) (t)| < , for each t [a, b]. Lemma 3.2.2 then
shows that for each n n
0
and almost all t E
C
, we have
L(
n
(t),
n
(t)) L((t), (t)) +
L
v
((t), (t))(
n
(t) (t)) ,
hence using this, together with the inequality L(
n
(t),
n
(t))
C(0) which holds almost everywhere, we obtain
L(
n
)
_
E
C
L((t), (t)) dt +C(0)[(b a) m(E
C
)]
+
_
E
C
L
v
((t), (t))(
n
(t) (t)) dt m(E
C
). (*)
Since | (t)| C, for t E
C
, the map t
E
C
(t)
L
v
((t), (t))
is bounded. But
n
for topology (L
1
, L

), therefore
_
E
C
L
v
((t), (t))(
n
(t) (t)) dt 0, when n .
Going to the limit in the inequality (*), we nd
= lim
n
L(
n
)

_
E
C
L((t), (t)) dt +C(0)[(b a) m(E
C
)] m(E
C
).
We can then let 0 to obtain

_
E
C
L((t), (t)) dt +C(0)[(b a) m(E
C
)]. (**)
Since the derivative (t) exists and is nite for almost all t [a, b],
we nd E
C
E

, when C +, with [a, b] E

of zero
Lebesgue measure. Since L((t), (t)) is bounded below by C(0),
we have by the monotone convergence theorem
_
E
C
L((t), (t)) dt
_
b
a
L((t), (t)) dt, when C .
95
If we let C + in (), we nd
= lim
n
L(
n
)
_
b
a
L((t), (t)) dt = L().
Corollary 3.2.3. Suppose L : TM R is a C
1
Lagrangian
convex in the bers, and superlinear above every compact sub-
set of M. Then the action L : (
ac
([a, b], M) R + is
lower semi-continuous for the topology of uniform convergence on
(
ac
([a, b], M). In particular, on any compact subset of (
ac
([a, b], M),
the action L achieves its inmum.
Proof. It is enough to see that if
n
uniformly, with all the

n
and in (
ac
([a, b], M), then, we have
liminf
n
L(
n
) L().
If liminf
n
L(
n
) = +, there is nothing to show. The case
where we have liminf
n
L(
n
) < + results from theorem
3.2.1.
3.3 Tonellis Theorem
Corollary 3.3.1 (Compact Tonelli). Let L : TM R be a C
1
Lagrangian convex in the bers, and superlinear above every com-
pact subset of the manifold M. If K M is compact, and C R,
then the subset

K,C
= (
ac
([a, b], M) [ ([a, b]) K, L() C
is a compact subset of (
ac
([a, b], M) endowed with the topology of
uniform convergence.
Proof. By the compactness of K, and theorem 3.2.1, the subset

K,C
is closed in the space (([a, b], M) of all the continuous curves,
for the topology of uniform convergence. Since ([a, b]) K, for
each
K,C
, by Ascolis theorem, it is enough to see that the
family of the
K,C
is equicontinuous. Let us then x some
Riemannian metric on M. We denote by d the distance induced
96
on M, and by ||
x
the norm induced on T
x
M, for x M. Using
the superlinearity of L above the compact the compact set K, for
each A 0, we can nd C(A) such that
q K, v T
q
M, C(A) +A|v|
x
L(q, v).
Applying this with (q, v) = ((s), (s)), and integrating, we see
that for each
K,C
, we have
C(A)(t

t) +A
_
t

t
| (s)|
(s)
ds
_
t

t
L((s), (s)) ds
C(0)[(b a) (t

t)]
_
[a,b]\[t,t

]
L((s), (s)) ds.
Adding these two inequalities and using the following one
d((t

), (t))
_
t

t
| (s)|
(s)
ds,
we obtain
Ad((t

), (t)) [L() C(0)(b a)] + (C(0) C(A))(t

t).
If > 0 is given, we choose A such that [CC(0)(ba)]/A /2.
It follows that for
K,C
and t, t

[a, b], we have


d((t

), (t))

2
+
C(0) C(A)
A
(t

t).
We therefore conclude that the family of curves in
K,C
is equicon-
tinuous.
Denition 3.3.2. We will say that the Lagrangian L : TM R
is bounded below by the Riemannian metric g on M, if we can
nd a constant C R such that
(x, v) TM, L(x, v) |v|
x
+C,
where ||
x
is the norm on T
x
M obtained from the Riemannian
metric.
This is a global condition which is relevant only when M is not
compact and the Riemannian metric g is complete.
97
Theorem 3.3.3 (Non Compact Tonelli). Let L be a C
1
Lagran-
gian on the manifold M. Suppose that the Lagrangian L is convex
in bers, superlinear above every compact subset of M, and that
it is bounded below by a complete Riemannian metric on M. If
K M is compact and C R, then the subset

K,C
= (
ac
([a, b], M) [ ([a, b]) K ,= , L() C
is a compact subset of (
ac
([a, b], M) for the topology of uniform
convergence.
Proof. We denote by d the distance on M obtained from the Rie-
mannian metric, and, for x M, by ||
x
the norm induced on
T
x
M by this same Riemannian metric. We rst show that there
exists a constant r < + such that

K,C
, ([a, b])

V
r
(K),
where

V
r
(K) = y M [ x K, d(y, x) r. Indeed, there
exists a constant C
0
R such that
(x, v) TM, L(x, v) |v|
x
+C
0
.
Therefore for every absolutely continuous curve : [a, b] M,
and every t, t

[a, b], with t t

, we have
C
0
(t

t) +
_
t

t
| (s)|
(s)
ds L(),
For
K,C
, it follows that
d((t

), (t)) C +[C
0
[(b a).
If we set r = C +[C
0
[,we get

K,C
, ([a, b])

V
r
(K).
Since the Riemannian metric on M is complete, the d-balls of M
of nite radius are compact, and thus

V
r
(K) is also a compact
subset of M. By the compact case of Tonellis Theorem 3.3.1, the
set
Vr(K),C
is compact in (
ac
([a, b], M). Since

K,C
is closed, and
contained in
Vr(K),C
, it is thus also compact in (
ac
([a, b], M).
98
Corollary 3.3.4 (Tonelli Minimizers). Let M be a connected
manifold. Suppose L : TM R is a C
1
Lagrangian convex in
bers, superlinear above every compact subset of M, and bounded
below by a complete Riemannian metric on M. Then, for each
x, y M, and each a, b R, with a < b, there exists an absolutely
continuous curve : [a, b] M with (a) = x, (b) = y which is
a minimizer for (
ac
([a, b], M).
Proof. Let us set C
inf
= inf L(
1
), where the inmum is taken on
the absolutely continuous curves
1
: [a, b] M with
1
(a) = x
and
1
(b) = y, C
inf
< +. This quantity makes sense since there
exists a C
1
curve between x to y. For each integer n 1, we dene
the subset (
n
of (
ac
([a, b], M) formed by the curves : [a, b] M
such that
(a) = x, (b) = y and L() C
inf
+
1
n
.
This set is by denition a nonempty subset of C
inf
. It is also
compact in (
ac
([a, b], M) by the previous theorem 3.3.3. Since the
sequence (
n
is decreasing, the intersection

n1
(
n
is nonempty.
Any curve : [a, b] M in this intersection is such that (a) =
x, (b) = y and L() = C
inf
.
3.4 Tonelli Lagrangians
Denition 3.4.1 (Tonelli Lagrangian). A Lagrangian L on the
manifold M is called a Tonelli Lagrangian if it satises the follow-
ing conditions:
(1) L : TM R is of class at least C
2
.
(2) For each (x, v) TM, the second partial derivative
2
L/v
2
(x, v)
is positive denite as a quadratic form.
(3) L is superlinear above compact subset of M.
Condition (2) is equivalent to:
(2) L non-degenerate and convex in the bers.
Theorem 3.4.2. If L is a C
r
Tonelli Lagrangian on the manifold
M, then we have the following properties:
99
(1) For each x M, the restriction L[T
x
M is strictly convex.
(2) The Legendre transform

L : TM T

M, (x, v)
_
x,
L
v
(x, v)
_
is a dieomorphism of class C
r1
.
(3) The Euler-Lagrange vector eld X
L
on TM is well dened,
of class C
r1
and uniquely integrable, and the(local) ow
L
t
of X
L
is of class C
r1
.
(4) The extremal curves are C
r
.
(5) The (continuous) piecewise C
1
minimizing curves are ex-
tremal curves, and therefore of class C
r
.
(6) The Hamiltonian associated with L, denoted H : T

M R,
is well-dened by
(x, v) TM, H(

L(x, v)) =
L
v
(x, v)(v) L(x, v).
It is of class C
r
. We have the have the Fenchel inequality
p(v) L(x, v) +H(x, p),
with equality if and only if p = L/v(x, v), or equivalently
(x, p) =

L(x, v). Therefore
(x, p) T

M, H(x, p) = sup
vTxM
p(v) L(x, v).
(7) The (local) Hamiltonian ow
H
t
of H is conjugated by

L to
the Euler-Lagrange ow
L
t
.
Proof. (1) This is clear since
2
L/v
2
(x, v) is positive denite, see
Proposition 1.1.2.
(2) Note that, by the superlinearity and the strict convex-
ity of L[T
x
M, for each x M the Legendre transtorm T
x
M
T

x
M, v L/v(x, v) is bijective, see 1.4.6. Therefore the global
Legendre transform

L : TM T

M is bijective, and we obtain


from Proposition 2.1.6 that it is a dieomorphism.
100
(3) See Theorem 2.6.5.
(4) In fact, if is an extremal, its speed curve t ((t), (t))
is a piece of an orbit of
L
t
, and is therefore of class C
r1
.
(5) See Proposition 2.3.7.
(6) See Proposition 2.6.3, and Fenchels Theorem 1.3.6.
(7) See Theorem 2.6.4.
Lemma 3.4.3. Suppose that L is a Tonelli Lagrangian on the
manifold M. Its associated Hamiltonian H : T

M R is also
superlinear above each compact subset of M. In particular, if
C R and K M is compact, the set
(x, p) T

M [ x K, H(x, p) C,
is a compact subset of T

M.
Proof. We know that in this case
H(x, p) = sup
vTxM
p, v) L(x, v).
Therefore we can apply theorem 1.3.12 to conclude that H is su-
perlinear above compact subset. In particular, if K M is com-
pact, and we x Riemannian metric g on M, we can nd a constant
C
1
(K) > such that
(x, p) T

x
M, H(x, p) |p|
x
+C
1
(K),
where ||
x
is the norm on T

x
M obtained from g. Therefore the
closed set (x, p) T

M [ x K, H(x, p) C is contained in
(x, p) T

M [ x K, |p|
x
C C
1
(K). But this last set is a
compact subset of T

M.
Corollary 3.4.4. Let L be a C
2
Lagrangian on the manifold
M which is non-degenerate, convex in the bers, and superlin-
ear above every compact subset of M. If (x, v) TM, denote
by ]
(x,v)
,
(x,v)
[ is the maximal interval on which t
L
t
(x, v) is
dened. If
(x,v)
< + (resp.
(x,v)
> ), then t
L
t
(x, v)
leaves every compact subset of M as t
(x,v)
(resp. t
(x,v)
).
In particular, if M is a compact manifold, the Euler-Lagrange vec-
tor eld X
L
is complete, i.e. the ow
L
t
: M M is dened for
each t R.
101
Proof. We know that the global Legendre transform

L : TM
T

M is a dieomorphism, and it conjugates


L
t
to the Hamil-
tonian ow
H
t
of the Hamiltonian associated to L. Therefore
H

L(
t
(x, v)) = H

L(x, v) by conservation of the Hamiltonian. Let
us suppose that there exists
i

(x,v)
such that
L

i
(x, v)
x

. The set K = x

i
(x, v) [ i N is then a compact
subset of M. The sequence

i
(x, v) is contained in the subset
(y, w) [ y K, H

L(y, w) H

L(x, v) of TM. But this last set is
compact by lemma 3.4.3, and the fact that

L is a dieomorphism.
Therefore, extracting a subsequence if necessary, we can suppose
that
L

i
(x, v) (x

, v

). By the theory of dierential equa-


tions, the solution
L
t
(x, v) can be extended beyond
(x,v)
, which
contradicts the maximality of the interval ]
(x,v)
,
(x,v)
[.
Denition 3.4.5 (Lagrangian Gradient). Let L : TM R be a
C
r
Tonelli Lagrangian, with r 2, on the manifold M. Suppose
: U R be a C
k
, k 1 function, dened on the open subset U
of M, we dene the Lagrangian gradient of as the vector eld
grad
L
on U given by
x U,
L
v
(x, grad
L
(x)) = d
x
.
Note that is well-dened, because the global Legendre transform

L : TM T

M is a C
r1
dieomorphism, and (x, grad
L
(x)) =

L
1
(x, d
x
u). It follows that grad
L
is of class C
min(r,k)1
.
More generally, for a function S :]a, b[U R of class C
k
, k
1, where ]a, b[ is an open interval in R, we dene its Lagrangian
gradient as the vector eld grad
L
S
t
, where we S
t
(x) = S(t, x).
It is a vector eld grad
L
S
t
dened on U and dependent on time
t ]a, b[. As a function dened on ]a, b[U, it is of class C
k1
.
3.5 Hamilton-Jacobi and Minimizers
Theorem 3.5.1 (Lagrangian Gradient and Hamilton-Jacobi). Let
L : TM R be a C
r
Tonelli Lagrangian. If S :]a, b[U R is a
C
1
solution of the Hamilton-Jacobi equation
S
t
+H(x,
S
x
) = 0,
102
then, for every absolutely continuous curve : [, ] U, with
[, ] ]a, b[, we have
L() =
_

L((s), (s)) ds S(, ()) S(, ()).


The inequality above is an equality if and only if is a solution of
the time dependent vector eld grad
L
S
t
.
The solutions of the vector eld grad
L
S
t
are necessarily ex-
tremal curves of L. If : [, ] U is such a solution, then
for every absolutely continuous curve
1
: (
ab
([, ], U), with

1
() = (),
1
() = (), and
1
,= , we have L(
1
) > L().
The vector eld grad
L
S
t
is uniquely integrable, i.e. if
i
: I
i

U are 2 solutions of grad
L
S
t
and
1
(t
0
) =
2
(t
0
), for some t
0

I
1
I
2
, then
1
=
2
on I
1
I
2
.
Proof. Let : [, ] U be an absolutely continuous curve. Since
S is C
1
, by lemma 3.1.3 the map [, ] R, t S(t, (t)) is
absolutely continuous and thus by theorem 3.1.2
S(, ()) S(, ()) =
_

_
S
t
(t, (t)) +
S
x
(t, (t))[ (t)]
_
dt.
By the Fenchel inequality, for each t where (t) exists, we have
S
x
(t, (t))[ (t)] H
_
(t),
S
x
(t, (t))
_
+L((t), (t)). ()
Since S satises the Hamilton-Jacobi equation, and (t) exists for
almost all t [a, b], adding S/t(t, (t)) to both sides in (), we
nd that for almost all t [, ]
S
t
(t, (t)) +
S
x
(t, (t))[ (t)] L((t), (t))
By integration, we then conclude that
S(, ()) S(, ())
_

L((t), (t)) dt. ()


We have equality in this last inequality if and only if we have
equality almost everywhere in the Fenchel inequality (), therefore
if and only if (t) = grad
L
S
t
((t)).
103
But the right-hand side of this last equality is continuous and
dened for each t [, ]. It follows that (t) can be extended by
continuity to the whole interval [, ], hence that is C
1
and that
it is a solution of the vector eld grad
L
S
t
. To sum up we have
shown that () is an equality if and only if is a (C
1
) solution
of the time dependent vector eld grad
L
S
t
.
If a curve : [, ] U satises the equality
S(, ()) S(, ()) = L(),
we have already shown that is C
1
. Moreover, for every curve

1
: [, ] U with
1
() = () and
1
() = (), we have
L(
1
) S(, ()) S(, ()) = L().
It follows that is an extremal curve (and thus of class C
r
).
Suppose now that for
1
, and as above we have L(
1
) =
L(). Therefore L(
1
) = S(, ()) S(, ()). By what we
already know, the curve
1
is also an extremal curve and
1
() =
grad
L
S

(()). However, we have () = grad


L
S

(()). The
two extremal curves and
1
go through the same point with the
same speed at t = . Since the Euler-Lagrange vector eld X
L
is
uniquely integrable, these two extremal curves are thus equal on
their common interval of denition [, ].
The last argument also shows the unique integrability of the
vector eld grad
L
S
t
.
In fact, it is possible to show that, under the assumptions
made above on L, a C
1
solution of the Hamilton-Jacobi equation
has a derivative which is Lipschitzian, see [Lio82, Theorem 15.1,
page 255] or [Fat03, Theorem 3.1]. The proof uses the Lagran-
gian gradient of the solution. Consequently, note that we are a
posteriori in the situation of uniqueness of solutions given by the
Cauchy-Lipschitz theorem, since grad
L
S
t
is Lipschitzian.
3.6 Small Extremal Curves Are Minimizers
In this section, we will suppose that our manifold M is provided
with a Riemannian metric g. We denote by d the distance on
M associated to g. If x M, the norm ||
x
on T
x
M is the one
104
induced by g. The projection TM M is denoted by . We
suppose that L : TM M is a C
2
bounded below, such that,
for each (x, v) TM, the second vertical derivative

2
L
v
2
(x, v) is
positive denite as a quadratic form, and that L is superlinear in
each ber of the tangent bundle : TM M.
Theorem 3.6.1. Suppose that L is a Tonelli Lagrangian on the
manifold M, and that inf
(x,v)TM
L(x, v) is nite. Then for each
compact subset K TM there exists a constant
0
> 0 such that
For (x, v) K, the local ow
t
(x, v) is dened for [t[
0
.
For each (x, v) K and for each ]0,
0
], the extremal
curve
(x,v,)
: [0, ] M, t
t
(x, v) is such that for any
absolutely continuous curve
1
: [0, ] M, with
1
(0) =
x,
1
() =

(x, v), and


1
,= we have L(
1
) > L(
(x,v,)
).
Proof. By the compactness of K, we can nd a
1
> 0, such that

t
(x, v) is dened for (x, v) K and [t[
1
.
Since

t[0,
1
]

t
(K) is compact, we can nd a constant C
0
,
which is an upper bound for L on the set

t[0,
1
]

t
(K). With
the notations of the statement, we see that
(x, v) K, [0,
1
], L(
(x,v,)
) C
0
.
In the sequel of the proof, we consider A
0
a compact neighborhood
of
_
t[0,
1
]

t
(K)
_
in M. Since L is superlinear above compact
subsets of M, we can nd C
1
> such that
x A
0
, v T
x
M, L(x, v) |v|
x
+C
1
. (*)
Therefore d(
(x,v,)
(), x) L(
(x,v,)
) C
1
(C
0
C
1
), and
consequently, applying this for each

, we nd that the diam-


eter diam
_

(x,v,)
([0, ])
_
of the image of
(x,v,)
satises
(x, v) K, [0, ], diam
_

(x,v,)
([0, ])
_
2(C
0
C
1
).
By the corollary 2.8.12 and the compactness of K, we can nd

2
]0,
1
] and > 0 such that for each (x, v) K
we have

B
d
(x, ) A
0
;
105
there exists a function S
(x,v)
:]
2
,
2
[

B
d
(x, ) R, which
is a solution of the Hamilton-Jacobi equation
S
(x,v)
t
+H(x,
S
(x,v)
x
) = 0,
with
S
(x,v)
x
(x, 0) =
L
v
(x, v).
In particular, we have v = grad
L
S
(x,v)
0
(x), where we set S
(x,v)
t
(x) =
S
(x,v)
(t, x). Consequently, the solution of the vector eld grad
L
S
t
,
going through x at time t = 0, is t
t
(x, v), see theorem
3.5.1. Since for
3

2
, such that (C
0
C
1
)
3
< , the curve

(x,v,)
takes its values in

B
d
(x, ) for each (x, v) K and all
]0,
3
], by the same Theorem 3.5.1 we obtain that for every

1
: [0, ]

B
d
(x, ) which is absolutely continuous and satises

1
(0) = x and
1
() =
(x,v,)
(), we have L(
1
) > L(
(x,v,)
).
It remains to check that a curve
1
: [0, ] M with
1
(0) = x
and
1
([0, ]) ,

B
d
(x, ) has a much too large action. This is where
we use that
C
2
= infL(x, v) [ (x, v) TM > .
In fact, if
1
([0, ]) ,

B
d
(x, ), there exists > 0 such that

1
([0, [)

B
d
(x, ) and d(
1
(0),
1
()) = . Since

B
d
(x, ) A
0
,
we then obtain from inequality () above
L(
1
[[0, ]) +C
1
.
We can use C
2
as a lower bound of L(
1
(s),
1
(s)) on [, ] to
obtain
L(
1
) +C
1
+C
2
( ),
from which, setting C
3
= min(C
1
, C
2
) > , it follows that
L(
1
) +C
3
,
as soon as
1
(0) = x and
1
([0, ]) ,

B
d
(, ). Since L(
(x,v,)
)
C
0
, to nish the proof of the theorem, we see that it is enough
to choose
0
with
0

3
, and (C
0
C
3
)
0
< .
106
Theorem 3.6.2. Suppose that L is a Tonelli Lagrangian on the
manifold M, and that inf
(x,v)TM
L(x, v) is nite. Fix d a distance
on M coming from a Riemannian metric. If K M is compact,
and C is a strictly positive constant, then there exists a constant

0
> 0 such that, if x K, y M, and ]0,
0
], satisfy d(x, y)
C, then there exists an extremal curve
(x,y,)
: [0, ] M with

(x,y,)
(0) = x,
(x,y,)
() = y, and for every curve absolutely con-
tinuous : [0, ] M which satises (0) = x, () = y, and
,=
(x,y,)
, we have L() > L(
(x,y,)
).
Proof. By the theorem of existence of local extremal curves 2.7.4,
we know that there exists a constant
1
> 0 such that, if d(x, y)
C with ]0,
1
], then there exists an extremal curve
(x,y,)
with

(x,y,)
(0) = x,
(x,y,)
= y, and |
(x,y,)
(0)|
x
2C. However the
set (x, v) TM [ x K, |v|
x
2C is compact in TM and we
can apply the previous theorem 3.6.1 to nd
0

1
satisfying the
conclusions of the theorem we are proving.
3.7 Regularity of Minimizers
In this section we will assume that the Lagrangian L : TM R is
C
r
, with

2
L
v
2
(x, v) denite positive as a quadratic form, for each
(x, v) TM, and L superlinear in the bers of the tangent bundle
TM. We still provide M with a Riemannian metric of reference.
Theorem 3.7.1 (Regularity). Suppose that L is a Tonelli Lag-
rangian on the manifold M. Let : [a, b] M be an absolutely
continuous curve such that L() L(
1
), for each other absolutely
continuous curve
1
: [a, b] M, with
1
(a) = (a),
1
(b) = (b),
then the curve is an extremal curve, and is therefore C
r
.
Proof. By an argument already used many times previously, it
suces to consider the case M = U is an open subset of of R
k
.
Let

W be a compact subset of U, containing ([a, b]) in its interior
W. By the compactness of

W and the uniform superlinearity of
L above compact subsets of M, we have
infL(x, v) [ x

W, v R
k
> .
We can then apply theorem 3.6.2 to the compact subset ([a, b])
contained in the manifold W R
k
(which then plays the role of
107
the M of theorem 3.6.2). Let us rst show that if the derivative
(t
0
) exists for some t
0
[a, b], then coincides with an extremal
curve in the neighborhood of t
0
. We have
lim
tt
0
|
(t) (t
0
)
t t
0
| = | (t
0
)|.
Choosing C > | (t
0
)|, we can then nd > 0 such that
0 < [t t
0
[ |(t) (t
0
)| < C[t t
0
[. ()
Let us apply theorem 3.6.2 with ([a, b]) as the compact subset
and C as the constant, to nd the
0
> 0 given by this theorem.
We can assume that
0
. We will suppose t
0
]a, b[, and let
the reader make the trivial changes in the cases t
0
= a or t
0
= b.
From () we get
|(t
0
+
0
/2) (t
0

0
/2)| < C
0
.
By theorem 3.6.2, the curve
1
: [0,
0
] M which minimizes
the action among the curves connecting (t
0

0
/2) to (t
0
+

0
/2) is an extremal curve. However, the curve [0,
0
] M, s
(s+t
0

0
/2) minimizes the action among the curves connecting
(t
0

0
/2) with (t
0
+
0
/2), since the curve : [a, b] M
minimizes the action for the curves connecting (a) to (b). We
conclude that the restriction of to the interval [t
0

0
/2, t
0
+
0
/2]
is an extremal curve.
Let then O [a, b] be the open subset formed by the points t
0
such that coincides with an extremal curve in the neighborhood
of t
0
. For every connected component I of O, the restriction [I
is an extremal because it coincides locally with an extremal, and
the Euler-Lagrange ow is uniquely integrable. Notice also that
(I) ([a, b]) which is compact therefore by corollary 3.4.4 the
extremal curve [I can be extended to the compact closure

I
[a, b], therefore by continuity [

I is an extremal. If O ,= [a, b],


let us consider a connected component I of O, then

I I is non-
empty and not contained in O. This component can be of one
of the following types ], [, [a, [, ], b]. We will treat the case
I =], [ O. We have , / O. Since [, ] is compact and
[[, ] is an extremal curve, its speed is bounded. Therefore
108
there exists a nite constant C, such that
s ], [, | (s)| < C.
Once again we apply theorem 3.6.2 to the compact subset ([a, b])
and the constant C, to nd
0
> 0 given by that theorem. We
have
|() (
0
/2)|
_

0
/2
| (s)| ds
< C

0
2
.
By continuity, for t > and close to , we do also have
|(t) (
0
/2)| < C[t (
0
/2)].
We can take t > close enough to so that t <
0
/2. By
theorem 3.6.2, the curve coincides with an extremal curve on
the interval [
0
/2, t]. This interval contains in its interior,
and thus O, which is absurd. Therefore we necessarily have
O = [a, b], and is an extremal curve.
Let us summarize the results obtained for M compact.
Theorem 3.7.2. Let L : TM R be a C
r
Tonelli Lagrangian,
with r 2, where M is a compact manifold. We have:
the Euler-Lagrange ow is well-dened complete and C
r1
;
the extremal curves are all of class C
r
;
for each x, y M, each and a, b R, with a < b, there
exists an extremal curve : [a, b] M with : [a, b] M
with (a) = x, (b) = y and such that for every absolutely
continuous curves
1
: [a, b] M, with
1
(a) = x,
1
(b) = y,
and
1
,= we have L(
1
) > L();
if : [a, b] M is an absolutely continuous curve which is
a minimizer for the class (
ac
([a, b], M, then it is an extremal
curve. In particular, it is of class C
r
;
if C R, the set
C
= (
ac
([a, b], M) [ L() C is
compact for the topology of uniform convergence.
Chapter 4
The Weak KAM
Theorem
In this chapter, as usual we denote by M a compact and connected
manifold. The projection of TM on M is denoted by : TM
M. We suppose given a C
r
Lagrangian L : TM R, with r
2, such that, for each (x, v) TM, the second partial vertical
derivative

2
L
v
2
(x, v) is denite > 0 as a quadratic form, and that
L is superlinear in each ber of the tangent bundle : TM M.
We will also suppose that M is provided with a xed Riemannian
metric. We denote by d the distance on M associated with this
Riemannian metric. If x M, the norm | |
x
on T
x
M is the one
induced by the Riemannian metric.
4.1 The Hamilton-Jacobi Equation Revis-
ited
In this section we will assume that we have a Tonelli Lagrangian
L of class C
r
, r 2, on the manifold M. The global Legendre
transform

L : TM T

M is a C
r1
dieomorphism, see Theorem
3.4.2. Its associated Hamiltonian H : T

M R given by
H

L(x, v) =
L
v
(x, v)(v) L(x, v),
is C
r
, and satises the Fenchel inequality
p(v) L(x, v) +H(x, p),
109
110
with equality if and only if p = L/v(x, v), or equivalently (x, p) =

L(x, v). The Hamiltonian ow


H
t
of H is conjugated by

L to the
Euler-Lagrange ow
L
t
of L.
Theorem 4.1.1 (Hamilton-Jacobi). Suppose that L is a C
r
Tonelli
Lagrangian, with r 2 on the manifold M. Call H : T

M R
the Hamiltonian associated to the Lagrangian L.
Let u : M R be a C
1
function. If for some constant c R
it satises the Hamilton-Jacobi equation
x M, H(x, d
x
u) = c
then the graph of du, dened by Graph(du) = x, d
x
u) [ x M,
is invariant under the Hamiltonian ow
H
t
of H.
Moreover, for each x M, the projection t

H
t
(x, d
u
) is
minimizing for the class of absolutely continuous curves.
Of course, if u is of class C
2
the rst part theorem follows from
2.5.10. The second part can be deduced from Theorem 3.5.1. In
fact, as we will see below, the main argument in proof of Theorem
4.1.1 is just a mere repetition of the main argument in the proof
of Theorem 3.5.1. Although this proof of Theorem 4.1.1 could be
rather short, we will cut it down in several pieces, because on doing
so we will be able to nd a notion of C
0
solution of the Hamilton-
Jacobi equation. With this notion we will prove, in contrast to
the C
1
case, that such a C
0
solution does always exist, see 4.7.1
below, and we will explain its dynamical signicance.
We start by studying the meaning for C
1
functions of the
Hamilton-Jacobi inequality.
Proposition 4.1.2. Suppose that L is a Tonelli Lagrangian on
the manifold M. Call H : T

M R the Hamiltonian associated


to the Lagrangian L.
Let c R be a constant, and let u : U R be a C
1
function
dened on the open subset U M. If u satises the inequality
x V, H(x, d
x
u) c ()
then for every absolutely continuous curve : [a, b] U, with
a < b, we have
u((b)) u((a))
_
b
a
L((s), (s)) ds +c(b a). ()
111
Conversely, if inequality () holds for every C

curve : [a, b]
M then () holds.
Proof. If : [a, b] U is absolutely continuous then by Lemma
3.1.3 the function u is also absolutely continuous, and therefore
u((b)) u((a))
_
b
a
d
(s)
u( (s)), ds. ( )
By Fenchels inequality, at each point s [a, b] where (s) exists
we can write
d
(s)
u( (s)) L((s), (s)) +H((s), d
(s)
u).
Suppose that () holds, we get
d
(s)
u( (s)) L((s), (s)) +c.
Integrating, and comparing with ( ), yields ().
Conversely, suppose that () for every C

curve : [a, b]
U. Fix x U. For a given v T
x
M we can nd a C

curve
: [, ] U, with > 0, (0) = x, and (0) = v. Then writing
condition () for every restriction [[0, t], t ]0, ], we obtain
u((t)) u((0))
_
t
0
d
(s)
u( (s)), ds +c(t 0).
Dividing both sides by t > 0, and letting t 0 yields
d
x
u(v) L(x, v) +c.
Since this is true for every v T
x
M, we conclude that H(x, d
x
u) =
sup
vTxM
d
x
u(v) L(x, v) c.
This suggests the following denition.
Denition 4.1.3 (Dominated Function). Let u : U R be a
function dened on the open subset U M. If c R, we say that
u is dominated by L + c on U, which we denote by u L + c, if
for each continuous piecewise C
1
curve : [a, b] U we have
u((b)) u((a))
_
b
a
L((s), (s)) ds +c(b a). (D)
If U = M, we will simply say that u is dominated by L +c.
We will denote by D
c
(U) the set of functions u : U R
dominated by L +c
112
In the denition above, we have used continuous piecewise C
1
curves in (D) instead of C

curves because if condition (D) holds


for C

curves it also holds for continuous piecewise C


1
curves, and
even for absolutely continuous curves, see the following exercise.
Exercise 4.1.4. Suppose L is a Tonelli Lagrangian on the man-
ifold M. Let u : U R be a function dened on the open subset
U of M such that the inequality (D) of Denition 4.1.3 holds for
every C

curve .
1) Show that (D) holds for C
1
curves. [Indication: Use a den-
sity argument.]
2) Show that (D) holds for a continuous piecewise C
1
curve.
[Indication: If : [a, b] U, there exists a
0
< a
1
< < a
n
= b
such that [[a
i
, a
i+1
] is C
1
.]
3) Show that (D) holds for absolutely continuous curves. For
this x such a curve : [a, b] U, if L() = +, there is nothing
to prove. Therefore we can assume that
_
b
a
L((s), (s)) ds < +.
We set
() = sup
_
t
t

L((s), (s)) ds [ t

t, t t

.
a) Show that () 0, when 0.
Fix K, K

U compact neighborhoods of ([a, b]) with K



K

.
b) Show that there exists
0
such that any absolutely continu-
ous curve : [c, d]

K

, with a c d b, c d
0
, (c) =
(c), (d) = (d), and L() (
0
), takes values only in K. [In-
dication: See the proof of Theorem 3.6.1. Notice that L is bounded
below on the subset (x, v) TM [ x K

.]
c) Show that for for every c, d [a, b], with c d and dc
0
,
there exists an absolutely continuous curve : [c, d] M which
satises:
(c) = (c), (d) = (d).
([c, d]) K.
113
is a minimizer for the class of absolutely continuous curves
with values in

K

.
d) Conclude.
Of course, the notion of dominated function does not use any
dierentiability assumption on the function. Therefore we can use
it as a notion of subsolution of the Hamilton-Jacobi equation. This
notion is equivalent to the notion of viscosity subsolution as we
will see in chapters 7 and 8.
The next step necessary to prove Theorem 4.1.1 is the intro-
duction of the Lagrangian gradient. Let us recall, see denition
3.4.5, that for u : U R is a C
1
function, its Lagrangian gradient
grad
L
u is the vector eld dened on U by
(x, d
x
u) =

L(x, grad
L
u(x)).
It follows that
Graph(du) =

L[Graph(grad
L
u)],
where Graph(grad
L
u) = (x, grad
L
u(x)) [ x U. Since
H
t
and
L
t
are conjugated by

L, invariance of Graph(du) under
H
t
is
equivalent to invariance of Graph(grad
L
u) under
L
t
. Therefore
the following proposition nishes the proof of Theorem 4.1.1
Proposition 4.1.5. Let L be a Tonelli Lagrangian on the man-
ifold M. If u : U R is a C
1
function which satises on U the
Hamilton-Jacobi equation
H(x, d
x
u) = c,
for some xed c R, then every solution : [a, b] U of the
vector eld grad
L
u satises
u((b)) u((a)) =
_
a
b
L((s), (s)) ds +c(b a).
It follows that solutions of grad
L
u are minimizing for the class of
absolutely continuous curves with values in U, and that it must
be an extremal of class C
2
.
Moreover, the graph of grad
L
u is locally invariant by the
Euler-Lagrange ow. If U = M and M is compact then the graph
Graph(grad
L
u) is invariant by the Euler-Lagrange ow.
114
Proof. Since (y, d
y
u) =

L(y, grad
L
u(y)), we have the following
equality in the Fenchel inequality
d
y
u(grad
L
u(y)) = L(y, grad
L
u(y)) +H(y, d
y
u).
Taking into account that H(y, d
y
u) = c, and using the equality
along a solution : [a, b] U of the vector eld grad
L
u, we get
d
(t)
u( (t)) = L((t), (t)) +c.
If we integrate we get
u((b)) u((a)) =
_
a
b
L((s), (s)) ds +c(b a).
This implies that is a minimizer for absolutely continuous curves,
and is therefore a C
2
extremal. In fact, if : [a, b] U is a
curve, by Proposition 4.1.2, we have u((b)) u((a)) L(). If
(a) = (a) and (b) = (a), we obtain
L() +c(b a) = u((b)) u((a))
= u((b)) u((a))
L().
We now show that the Graph(grad
L
u) is locally invariant by
the Euler-Lagrange ow. Given x U, since the vector eld
grad
L
u is continuous, we can apply the Cauchy-Peano Theorem,
see [Bou76], to nd a map : [, ] V , with > 0 and V an
open neighborhood of x, such that for every y V the curve t

y
(t) = (y, t) is a solution of grad
L
u with
y
(0) = y. Therefore

y
(t) = grad
L
u(
y
(t)).
But we know that
y
is an extremal, hence its speed curve is
t
L
t
(y,

y
(0)). Therefore, for every (t, y) [, ] V , we have

L
t
(y,

y
(0)) = (
y
(t), grad
L
u(
y
(t))) Graph(grad
L
u). ()
If U = M is compact we can nd a nite family of [
i
,
i
]
V
i
, i = 1, . . . , satisfying (), and such that M =

i=1
V
i
. Setting
= min

i=1

i
> 0, we obtain
t [, ],
L
t
(Graph(grad
L
u)) Graph(grad
L
u).
115
Since
L
t
is a ow dened for all t R, we obtain
t R,
L
t
(Graph(grad
L
u)) Graph(grad
L
u).
Using the inclusion above for t and t yields
L
t
(Graph(grad
L
u)) =
Graph(grad
L
u), for all t R.
Exercise 4.1.6. Under the hypothesis of Proposition 4.1.5, if x
U, show that the orbit
L
t
(x, grad
L
u(x)) is dened at least on an
interval ], [ such that (t) =
L
t
(x, grad
L
u(x)) U for every
t ], [, and t (t) leaves every compact subset of U as t tends
to either or . Show that this curve is a solution of grad
L
u
on ], [. Show that, any other solution : I U of , grad
L
u,
with (0) = x, satises I ], [ and = on I.
Proposition 4.1.5 suggests the following denition
Denition 4.1.7 (Calibrated Curve). Let u : U R be a func-
tion and let c R be a constant, where U is an open subset of
M. We say that the (continuous) piecewise C
1
curve : I U,
dened on the interval I R is (u, L, c)-calibrated, if for every
t t

I, with t t

, we have
u((t

)) u((t)) =
_
t

t
L((s), (s)) ds +c(t

t).
Although the following proposition is an immediate conse-
quence of the denition of calibrated curve, it will be used often.
Proposition 4.1.8. Let L be a Tonelli Lagrangian dened on the
manifold M. Suppose u : U R is a C
1
function dened on
the open subset U M, and c R. If the curve : I U is
(u, L, c)-calibrated, then for any subinterval I

I the restriction
[I

is also (u, L, c)-calibrated.


The following theorem explains why calibrated curves are spe-
cial.
Theorem 4.1.9. Suppose that L is a Tonelli Lagrangian on the
manifold M. Let u : U R be a function dened on the open
subset U M. Assume that u L + c, where c R. Then any
continuous piecewise C
1
(u, L, c)-calibrated curve : I U is
116
necessarily a minimizing curve for the class of continuous piecewise
C
1
on U. Therefore it is an extremal curve and it is as smooth as
the Lagrangian L.
Proof. We x a compact interval [t, t

] I, with t t

. If :
[t, t

] U is a (continuous) piecewise C
1
, from u L + c, it
follows that
u((t

)) u((t)) L() +c(t

t).
Moreover, since is (u, L, c)-calibrated, we have equality when
= [[t, t

]. If (t

) = (t

) and (t) = (t), we obtain


L([[t, t

]) +c(t

t) = u((t

)) u((t)) L() +c(t

t),
hence L([[t, t

]) L(), and is therefore a minimizing curve.


This implies that is an extremal and is as smooth as L, see
Proposition 2.3.7.
We now can give a characterization of C
1
solutions of the
Hamilton-Jacobi equation which does not involve the derivative.
Proposition 4.1.10. Let L be a Tonelli Lagrangian dened on
the manifold M. If u : U R is a C
1
function dened on the open
subset U M, and c R, the following conditions are equivalent:
(i) The function u satises Hamilton-Jacobi equation
x U, H(x, d
x
u) = c.
(ii) The function u is dominated by L +c, and for every x U
we can nd > 0 and a C
1
curve : [, ] U which is
(u, L, c)-calibrated, and satises (0) = x.
(iii) The function u is dominated by L +c, and for every x U
we can nd > 0 and a C
1
curve : [, 0] U which is
(u, L, c)-calibrated, and satises (0) = x.
(iv) The function u is dominated by L +c, and for every x U
we can nd > 0 and a C
1
curve : [0, ] U which is
(u, L, c)-calibrated, and satises (0) = x.
117
Proof. We prove that (i) implies (ii). The fact that u L + c
follows from Proposition 4.1.2. By Proposition 4.1.5, we can take
for any solution of grad
L
u with (0) = x. Again such a solution
exists by the Cauchy-Peano Theorem, see [Bou76], since grad
L
u
is continuous.
Obviously (ii) implies (iii) and (iv).
It remains to prove that either (iii) or (iv) implies (i). We
show that (iii) implies (i), the other implication being similar.
From Proposition 4.1.2, we know that
x U, H(x, d
x
u) c.
To show the reversed inequality, we pick a (u, L, c)-calibrated C
1
curve : [, 0] M with (0) = x. For every t [0, ], we have
u((0)) u((t)) =
_
0
t
L((s), (s)) ds +ct.
If we divide by t > 0, after changing signs in the numerator and
denominator of the left hand side, we get
u((t)) u((0))
t
=
1
t
_
0
t
L((s), (s)) ds +c.
If we let t 0 taking into account that (0) = x, and that both
u and are C
1
, we obtain
d
x
u( (0)) = L(x, (0)) +c.
But by Fenchels inequality H(x, d
x
u) d
x
u( (0)) L(x, (0)),
therefore H(x, d
x
u)) c.
Therefore we could take any one of condition (ii), (iii) or (iv)
above as a denition of a continuous solution of the Hamilton-
Jacobi equation. In fact, a continuous function satisfying (ii) is
necessarily C
1
, see ?? below. Condition (iii) and (iv) lead both to
the notion of continuous solutions of the Hamilton-Jacobi equa-
tion. The two sets of solutions that we obtain are in general dif-
ferent, and they both have a dynamical meaning as we will see
later.
118
We will be mainly using continuous solution of the Hamilton-
Jacobi equation which are dened on a compact manifold M. No-
tice that for C
1
solutions of the Hamilton-Jacobi equation dened
on M, by the last part of Proposition 4.1.5, in condition (ii), (iii),
and (iv) of Proposition 4.1.10 we can impose = +. This justi-
es the following denition (see also ?? below).
Denition 4.1.11. Let L be a Tonelli Lagrangian on the compact
manifold M. A weak KAM solution of negative type (resp. of
positive type) is a function u : M R for which there exists
c R such that
(1) The function u is dominated by L +c.
(2) For every x M we can nd a (u, L, c)-calibrated C
1
curve
:] , 0] M (resp. : [0, +[ M) with (0) = x.
We denote by o

(resp. o
+
) the set of weak KAM solutions of
negative (resp. positive) type.
We will usually use the notation u

(resp. u
+
) to denote an
element of o

(resp. o
+
).
4.2 Dominated Functions and the Ma ne Crit-
ical Value
We now establish some properties of dominated function. Before
doing that let us recall that the notion of locally Lipschitz function
makes perfect sense in a manifold M. In fact, a function u :
M R is said to be locally Lipschitz if for every coordinate
chart : U M, the function u is locally Lipschitz on the
open subset U of some Euclidean space. Since all Riemannian
metrics are equivalent above compact subsets, it is equivalent to
say that u is locally Lipschitz for one distance (or for all distances)
d coming from a Riemannian metric. If u : X Y is a Lipschitz
map between the metric spaces X, Y we will denote by Lip(u) its
smallest Lipschitz constant
Lip(u) = sup
d(u(x), u(x

))
d(x, x

)
where the supremum is taken over all x, x

X, with x ,= x

.
119
Proposition 4.2.1. Suppose L is a Tonelli Lagrangian on the
manifold M. Endow M with a distance d and TM a norm (x, v)
|V |
x
both coming from the same Riemannian metric on M (Note
that d well dened and nite on each connected component of M).
Let U be an open subset of M, and c R. We have the
following properties:
(i) The set D
c
(U) of functions u : U R dominated by L + c
is a closed convex subset of the set of functions U R for
the topology of point-wise convergence. Moreover, if k R,
we have u D
c
(U) if and only if u +k D
c
(U).
(ii) Every function in D
c
(U) is locally Lipschitz. More precisely,
for every x
0
U, we can nd a compact neighborhood V
x
0
such that for every u D
c
(U) the Lipschitz constant of u[V
x
0
is A
Vx
0
+c, where
A
Vx
0
= supL(x, v) [ (x, v) TM, x V
x
0
, |v|
x
= 1 < +.
In particular the family of functions in D
c
(U) is locally equi-
Lipschitzian.
(iii) If M is compact and connected, and u : M R is dened
on the whole of M and is dominated by L + c, then u is
Lipschitz. More precisely, then Lipschitz constant Lip(u) of
u is A+c, where
A = supL(x, v) [ (x, v) TM, |v|
x
= 1.
In particular the family of functions in D
c
(M) is l equi-
Lipschitzian.
(iv) Moreover, if M is compact and connected, then every Lip-
schitz function u : M R is dominated by L + c for some
c R. More precisely, given K [0, +[, we can nd c
K
such that every u : M R, with Lip(u) K, satises
u L +c
K
.
(v) Suppose that c, k R, and that U is a connected open subset
of M. If x
0
U is xed, then the subset u D
c
(U) [
[u(x
0
)[ k is a compact convex subset of (
0
(U, R) for the
topology of uniform convergence on compact subsets.
120
Proof. Its is obvious from the denition of domination that D
c
(U)
is convex and that u D
c
(U) if and only if u +k D
c
(U). For a
xed continuous piecewise C
1
curve : [a, b] U the set T
,c
of
function u : U R such that
u((b)) u((a)) =
_
b
a
L((s), (s)) ds +c(b a)
is clearly closed in the topology of point-wise convergence. Since
D
c
(U) is the intersection of the T
,c
for all s with values in U,
it is also closed.
To prove (ii), let u D
c
(U). Fix x
0
U. We can nd a
compact neighborhood V
x
0
U of x
0
, such that for every x, y

V
x
0
we can nd a geodesic : [a, b] M parametrized by unit
length such that (a) = x, (b) = y and length() = b a =
d(x, y). Since V
x
0
is compact, the constant
A
Vx
0
= supL(x, v) [ (x, v) TM, x V
x
0
, |v|
x
= 1
is nite. Since | (s)| = 1, for every s [a, b], we have L((s), (s))
A
Vx
0
, therefore we obtain
u((b)) u((a))
_
b
a
L((s), (s)) ds +c(b a)

_
b
a
A
Vx
0
ds +c(b a)
= (A
Vx
0
+c) length()
= (A
Vx
0
+c)d(x, y).
To prove (iii), it suces to observe that, when M is compact
and connected, we can take V
x
0
= M in the argument above.
Suppose now that M is compact and connected. If we x
K 0, by the superlinearity of L we can nd A(K) > such
that
(x, v) TM, L(x, v) K|v|
x
+A(K).
therefore if : [a, b] M is a continuous piecewise C
1
curve,
applying this inequality for (x, v) = ((s), (s)) and integrating
we get
K length() L() A(K)(b a),
121
and hence
Kd(((b), (a)) L() A(K)(b a).
If u : M R has Lipschitz constant K, we therefore obtain
u((b)) u((b)) L() A(K)(b a).
This proves (iv) with c
K
= A(K) < +.
It remains to prove (v). Set c = u D
c
(U) [ [u(x
0
)[ k.
By (1) this set is clearly closed for the topology of point-wise
convergence. It suces to show that for each compact subset
K U, the set of restrictions c[K = u[K [ u c is relatively
compact in (
0
(K, R). We apply Ascolis Theorem, since this the
family c[K is locally equi-Lipschitz by (ii), it suces to check that
sup[u(x)[ [ x K, u c < +.
Since U is connected locally compact and locally connected, en-
larging K if necessary, we can assume that K is connected and
contains x
0
. Again by (ii), we can cover K by a nite number of
open sets V
1
, . . . , V
n
, and nd nite numbers k
i
0, i = 1, . . . , n
such Lip(u[V
i
) k
i
for every u D
c
(U), and every i = 1, . . . , n. If
x K, by connectedness of K, we can nd i
1
, . . . , i

1, . . . , n
such that V
i
j
V
i
j+1
K ,= , j = 1, . . . , , x
0
V
i
1
, x V
i

. By
assuming minimal with this properties, we get that the i
j
are all
distinct, therefore n. We can choose x
j
V
i
j
V
i
j+1
K, for
i = 1, . . . , 1. Therefore setting x

=, for u D
c
(U), we obtain
[u(x) u(x
0
)[
1

j=0
[u(x
j+1
) u(x
j
)[

j=0
k
i
j
d(x
j+1
, x
j
)
diam(K)
n
max
i=1
k
i
ndiam(K)
n
max
i=1
k
i
.
Therefore [u(x)[ k +ndiam(K) max
n
i=1
k
i
, for every x K and
every u c.
122
Proposition 4.2.2. Let L be a Tonelli Lagrangian on the man-
ifold M. Suppose that u : U R is dened on the open subset
U M and that it is dominated by L + c. Then at each point
x U where the derivative d
x
u exists, we have
H(x, d
x
u) c.
By Rademachers Theorem 1.1.10, the derivative d
x
u exists almost
everywhere on U, the above inequality is therefore satised almost
everywhere.
Proof. Suppose that d
x
u exists. We x v T
x
(M). Let : [0, 1]
U be a C
1
curve such that (0) = x and (0) = v. Since u L+c,
we have
t [0, 1], u((t)) u((0))
_
t
0
L((s), (s)) ds +ct.
By dividing this inequality by t > 0 and letting t tend to 0, we
nd d
x
u(v) L(x, v) +c and hence
H(x, d
x
u) = sup
vTxM
d
x
u(v) L(x, v) c.
We now prove the converse of Proposition 4.2.2.
Proposition 4.2.3. Let L be a Tonelli Lagrangian on the man-
ifold M. Call H the Hamiltonian associated to L. Suppose that
u : U R is a locally Lipschitz function dened on the open
subset U of M. By Rademachers Theorem 1.1.10, the derivative
d
x
u exists for almost all x U. If there exists a c such that
H(x, d
x
u) c, for almost all x U, then u L +c.
Proof. Using a covering of a curve by coordinates charts, it is not
dicult to see that we can assume that U is an open convex set
in R
k
. We call R the set of points x U were d
x
u exists and
H(x, d
x
u) c. By assumption U R is negligible for Lebesgue
measure.
We show that rst that u((b)) u((a)) L() + c(b a),
for an ane segment : [a, b] U. To treat the case where is
constant we have to show that L(x, 0) + c 0 for every x U.
123
If x R, this is true since L(x, 0) + c L(x, 0) + H(x, d
x
u)
d
x
u(0) = 0, where the second inequality is a consequence of part (i)
of Fenchels Theorem 1.3.6. Since R is dense in U, the inequality
L(x, 0) + c 0 is therefore true on the whole of U. We now
assume that the ane segment is not constant. We can then write
(s) = x + (t a)v, with |v| = r > 0. We call S the set of
vectors w such that |w| = r, and the line D
w
= x +tw [ t R
intersects R in a set of full linear measure in U D
w
. By Fubinis
Theorem, the set S itself is of full Lebesgue measure in the sphere
w R
k
[ |w| = r. Hence we can nd a sequence v
n
S with
v
n
v, when n . Dropping the rst ns, if necessary, we can
assume that the ane curve
n
: [a, b] R
k
, t x + (t a)v
n
is
contained in fact in U. By the denition of the set S, for each n,
the derivative d
n(t)
u exists and veries H(
n
(t), d
n(t)
u) c for
almost every t [a, b]. It follows that the derivative of u
n
is
equal to d
n(t)
u(
n
(t)) at almost every t [a, b]. Using again part
(i) of Fenchels Theorem 1.3.6, we see that
du
n
dt
(t) = d
n(t)
u(
n
(t))
H(
n
(t), d
n(t)
u) +L(
n
(t),
n
(t))
c +L(
n
(t),
n
(t)),
Since u
n
is Lipschitz, we obtain
u(
n
(b)) u(
n
(a)) =
_
b
a
du
n
dt
(t) dt

_
b
a
c +L(
n
(t),
n
(t)) dt
= L(
n
) +c(b a).
Since
n
converges in the C
1
topology to we obtain
u((b)) u((a)) L() +c(b a).
Of course, we now have the same inequality for any continuous
piecewise ane segment : [a, b] U. To show the inequality
u((b)) u((a)) L() + c(b a) for an arbitrary C
1
curve
: [a, b] U, we introduce piecewise ane approximation
n
:
124
[a, b] U in the usual way. For each integer n 1, the curve
n
is ane on each of the intervals [a+i(ba)/n, a+(i+1)(ba)/n],
for i = 0, . . . , n 1, and
n
(a + i(b a)/n) = (a + i(b a)/n),
for i = 0, . . . , n. The sequence
n
converges uniformly on [a, b] to
. Since
n
(a) = (a) and
n
(b) = (b), we obtain from what we
just proved
u((b)) u((a)) L(
n
) +c(b a). ()
The derivative
n
(t) exists for each n at each t in the complement
A of the countable set a+i(b a)/n [ n 1, i = 0, . . . , n. More-
over, using the Mean Value Theorem, the sequence
n
[A converges
uniformly to [A, as n , since is C
1
. Using the fact the the
set ((t), (t)) [ t [a, b] is compact and the continuity of L, it
follows that the sequence of maps A R, t L(
n
(t),
n
(t)) con-
verges uniformly to A R, t L((t), (t)), therefore L(
n
)
L(), since [a, b] A is countable. passing to the limit in the above
inequality () we indeed obtain
u((b)) u((a)) L() +c(b a).
Denition 4.2.4 (Hamiltonian constant of a function). If u :
U R is a locally Lipschitz function, we dene H
U
(u) as the es-
sential supremum on U of the almost everywhere dened function
x H(x, d
x
u).
We summarize the last couples of Propositions 4.2.2 and 4.2.3
in the following theorem.
Theorem 4.2.5. Suppose that L is a Tonelli Lagrangian on the
manifold M. Let U be an open subset of M. A function u : U R
is dominated by L + c on U, for some c R, if and only if it is
locally Lipschitz and c H
U
(u).
Denition 4.2.6 (Ma nes Critical Value). If L is a Tonelli Lag-
rangian on the connected compact manifold M, the Ma ne critical
value of L is the constant c[0] (or c
L
[0], if we need to precise the
Lagrangian) dened by
c[0] = infc R [ u : M R, u L +c.
125
Theorem 4.2.7. The Ma ne critical value c[0] of L is nite. In
fact we have c[0] inf
xM
L(x, 0).
If u : M R is dominated by L + c for some c R, then
c c[0]. Moreover, there exists a function u
0
: M R such that
u
0
L +c.
Proof. For a given x, the constant curve
x
: [0, 1] x has action
equal to L(x, 0). Therefore, if u L+c,we have 0 = u(x)u(x)
L(x, 0)+c1 hence c L(x, 0). Since this is true for every x M,
we obtain c inf
xM
L(x, 0), for every c for which we can nd
u : M R with u L + c.taking the inmum over all such
c yields c[0] inf
xM
L(x, 0), which is of course nite by the
compactness of M.
By denition of c[0], if c is such that there exists u : M R
with u L +c we have c c[0].
It remains to nd u : M R such that u L + c[0]. By
denition of c[0] we can nd a sequence c
n
c[0] of numbers, and
a sequence of functions u
n
: M R, with u
n
L + c
n
, for each
n. We now x x
0
M. By (i) of Proposition 4.2.1, the function
u
n
u
n
(x
0
) is also dominated by L+c
n
. therefore we can assume
u
n
(x
0
) = 0, for every n. We could apply part (v) of Proposition
4.2.1, to nish the proof by having a subsequence of u
n
converge.
In fact, it is easier to argue directly. We dene u : M [, +]
by
x M, u(x) = liminf
n
u
n
(x).
Since u
n
(x
0
) = 0, we have u(x
0
) = 0. Given a continuous piecewise
C
1
curve : [a, b] M, since u
n
L +c
n
, we have
u
n
((b)) u
n
((a)) +
_
b
a
L((s), (s)) ds +c
n
(b a).
Since c
n
c[0], by taking the liminf in the equality above, we
obtain
u((b)) u((a)) +
_
b
a
L((s), (s)) ds +c[0](b a). ((*))
Since u(x
0
) = 0, using a continuous piecewise C
1
curve starting
at a given point x M and ending at x
0
, we obtain from (*)
that u(x) > . Using a curve starting at x
0
and ending at x
126
we conclude u(x) < +. Finally, the function u is nite valued
everywhere. since (*) is true for an arbitrary continuous piecewise
C
1
curve : [a, b] M. We obtain u L +c[0].
4.3 Defect and Calibration of Curves
We now study the properties of calibrated curves. It is convenient
to introduce the notion of defect of a curve on an interval with
respect to a function u L +c.
Denition 4.3.1 (Defect of a Curve). Let U be an open subset of
M, and c R. If u : U R is dominated by L+c and : I U
is a continuous piecewise C
1
curve, for [, ] I, with ,
we dene the defect T(, u, c; , ) of the curve on the interval
[, ] for the (L +c)-dominated function u by
T(, u, c; , ) =
_

L((s), (s)) dsc()(u(()u(()).


Of course, in the denition of defect, there is no need to
assume u L + c. However, this denition is useful only when
u L +c, as we will presently see.
Proposition 4.3.2. Let L be a Tonelli Lagarangian on the man-
ifold M. Suppose that u : U R is a continuous function dened
on the open subset U M, with u L+c, and that : I U is a
continuous piecewise C
1
curve. We have the following properties:
(1) If [, ] I, with , then
T(, u, c; , ) 0.
(2) For every k R, we have T(, u, c; , ) = T(, u+k, c; , ).
(3) If we dene the curve
t
0
by
t
0
(s) = (t
0
+ s), then its
interval of denition is I t
0
= s t
0
[ s , and for every
[, ] I t
0
, with , then
T(
t
0
, u, c; , ) = T(, u, c; +t
0
, +t
0
).
127
(4) [Chasles Property] If a
,
a
2
, a
3
I, with a
1
a
2
a
3
,
then
T(, u, c; a
1
, a
3
) = T(, u, c; a
1
, a
2
) +T(, u, c; a
2
, a
3
).
(5) If
1
,
1
, , I, with
1

1
, then
T(, u, c; , ) T(, u, c;
1
,
1
) 0.
(6) The function (, ) T(, u, c; , ) is continuous on the
set (, ) I I [ .
(7) If
n
: [a, b] U is a sequence of C
1
curves which converges
to the C
1
curve

: [a, b] U in the C
1
topology then
T(
n
, u, c; a, b) T(

, u, c; a, b).
(8) Let u
n
: U R is a sequence of functions, with u
n
L+c
n
,
where c
n
R. If c
n
c and u
n
(x) u(x) at every point
x U, then T(, u
n
, c
n
; a, b) T(, u, c; a, b).
Proof. Claim (1) is easy since U L +c on U implies
u(() u(()
_

L((s), (s)) ds +c( ).


Claims (2,3,4) follow easily from the denition of the defect.
For claim (5), by Chasles Property (4), we have
T(, u, c; , ) = T(, u, c; ,
1
)+T(, u, c;
1
,
1
)+T(, u, c;
1
, ).
But by claim (1), we have T(, u, c; ,
1
) 0 and T(, u, c;
1
, )
0.
Claims (6,7) and (8) can also be obtained from the denition
of the defect since a dominated function is locally Lipschitz, and
the action of the Lagrangian for C
1
curves is continuous in the C
1
topology on the space of C
1
curves.
The rst corollary we obtain is a simplication of the denition
of a (u, L, c)-calibrated curve on a compact interval when u L+c.
128
Corollary 4.3.3. Let L be a Tonelli Lagarangian on the manifold
M. Suppose that u : U R is a continuous function dened on
the open subset U M, with u L + c, and that : [a, b] U
is a continuous piecewise C
1
curve. the following properties are
equivalent
(i) The curve is (u, L, c)-calibrated.
(ii) We have
u((b)) u((a) =
_
b
a
L((s), (s)) ds +c(b a).
(iii) The defect T(, u, c; a, b) is equal to 0.
Proof. It is clear that (ii) and (iii) are equivalent. Obviously if
is (u, L, c)-calibrated we have
u((b)) u((a) =
_
b
a
L((s), (s)) ds +c(b a).
It remains to prove that (iii) implies (i). If T(, u, c; a, b) = 0,
from claims (1) and (5) of Proposition 4.3.2 above, we get that
T(, u, c; , ) = 0, for every subinterval [, ] [a, b], and hence
u(() u(() =
_

L((s), (s)) ds +c( ),


hence is (u, L, c)-calibrated on the interval [a, b].
We now state some of the properties of calibrated curves.
Corollary 4.3.4. Let L be a Tonelli Lagarangian on the manifold
M. Suppose that u : U R is a continuous function dened on
the open subset U M, that : I U is a continuous piecewise
C
1
curve. We have the following properties:
(1) If : I M is (u, L, c)-calibrated, then for every subinterval
I

I the restriction [I

is also (u, Lc)-calibrated.


(2) If I

is a subinterval of I and the restriction [I

is (u, L, c)-
calibrated, then is (u, L, c)-calibrated on the interval

I

I.
129
(3) Suppose that I is a nite union of subintervals I
1
, . . . , I
n
. If
[I
i
is (u, L, c)-calibrated, for i = 1, . . . , n, then is (u, L, c)-
calibrated (on I).
(4) Suppose that I =
nN
I
i
, where each I
i
is an interval and
I
i
I
i+1
. If each [I
i
is (u, L, c)-calibrated, then is
(u, L, c)-calibrated (on I).
(5) For every t
0
I, there exists a largest subinterval I
t
0
I
containing x
0
on which is (u, L, c)-calibrated. Moreover
I
t
0
=

I
t
0
I.
(6) If k R, then is (u, L, c)-calibrated if and only if it is
(u +k, L, c)-calibrated.
(7) If t
0
R, then is (u, L, c)-calibrated on I if and only if
the curve s (s +t
0
) is (u, L, c)-calibrated on the interval
I t
0
= t t
0
[ t I.
(8) If
n
: [a, b] U is a sequence of C
1
curves which converges
to the C
1
curve

: [a, b] U in the C
1
topology, and

n
is (u, L, c)-calibrated for every n then the curve is also
(u, L, c)-calibrated.
Proof. Claim (1) was already given as Proposition 4.1.8. It is a
simple consequence of the denition of a calibrated curve.
To prove claim (2), consider a compact subinterval [a, b]
I

I

, with a < b. For n large enough we have a +1/n < b 1/n,


and [a+1/n, b1/n] I

, therefore by by claim (1) and Corollary


4.3.3 above we get T(, u, c; a + 1/n, b 1/n) = 0. By claim (6)
of Proposition 4.3.2 above T(, u, c; a, b) = 0. Therefore [[a, b] is
(u, L, c)-calibrated for every compact subinterval [a, b] of

I

. Hence
[

is (u, L, c)-calibrated.
To prove (3), we x [a, b] I. By (2) we can assume that each
I
i
[a, b] is a compact interval [a
i
, b
i
]. Extracting a minimal cover
and reindexing, we can assume
a = a
1
a
2
b
1
a
3
b
2

a
i+1
b
i
a
n
b
n1
b
n
= b. ()
130
It suces to prove by induction on i that T(, u, c; a
1
, b
i
) = 0
For i = 1, this follows from the hypothesis that [I
1
is (u, L, c)-
calibrated. Let us now do the induction step from i to i + 1.
By (), we have [b
i
, b
i+1
] [a
i+1
, b
i+1
] I
i+1
. Since [I
i+1
is (u, L, c)-calibrated, we therefore obtain T(, u, c; b
i
, b
i+1
) = 0.
From Chasles Property, claim (4) of Proposition 4.3.2, it follows
that T(, u, c; a
1
, b
i+1
) = 0.
To prove (4), it suces to observe that if [a, b] I is a compact
subinterval, then for n large enough we have [a, b] I
n
.
To prove (5), call the family of subinterval J I such that
t
0
J and [J is (u, L, c)-calibrated. Notice that is not empty
since [t
0
, t
0
] . Since t
0


J
J, the union I
t
0

J
J is an
interval. We have to show that is (u, L, c)-calibrated on I
t
0
. Let
[a, b] I
t
0
. We can nd J
1
, J
2
with a J
1
, b J
2
. the union
J
3
= J
1
J
2
is an interval because t
0
J
1
J
2
. Therefore by (3),
the restriction [J
3
is (u, L, c)-calibrated. This nishes the proof
since [a, b] J
3
.
Claim (6), follows from Corollary 4.3.3 above and claim (2)
of Proposition 4.3.2. In the same way claim (7), follows from
Corollary 4.3.3 above and claim (3) of Proposition 4.3.2. Claim
(8) follows from Corollary 4.3.3 above and claim (7) of Proposition
4.3.2.
Exercise 4.3.5. Let L be a Tonelli Lagarangian on the manifold
M. Suppose that u : U R is a continuous function dened on
the open subset U M, that : I U is a continuous piecewise
C
1
curve. Suppose that I =

nN
I
n
, with I
n
a subinterval I of I
on which is (u, L, c)-calibrated. (Do not assume that the family
I
n
, n N is increasing.) Show that is (u, L, c)-calibrated on I.
[Indication: Reduce to the case I = [a, b] and each I
n
compact.
Show, using part (5) of Proposition 4.3.4 above, that you can as-
sume that the I
n
are pairwise disjoint. Call F the complement in
[a, b] of the union

nN

I
n
. Show that F is countable. Use Baires
Category Theorem to show that if F a, b is not empty then it
must has an isolated point.]
Let us now show that innite calibrated curves can only exist
for the Ma ne critical value c[0].
131
Proposition 4.3.6. Suppose u : M R is L + c. If it admits
a (u, L, c)-calibrated curve : I M, with I an innite interval
then necessarily c is equal to the Ma ne critical value c[0].
Proof. By the denition of the Ma ne critical value, since u L+c
and u is dened on the whole of M, we have c c[0]. To prove
the converse inequality, we pick a continuous function u
0
: M R
with u
0
L +c[0]. For any a, b I with a b, we have
u((b)) u((a)) =
_
b
a
L((s), (s)) ds +c(b a)
u
0
((b)) u
0
((a))
_
b
a
L((s), (s)) ds +c[0](b a).
Subtracting the rst equality from the second inequality, we get
u
0
((b)) u
0
((b)) u((b)) +u((a)) (c[0] c)(b a).
Since both u and u
0
are continuous functions on the compact space
M, the constant
K = supx M[u
0
(x)[ + supx M[u(x)[
is nite, and we obtain from the inequality above
2K (c[0] c)(b a),
for all a, b I with a b. Since I is an innite interval, we can
nd sequences a
n
, b
n
I with a
n
< lb
n
such that b
n
a
n
,
as n . This yields c[0] c 2L/(b
n
a
n
) 0. Hence we
obtain c[0] c.
The following corollary is now a consequence of Denition
4.1.11 of a weak KAM solution and Proposition 4.3.6 above.
Corollary 4.3.7. If u : M R is a negative (resp. positive) weak
KAM solution with the constant c, on the compact manifold M,
then c is necessarily c is equal to the Ma ne critical value c[0].
Theorem 4.3.8. Suppose that L is a Tonelli Lagrangian on the
manifold M. Let u : U R be a function dened on the open
subset U M. Assume that u L + c, where c R, and that
: [a, b] M is a (u, L, c)-calibrated curve, with a < b, then we
have:
132
(i) If for some t, the derivative of u at (t) exists then
d
(t)
u =
L
v
((t), (t)) and H((t), d
(t)
u) = c,
where H is the Hamiltonian associated to L.
(ii) For every t ]a, b[ the derivative of u at (t) exists.
Proof. We prove (i). We will assume t < b (for the case t = b
use a t

< t in the argument). For t

[a, b] satisfying t

> t, the
calibration condition implies
u((t

)) u((t)) =
_
t

t
L((s), (s)) ds +c(t

t).
Dividing by t

t and letting t

t, we obtain
d
(t)
u( (t)) = L((t), (s)) +c.
Combining with the Fenchel Inequality 1.3.1 we get
c = d
(t)
u( (t)) L((t), (s)) H((t), d
(t)
u). ()
But by Proposition 4.2.2, we know that H((t), d
(t)
u) c. This
yields equality in (). Therefore H((t), d
(t)
u) = c. But also the
equality
d
(t)
u( (t)) L((t), (s)) = H((t), d
(t)
u)
means that we have equality in the Fenchel inequality, therefore
we conclude
d
(t)
u =
L
v
((t), (t))
To prove (ii), we choose a open C

chart : U

R
k
on M,
such that (U

) = R
k
and x = (t) U

U. We can nd a

, b

such that a a

< t < b

b and ([a

, b

] U. To simplify
notations, we identify U

and R
k
via . For every y U

= R
k
,
we dene the curve
y
: [a

, t] U

by

y
(s) = (s) +
s a

t a

(y x).
133
We have
y
(a

) = (a

), and
y
(t) = y, since x = (t). Moreover,
we also have
x
= [[a

, t]. Since u L +c, we obtain


u(y) u((a

)) +
_
t
a

L(
y
(s),
y
(s)) ds +c(t a

),
with equality at x = (t) since
x
= is (u, L, c)-calibrated. If we
dene
+
: U

R by

+
(y) = u((a

)) +
_
t
a

L(
y
(s),
y
(s)) ds +c(t a

)
= u((a

))+
_
t
a

L((s) +
s a

t a

(y x), (s) +
y x
t a

) ds
+c(t a

),
we easily see that
+
is as smooth as L (note that is a minimizer,
and is therefore as smooth as L). Moreover we have u(y)
+
(y),
with equality at x.
We now will nd a function

: U

R satisfying

u
with equality at x = (t). For this, given y U

= R
k
we dene

y
: [t, b

] U

by

y
(s) = (s) +
b

s
b

t
(y x).
Again we get
y
(b

) = (b

),
y
(t) = y and
x
= [[t, b

]. Since
u L +c, we obtain
u((b

)) u(y)
_
b

t
L(
y
(s),


y
(s)) ds +c(b

t),
with equality at x since
x
= is (u, L, c)-calibrated. If we dene

: U

R by

(y) = u((b

))
_
b

t
L(
y
(s),


y
(s)) ds c(b

t)
= u((b

)))
_
b

t
L((s) +
b

s
b

t
(y x), (s)
y x
b

t
) ds
c(b

t).
134
Again we easily see that

is as smooth as L. Moreover we have


u(y)

(y), with equality at x.


Since

(y) u(y)
+
(y), with equality at 0, the C
1
func-
tion
+

is non negative and is equal to 0 at x therefore


its derivative at x is 0. Call p the common value d
x

+
= d
x

.
By the denition of the derivative, using

(x) = u(x) =
+
(x),
we can write

(y) = u(x) + p(y x) + |y x|

(y x), with
lim
h0

(h) = 0. The inequality

(y) u(y)
+
(y) now
gives
u(x) +p(y x) +|y x|

(y x) u(y)
u(x) +p(y x) +|y x|
+
(y x).
This obviously implies that p is the derivative of u at x = (t).
Another important property of calibrated curves and domi-
nated functions is given in the following theorem. We will call this
result the Lyapunov property for reasons that will become clear
later, see ????? below.
Theorem 4.3.9 (Lyapunov Property). Let L be a Tonelli Lag-
rangian on the manifold M. Suppose that : [a, b] M is a
continuous piecewise C
1
curve, and that u
1
, u
2
are two real-valued
functions dened on a neighborhood of ([a, b]). If, for some c R,
the curve is (u
1
, L, c)-calibrated and u
2
L + c on a neighbor-
hood of ([a, b]), then the function t u
2
((t)) u
1
((t)) is
non-increasing on [a, b].
Proof. Since is (u
1
, L, c)-calibrated, for t, t

[a, b] with t t

we have
u
1
((t

)) u
1
((t)) =
_
t

t
L((s), (s)) ds +c(t

t).
Using that u
2
L +c on a neighborhood of ([a, b]), we get
u
2
((t

)) u
2
((t))
_
t

t
L((s), (s)) ds +c(t

t).
Comparing we obtain u
2
((t

)) u
2
((t)) u
1
((t

)) u
1
((t)),
therefore u
2
((t

)) u
1
((t

)) u
2
((t)) u
1
((t)).
135
4.4 Minimal Action for a Given Time
As we said minimizers are the important object of the theory.
Therefore if we x a pair of points x, y and a time t > 0, the
minimal action of a curve joining x to y in time t will enjoy some
special properties.
Denition 4.4.1 (Minimal Action). If L is a Tonelli Lagrangian
on the compact connected manifold M, for t > 0 xed, we dene
the function h
t
: M M R by
h
t
(x, y) = inf

_
t
0
L((s), (s)) ds,
where the inmum is taken over all the (continuous) piecewise C
1
curves : [0, t] M with (0) = x and (t) = y.
The quantity h
t
(x, y) is called the minimal action to go from
x to y in time t.
Note that h
t
is well-dened since we are assuming that M is
connected, therefore any pair of points in M can be joined by a
smooth path. Moreover, the function h
t
is nite valued since L is
bounded from below, by superlinearity and compactness of M.
Of course, in the denition of h
t
(x, y), we could have taken the
inmum on all absolute continuous paths. this would have note
changed the value of h
t
(x, y), since minimizers for a given posi-
tive time between two points do always exist by Tonellis theorem
3.3.4 and are in fact as smooth as the Lagragian by the regularity
theorem 3.7.1.
Here are some of the important properties of h
t
.
Proposition 4.4.2 (Properties of h
t
). The properties of h
t
are
(1) For each x, y M, and each t > 0, we have
h
t
(x, y) t inf
TM
L.
(2) For each x, z M and each t, t

> 0, we have
h
t+t
(x, z) = inf
yM
h
t
(x, y) +h
t
(y, z).
136
(3) If u : M R is a function dened on the whole of M, then
u L +c if and only if
x, y M, t > 0, u(y) u(x) h
t
(x, y) +ct.
(4) A continuous piecewise C
1
curve : [a, b] M, with a < b,
is minimizing if and only if
h
ba
((a), (b)) =
_
b
a
L((s), (s)) ds.
(5) For each t > 0 and each x, y M, there exists an extremal
curve : [0, t] M with (0) = x, (t) = y and h
t
(x, y) =
_
t
0
L((s), (s)) ds.
Proof. Property (1) is obvious since for any continuous piecewise
C
1
curve : [0, t] M, with t > 0, we have
_
t
0
L((s), (s)) ds
_
t
0
inf
TM
Lds = t inf
TM
L.
To prove property (2), let us consider two continuous piecewise
linear curves
1
: [0, t] M, with
1
(0) = x,
1
(t) = y, and

2
: [0, t

] M, with
2
(0) = y,
2
(t

) = y. We can dene the


curve
3
=
1

2
: [0, t +t

] M by

3
(s) =
1
(s), for s [0, t]
=
2
(s t), for s [t, t +t

].
The curve
3
is continuous piecewise C
1
, with
3
(0) = x and

3
(t +t

) = z. Moreover, its action L(


3
) is equal to L(
1
)+L(
2
).
Therefore, we have
h
t+t
(x, z) L(
1
) +L(
2
).
Taking the inmum over all possible
1
and
2
gives
h
t+t
(x, z) h
t
(x, y) +h
t
(y, z).
Taking now the inmum over y M, we obtain
h
t+t
(x, z) inf
yM
h
t
(x, y) +h
t
(y, z).
137
We now prove that the inequality above is an equality. Given >
0, we can nd a continuous piecewise C
1
curve : [0, t +t

] M,
with (0) = x, (t +t

) = z, and
_
t+t

0
L((s), (s)) ds h
t+t
(x, z) +.
Using the curve [[0, t], we obtain that
h
t
(x, (t))
_
t
0
L((s), (s)) ds.
Using a reparametrization of [[t, t +t

] by [0, t

], we obtain
h
t
((t), z)
_
t+t

t
L((s), (s)) ds.
Adding these inequalities, we get
h
t
(x, (t)) +h
t
((t), z)
_
t+t

0
L((s), (s)) ds h
t+t
(x, z) +.
Therefore inf
yM
h
t
(x, y) + h
t
(y, z) h
t+t
(x, z) + , for every
> 0. We conclude by letting 0.
For property (3), we observe that if : [a, b] M is a con-
tinuous piecewise C
1
curve, then the reparametrized curve :
[0, b a] M dened by
(s) = (a +s),
has the same endpoints as , and also the same action, since L is
time-independent. In particular we could have dened h
t
by
h
t
(x, y) = inf

_
b
a
L((s), (s)) ds,
where the inmum is taken over all the continuous piecewise C
1
curves : [a, b] M with (a) = x, (b) = y, and b a = t. With
this observation, we get that for a given t > 0, the inequality
u(y) u(x) h
t
(x, y) +ct
138
is equivalent to
u((b)) u((a))
_
b
a
L((s), (s)) ds,
for every continuous piecewise C
1
curves : [a, b] M with
(a) = x, (b) = y, and b a = t.
For property (4), we observe that, by denition of h
t
, we have
h
t
((0), (t)) =
_
t
0
L((s), (s)) ds,
for a curve : [0, t] M if and only if is a minimizer. It remains
to observe as indicated above that we can reparametrize any curve
by an interval starting at 0 without changing neither its endpoints
nor its action.
Property (5) results from Tonellis Theorem 3.3.1.
It is probably dicult to nd out when the next theorem ap-
peared, in some of its forms, for the rst time in the literature. It
has certainly been known for some time now, at least in its equiv-
alent form given as Lemma 4.6.3 below, see for example [Fle69,
Theorem1, page 518]. It has of course been, in a form or another,
been rediscovered by several people, including the author himself,
for whom it started weak KAM Theory since it has as an obvi-
ous consequence the existence of a xed point (up to a constant)
for the LaxOleinik semi-group, see section 4 below.
Theorem 4.4.3 (Flemings Lemma). For each t
0
> 0, there exists
a constant
t
0
[0, +[ such that, for each t t
0
the function
h
t
: M M R is Lipschitzian with a Lipschitz constant
t
0
.
Before proving the theorem, we need to prove some preliminary
results.
Proposition 4.4.4. Let L be a Tonelli Lagrangian on the compact
connected manifold M. For every given t > 0, there exists a
constant C
t
< +, such that, for each x, y M, we can nd a
C

curve : [0, t] M with (0) = x, (t) = y and


L() =
_
t
0
L((s), (s)) ds C
t
.
139
Proof. Choose a Riemannian metric g on M. By the compactness
of M, we can nd a geodesic (for the metric g) between x and y
and whose length is d(x, y). Let us parametrize this geodesic by
the interval [0, t] with a speed of constant norm and denote by
: [0, t] M this parametrization, with (0) = x, (t) = y. As
the length of this curve is d(x, y), we nd that
s [0, t], | (s)|
(s)
=
d(x, y)
t
.
Since the manifold M is compact, the diameter diam(M) of M for
the metric d is nite, consequently, the set
A
t
= (x, v) TM [ |v|
x

diam(M)
t

is compact. We have ((s), (s)) A
t
, for all s [0, t]. By
compactness of A
t
, we can nd a constant

C
t
< + such that
(x, v) A
t
, L(x, v)

C
t
.
If we set C
t
= t

C
t
, we do indeed have L() C
t
.
Corollary 4.4.5 (A Priori Compactness). Let L be a Tonelli Lag-
rangian on the compact manifold M. If t > 0 is xed, there exists
a compact subset K
t
TM such that for every minimizing ex-
tremal curve : [a, b] M, with b a t, we have
s [a, b], ((s), (s)) K
t
.
Proof. Let us recall that we are assuming in this chapter that M
is compact and connected. We rst observe that it is enough to
show the corollary if [a, b] = [0, t]. Indeed, if t
0
[a, b], we can
nd an interval of the form [c, c + t], with t
0
[c, c + t] [a, b].
The curve
c
: [0, t] M, s (c + s) satises the assumptions
of the corollary with [0, t] in place of [a, b].
Thus let us give the proof of the corollary with [a, b] = [0, t].
With the notations of the previous Proposition 4.4.4, we necessar-
ily have
L() =
_
t
0
L((s), (s)) ds C
t
.
140
Since s L((s), (s)) is continuous on [0, t], by the Mean Value
Theorem, we can nd s
0
[0, t] such that
L((s
0
), (s
0
))
C
t
t
. (*)
The set B = (x, v) TM [ L(x, v)
Ct
t
is a compact subset
of TM. By continuity of the ow
t
, the set K
t
=

|s|t

s
(B)
is also compact subset of TM. As is an extremal curve, the
inequality () shows that
s [0, t], ((s), (s))
ss
0
(B) K
t
.
Proof of Theorem 4.4.3. We x some t
0
> 0, and we will study h
t
only for t t
0
.
Let us consider

B(0, 3) the closed ball of center 0 and radius
3 in the Euclidean space R
k
, where k is the dimension of M. By
compactness of M, we can nd a nite number of coordinate charts

i
: R
k
M, i = 1, . . . , p, such that M =

p
i=1

i
(

B(0, 1)). We
denote by > 0, a constant such that
i = 1, . . . , p, x, x

M, d(x, x

) and x
i
_

B(0, 1)
_

x


i
_

B(0, 2)
_
and |
1
i
(x

)
1
i
(x)| 1,
where the norm | | is the Euclidean norm. Let us denote by
K
t
0
the compact set obtained from corollary 4.4.5. We can nd a
constant A < + such that
(x, v) K
t
0
, |v|
x
A.
In the remaining part of the proof, we set
= min
_
t/2,

A
_
.
In the same way by the compactness of K
t
0
, we can nd a constant
B < + such that
x

B(0, 3), v R
k
, i = 1, . . . , p, T
i
(x, v) K
t
0
|v| B.
141
Consider two points x, y M, and suppose that i, j 1, . . . , p
are such that x
i
_

B(0, 1)
_
and y
j
_

B(0, 1)
_
. For a xed
t t
0
, we can nd a minimizing extremal curve : [0, t] M
such that (0) = x, (t) = y and
h
t
(x, y) =
_
t
0
L((s), (s)) ds. (*)
By the a priori compactness given by Corollary 4.4.5, since t t
0
,
we have ((s), (s)) K
t
0
, for each s [0, t]. Consequently,
by the choice of A, we obtain d((s), (s

)) A[s s

[, for all
s, s

[0, t]. In particular, since x = (0)


i
_

B(0, 1)
_
and
y = (t)
j
_

B(0, 1)
_
by the choice of , we nd that
([0, ])
i
_

B(0, 2)
_
([t , t])
j
_

B(0, 2)
_
.
We can then dene the two curves
0
: [0, ]

B(0, 2) and
1
:
[t , t]

B(0, 2) by
i
(
0
(s)) = (s) and
j
(
1
(s)) = (s) .
If d(x, x

) and d(y, y

) , there are unique x

, y


B(0, 2)
such that
i
( x

) = x

and
j
( y

) = y

. By the denition of , we
also have | x

x| 1 and | y

y| 1. Let us dene curves



0
x

,y

: [0, ]

B(0, 3) and
1
x

,y

: [t , t]

B(0, 3) by

0
x

,y
(s) =
s

( x

x) + (s), for s 0, ],

1
x

,y
(s) =
s (t )

( y

y) + (s), for s [t , t].


The curve
0
x

,y

[[0, ] connects the point x

to the point (), and


the curve
1
x

,y

[[t , t] connects the point (t ) to the point


y

. Since T/2 we can, then, dene the curve


x

,y
: [0, t]
M by
x

,y
= on [, t ],
x

,y
=
i

0
x,

on [0, ], and

,y
=
j

1
x,

on [t , t]. The curve


x

,y
is continuous on the
interval [0, t], moreover, it of class C
1
on each of each the intervals
[0, ], [, t ] and [t , t]. We of course have
x,y
= . As
x

,y
(s)
is equal to x

, for s = 0, and to y

, for s = t, we have
h
t
(x

, y

)
_
t
0
L(
x

,y
(s),
x

,y
(s)) ds.
142
Subtracting the equality () above from this inequality and using
the fact that
x

,y
= on [, t ], we nd
h
t
(x

, y

) h
t
(x, y)
_

0
L(
x

,y
(s),
x

,y
(s)) ds
_

0
L((s), (s)) ds
+
_
t
t
L(
x

,y
(s),
x

,y
(s)) ds
_
t
t
L((s), (s)) ds.
We will use the coordinate charts
i
and
j
to estimate the right-
hand side of this inequality. For that, it is convenient, for =
1, . . . , p, to consider the Lagrangian

L

:

B(0, 3) R
k
R dened
by

(z, w) = L
_

(z), D

(w)[w]
_
.
This Lagrangian

L

is of class C
r
on

B(0, 3) R
k
. Using these
Lagrangian for = i, j, we have
h
t
(x

, y

) h
t
(x, y)
_

0

L
i
(
0
x

,y
(s),


0
x

,y
(s)) ds
_

0

L
i
(
0
(s),


0
(s)) ds
+
_
t
t

L(
1
x

,y
(s),


1
x

,y
(s)) ds
_
t
t
L(
1
(s),


1
(s)) ds. (*)
By the choice of B, we have |

(s)| B, for each s [0, ]. By


the denition of
0
x

,y

and the fact that | x

x| 1, we obtain
s [0, ], |


0
x

,y
(s)|
| x

x|

+B
B +
1
.
Since

L
i
is C
1
and the set
E
B,
= (z, v)

B(0, 3) R
k
[ |v| B +
1

is compact, we see that there exists a constant C


i
, which depends
only on the restriction of the derivative of

L
i
on this set E
B,
, such
that
(z, v), (z

, v

) E
B,
,
[

L
i
(z, v)

L
i
(z

, v

)[ C
i
max(|z z

|, |v v

|).
143
Since the two points (
0
x

,y

(s),


0
x

,y
(s)), (
0
(s),


0
(s)) are in E
B,
,
and
0
x

,y

(s)
0
(s) =
s

( x

x), we nd
[

L
i
(
0
x

,y
(s),


0
x

,y
(s))

L(
0
(s),


0
(s))[
C
i
max
_
|
s

( y x)|, |
1

( y x)|

C
i
max(1,
1

)| y x|.
By integration on the interval [0, ], it follows that
_

0

L
i
(
0
x

,y
(s),


0
x

,y
(s)) ds
_

0

L
i
(
0
(s),

0
(s)) ds
C
i
max(, 1)| y x|.
Since
i
is a dieomorphism of class C

, its inverse is Lipschitzian


on the compact subset
i
_

B(0, 3)
_
. We then see that there exists
a constant

C
i
, independent of x, x

, y, y

and t t
0
, and such that
x
i
_

B(0, 1)
_
and d(x

, x)
_

0

L
i
(
0
x

,y
(s),


0
x

,y
(s)) ds
_

0

L
i
(
0
(s),

0
(s)) ds

C
i
d(x

, x).
In the same way we can prove the existence of a constant

C

j
,
independent of x, x

, y, y

and t t
0
, and such that
y
j
_

B(0, 1)
_
and d(y

, y)
_
t
t

L
j
(
1
x

,y
(s),


1
x

,y
(s)) ds
_
t
t

L
j
(
1
(s),

1
(s)) ds

C

j
d(y

, y).
Therefore by the inequality (*) above, we obtain
x
i
_

B(0, 1)
_
, y
j
_

B(0, 1)
_
, d(x

, x) and d(y

, y)
h
t
(x

, y

) h
t
(x, y)

C
i
d(x

, x) +

C

j
d(y

, y).
Setting
t
0
= max
p
i=1
max(

C
i
,

C

i
), we nd that we have
d(x

, x) , d(y

, y) and t t
0

h
t
(x

, y

) h
t
(x, y)
t
0
[d(x

, x) +d(y

, y)].
144
If x, x

, y, y

are arbitrary points, and


0
: [0, 1] M,
1
: [0, 1]
M are geodesics of length respectively d(x, y) and d(x

, y

), param-
eterized proportionally arclength and connecting respectively x to
y and x

to y

, we can nd a nite sequence t


0
= 0 t
1

t

= 1 such that d(
0
(t
i+1
),
0
(t
i
)) and d(
1
(t
i+1
),
1
(t
i
)) .
Applying what we did above, we obtain
i 0, 1, . . . , 1, h
t
(
0
(t
i+1
),
1
(t
i+1
)) h
t
(
0
(t
i
),
1
(t
i
))

t
0
[d(
0
(t
i+1
),
0
(t
i
)) +d(
1
(t
i+1
),
1
(t
i
))].
Adding these inequalities, we nd
h
t
(x

, y

) h
t
(x, y)
t
0
[d(x

, x) +d(y

y)].
We nish the proof by exchanging the roles of (x, y) and (x

, y

)
The following theorem is a consequence of Corollary 4.4.5 and
Theorem 4.4.6. Let L be a C
r
Tonelli Lagrangian, with r 2, on
the compact connected manifold M. Given a, b R with a < b,
call M
a,b
the set of curves : [a, b] M which are minimizers
for the class of continuous piecewise C
1
curves. Then M
a,b
is a
compact subset of (
r
([a, b], M) for the C
r
topology.
Moreover, For every t
0
[a, b] the map M
a,b
TM,
((t
0
), (t
0
)) is a homeomorphism on its image (which is therefore
a compact subset of TM.)
4.5 The Lax-Oleinik Semi-group.
We call T(M, [, +]) the set of arbitrary functions from the
manifold M to the set [, +] of extended real numbers. We
will also use the notation T(M, R) for the set of arbitrary functions
M :R.
4.6 The Lax-Oleinik semi-group
We introduce a semi-group of non-linear operators (T

t
)
t0
from
T(M, [, +]) into itself. This semigroup is well-known in PDE
and in Calculus of Variations, it is called the Lax-Oleinik semi-
group. Again it has been rediscovered many times in dierent
145
forms; For the author, it came as a natural by-product of the
proof of the hamilton-Jacobi Theorem for C
1
functions, see the
proof of Theorem 4.1.1 and the discussion in section 1.
Denition 4.6.1 (Lax-Oleinik semi-group). Fix u T
0
(M, [, +])
and t > 0. The function T

t
u : M [, +] is dened at
x M by
T

t
u(x) = inf

u((0)) +
_
t
0
L((s), (s)) ds,
where the inmum is taken over all the absolutely continuous
curves : [0, t] M such that (t) = x. By the denition of
the minimal action h
t
, see Denition 4.4.1, for t > 0, we have
T

t
u(x) = inf
yM
u(y) +h
t
(y, x).
We will also set T

0
u = u. The family of maps T

t
: T(M, [, +])
T(M, [, +]), t [0, +[ is called the Lax-Oleinik semi-group.
Here are some properties of T

t
on T(M, [, +]).
Proposition 4.6.2. Consider u T(M, [, +]).
(1) For x M and t > 0, we have
inf
M
u +t inf
TM
L T
t
u(x) inf
M
u + max
MM
h
t
.
Since inf
TM
L > by the superlinearity of L and h
t
is contin-
uous on the compact space M M, we obtain that the following
properties are equivalent:
(a) there exists a t > 0 and and an x M such that T

t
u(x) is
nite;
(b) we have inf
M
u ] , +[;
(c) for every t > 0, the function T

t
u is nite, valued.
(2) (Semigroup Property) We have T

t+t

= T

t
T

, for
each t, t

0.
(3) for every c R, we have T

t
(c +u) = c +T

t
u.
146
(4) (Inf Commutativity) If u
i
, i I is a family of functions
in u T(M, [, +]), we have
T

t
(inf
iI
u
i
) = inf
iI
T

t
(u
i
).
(5) (Monotony) For each u, v T(M, [, +])) and all
t > 0, we have
u v T

t
u T

t
v.
(6) If c R, the function u T(M, [, +]) satises u
T

t
u +ct if and only if one of the following three things happens:
(i) the function u is identically ;
(ii) the function u is identically +;
(iii) the function u is nite everywhere and u L +c.
(7) Suppose that c R, and u : M R are such that u
L+c, then, for every t 0, the function T

t
u is nite valued, and
T

t
u L +c.
Proof. To prove assertion (1), we juste notice that T

t
u(x) =
inf
yM
u(y) + h
t
(y, x) inf
yM
u(y) + max
MM
h
t
= inf
M
u +
max
MM
h
t
. Moreover by part (1) of Proposition 5.3.2, we have
h
t
(y, x) t inf
TM
L, from which it follows that
T

t
u(x) inf
yM
u(y) +t inf
TM
L
= t inf
TM
L + inf
M
u.
The rest of assertion (1) follows easily from this double inequality.
Assertion (2) follows from part (2) of Proposition 5.3.2, which
147
states that h
t

+t
(y, x) = inf
zM
h
t
(y, z) +h
t
(z, x). Therefore
T

t+t

u(x) = inf
yM
u(y) +h
t

+t
(x, y)
= inf
yM
[u(y) + inf
zM
h
t
y, z) +h
t
(z, x)]
= inf
yM
inf
zM
[u(y) +h
t
(y, z) +h
t
(z, x)]
= inf
zM
inf
yM
[u(y) +h
t
(y, z) +h
t
(z, x)]
= inf
zM
[ inf
yM
[u(y) +h
t
(y, z)] +h
t
(z, x)
= inf
zM
T

u(z) +h
t
(z, x)
= T

t
[T

u].
Assertion (3) is obvious from the denition of T

t
.
For assertion (4), we notice that
T

t
(inf
iI
u
i
)(x) = inf
yM
inf
iI
u
i
(y) +h
t
(y, x)
= inf
iI
inf
yM
u
i
(y) +h
t
(y, x)
= inf
iI
T

t
(u
i
)(x).
Assertion (5) is also an immediate consequence of the def-
inition of T

t
. It also easily follows from assertion (4), since
we have u = inf(u, v), which yields T

t
(u) = T

t
(inf(u, v)) =
inf(T

t
(u), T

t
(v)) T

t
(v).
For assertion (6), assume that u T

t
u + ct for every t. Ob-
viously, by assertion (1), if inf
M
u = , then T

t
u hence
u T

t
u + ct is also . If inf
M
u = +, of course u +.
In the remaining case inf
M
u ] , +[, we obtain from (1)
above that T

t
u is nite everywhere therefore u, which satises
u T

t
u+ct and u inf
M
u > , is also nite valued. The con-
dition u T

t
u+ct yields u(x) inf
yM
u(y)+h
t
(y, x)+ct. There-
fore u(x) u(y)+h
t
(y, x)+ct, for every x, y M and t > 0. Since
u is nite valued this is equivalent to u(x)u(y) h
t
(y, x)+ct, for
every x, y M and t > 0. By Assertion (3) of Proposition 5.3.2
this is equivalent to u L +c. Note that reversing the reasoning
just done we can show that if u is nite valued and u L + c
then u T

t
u + ct, for every t 0. Moreover, if u (resp.
148
u +) we have T

t
(u) (resp. T

t
(u) +) which of
course yields u = T

t
u +ct.
For assertion (7), we rst observe that by part (iii) of Propo-
sition 4.2.1, then function u is continuous (it is even Lipschitz)
on the compact manifold M. Therefore it is bounded from below
and by Assertion (1, the function T

t
u is nite everywhere. By
Assertion (6), we have
u T

u +ct

,
for every t

0. Using Assertions (2), (3), and (5), we get


T

t
u T

t
[T

u +ct

]
= T

t
[T

u] +ct

= T

t+t

u = T

[T

t
u] +ct

Therefore T

t
u T

[T

t
u] + ct

, for every t

0. Since T

t
u is
nite, from Assertion (6), we get T

t
u L +c.
Let us now introduce the space B(M, R) of bounded func-
tions u : M R. As usual, we endow this space of the norm ||
dened by
|u|

= sup
xM
[u(x)|.
By assertion (1) of Proposition 4.6.2, if u T(M, R) satises
inf
M
u > , then T

t
u B(M, R). Flemings Lemma 4.4.3
yields the following much stronger property on the Lax-Oleinik
semi-group.
Lemma 4.6.3. For each t
0
> 0, there exists a constant
t
0
such
that for every u

T(M, R) with inf
M
u > , and every t t
0
the functionT

t
u : M R is
t
0
-Lipschitzian.
In particular for every t > 0, we have T

t
(B(M, R) (
0
(M, R).
Proof. By Flemings Lemma 4.4.3, we can nd a constant
t
0
such
that for every t t
0
, the function h
t
: M M : toR is Lipschitz
with Lipschitz constant
t
0
. If follows that for any u in(
0
(M, R)
and any t t
0
, the family of function u(y) +h(y, ), y M is equi-
Lipschitzian with Lipschitz constant
t
0
. Since, by the condition
inf
M
u > , its inmum T

t
u(x) = inf
yM
u(y)+h(y, x) is nite
everywhere, it is also Lipschitz with Lipschitz constant
t
0
.
149
Before giving further properties of the semi-group T

t
, we recall
the following well-known denition
Denition 4.6.4 (Non-expansive Map). A map : X Y ,
between the metric spaces X and Y , is said to be non-expansive
if it is Lipschitzian with a Lipschitz constant 1.
Proposition 4.6.5 (Non-expansiveness of the Lax Oleinik semi
group). The maps T

t
: are non-expansive for the norm ||

, i.e.
u, v B(M, R), t 0, |T

t
u T

t
v|

|u v|

.
Proof. If u, v B(M, R), we have
|u v|

+v u |u v|

+v.
By parts (5) and (3) of Proposition 4.6.2, we get
|u v|

+T

t
v T

t
u |u v|

+T

t
v.
This clearly implies |T

t
u T

t
v|

|u v|

.
We now turn to the properties of the semi-group on the space
(
0
(M, R) of continuous functions, endowed with the topology of
uniform convergence, i.e. the topology induced by the norm ||.
Proposition 4.6.6. The semi-group T

t
sends (
0
(M, R) to itself.
It satises the following properties:
(1) For each u (
0
(M, R), we have lim
t0
T

t
u = u.
(2) For each u (
0
(M, R), the map t T

t
u is uniformly
continuous.
(3) For each u (
0
(M, R), the function (t, x) T

t
u(x) is con-
tinuous on [0, +[M and locally Lipschitz on ]0, +[M. In
fact, for each t
0
> 0, the family of functions (t, x) T

t
u(x), u
(
0
(M, R), is equi-Lipschitzian on [t
0
, +[M.
(4) For each u (
0
(M, R), each x M, and each t > 0 we
can nd a a continuous piecewise C
1
curve
x,t
: [0, t] M with

x,t
(t) = x and
T

t
(x) = u(
x,t
(0))+h
t
(
x,t
(0), x)) = u(
x,t
(0))+
_
t
0
L(
x,t
(s),
x,t
(s)) ds.
150
Proof. As the C
1
maps form a dense subset of (
0
(M, R) in the
topology of uniform convergence, it is not dicult to see, using
Proposition 4.6.5 above, that it is enough to show property (1)
when u is Lipschitz. We denote by K the Lipschitz constant of u.
By the compactness of M and the superlinearity of L, there is a
constant C
K
such that
(x, v) TM, L(x, v) K|v|
x
+C
K
.
It follows that for every curve : [0, t] M, we have
_
t
0
L((s), (s)) ds Kd((0), (t)) +C
K
t.
Since the Lipschitz constant of u is K, we conclude that
_
t
0
L((s), (s)) ds +u((0)) u((t)) +C
K
t,
which gives
T

t
u(x) u(x) +C
K
t.
Moreover, using the constant curve
x
: [0, t] M, s x, we
obtain
T

t
u(x) u(x) +L(x, 0)t.
Finally, if we set A
0
= max
xM
L(x, 0), we have obtained
|T

t
u u|

t max(C
K
A
0
),
which does tend to 0, when t tends to 0.
To show (2), we notice that by the semi-group property of T

t
,
see part (2) of Proposition 4.6.2, we have
|T

u T

t
u|

|T

|t

t|
u u|

,
and we apply (1) above.
To prove (3), we remark that the continuity of (t, x) T

t
u(x)
follows from (1) and the fact that T

t
u is continuous for ever t 0.
To prove the equi-Lipschitzianity, we x t
0
> 0. By Lemma 4.6.3,
we know that there exists a nite constant K(t
0
) such that T

t
u is
Lipschitz with Lipschitz constant K(t
0
), for each u (
0
(M, R)
151
and each t t
0
. It follows from the semi-group property and the
proof of (1) that
|T

u T

t
u|

(t

t) max(C
K(t
0
)
, A
0
),
for all t

t t
0
. It is then easy to check that (t, x) T

t
u(x) is
Lipschitz on [t
0
, +[M, with Lipschitz constant max(C
K(t
0
)
, A
0
)+
K(t
0
), for anyone of the standard metrics on the product RM.
To prove (4), we recall that T

t
(x) = inf
yM
u(y) +h
t
(y, x). since
the function y u(y)+h
t
(y, x) is continuous on the compact space
M, we can nd y
x
M such that T

t
(x) = u(y
x
) + h
t
(y
x
, x). We
can apply part (5) of Proposition 5.3.2 to nd a continuous piece-
wise C
1
curve
x,t
: [0, t] M with
x,t
(0) = y
x
,
x,t
(t) = x and
h
t
(y
x
, x) =
_
t
0
L(
x,t
(s),
x,t
(s)) ds. Therefore T

t
(x) = u(y
x
) +
h
t
(y
x
, x)) = u(y
x
) +
_
t
0
L(
x,t
(s),
x,t
(s)) ds.
We now give the connection of the semi-group T

t
with weal
KAM solutions of the negative type.
Proposition 4.6.7. Suppose that u : M R s a function and
c R. We have T

t
u + ct = u, for each t [0, +[, if and only if
u is a negative weak KAM solution, i.e.we have
(i) u L +c;
(ii) for each x M, there exists a (u, L, c)-calibrated curve
x

:
] , 0] M such that
x

(0) = x.
Proof. We suppose that T

t
u + ct = u, for each t [0, +[. In
particular, we have By part (6) of Proposition 4.6.2 above, since u
is nite-valued, we obtain u L+c. In particular, the function u is
continuous. It remains to show the existence of
x

: ], 0] M,
for a given x M. We already know by part (4) of Proposition
4.6.6 that, for each t > 0, there exists a continuous piecewise C
1
curve
t
: [0, t] M, with
t
(t) = x and
u

(x)ct = T
t
u

(x) = u

(
t
(0))+h
t
(
t
(0),
t
(t)) = u

(
t
(0))+
_
t
0
L(
t
(s),
t
(s)) ds.
We set
t
(s) =
t
(s +t), this curve is parametrized by the interval
[t, 0], is equal to x at 0, and satises
u

(
t
(0))u

(
t
(t)) = h
t
(
t
(t),
t
(0))+ct
_
0
t
L(
t
(s),


t
(s)) ds+ct.
152
It follows from Proposition ?? above that
t
is (u, L, c)-calibrated.
In particular, the curve
t
, and we have
t

[t, 0], u

(x)u

(
t
(t

)) = ct

+
_
0
t

L(
t
(s),


t
(s)) ds. (*)
As the
t
are minimizing extremal curves, by the a priori com-
pactness given by corollary 4.4.5, there exists a compact subset
K
1
TM, such that
t 1, s [t, 0], (
t
(s),


t
(s)) K
1
.
Since the
t
are extremal curves, we have (
t
(s),


t
(s)) =
s
(
t
(0),


t
(0)).
The points (
t
(0),


t
(0)) are all in the compact subset K
1
, we can
nd a sequence t
n
+such that the sequence (
tn
(0),


tn
(0)) =
(x,


tn
(0)) tends to (x, v

), when n +. The negative orbit

s
(x, v

), s 0 is of the form (
x

(s),
x

(s)), where
x

:], 0]
M is an extremal curve with
x
(0) = x. If t

], 0] is xed, for n
large enough, the function s (
tn
(s),


tn
(s)) =
s
(
tn
(0),


tn
(0))
is dened on [t

, 0], and, by continuity of the Euler-Lagrange ow,


this sequence converges uniformly on the compact interval [t

, 0]
to the map s
s
(x, v

) = (
x
(s),
x
(s)). We can then pass to
the limit in the equality () to obtain
u

(x) u

(t

)) = ct

+
_
0
t

L(
x

(s),
x

(s)) ds.
Conversely, let us suppose that u L+c and that, for each x M,
there exists a curve
x

: ] , 0] M, with
x

(0) = x, and such


that for each t [0, +[
u

(x) u

(
x

(t)) = ct +
_
0
t
L(
x

(s),
x

(s)) ds.
Let us show that T

t
u + ct = u, for each t [0, +[. If x M
and t > 0, we dene the curve : [0, t] M by (s) =
x

(s t).
It is not dicult to see that (t) = x and that
u

(x) = ct +
_
t
0
L((s), (s)) ds +u

((0)).
It follows that T

t
u(x) + ct u(x) and thus T

t
u + ct u. The
inequality u T

t
u +ct results from u L +c.
153
4.7 Existence of Negative Weak KAM So-
lutions
Our goal in this section is to prove the existence of negative weak
KAM solutions.
Theorem 4.7.1 (Weak KAM). There exists a function u

: M
R which is a negative weak KAM solution with constant c[0].
Note that we already know by Corollary 4.3.7 that a weak
KAM solution can only have c[0] as a constant. By Proposition
the weak KAM theorem above is equivalent to
Theorem 4.7.2. There exists a function u

: M R such that
T

t
u

+tc[0] = u

, for each t [0, +[.


We will give two proofs of this theorem.
For the rst proof we need some lemmas.
Lemma 4.7.3. Let u : M R be a function (not necessarily
continuous or even bounded), and c R. If the function u dened
on M by
u(x) = inf
t0
T

t
u(x) +ct
is nite at some point, then it is nite everywhere and u L +c.
Proof. We use parts (4), (3), (2) of Proposition 4.6.2 to obtain
T

u = T

[inf
t0
T

t
u +ct]
= inf
t0
T

[T

t
u +ct]
= inf
t0
T

[T

t
u] +ct
= inf
t0
T

+t
T

t
u +ct.
Therefore
T

u +ct

= ct

+ inf
t0
T

+t
T

t
u +ct
= inf
t0
T

+t
T

t
u +ct +ct

= inf
t0
T

+t
T

t
u +c(t

+t)
geq inf
t0
T

t
u +ct = u.
154
Since u is nite at some point it follows from part (6) of Proposi-
tion 4.6.2 that it is nite everywhere and u L +c.
Lemma 4.7.4. Let u : M R be a function with inf
M
u > ,
but not necessarily continuous or even bounded. Suppose t > 0
and c R are such that u T

t
u +c, then c/t c[0].
Proof. Using parts (2), (3) and (5) of Proposition 4.6.2, we obtain
for n N
u T

t
u +c T

2t
u + 2c T

nt
u +nc.
This implies that for s 0
T

s
u T

nt+s
u +nc.
If we set c = c/t, this yields
T

s
u +s c T

nt+s
u + (nt +s) c,
for all s 0, and all n N. In particular, we have
u = inf
s0
T

s
u +s c = inf
0st
T

s
u +s c.
We now note that by part (1) of Proposition 4.6.2 we have
T
s
u inf
M
u +s inf
TM
L inf
M
u s[ inf
TM
L[.
Therefore
u = inf
0st
T

s
u +s c
inf
0st
inf
M
u s[ inf
TM
L[ +s c
inf
M
u s([ inf
TM
L[ +[ c[ > .
Hence we can apply Lemma 4.7.3 to obtain that u L+ c. Hence
c = c/t c[0].
Lemma 4.7.5. Let u : M R be dominated by L + c[0] on the
compact manifold M. then for every t 0, we can nd a point
x
t
M such that u(x
t
) = T

t
u(x
t
) +tc[0].
In particular, we have sup
t0
|T

t
u +tc[0]|

< +.
155
Proof. Let us observe that, part (iii) of Proposition 4.2.1 and part
(7) of Proposition 4.6.2, the family of functions T

t
u, t 0, is
equi-Lipschitzian. We call L < + a common Lipschitz constant
for T

t
u, t 0.
To prove the rst part we can assume t > 0. Suppose that
no such x
t
exists, since u T

t
u + tc[0], we obtain that T

t
u +
tc[0] u > 0 everywhere. Since both u and T

t
u are continuous,
the compactness of M now implies that there exists > 0 such
that T

t
u + tc[0] u everywhere. This can be rewritten as
u T

t
u +tc[0] . By Lemma 4.7.4 above, this implies (tc[0]
)/t c[0]. Which is obviously false. This proves the existence of
x
t
.
For each t 0 the function T

t
u+tc[0] u is Lipschitzian with
Lipschitz constant 2L. Since it vanishes at some point x
t
M,
we have for every x M
[T

t
u(x) +tc[0] u(x)[ = [(T

t
u +tc[0] u(x)) (T

t
u(x
t
) +tc[0] u(x))[
2Ld(x, x
t
)
2Ldiam(M).
Therefore we, obtain sup
t0
|T

t
u+tc[0]|

|u|

+2Ldiam(M) <
+.
We can now prove the following theorem which yields imme-
diately the weak KAM Theorem 4.7.1.
Theorem 4.7.6. Let L be a Tonelli Lagrangian on the compact
connected manifold M. Suppose uIf u : M R is dominated by
L + c[0]. Then T

t
u + tc[0] converges uniformly to a continuous
function u

: M R which is a negative weak KAM solution.


Proof. We note that by parts (6), (5) and (2) of Proposition 4.6.2
we have for s, t 0
T

s
u T

s
(T

t
u +tc[0])
= T

t+s
u +tc[0].
Hence T

s
u +sc[0] T

t+s
u + (t +s)c[0], this implies that T

t
u +
tc[0] T

u + t

c[0], for every t, t

0, with t t

. Since by
Lemma 4.7.5, the family T

t
u + tc[0], t 0 is equibounded, it
156
follows that the point-wise limit u

(x) = lim
t+
T

t
u(x) +tc[0]
exists everywhere and is nite. We now remark as at the beginning
of the proof of Lemma 4.7.5 above that the family T

t
u+tc[0, t 0
is equicontinuous, to conclude that the limit u

= lim
t+
T

t
u+
tc[0] is uniform.
It remains to prove that u

: M R which is a negative
weak KAM solution. By Proposition 4.6.7, we have to check that
T

s
u + sc[0] = u, for each s . Using the non-expansiveness of
the Lax Oleinik semi-group Proposition 4.6.5, we obtain T

s
u

=
lim
t+
T

s
[T

t
u + tc[0]] = lim
t+
T

t+s
u + tc[0]. Therefore
T

s
u

+sc[0] = lim
t+
T

t+s
u + (t +s)c[0] = u

.
Before giving a second proof of the weak KAM Theorem 4.7.1,
we will need to recall some xed point theorems. We leave this as
an exercise, see also [GK90, Theorem 3.1, page 28].
Exercise 4.7.7. 1) Let E be a normed space and K E a com-
pact convex subset. We suppose that the map : K K is
non-expansive. Show that has a xed point. [Hint: Reduce rst
to the case when 0 K, and then consider x (x), with
]0, 1[.]
2) Let E be a Banach space and let C E be a compact subset.
Show that the closed convex envelope of C in E is itself compact.
[Hint: It is enough to show that, for each > 0, we can cover the
closed convex envelope of C by a nite number of balls of radius
.]
3) Let E be a Banach space. If : E E is a non-expansive
map such that (E) has a relatively compact image in E, then the
map admits a xed point. [Hint: Take a compact convex subset
containing the image of .]
4) Let E be a Banach space and
t
: E E be a family of maps
dened for t [0, [. We suppose that the following conditions
are satised
For each t, t

[0, [, we have
t+t
=
t

t
.
For each t [0, [, the map
t
is non-expansive.
For each t > 0, the image
t
(E) is relatively compact in E.
For each x E, the map t
t
(x) is continuous on [0, +[.
157
Show then that the maps
t
have a common xed point. [Hint:
A xed point of
t
is a xed point of
kt
for each integer k 1.
Show then that the maps
1/2
n, with n N, admit a common xed
point.]
We notice that we can use Brouwers Fixed Point Theorem
(instead of Banachs Fixed Point Theorem) and an approximation
technique to show that the result established in part 1) of the
exercise above remains true if is merely continuous. This is the
Schauder-Tykhonov Theorem, see [Dug66, Theorems 2.2 and 3.2,
pages 414 and 415] or [DG82, Theorem 2.3, page 74]. It follows
that the statements in parts 3) and 4) are also valid when the
involved maps are merely continuous.
Second proof of Weak KAM Theorem 4.7.1. Let us denote by 1
the constant function equal to 1 everywhere on M and consider the
quotient E = (
0
(M, R)/R.1. This quotient space E is a Banach
space for the quotient norm
|[u]| = inf
aR
|u +a1|

,
where [u] is the class in E of u (
0
(M, R). Since T

t
(u + a1) =
T

t
(u) + a1, if a R, the maps T

t
pass to the quotient to a
semigroup

T

t
: E E consisting of non-expansive maps. Since,
for each t > 0, the image of T

t
is an equi-Lipschitzian family
of maps, Ascolis Theorem, see for example [Dug66, Theorem 6.4
page 267], then shows that the image of

T

t
is relatively compact
in E (exercise). Using part 4) in the exercise above, we nd a
common xed point for all the

T

t
. We then deduce that there
exists u

(
0
(M, R) such that T

t
u

= u

+ c
t
, where c
t
is a
constant. The semigroup property gives c
t+t
= c
t
+ c
t
; since
t T

t
u is continuous, we obtain c
t
= tc with c = c
1
. We
thus have T

t
u

+ct = u

.
FROM HERE ON NOTES HAVE TO REVISED.
158
4.8 Invariant Measures and Ma nes Criti-
cal Value
Corollary 4.8.1. If T

t
u

= u

ct, then, we have


c = inf

_
TM
Ld,
where varies among Borel probability measures on TM invariant
by the Euler-Lagrange ow
t
. This lower bound is in fact achieved
by a measure with compact support. In particular, the constant c
is unique.
Proof. If (x, v) TM, then : [0, +[ M, dened by (s) =

s
(x, v), satises ((s), (s)) =
s
(x, v). Since u

L + c, we
nd
u

(
1
(x, v)) u

((x, v))
_
1
0
L(
s
(x, v)) ds +c.
If is a probability measure invariant by
t
, the function u


is integrable since it is bounded. Invariance of the measure by
t
gives
_
TM
[u

(
1
(x, v)) u

((x, v))] d(x, v) = 0,


from where, by integration of the inequality above, we obtain
0
_
TM
[
_
1
0
L(
s
(x, v)) ds +c] d(x, v).
Since L is bounded below and is a probability measure, we can
apply Fubini Theorem to obtain
0
_
1
0
_
_
TM
(L(
s
(x, v)) +c) d(x, v)

ds.
By the invariance of under
s
, we nd that 0
_
(L + c)d.
Since is a probability measure, this yields
c
_
TM
Ld.
159
It remains to see that the value c is attained. For that, we x
x M, and we take a curve

x
:] , 0] M with

x
(0) = x
and such that
t 0, u

x
(0)) u

x
(t)) =
_
0
t
L(

x
(s),

x
(s)) ds ct.
The curve
x
is a minimizing extremal curve with

x
(0) = x,
therefore, we have

s
(x,

x
(0)) = (

x
(s),

x
(s))
and the curve (

x
(s),

x
(s)), s 0 is entirely contained in a com-
pact subset K
1
of TM as given by corollary 4.4.5. Using Riesz Rep-
resentation Theorem [Rud87, Theorem 2.14, page 40], for t > 0,
we dene a Borel probability measure
t
on TM by

t
() =
1
t
_
0
t
(
s
(x,

x
(0))) ds,
for : TM R a continuous function. All these probability
measures have their supports contained in the compact subset K
1
,
consequently, we can extract a sequence t
n
+ such that
tn
converges weakly to a probability measure with support in K
1
.
Weak convergence means that for each continuous function :
TM R, we have
_
TM
d = lim
n
1
t
n
_
0
tn
(
s
(x,

x
(0)) ds.
We have
_
TM
(L +c) d = 0, because
_
TM
(L +c) d = lim
n
1
t
n
_
0
tn
_
L(

x
(s),

x
(s)) +c
_
ds
and
_
0
tn
L(

n
(s),

x
(s)) ds +ct
n
= u

x
(0)) u

x
(t
n
))
which is bounded by 2|u

. This does indeed show that for the


limit measure , we have
_
TM
Ld = c.
160
In the second part of the previous proof, we have in fact shown
the following proposition.
Proposition 4.8.2. If T

t
u

= u

+ ct, for every t 0, and

x
:] , 0] M is an extremal curve, with

x
(0) = x, and such
that
t 0, u

(x) u

x
(t)) =
_
0
t
L(

x
(s),

x
(s)) ds. +ct,
the following properties are satised
for each s 0, we have
s
(x,

x
(0)) = (

x
(s),

x
(s));
the -limit set of the orbit of (x,

x
(0)) for
s
is compact;
there exists a Borel probability measure on TM, invariant
by
t
, carried by the -limit set of the orbit of (x,

x
(0)),
and such that
_
Ld = c.
We dene c[0] R by
c[0] = inf

_
TM
Ld,
where the lower bound is taken with respect to all Borel probability
measures on TM invariant by the Euler-Lagrange ow. We will
use the notation c
L
[0], if we want to specify the Lagrangian.
Denition 4.8.3 (Minimizing Measure). A measure on TM
is said to be minimizing if it is a Borel probability measure ,
invariant by the Euler-Lagrange ow, which satises
c[0] =
_
TM
Ld.
Exercise 4.8.4. Show that each Borel probability measure , in-
variant by the Euler-Lagrange ow, and whose support is contained
in the -limit set of a trajectory of the form t (
x

(t),
x

(t)),
must be minimizing.
Corollary 4.8.5. If a weak KAM solution u

is dierentiable at
x M, we have
H(x, d
x
u

) = c[0].
161
Proof. If u

is a weak KAM solution, we have u

L +c[0] and
thus
H(x, d
x
u

) = sup
vTxM
d
x
u

(v) L(x, v) c[0].


It remains to nd v
0
T
x
M such that d
x
u

(v
0
) = L(x, v
0
) +c[0].
For that, we pick extremal curves
x

:], 0] such that


x

(0) = x
and
t 0, u

(
x

(0)) u

(
x

(t)) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t.
By dividing this equality by t > 0 and letting t tend to 0, we nd
d
x
u

(
x

(0)) = L(x,
x

(0)) +c[0].
4.9 The Symmetrical Lagrangian
Denition 4.9.1 (Symmetrical Lagrangian). If L : TM R is
a Lagrangian we dene its symmetrical Lagrangian

L : TM R
by

L(x, v) = L(x, v).


If : [a, b] M is a curve, we dene the curve : [a, b] M
by (s) = (a + b s). It is immediate to check that

(s) =
(a +b s) and thus

L( ) = L(),
where

L is the action associated with

L, i.e.

L( ) =
_
b
a

L( (s),

(s)) ds.
It clearly results that is an extremal curve of L if and only if
is an extremal curve of

L. We want to express the Lax-Oleinik
semigroup

T

t
: (
0
(M, R) (
0
(M, R) associated with

L in terms
of L only. For that, we notice that when : [0, t] M varies
among all the curves such that (0) = x, then varies among all
162
all the curves such that (t) = x, and we have (0) = (t). We
thus nd

t
u(x) = inf

_
t
0

L( (s),

(s)) ds +u( (0))
= inf

_
t
0
L((s), (s)) ds +u((t)),
where the lower bound is taken all the piecewise C
1
curves :
[0, t] M such that (0) = x.
We then introduce the semigroup T
+
t
: (
0
(M, R) (
0
(M, R)
dened by T
+
t
(u) =

t
(u). We nd
T
+
t
u(x) = sup

u((t))
_
t
0
L((s), (s)) ds,
where the upper bound is taken on the (continuous) piecewise C
1
curves : [0, t] M such that (0) = x.
The following lemma is easily veried.
Lemma 4.9.2. We have u L +c if and only if T
+
t
u ct u.
By the Weak KAM Theorem 4.7.1, we can nd u

(
0
(M, R)
and c such that

T

t
u + ct = u

. If we set u
+
= u

, and we nd
T
+
t
u
+
= u
+
+ ct. If : [a, b] M is an arbitrary (continuous)
piecewise C
1
curve of C
1
, we see that
u
+
((a))+ c(ba) = T
+
(ba)
u
+
((a)) u
+
((b))
_
b
a
L((s), (s))ds,
which gives u
+
L + c.
By arguments similar to the ones we made for u

, we obtain:
Theorem 4.9.3 (Weak KAM). There exists a Lipschitzian func-
tion u
+
: M R and a constant c such that T
+
t
u
+
ct = u
+
.
This function u
+
satises the following properties
(a) u
+
L +c.
(b) For each x M, there exists a minimizing extremal curve

x
+
: [0, +[ M with
x
+
(0) = x and such that
t [0, +[, u
+
(
x
+
(t)) u
+
(x) =
_
t
0
L(
x
+
(s),
x
+
(s))ds +ct.
163
Conversely, if u
+
(
0
(M, R) satises the properties (a) and (b)
above, then, we have T
+
t
u
+
ct = u
+
.
We also have
c = inf

_
Ld,
where the lower bound is taken over all the Borel probability mea-
sures on TM invariant by the Euler-Lagrange ow
t
. It follows
that c = c[0]. For a curve
x
+
of the type above, we have
s 0, (
x
+
(s),
x
+
(s)) =
s
(x,
x
+
(0)),
since it is a minimizing extremal curve. This implies that
s
(x,
x
+
(0)), s
0 is relatively compact in TM. Moreover, we can nd a Borel prob-
ability measure on TM, invariant by the Euler-Lagrange ow

t
, such that c[0] =
_
Ld, and whose support is contained in
the -limit set of the orbit
s
(x,
x
+
(0)). At each point x M,
where u
+
has a derivative, we have H(x, d
x
u
+
) = c[0].
4.10 The Mather Function on Cohomology.
Let us rst give several characterizations of c[0].
Theorem 4.10.1. Suppose that u (
0
(M, R), and that c R.
If c > c[0] then T

t
u +ct tends uniformly to +, as t +,
and T
+
t
u ct tends uniformly to , as t +.
If c < c[0], then T

t
u + ct tends uniformly to , when t
+, and T
+
t
u ct tends uniformly to +, as t +.
Moreover, we have sup
t0
|T

t
u+c[0]t| < +and sup
t0
|T
+
t
u
c[0]t| < +.
Proof. By the Weak KAM Theorem 4.7.1, there exists u

(
0
(M, R)
with T

t
u

+ c[0]t = u

, for each t 0. As the T

t
are non-
expansive maps, we have
|T

t
u T

t
u

| |u u

and thus
|u u

|u

t
u +c[0]t |u u

+|u

.
164
Theorem 4.10.2. We have the following characterization of the
constant c[0]:
The constant c[0] is the only constant c such that the semi-
group u T

t
u +ct (resp. u T
+
t
u ct) has a xed point
in (
0
(M, R).
The constant c[0] is the greatest lower bound of the set of
the numbers c R for which there exists u (
0
(M, R) with
u L +c.
The constant c[0] is the only constant c R such that there
exists u (
0
(M, R) with sup
t0
|T

t
u + ct|

< + (resp.
sup
t0
|T
+
t
u ct|

< +.)
Proof. The rst point results from the Weak KAM Theorem. The
last point is a consequence of the previous Theorem 4.10.1. The
second point also results from the previous theorem, because we
have u L+c if and only if u T

t
u+ct and in addition, by the
weak KAM theorem, there exists u

with u

= T
t
u

+c[0]t.
If is a C

dierential 1-form, it is not dicult to check the


Lagrangian L

: TM R, dened by
L

(x, v) = L(x, v)
x
(v),
is C
r
like L, with r 2, that

2
L
v
2
=

2
L
v
2
is thus also > 0 denite
as a quadratic form, and that L

is superlinear in the bers of the


tangent bundle TM. We then set
c[] = c
L
[] = c
L
[0].
Proposition 4.10.3. If : M R is a dierentiable function,
then we have
c[ +d] = c[].
In particular, for closed forms , the constant c[] depends only
on the cohomology class.
Proof. We have u (L [ + d]) + c if and only if u +
(L ) +c.
165
The following denition is due to Mather, see [Mat91, page
177].
Denition 4.10.4 (Mathers Function). The function of
Mather is the function : H
1
(M, R) R dened by
() = c[],
where is a class C

dierential 1-form representing the class of


cohomology .
The next theorem is due to Mather, see [Mat91, Theorem 1,
page 178].
Theorem 4.10.5 (Mather). The function is convex and super-
linear on the rst cohomology group H
1
(M, R).
Proof. Let
1
and
2
be two dierential 1-forms of class C

. By
the weak KAM Theorem applied to L

1
and L

2
, we can nd
u
1
, u
2
(
0
(M, R) such that
u
i
(L
i
) +c[
i
].
If t [0, 1], it is not dicult to conclude that
tu
1
+ (1 t)u
2
(L [t
1
+ (1 t)
2
]) + (tc[
1
] + (1 t)c[
2
].
It follows that
c[t
1
+ (1 t)
2
] tc[
1
] + (1 t)c[
2
].
Let us show the superlinearity of . Let us recall that by com-
pactness of M, the rst group of homology is a vector space of
nite dimension. Let us x a nite family
1
, ,
n
: [0, 1] M
of C

closed (i.e.
i
(0) =
i
(1) curves such that the homology
H
1
(M, R) is generated by the homology classes of
1
, ,

. We
can then dene a norm on H
1
(M, R) by
|| = max
_

, . . . ,

_
,
where is a C

closed dierential 1-form representing the coho-


mology class . Let k be an integer. Let us note by
k
i
the closed
166
curve
k
i
: [0, 1] M obtained by going k times through
i
in the
direction of the increasing t and reparametrizing it by the interval
[0, 1]. Let us also note by
k
i
: [0, 1] M the curve opposite to
i
,
i.e.
k
i
(s) =
k
i
(1 s). We then have
_

k
i
= k
_

_

k
i
= k
_

i
,
from which we obtain the equality
k|| = max
_

k
1

, . . . ,

_

k
1

, . . . ,

_

k

_
,
where is a C

closed dierential 1-form representing . By the


Weak KAM Theorem 4.7.1, there exists u

(
0
(M, R) such that
u

(L) +c[]. We deduce from it that for every closed (i.e.


(b) = (a)) curve : [a, b] M, we have
L()
_

+c[](b a) 0.
In particular, we nd
c[]
_

i
k
L(
k
i
),
c[]
_

i
k
L(
k
i
).
Hence, if we set C
k
= max(L(
k
1
), . . . , L(
k

), L(
k
1
), . . . , L(
k

)),
which is a constant which depends only on k, we nd
() = c[] k|| C
k
.
Theorem 4.10.6. If is a closed 1-form, the Euler-Lagrange
ows
L
t
and
L
t
coincide.
167
Proof. Indeed, if
1
,
2
: [a, b] M are two curves with the same
ends and close enough, they are homotopic with xed ends and
thus
_

1
=
_

2
. It follows that the actions for L and L

are
related by
L

(
i
) = L(
i
)
_

1
.
Hence, the critical points of L

and of L on the space of curves with


xed endpoints are the same ones. Consequently, the Lagrangians
L and L

have same the extremal curves.


Corollary 4.10.7. We have c[] = inf

_
TM
(L )d, where
varies among the Borel probability measures on TM invariant
under the Euler-Lagrange ow
t
of L. Moreover, there exists
such a measure with compact support and satisfying c[] =
_
TM
(L )d.
4.11 Dierentiability of Dominated Func-
tions
In the sequel we denote by

B(0, r) (resp.

B(0, r)) the open ball
(resp. closed) of center 0 and radius r in the Euclidean space R
k
,
where k is the dimension of M.
Proposition 4.11.1. Let :

B(0, 5) M be a coordinate chart
and t
0
> 0 be given. There is a constant K 0 such that for each
function u (
0
(M, R), for each x

B(0, 1), and for each t t
0
,
we have:
(1) For each y

B(0, 1) and each extremal curve : [0, t] M
with (t) = (x) and
T

t
u[(x)] = u((0)) +
_
t
0
L((s), (s))ds,
we have
T

t
u[(y)]T

t
u[(x)]
L
v
((x), (t))(D(x)[yx])+K|yx|
2
.
In particular, if T

t
u is dierentiable at (x) then
d
(x)
T

t
u = L/v((x), (t))
168
and the curve is unique.
(2) For each y

B(0, 1) and each curve : [0, t] M with
(0) = (x) and
T
+
t
u[(x)] = u((t))
_
t
0
L((s), (s))ds,
we have
T
+
t
u[(y)]T
+
t
u[(x)]
L
v
((x), (0))(D(x)[yx])K|yx|
2
.
In particular, if T
+
t
u is dierentiable at (x) then
d
(x)
T
+
t
u = L/v((x), (0))
and the curve is unique.
Proof. We use some auxiliary Riemannian metric on M to have a
norm on tangent space and a distance on M.
Since T

t
u = T

t
0
T

tt
0
u by the semigroup property, we have
only to consider the case t = t
0
. By corollary 4.4.5, we can nd a
nite constant A, such that any minimizer dened on an interval
of time at least t
0
> 0 has speed uniformly bounded in norm
by A.This means that such curves are all A-Lipschitz. We then
pick > 0 such that for each ball

B(y, A), for y (

B(0, 1)) is
contained in (

B(0, 2)). Notice that this does not depend on u.


Since any curve , as in part (1) of the proposition is necessary a
minimizer, we therefore have
([t
0
, t
0
]) (

B(0, 2)).
We then set =
1
and

L(x, w) = L((x), D(x)w) for
(x, w)

B(0, 5) R
k
. Taking derivatives, we obtain

L(x, w)
w
=
L
v
((x), D(x)(w))[D(x)()].
The norm of the vector h = y x is 2, hence if we dene

h
(s) =
s(t
0
)

h+ (s), for s [t
0
, t
0
], we have
h
(s)

B(0, 4),
and
T

t
0
u[(x+h)]T

t
0
u[(x)]
_
t
0
t
0

L(
h
(s),


h
(s))L( (s),

(s))] ds.
169
Since the speed of is bounded in norm by A, we see that ( (s),

(s))
is contained in a compact subset of

B(0, 2) R
k
independent of
u, x and . Moreover, we have
h
(s) (s) =
s(t)

h, which
is of norm 2, and


h
(s)

(s) =
1

h, which is itself of norm


2/. We then conclude that there exists a compact convex sub-
set C

B(0, 4) R
k
, independent of u, x and , and containing
( (s),

(s)) and (
h
(s),


h
(s)), for each s [t
0
, t
0
]. Since we
can bound uniformly on the compact subset C the norm of the
second derivative of

L, by Taylors formula at order 2 applied to

L, we see that there exists a constant



K independent of u, x and
of such that for each s [t
0
, t
0
]
[

L(
h
(s),


h
(s))

L( (s), (s))

L
x
( (s), (s))
_
s (t )

h
_

L
w
( (s),

(s))
_
h

_
[


K max
_
[s (t )[

|h|,
|h|

_
2

2
|h|
2
,
supposing that < 1. As is an extremal curve for the Lagrangian

L, it satises the Euler-Lagrange equation

L
x
( (s), (s)) =
d
ds
_

L
w
((s), (s))
_
hence
[

L(
h
(s),


h
(s))

L( (s), (s))
1

d
ds
_
[s(t)]

L
w
( (s), (s))(h)
_
[

2
|h|
2
.
If we integrate on the interval [t
0
, t
0
], we obtain
T

t
0
u[(x +h)] T

t
0
u[(x)]

L
w
((x),

(t
0
))(h) +

|h|
2
.
We just have to take K

K/.
If T

t
u is dierentiable at (x), for v R
n
and small enough,
we can write
T

t
u[(x+v)]T

t
u[(x)]
L
v
((x), (t))(D(x)[v])+K|v|
2
.
170
If we divide by > 0 and we let go to 0, we obtain
v R
n
d
(x)
T

t
uD(x)[v]
L
v
((x), (t))(D(x)[v]).
Since we can also apply the inequality above with v instead of
v, we conclude that If we divide by > 0 and we let go to 0, we
obtain
v R
n
d
(x)
T

t
uD(x)[v] =
L
v
((x), (t))(D(x)[v]),
by the linearity in v of the involved maps. Since is a dieomor-
phism this shows that d
(x)
T

t
uL/v((x), (t)). By the bijec-
tivity of the Legendre Transform the tangent vector (t) is unique.
Since is necessarily a minimizing extremal, it is also unique, since
both its position x and its speed (t) at t are uniquely determined.
To prove (2), we can make a similar argument for T
+
t
, or,
more simply, apply what we have just done to the symmetrical
Lagrangian

L of L.
Exercise 4.11.2. 1) Let u

: M R be a weak KAM solution.


Show that u

has a derivative at x if and only if there is one and


only one curve
x

:] , 0] M such that
x

(0) = x and
t 0, u

(
x

(0)) u

(
x

(t)) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t.
In that case we have d
x
u

=
L
v
(x,
x

(0)).
2) Suppose that x M, and that
x

:] , 0] M satises

(0) = x and
t 0, u

(
x

(0)) u

(
x

(t)) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t.
Show that necessarily
H(x,
L
v
(x,
x

(0)) = c[0].
We will need a criterion to show that a map is dierentiable
with a Lipschitz derivative. This criterion has appeared in dif-
ferent forms either implicitly or explicitly in the literature, see
171
[CC95, Proposition 1.2 page 8], [Her89, Proof of 8.14, pages 63
65], [Kni86], [Lio82, Proof of Theorem 15.1, pages 258-259], and
also [Kis92] for far reaching generalizations. The simple proof
given below evolved from discussions with Bruno Sevennec.
Proposition 4.11.3 (Criterion for a Lipschitz Derivative). Let

B
be the open unit ball in the normed space E. Fix a map u :

B R.
If K 0 is a constant, denote by A
K,u
the set of points x

B, for
which there exists
x
: E R a continuous linear form such that
y

B, [u(y) u(x)
x
(y x)[ K|y x|
2
.
Then the map u has a derivative at each point x A
K,u
,
and d
x
u =
x
. Moreover, the restriction of the map x d
x
u
to x A
K,u
[ |x| <
1
3
is Lipschitzian with Lipschitz constant
6K.
More precisely
x, x

A
K,u
, |xx

| < min(1|x|, 1|x

|) |d
x
ud
x
u| 6K|xx

|.
Proof. The fact that d
x
u =
x
, for x A
K,u
, is clear. Let us x
x, x

A
K,u
, with |x x

| < min(1 |x|, 1 |x

|). If x = x

there is nothing to show, we can then suppose that |x x

| > 0.
If h is such that |h| = |xx

|, then the two points x+h et x

+h
are in

B. This allows us to write
[u(x +h) u(x)
x
(h)[ K|h|
2
[u(x) u(x

)
x
(x x

) K|x x

|
2
[u(x +h) u(x

)
x
(x x

+h)[ K|x x

+h|
2
.
As |h| = |x x

|, we obtain from the last inequality


[u(x

) u(x +h) +
x
(x x

+h)[ 4K|x x

|
2
.
Adding this last inequality with the rst two above, we nd that,
for each h such that |h| = |x x

|, we have
[
x
(h)
x
(h)[ 6K|x x

|
2
,
hence
|
x

x
| = sup
h=xx

[
x
(h)
x
(h)[
|x x

|
6K|x x

|.
172
Exercise 4.11.4. If A
k,u
is convex, for example if A
K,U
=

B,
show that the derivative is Lipschitzian on A
K,u
with Lipschitz
constant 6K.
Theorem 4.11.5. If > 0 is given, then there are constants
A 0 and > 0, such that any map u : M R, with u
L+c, is dierentiable at every point of the set /
,u
formed by the
x M for which there exists a (continuous) piecewise C
1
curve
: [, ] M with (0) = x and
u(()) u(()) =
_

L((s), (s)) ds + 2c.


Moreover we have
(1) Such a curve is a minimizing extremal and
d
x
u =
L
x
(x, (0));
(2) the set /
,u
is closed;
(3) the derivative map /
,u
T

M, x (x, d
x
u) is Lipschitzian
with Lipschitz constant A on each subset /
,u
with diam-
eter .
Proof. The fact that is a minimizing extremal curve results from
u L +c. This last condition does also imply that
u(()) u(x) =
_

0
L((s), (s)) ds +c,
u(x) u(()) =
_
0

L((s), (s)) ds +c. (*)


Since > 0 is xed, by the Corollary of A Priori Compactness, we
can nd a compact subset K

TM such that for each minimizing


extremal curve : [, ] M, we have ((s), (s)) K

. It is
not then dicult to deduce that /
,u
is closed. It also results from
() that
T
+

u(x) u(x) +c
T

u(x) u(x) c. (**)


173
As u L +c, we have
T
+

u u +c,
T

u u c.
We then obtain equality in (). Subtracting this equality from
the inequality above, we nd
y M, T
+

u(y) T
+

u(x) u(y) u(x) T

u(y) T

u(x).
()
Let us then cover the compact manifold M by a nite number of
open subsets of the form
1
(

B(0, 1/3)), ,

B(0, 1/3)), where

i
:

B(0, 5) M, i = 1, . . . , , is a C

coordinate chart. By
Proposition 4.11.1, there exists a constant K, which depends only
on and the xed
p
, p = 1, . . . , such that, if x
i
(

B(0, 1/3)),
setting x =
i
( x), for each y

B(0, 1), we have
T

u(
i
(y)) T

u(
i
( x))
L
v
(x, (0)) D
i
( x)(y x) +K|y x|
2
T
+

u(
i
(y) T
+

u(
i
( x))
L
v
(x, (0)) D
i
( x)(y x) K|y x|
2
,
Using the inequalities () we get
[u(
i
(y)) u(
i
( x))
L
v
(x, (0)) D
i
( x)(y x)[ K|y x|
2
.
By the Criterion for a Lipschitz Derivative 4.11.3, we nd that
d
x
u exists and is equal to
L
v
(x, (0)). Moreover the restriction of
x d
x
u on /
,U

i
(

B(0, 1/3)) is Lipschitzian with a constant of


Lipschitz which depends only on . It is then enough to choose for
> 0 a Lebesgue number for the open cover (
i
(

B(0, 1/3))
i=1, ,
of the compact manifold M.
Denition 4.11.6 (The sets o

and o
+
). We denote by o

(resp.
o
+
) the set of weak KAM solutions of the type u

(resp. u
+
), i.e.
the continuous functions u : M R such that T

t
u + c[0]t = u
(resp. T
+
t
u c[0]t = u).
Exercise 4.11.7. If c R, show that u o

(resp. u o
+
) if
and only if u+c o

(resp. u+c o
+
). If x
0
M is xed, show
that the set u o

[ u(x
0
) = 0 (resp. u o
+
[ u(x
0
) = 0) is
compact for the topology of uniform convergence.
174
Theorem 4.11.8. Let u : M R be a continuous function. The
following properties are equivalent
(1) the function u is C
1
and belongs to o

,
(2) the function u is C
1
and belongs to o
+
.
(3) the function u belongs to the intersection o

o
+
. (4)] the
function u is C
1
and there exists c R such that H(x, d
x
u) =
c, for each x M.
In all the cases above, the derivative of u is (locally) Lipschitzian.
Proof. Conditions (1) or (2) imply (4). It is enough, then to show
that (4) implies (1) and (2) and that (3) implies that u is C
1
, and
that its derivative is (locally) Lipschitzian. Thus let us suppose
condition (3) satised. Indeed, in this case, if x M, we can nd
extremal curves
x

:] , 0] M and
x
+
: [0, [ M, with

(0) =
x
+
(0) = x and
t 0,u(
x
+
(t)) u(x) =
_
t
0
L(
x
+
(s),
x
+
(s)) ds +c[0]t,
u(x) u(
x

(t)) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t.
The curve : [1, 1] M, dened by [[1, 0] =
x

and [[0, 1] =

x
+
, shows that x /
1,u
. By the previous theorem, the function
u is of class C
1
and its derivative is (locally) Lipschitzian.
Let us suppose that u satises condition (4). By the Fenchel
inequality, we have
(x, v) TM, d
x
u(v) H(x, d
x
u) +L(x, v)
= c +L(x, v).
Consequently, if : [a, b] M is a C
1
curve, we obtain
s [a, b], d
(s)
u( (s)) c +L((s), (s)), (*)
and by integration on the interval [a, b]
u((b)) u((a)) c(b a) +
_
b
a
L((s), (s)) ds, ()
175
which gives us u L + c. If : [a, b] M is an integral curve
of grad
L
u, we have in fact equality in () and thus in (). Since
u L+c, it follows that is a minimizing extremal curve. If two
solutions go through the same point x at time t
0
, they are, in fact,
equal on their common interval of denition, since they are two ex-
tremal curves which have the same tangent vector (x, grad
L
u(x))
at time t
0
. As we can nd local solutions by the Cauchy-Peano
Theorem, we see that grad
L
u is uniquely integrable. Since M
is compact, for any point x M, we can nd an integral curve

x
:] , +[ M of grad
L
u with
x
(0) = x. This curve gives
at the same time a curve of the type
x

and one of the type


x
+
for u. This establishes that u o

o
+
.
4.12 Mathers Set.
The denition below is due to Mather, see [Mat91, page 184].
Denition 4.12.1 (Mather Set). The Mather set is

/
0
=
_

supp() TM,
where supp() is the support of the measure , and the union is
taken over the set of all Borel probability measures on TM invari-
ant under the Euler-Lagrange ow
t
, and such that
_
TM
Ld

=
c[0].
The projection /
0
= (

/
0
) M is called the projected
Mather set.
As the support of an invariant measure is itself invariant under
the ow, the set

/
0
is invariant by
t
.
Lemma 4.12.2. If (x, v)

/
0
and u L + c[0], then, for each
t, t

R, with t t

, we have
u(
t
(x, v)) u(
t
(x, v)) =
_
t

t
L(
s
(x, v)) ds +c[0](t

t).
Proof. By continuity, it is enough to see it when (x, v) supp()
with a Borel probability measure on TM, invariant by
t
and
176
such that
_
TM
Ld

= c[0] Since u L + c[0], for each (x, v)


TM, we have
u(
t
(x, v)) u(
t
(x, v))
_
t

t
L(
s
(y, w)) ds +c[0](t

t).
()
If we integrate this inequality with respect to , we nd by the
invariance of
_
TM
u d
_
TM
u d (t

t)
_
_
L
TM
d +c[0]
_
,
which is in fact the equality 0 = 0. It follows that the inequality
() is an equality at any point (x, v) contained in supp().
Theorem 4.12.3. A function u (
0
(M, R), such that u L +
c[0], is dierentiable at every point of the projected Mather set
/
0
= (

/
0
). Moreover, if (x, v)

/
0
, we have
d
x
u =
L
v
(x, v)
and the map /
0
T

M, x (x, d
x
u) is locally Lipschitzian
with a (local) Lipschitz constant independent of u.
Proof. If (x, v)

/
0
, we set
x
(s) =
s
(x, v). We then have

x
(0) = x,
x
(0) = v, and (
x
(s),
x
(s)) =
s
(x, v). In particular,
by Lemma 4.12.2 above
u(
x
(1)) u(
x
(1)) =
_
1
1
L(
x
(s),
x
(s)) ds + 2c[0],
thus x /
1,u
, where /
1,u
is the set introduced in 4.11.5. Conse-
quently, the derivative d
x
u exists and is equal to
L
v
(x, v). More-
over, the map x (x, d
x
u) is locally Lipschitzian with a Lipschitz
constant independent of u.
Corollary 4.12.4 (Mather). The map [

/
0
:

/
0
/
0
is
injective. Its inverse is Lipschitzian.
Proof. Let u L + c[0] be xed, for example a weak KAM so-
lution. By the previous Theorem ??, the inverse of on /
0
is
x

L
1
(x, d
x
u), which is Lipschitz as a composition of Lipschitz
functions.
177
The following corollary is due to Carneiro, see [Car95, Theorem
1, page 1078].
Corollary 4.12.5 (Carneiro). The Mather set

/
0
is contained
in the energy level c[0], i.e.
(x, v)

/
0
, H(x,
L
v
(x, v)) = c[0].
Proof. Let u

be a weak KAM solution. It is known that for


(x, v)

/
0
, the function u

is dierentiable at x and d
x
u

=
L
v
(x, v).
The functions of o

and o
+
are completely determined by their
values on /
0
as we show in the following theorem.
Theorem 4.12.6 (Uniqueness). Suppose that u

, u

are both in
o

(resp. u
+
, u
+
are both in o
+
). If u

= u

(resp. u
+
= u
+
) on
/
0
, then, we have u

= u

(resp. u
+
= u
+
) everywhere on M.
Proof. Let us x x M, we can nd an extremal curve
x

:
] , 0] M, with
x

(0) = x and such that


t 0, u

(x) u

(
x

(t)) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t.
Since u

L +c[0], we have
t 0, u

(x) u

(
x

(t))
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t.
It follows that
t 0, u

(x) u

(
x

(t)) u

(x) u

(
x

(t)). (*)
In addition, we know that s (
x

(s),
x

(s)) is a trajectory of
the Euler-Lagrange ow
t
and that the -limit set of this tra-
jectory carries a Borel probability measure invariant by
t
and
such that
_
TM
Ld = c[0]. The support of this measure is thus
contained in

/
0
. We conclude from it, that there exists a se-
quence t
n
+ such that (
x

(t
n
),
x

(t
n
)) converges to a
point (x

, v

)

/
0
, in particular, we have
x

(t
n
) x

which is in /
0
. It follows that u

(
x

(t
n
)) u

(
x

(t
n
)) tends
to u

(x

) u

(x

), which is 0 since x

/
0
. From (), we
then obtain the inequality u

(x) u

(x) 0. We can of course


exchange the role of u

and that of u

to conclude.
178
4.13 Complements
If u : M R is a Lipschitz function, we will denote by dom(du)
the domain of denition of du, i.e. the set of the points x where
the derivative d
x
u exists. The graph of du is
Graph(du) = (x, d
x
u) [ x dom(u) T

M.
Let us recall that Rademachers theorem 1.1.10 says that M
dom(u) is negligible (for the Lebesgue class of measures). Since
|d
x
u|
x
is bounded by the Lipschitz constant of u, it is not di-
cult to use a compactness argument to show that the projection

_
Graph(du)
_
is the whole of M.
Lemma 4.13.1. Suppose that u

. If x M and
x

:
] , 0] M is such that
x

(0) = x and
t 0, u

(x) u

(
x

(t)) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t,
then, the function u

has a derivative at each point


x

(t) with
t > 0, and we have
t > 0, d

(t)
u

=
L
v
(
x

(t),
x

(t)).
It follows that
t, s > 0, (
x

(t s), d

(ts)
u

) =

s
(
x

(t), d

(t)
u

).
We also have
t 0, H
_

(t),
L
v
(
x

(t),
x

(t))

= c[0].
Moreover, if u

has a derivative at x, we have


d
x
u

=
L
v
(x,
x

(0))
and d

(t)
u

t
(x, d
x
u

), for each t 0.
There is a similar statement for the functions u
+
o
+
.
179
Proof. For the rst part, we notice that the curve : [t, t] M
dened by (s) =
x

(s t) shows that (0) =


x

(t) is in /
t,u

,
for each t > 0. It follows that d

(t)
u

=
L
v
(
x

(t),
x

(t)).
The fact that (
x

(t s), d

(ts)
u

) =

s
(
x

(t), d

(t)
u

)
is, by Legendre transform, equivalent to
s
(
x

(t),
x

(t)) =
(
x

(t s),
x

(t s)), which is true since


x

is an extremal
curve. Let us suppose that u

has a derivative at x. The relation


u

L +c[0] implies that


v T
x
M, d
x
u

(v) c[0] +L(x, v).


Moreover, taking the derivative at t = 0 of the equality
t 0, u

(x) u

(
x

(t) =
_
0
t
L(
x

(s),
x

(s)) ds +c[0]t,
we obtain the equality d
x
u

(
x

(0)) = c[0] +L(x,


x

(0)). We then
conclude that H(x, d
x
u

) = c[0] and that d


x
u

=
L
v
(x,
x

(0)).
In particular, it follows that we have H(y, d
y
u

) = c[0] at any
point y M where u

has a derivative. By the rst part, with-


out any assumption on the dierentiability of u

at x, we nd
that H
_

(t),
L
v
(
x

(t),
x

(t))

= c[0], for each t > 0. By


continuity this equality is also true for t = 0.
Theorem 4.13.2. Let u

. The derivative map x (x, d


x
u

)
is continuous on its domain of denition dom(du

). The sets
Graph(du

) and Graph(du

) are invariant by

t
, for each t 0.
Moreover, for each (x, p) Graph(du

), we have H(x, p) = c[0].


The closure Graph(du

) is the image by the Legendre trans-


form

L : TM T

M of the subset of TM formed by the (x, v)


TM such that
t 0, u

(x) u

[(
t
(x, v))] =
_
0
t
L(
s
(x, v)) ds +c[0]t,
i.e. the set of the (x, v) such that the extremal curve : ], 0]
M, with (0) = x and (0) = v, is a curve of the type
x

for u

.
Proof. The above Lemma 4.13.1 shows that Graph(du

) is invari-
ant by

t
, for each t 0 and that Graph(du

) H
1
(c[0]).
Since the ow

t
is continuous, the closure Graph(du

) is also
180
invariant by

t
, for each t 0. In the same way, the inclu-
sion Graph(du

) H
1
(c[0]) results from the continuity of H.
If we denote by D

the subset of TM dened in the last part of


the theorem, Lemma ?? also shows that Graph(du

)

L(D

)
Graph(du

). If x
n
is a sequence in dom(du

) and (x
n
, d
xn
u

)
(x, p), let us then show that
t 0, u

(x) u

[(
t
(x, v))] =
_
0
t
L(
s
(x, v)) ds +c[0]t, ()
where v T
x
M is dened by p =
L
v
(x, v). For that we dene
v
n
T
xn
M by d
xn
u

=
L
v
(x
n
, v
n
). We have (x
n
, v
n
) (x, v).
By Lemma ??, the extremal curve
xn
: ] , 0] M is s
(
s
(x
n
, v
n
)), hence we obtain
t 0, u

(x
n
) u

[(
t
(x
n
, v
n
))] =
_
0
t
L(
s
(x
n
, v
n
)) ds+c[0]t.
When we let n tend to +, we nd (). We conclude that (x, v)
D

and thus

L(D

) = Graph(du

). Moreover by Lemma ??, if


x dom(du

) we necessarily have d
x
u

=
L
v
(x, v) = p. As
Graph(du

) is contained in the compact subset H


1
(c[0]), we then
obtain the continuity of x (x, d
x
u

) on dom(du

).
We have of course a similar statement for the functions in o
+
.
Theorem 4.13.3. If u
+
o
+
, the derivative map x (x, d
x
u
+
)
is continuous on its domain of denition dom(du
+
).
The sets Graph(du
+
) and Graph(du
+
) are invariant by

t
, for
each t 0. Moreover, for each (x, p) Graph(du
+
), we have
H(x, p) = c[0]. The closure Graph(du
+
) is the image by the Le-
gendre transform

L : TM T

M of the subset of TM formed by


the (x, v) TM such that
t 0, u
+
(
t
(x, v)) u
+
(x) =
_
t
0
L(
s
(x, v)) ds +c[0]t,
i.e. the set of (x, v) such that the extremal curve : [0, +[ M,
with (0) = x and (0) = v, is a curve of the type
x
+
for u
+
.
181
4.14 Examples
Denition 4.14.1 (Reversible Lagrangian). The Lagrangian L
is said to be reversible if it satises L(x, v) = L(x, v), for each
(x, v) TM.
Example 4.14.2. Let g be a Riemannian metric on M, we denote
by | |
x
the norm deduced from g on T
x
M. If V : M R is C
2
,
the Lagrangian L dened by L(x, v) =
1
2
|v|
2
x
V (x) is reversible.
Proposition 4.14.3. For a reversible Lagrangian L, we have
c[0] = inf
xM
L(x, 0) = inf
(x,v)TM
L(x, v).
Moreover

/
0
= (x, 0) [ L(x, 0) = c[0].
Proof. By the strict convexity and the superlinearity of L in the
bers of the tangent bundle TM, we have L(x, 0) = inf
vTxM
L(x, v),
for all x M. Let us set k = inf
xM
L(x, 0) = inf
(x,v)TM
L(x, v).
Since c[0] = inf
_
Ld, where the inmum is taken over all Borel
probability measures on TM invariant under the ow
t
, we ob-
tain k c[0]. Let then x
0
M be such that L(x
0
, 0) = k,
the constant curve ] , +[ M, t x
0
is a minimizing ex-
tremal curve. Consequently
t
(x
0
, 0) = (x
0
, 0) and the Dirac mass

(x
0
,0)
is invariant by
t
, but
_
Ld
(x
0
,0)
= k. Therefore c[0] = k
and (x
0
, 0)

/
0
. Let be a Borel probability measure on TM
such that
_
TM
Ld = c[0]. Since c[0] = inf
TM
L, we necessar-
ily have L(x, v) = inf
TM
L on the support of . It follows that
supp() (x, 0) [ L(x, 0) = c[0].
We then consider the case where M is the circle T = R/Z.
We identify the tangent bundle TT with T R. As a Lagran-
gian L we take one dened by L(x, v) =
1
2
v
2
V (x), where
V : T R is C
2
. We thus have c[0] = inf
TR
L = supV ,
hence c[0] = supV . Let us identify T

T with T R. The Hamil-


tonian H is given by H(x, p) =
1
2
p
2
+V (x). The dierential equa-
tion on T

T which denes the ow


t
is given by x = p and
p = V

(x). If u

, the compact subset Graph(du

) is con-
tained in the set H
1
(c[0]) = (x, p) [ p =
_
supV V (x).
We strongly encourage the reader to do some drawings FAIRE
182
DES DESSINS of the situation in R R, the universal cover
of T R. To describe u

completely let us consider the case


where V reaches its maximum only at 0. In this case the set
H
1
(c[0]) consists of three orbits of

t
, namely the xed point
(0, 0), the orbit O
+
= (x,
_
supV V (x)) [ x ,= 0 and the orbit
O

= (x,
_
supV V (x)) [ x ,= 0. On O
+
the direction of the
increasing t is that of the increasing x (we identify in a natural way
T0 with ]0, 1[). On O

the direction of the increasing t is that of


the decreasing x. Since Graph(du

) is invariant by the maps

t
,
for t 0, if (x,
_
supV V (x)) Graph(du

), then we must have


(y,
_
supV V (y)) Graph(du

), for each y ]0, x]. By symme-


try we get (y,
_
supV V (y)) Graph(du

), for each y [x, 1[.


It follows that there is a point x
0
such that Graph(du

) is the
union of (0, 0) and the two sets (y,
_
supV V (y)) [ y ]0, x
0
]
and (y,
_
supV V (y)) [ y [x
0
, 1[. Moreover, since the
function u

is dened on T, we have lim


x1
u

(x) = u

(0) and
thus the integral on ]0, 1[ of the derivative of u

must be 0. This
gives the relation
_
x
0
0
_
supV V (x) dx =
_
1
x
0
_
supV V (x) dx.
This equality determines completely a unique point x
0
, since supV
V (x) > 0 for x ]0, 1[. In this case, we see that u

is unique up
to an additive constant and that
u

(x) =
_
u

(0) +
_
x
0
_
supV V (x) dx, if x [0, x
0
];
u

(0) +
_
1
x
_
supV V (x) dx, if x [x
0
, 1].
Exercise 4.14.4. 1) If V : T R reaches its maximum exactly
n times, show that the solutions u

depend on n real parameters,


one of these parameters being an additive constant.
2) Describe the Mather function : H
1
(T, R) R, c[].
3) If is a closed dierential 1-form on T, describe the func-
tion u

for the Lagrangian L

dened by L(x, v) =
1
2
v
2
V (x)

x
(v).
Chapter 5
Conjugate Weak KAM
Solutions
In this chapter, as in the previous ones, we denote by M a com-
pact and connected manifold. The projection of TM on M is
denoted by : TM M. We suppose given a C
r
Lagrangian
L : TM R, with r 2, such that, for each (x, v) TM, the
second vertical derivative

2
L
v
2
(x, v) is denite > 0 as a quadratic
form, and that L is superlinear in each ber of the tangent bundle
: TM M. We will also endow M with a xed Riemannian
metric. We denote by d the distance on M associated with this
Riemannian metric. If x M, the norm | |
x
on T
x
M is the one
induced by this same Riemannian metric.
5.1 Conjugate Weak KAM Solutions
We start with the following lemma
Lemma 5.1.1. If u L +c[0], then we have
x /
0
, t 0, u(x) = T

t
u(x) +c[0]t = T
+
t
u(x) c[0]t.
Proof. Since u L + c[0], we have u T

t
u + c[0]t and u
T
+
t
u c[0]t. We consider the point (x, v)

/
0
above x. Let us
note by : ] , +[ M the extremal curve s (
s
(x, v)).
183
184
By lemma 4.12.2, for each t 0, we have
u((0)) u((t)) =
_
0
t
L((s), (s)) ds +c[0]t,
u((t)) u((0)) =
_
t
0
L((s), (s)) ds +c[0]t.
Since (0) = x, we obtain the inequalities u(x) T

t
u(x) + c[0]t
and u(x) T
+
t
u(x) c[0]t.
Theorem 5.1.2 (Existence of Conjugate Pairs). If u : M R
is a function such that u L + c[0], then, there exists a unique
function u

(resp. u
+
o
+
) with u = u

(resp. u = u
+
) on
the projected Mather set /
0
. These functions verify the following
properties
(1) we have u
+
u u

;
(2) if u
1

(resp. u
1
+
o
+
) veries u u
1

(resp. u
1
+
u),
then u

u
1

(resp. u
1
+
u
+
);
(3) We have u

= lim
t+
T

t
u+c[0]t and u
+
= lim
t+
T
+
t
u
c[0]t, the convergence being uniform on M.
Proof. It will be simpler to consider the modied semigroup

T

t
v =
T

t
v + c[0]t. The elements of o

are precisely the xed points of


the semigroup

T

t
. The condition u L + c[0] is equivalent to
u

T

t
u. As

T

t
preserves the order, we see that

T

t
u u
1

for
each u
1

satisfying u u
1

. As

T

t
u = u on the projected
Mather set /
0
, it then remains to show that

T

t
u is uniformly
convergent for t . However, we have

T

t
u

T

t+s
u, if s 0,
because this is true for t = 0 and the semigroup

T

t
preserves the
order. Since for t 1 the family of maps

T

t
u is equi-Lipschitzian,
it is enough to see that this family of maps is uniformly bounded.
To show this uniform boundedness, we x u
0

, by compact-
ness of M, there exists k R such that u u
0

+k. By what was


already shown, we have

T

t
u u
0

+k.
Corollary 5.1.3. For any function u

(resp. u
+
o
+
),
there exists one and only one function of u
+
o
+
(resp. u

)
satisfying u
+
= u

on /
0
. Moreover, we have u
+
u

on all M.
185
Denition 5.1.4 (Conjugate Functions). A pair of functions (u

, u
+
)
is said to be conjugate if u

, u
+
o
+
and u

= u
+
on /
0
.
We will denote by T the set formed by the dierences u

u
+
of
pairs (u

, u
+
) of conjugate functions.
The following lemma will be useful in the sequel.
Lemma 5.1.5 (Compactness of the Dierences). All the functions
in T are 0. Moreover, the subset T is compact in (
0
(M, R) for
the topology of uniform convergence.
Proof. If u

and u
+
are conjugate, we then know that u
+
u

and thus u

u
+
0. If we x x
0
M, the set o
x
0

= u

[
u

(x
0
) = 0 (resp. o
x
0
+
= u
+
[ u
+
(x
0
) = 0) is compact, since it
is a family of equi-Lipschitzian functions on the compact manifold
M which all vanishes at the point x
0
. However, for c R, it is
obvious that the pair (u

, u
+
) is conjugate if and only if the pair
(u

+ c, u
+
+ c) is conjugate. We conclude that T is the subset
of the compact subset o
x
0

o
x
0
+
formed by the functions which
vanish on /
0
.
Corollary 5.1.6. Let us suppose that all the functions u

are C
1
(what is equivalent to o

= o
+
). Then, two arbitrary
functions in o

dier by a constant.
Proof. Conjugate functions are then equal, because the C
1
func-
tions contained in o

or o
+
are also in o

o
+
by 4.11.8. Sup-
pose then that u
1

and u
2

are two functions in o

. We of course
do have u = (u
1

+ u
2

)/2 L + c[0]. By the Theorem of Ex-


istence of Conjugate Pairs 5.1.2, we can nd a pair of conju-
gate functions (u

, u
+
) with u
+
u u

. As conjugate func-
tions are equal, we have u = (u
1

+ u
2

)/2 o

. The three
functions u, u
1

and u
2

are C
1
and in o

, we must then have


H(x, d
x
(u
1

+u
2

)/2) = H(x, d
x
u
1

) = H(x, d
x
u
2

) = c[0], for each


x M. This is compatible with the strict convexity of H in bers
of T

M only if d
x
u
1

= d
x
u
2

, for each x M.
186
5.2 Aubry Set and Ma ne Set.
Denition 5.2.1 (The Set 1
(u

,u
+
)
). Let us consider a pair (u

, u
+
)
of conjugate functions. We denote by 1
(u

,u
+
)
, the set
1
(u

,u
+
)
= x M [ u

(x) = u
+
(x).
We have 1
(u

,u
+
)
/
0
.
Theorem 5.2.2. For each x 1
(u

,u
+
)
, there exists an extremal
curve a
x
: ] , +[ M, with
x
(0) = x,
x
(0) = v and such
that, for each t R, we have u

([
t
(x, v)]) = u
+
([
t
(x, v)]) and
t t

R, u

(
x
(t

))u

(
x
(t)) =
_
t

t
L(
x
(s),
x
(s)) ds+c[0](t

t).
It follows that the functions u

and u
+
are dierentiable at every
point of 1
(u

,u
+
)
with the same derivative. Moreover, there exists
a constant K which depends only on L and such that the section
1
(u

,u
+
)
T

M, x d
x
u

= d
x
u
+
is Lipschitzian with Lipschitz
constant K.
Proof. Let us x x 1
(u

,u
+
)
. There exists extremal curves
x

:
] , 0] M and
x
+
: [0, +[ M with
x

(0) =
x
+
(0) = x and
for each t [0, +[
u
+
(
x
+
(t)) u
+
(x) = c[0]t +
_
t
0
L(
x
+
(s),
x
+
(s)) ds,
u

(x) u

(
x

(t)) = c[0]t +
_
0
t
L(
x

(s),
x

(s)) ds.
But, since u

L + c[0], u
+
L + c[0], u
+
u

and u

(x) =
u
+
(x), we have
u
+
(
x
+
(t)) u
+
(x) u

(
x
+
(t)) u

(x)
c[0]t +
_
t
0
L(
x
+
(s),
x
+
(s)) ds,
and
u

(x) u

(
x

(t)) u
+
(x) u
+
(
x

(t))
c[0]t +
_
0
t
L(
x

(s),
x

(s)) ds.
187
We thus have equality everywhere. It is not dicult to deduce
that the curve
x
which is equal to
x

on ] , 0] and to
x
+
on
[0, +[ satises
t t

R, u

(
x
t

)u

(
x
t) =
_
t

t
L(
x
(s),
x
(s)) ds+c[0](t

t).
It follows that
x
is the sought extremal curve. The existence
of
x
shows that x /
1,u

and x /
1,u
+
and thus u

and u
+
are dierentiable at x with d
x
u

= d
x
u
+
=
L
v
(x,
x
(0)). The
existence of K also results from x /
1,u

(or x /
1,u
+
).
Denition 5.2.3 (The set

1
(u

,u
+
)
). If (u

, u
+
) is a pair of con-
jugate functions, we dene the set

1
(u

,u
+
)
by

1
(u

,u
+
)
= (x, v) [ x 1
(u

,u
+
)
, d
x
u

= d
x
u
+
=
L
v
(x, v).
Theorem 5.2.4. If (u

, u
+
) is a pair of conjugate functions, the
projection : TM induces a bi-Lipschitzian homeomorphism

1
(u

,u
+
)
on 1
(u

,u
+
)
. The set

1
(u

,u
+
)
is compact and invariant by
the Euler-Lagrange ow
t
. It contains

/
0
. If (x, v)

1
(u

,u
+
)
,
for each t R, we have u

([
t
(x, v)]) = u
+
([
t
(x, v)]) and for
all t t

R
u

([
t
(x, v)]) u

([
t
(x, v)]) =
_
t

t
L(
s
(x, v)) ds+c[0](t

t).
Proof. By the previous theorem 5.2.2 the projection restricted
to

1
(u

,u
+
)
is surjective onto 1
(u

,u
+
)
with a Lipschitzian inverse.
In particular, the set

1
(u

,u
+
)
is compact. Moreover, if (x, v)

1
(u

,u
+
)
and
x
is the extremal curve given by the previous theo-
rem, we have v =
x
(0), and, for each s R, we have (
x
(s),
x
(s))

1
(u

,u
+
)
, because
x
(s) 1
(u

,u
+
)
and the extremal curve t

x
(s + t) can be used as

x
(s)
. Since (
x
(s),
x
(s)) =
s
(x, v),
it follows that

1
(u

,u
+
)
is invariant by the ow
t
. From theorem
4.12.3, if x /
0
, we have d
x
u = L/v(x, v), where (x, v) is
the point

/
0
above x. Since /
0
1
(u

,u
+
)
, the denition of

1
(u

,u
+
)
implies (x, v)

1
(u

,u
+
)
.
188
The Ma ne set was introduced in [Mn97, page 144], where it is
denoted (L).
Denition 5.2.5 (Ma ne Set). The Ma ne set is

^
0
=
_

1
(u

,u
+
)
,
where the union is taken over all pairs (u

, u
+
) of conjugate func-
tions.
Proposition 5.2.6. The Ma ne set

^
0
is a compact subset of TM
which is invariant by
t
. It contains

/
0
.
Denition 5.2.7 (Aubry Set). The Aubry Set in TM is

/
0
=


1
(u

,u
+
)
,
where the intersection is taken on the pairs (u

, u
+
) of conjugate
functions. The projected Aubry set in M is /
0
= (

/
0
).
Theorem 5.2.8. The Aubry sets /
0
and

/
0
are both compact,
and satisfy

/
0


/
0
and /
0
/
0
. The compact set

/
0
TM
is invariant by the Euler-Lagrange ow
t
.
Moreover, there is a pair (u

, u
+
) of conjugate functions such
that /
0
= 1
(u

,u
+
)
and

/
0
=

1
(u

,u
+
)
.
Therefore the projection : TM M induces a bi-Lipschitz
homeomorphism

/
0
on /
0
= (

/
0
).
Proof. The rst part of the theorem is a consequence of the same
properties which hold true for

1
(u

,u
+
)
and 1
(u

,u
+
)
which are true
by theorem 5.2.4. The last part of the theorem is, again by 5.2.4,
a consequence of the second part.
It remains to prove the second part. We x a base point x
0

/
0
, any pair of conjugate function is of the form (u

+c, u
+
+c),
where (u

, u
+
) is a pair of conjugate functions with u

(x
0
) =
u
+
(x
0
) = 0, and c R. Using the fact that ((M, R) is metric
and separable (i.e. contains a dense sequence) for the topology of
uniform convergence, we can nd a sequence of pairs of conjugate
functions (u
n

+ c
n
, u
n
+
c
n
), dense in the set of pairs of conjugate
functions, and such that u
n

(x
0
) = u
n
+
(x
0
) = 0, c
n
R. Since
the sets o

and o
+
form equi-Lipschitzian families of functions on
189
the compact space M, and u
n

(x
0
) = u
n
+
(x
0
) = 0, we can nd a
constant C < + such that |u
n

C and |u
n
+
|

C, for
each n 0. It follows that the series

n0
2
n1
u
n

converges to
a continuous function. The sum is dominated by L+c[0], because
this is the case for each u
n

and

n0
2
n1
= 1. By theorem
5.1.2, we can thus nd u

with u



n0
2
n
u
n

and
u

n0
2
n
u
n

on /
0
. In the same way, we can nd u
+
o
+
with u
+


n0
2
n
u
n
+
and u
+
=

n0
2
n
u
n
+
on /
0
. Since
u
n
+
u
n

with equality on /
0
, we see that
u
+

n0
2
n
u
n
+

n0
2
n
u
n

,
with equalities on /
0
. It follows that functions u

and u
+
are
conjugate. Moreover, if u

(x) = u
+
(x), we necessarily have u
n

(x) =
u
n
+
(x) for each n 0. By density of the sequence (u
n

+c
n
, u
n
+
c
n
)
we conclude that for each pair v

, v
+
) of conjugate functions, we
have 1
(u

,u
+
)
1
(v

,v
+
)
. Therefore shows that 1
(u

,u
+
)
= /
0
.
If (x, v)

1
(u

,u
+
)
, the curve (s) = (
s
(x, v)) is contained in
1
(u

,u
+
)
= /
0
, and (u

, L, c[0])-calibrated. Therefore, for example


t t

R, u

((t

))u

((t)) =
_
t

t
L((s), (s)) ds+c[0](t

t)..
(*)
Since u
n
L +c[0], for each n 1, we also have
t t

R, u
n

((t

))u
n

((t)) int
t

t
L((s), (s)) ds+c[0](t

t)..
(**)
From what we established we have
y 1
(u

,u
+
)
, u

(y) =

n1
u
n

(y)
2
n
.. (***)
The equalities (*) and (***), taken with the fact that the image
of is contained in 1
(u

,u
+
)
, do imply that the inequality (**) is
in fact an equality, which means that is (u
n

, L, c[0])-calibrated,
for every n 1. By denseness of the sequence (u
n

+ c
n
) in o

,
we obtain that is (v

, L, c[0])-calibrated, for every v

.
Therefore d
x
v

is the Legendre transform of (x, v) = ((0), (0)).


Since x /
0
1
(v

,v
+
)
, this implies that (x, v)

1
(v

,v
+
)
. It
follows easily that

1
(u

,u
+
)
=

/
0
.
190
5.3 The Peierls barrier.
This denition of the Peierls barrier is due to Mather, see [Mat93,
7, page 1372].
Denition 5.3.1 (Peierls Barrier). The Peierls barrier is the func-
tion h : M R dened by
h(x, y) = liminf
t+
h
t
(x, y) +c[0]t.
It is not completely clear that h is nite nor that it is contin-
uous. We start by showing these two points.
Lemma 5.3.2 (Properties of h
t
). The properties of h
t
are
(1) for each x, y, z M and each t, t

> 0, we have
h
t
(x, y) +h
t
(y, z) h
t+t
(x, z);
(2) if u L +c, we have h
t
(x, y) +ct u(y) u(x);
(3) for each t > 0 and each x M, we have h
t
(x, x) +c[0]t 0;
(4) for each t
0
> 0 and each u

, there exists a constant


C
t
0
,u

such that
t t
0
, x, y M, 2|u

h
t
(x, y)+c[0]t 2|u

+C
t
0
,u

;
(5) for each t > 0 and each x, y M, there exists an ex-
tremal curve : [0, t] M with (0) = x, (t) = y and
h
t
(x, y) =
_
t
0
L((s), (s)) ds. Moreover, an extremal curve
: [0, t] M is minimizing if and only if h
t
((0), (t)) =
_
t
0
L((s), (s)) ds;
(6) for each t
0
> 0, there exists a constant K
t
0
[0, +[ such
that, for each t t
0
the function h
t
: M M R is
Lipschitzian with a Lipschitz constant K
t
0
.
Proof. Properties (1) and (2) are immediate, and property (3)
results from (2) taking for u a function in o

.
To prove property (4), we rst remark that the inequality
2|u

h
t
(x, y) +c[0]t also results from (2). By compactness
191
of M, we can nd a constant C
t
0
such that for each x, z M, there
exists a C
1
curve
x,z
: [0, t
0
] M with
x,z
(0) = x,
x,z
(t
0
) = z,
and
_
t
0
0
L(
x,z
(s),
x,z
(s)) ds C
t
0
. By the properties of u

, we
can nd an extremal curve
y

:] , 0] M, with
y

(0) = y,
and
t 0, u

(y) u

(
y

(t)) =
_
0
t
L(
y

(s),
y

(s)) ds +c[0]t.
If t t
0
, we can dene a (continuous) piecewise C
1
curve : [0, t]
by (s) =
x,
y

(t
0
t)
(s), for s [0, t
0
], and (s) =
y

(s t), for
s [t
0
, t]. This curve joins x with y, and we have
_
t
0
L((s), (s)) ds +c[0]t C
t
0
+c[0]t
0
+u

(y) u

(
y

(t
0
t)).
It is then enough to set C
t
0
,u

= C
t
0
+c[0]t
0
to nish the proof of
(4).
The rst part of the property (5) results from Tonellis Theo-
rem 3.3.1. The second part is immediate starting from the deni-
tions.
To prove property (6), suppose that : [0, t] M is an
extremal curve such that (0) = x, (t) = y, and h
t
(x, y) =
_
t
0
L((s), (s)) ds. Since t t
0
, we know by the Compactness
Lemma that there exists a compact subset K of TM with ((s), (s))
K for each x, y M, each t t
0
and each s [0, t]. It is then
enough to adapt the ideas which made it possible to show that the
family T

t
u [ t t
0
, u (
0
(M, R) is equi-Lipschitzian.
Corollary 5.3.3 (Properties of h). The values of the map h are
nite. Moreover, the following properties hold
(1) the map h is Lipschitzian;
(2) if u L +c[0], we have h(x, y) u(y) u(x);
(3) for each x M, we have h(x, x) 0;
(4) h(x, y) +h(y, z) h(x, z);
(5) h(x, y) +h(y, x) 0;
192
(6) for x /
0
, we have h(x, x) = 0;
(7) for each x, y M, there exists a sequence of minimizing
extremal curves
n
: [0, t
n
] M with t
n
,
n
(0) =
x,
n
(t
n
) = y and
h(x, y) = lim
n+
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
;
(8) if
n
: [0, t
n
] M is a sequence of (continuous) piecewise
C
1
curves with t
n
,
n
(0) x, and
n
(t
n
) y, then
we have
h(x, y) liminf
n+
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
.
Proof. Properties (1) to (5) are easy consequences of the lemma
giving the properties of h
t
5.3.2. Let us show the property (6). By
the continuity of h, it is enough to show that if is a Borel proba-
bility measure on TM, invariant by
t
and such that
_
T
MLd =
c[0], then, for each (x, v) supp(), the support of , we have
h(x, x) = 0. By Poincares Recurrence Theorem, the recurrent
points for
t
contained in supp() form a dense set in supp().
By continuity of h, we can thus assume that (x, v) is a recurrent
point for
t
. Let us x u

. We have
u

(
t
(x, v)) u

(x) =
_
t
0
L(
s
(x, v)) ds +c[0]t
By the denition of a recurrent point, there exists a sequence t
n

with
tn
(x, v) (x, v), it is not dicult, for each > 0 and each
t

0, to nd a (continuous) piecewise C
1
curve : [0, t] M,
with t t

, (0) = (t) = x, and such that


_
t
0
L((s), (s)) ds +
c[0]t . Consequently, we obtain h(x, x) 0. The inequality
h(x, x) 0 is true for each x M.
Property (7) results from part (5) of the lemma giving the
properties of h
t
, since there is a sequence t
n
+ such that
h(x, y) = lim
tn+
h
tn
(x, y) +c[0]t
n
.
Let us show property (8). By the previous lemma, there is a
constant K
1
such that
t 1, x, x

, y, y

M, [h
t
(x, y)h
t
(x

, y

)[ K
1
(d(x, x

)+d(y, y

)).
193
In addition, we also have
h
tn
(
n
(0),
n
(t
n
))
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
.
For n large, it follows that
h
tn
(x, y) +c[0]t
n
h
tn
(
n
(0),
n
(t
n
)) +c[0]t
n
+K
1
(d(x,
n
(0)) +d(y,
n
(t
n
)))
c[0]t
n
+K
1
(d(x,
n
(0)) +d(y,
n
(t
n
))) +
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
.
Since d(x,
n
(0)) +d(y,
n
(t
n
)) 0, we obtain the sought inequal-
ity.
The following lemma is useful.
Lemma 5.3.4. Let V be an open neighborhood of

/
0
in TM.
There exists t(V ) > 0 with the following property:
If : [0, t] M is a minimizing extremal curve, with t t(V ),
then, we can nd s [0, t] with ((s), (s)) V .
Proof. If the lemma were not true, we could nd a sequence of
extremal minimizing curves
i
: [0, t
i
] M, with t
i
, and
such that (
i
(s),
i
(s)) / V , for each s [0, t
i
]. Since t
i
+,
by corollary 4.4.5, there exists a compact subset K TM with
(
i
(s),
i
(s)) K, for each s [0, t
i
] and each i 0. We then
consider the sequence of probability measures
n
on TM dened
by
_
TM
d
n
=
1
t
n
_
tn
0
(
n
(s),
n
(s)) ds,
for : TM R continuous. All the supports of these mea-
sures are contained in the compact subset K of TM. Extract-
ing a subsequence, we can assume that
n
converge weakly to a
probability measure . Since (
n
(s),
n
(s)), s [0, t
n
] are pieces
of orbits of the ow
t
, and since t
n
+, the measure is
invariant by
t
. Moreover, its support supp() is contained in
TMV , because this is the case for all supp(
n
) = (
n
(s),
n
(s)) [
s [0, t
n
. Since the
n
are minimizing extremals, we have
_
Ld
n
= h
tn
(
n
(0), (t
n
))/t
n
. By the lemma giving the prop-
erties of h
t
5.3.2, if u

, we can nd a constant C
1
such
that
t 1, x, y M, 2|u

|
0
h
t
(x, y) +c[0]t 2|u

|
0
+C
1
.
194
It follows that lim
n+
_
Ld
n
= c[0]. Hence
_
TM
Ld = c[0]
and the support of is included in Mathers set

/
0
. This is
a contradiction, since we have already observed that supp() is
disjoint from the open set V which contains

/
0
.
Corollary 5.3.5. For each pair u

, u
+
o
+
of conjugate
functions, we have
x, y M, u

(y) u
+
(x) h(x, y).
Proof. We pick a sequence of extremals
n
: [0, t
n
] M joining x
to y, and such that
h(x, y) = lim
n
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
.
By the previous lemma 5.3.4, extracting a subsequence if neces-
sary, we can nd a sequence t

n
[0, t
n
] such that
n
(t

n
) z
/
0
. If u

, and u
+
o
+
, we have
u
+
(
n
(t

n
)) u
+
(x)
_
t

n
0
L(
n
(s),
n
(s)) ds +c[0]t

n
,
u

(y) u

(
n
(t

n
))
_
tn
t

n
L(
n
(s),
n
(s)) ds +c[0](t
n
t

n
).
If we add these inequalities, and we let n go to +, we nd
u

(y) u

(z) +u
+
(z) u
+
(x) h(x, y).
But the functions u

and u
+
being conjugate, we have u
+
(z) =
u

(z), since z /
0
.
Theorem 5.3.6. For x M, we dene the function h
x
: M R
(resp. h
x
: M R) by h
x
(y) = h(x, y) (resp. h
x
(y) = h(y, x)). For
each x M, the function h
x
: M R (resp. h
x
) is in o

(resp.
o
+
). Moreover, its conjugate function u
x
+
o
+
(resp. u
x

)
vanishes at x.
Proof. We rst show that the function h
x
is dominated by L+c[0].
If : [0, t] M is a (continuous) piecewise C
1
curve, we have
195
h
t
((0), (t))
_
t
0
L((s), (s)) ds and thus, by part (1) of the
lemma giving the properties of h
t
5.3.2, we obtain
h
t

+t
(x, (t)) h
t
(x, (0)) +
_
t
0
L((s), (s)) ds,
which gives by adding c[0](t +t

) to the two members


h
t

+t
(x, (t))+c[0](t+t

) h
t
(x, (0))+c[0]t

+
_
t
0
L((s), (s)) ds+c[0]t.
By taking the liminf for t

+, we nd
h(x, (t)) h(x, (0)) +
_
t
0
L((s), (s)) ds +c[0]t,
which we can write as
h
x
((t)) h
x
((0))
_
t
0
L((s), (s)) ds +c[0]t.
To nish showing that h
x
o

, it is enough to show that for


y M, we can nd an extremal curve

: ] , 0] such that

(0) = y and
t 0, h(x, y) h(x,

(t)) +
_
0
t
L(

(s),

(s)) ds c[0]t.
We take a sequence of extremal curves
n
: [0, t
n
] M connecting
x to y, and such that
h(x, y) = lim
n
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
.
Since t
n
and the
n
are all minimizing extremal curves, by
extracting a subsequence if necessary, we can suppose that the
sequence of extremal curves

n
: [t
n
, 0] M, t
n
(t
n
+ t)
converges to an extremal curve

: ] , 0] M. We have

(0) = lim
n

n
(t
n
) = y. Let us x t ] , 0], for n big
enough, we have t
n
+t 0 and we can write
_
tn
0
L(
n
(s),
n
(s)) ds+c[0]t
n
=
_
tn+t
0
L(
n
(s),
n
(s)) ds+c[0](t
n
+t)
_
tn+t
0
L(
n
(s),
n
(s)) ds+c[0](t
n
+t)+
_
0
t
L(

n
(s),

n
(s)) dsc[0]t.
()
196
By convergence of the

n
, we have
_
0
t
L(

n
(s),

n
(s)) ds
_
0
t
L(

(s),

(s)) ds.
Since lim
n
t
n
+t = , and lim
n

n
(t
n
+ t) = lim
n

n
(t) =

(t), by part (8) of the corollary giving the properties of h, we


obtain h(x,

(t)) liminf
n
_
tn+t
0
L(
n
(s),
n
(s)) ds+c[0](t
n
+
t). By taking the liminf in the equality (), we do indeed nd
h(x, y) h(x,

(t)) +
_
0
t
L(

(s),

(s)) ds c[0]t.
It remains to be seen that u
x
+
o
+
, the conjugate function of h
x
,
vanishes at x. For that, we dene
a(t) = c[0]t + sup

h(x, (t))
_
t
0
L((s), (s)) ds,
where : [0, t] M varies among C
1
curves with (0) = x. This
quantity a(t)is nothing but T
+
t
(h
x
)(x) c[0]t, and thus u
x
+
(x) =
lim
t
a(t). For each t > 0, we can choose an extremal curve

t
: [0, t] M, with
t
(0) = x and
a(t) = c[0]t +h(x,
t
(t))
_
t
0
L(
t
(s),
t
(s)) ds.
We then choose a sequence t
n
+ such that
tn
(t
n
) converges
to a point of M which we will call y. By continuity of h we have
h(x, y) = lim
n+
h(x,
tn
(t
n
)). Moreover, by part (8) of the
corollary giving the properties of h, we have
h(x, y) liminf
n+
_
t
0
L(
t
(s),
t
(s)) ds +c[0]t.
It follows that u
x
+
(x) = lima(t
n
) 0. Since we already showed
the inequality h(x, y) u

(y) u
+
(x), for any pair of conjugate
functions u

, u
+
o
+
, we have h(x, x) h
x
(x) u
x
+
(x).
However h(x, x) = h
x
(x), which gives u
x
+
(x) 0.
Corollary 5.3.7. For each x, y M, we have the equality
h(x, y) = sup
(u

,u
+
)
u

(y) u
+
(x),
197
the supremum being taken on pairs (u

, u
+
) of conjugate functions
u

, u
+
o
+
.
We can also give the following characterization for the Aubry
set /
0
.
Proposition 5.3.8. If x M, the following conditions are equiv-
alent
(1) x /
0
;
(2) the Peierls barrier h(x, x) vanishes;
(3) there exists a sequence
n
: [0, t
n
] M of (continuous)
piecewise C
1
curves such that
for each n, we have
n
(0) =
n
(t
n
) = x;
the sequence t
n
tends to +, when n ;
for n , we have
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
0;
(4) there exists a sequence
n
: [0, t
n
] M minimizing ex-
tremal curves such that
for each n, we have
n
(0) =
n
(t
n
) = x;
the sequence t
n
tends to +, when n ;
for n , we have
_
tn
0
L(
n
(s),
n
(s)) ds +c[0]t
n
0.
Proof. Equivalence of conditions (1) and (2) results from the pre-
vious corollary. Equivalence of (2), (3) and (4) results from the
denition of h.
5.4 Chain Transitivity
Proposition 5.4.1. Let (u

, u
+
) be a given pair of conjugate
functions. If t
0
> 0 given, then for each > 0, there exists
> 0 such that if : [0, t] M is an extremal curve, with
t t
0
, u

((0)) u
+
((0)) + and
_
t
0
L((s), (s)) ds + c[0]t
u
+
((t)) u
+
((0)) + , then for each s [0, t], we can nd a
point in

1
(u

,u
+
)
at distance at most from ((s), (s)).
Proof. Since u
+
L +c[0], for 0 a b t, we have
u
+
((a)) u
+
((0))
_
a
0
L((s), (s)) ds +c[0]a,
u
+
((t)) u
+
((b))
_
t
b
L((s), (s)) ds +c[0](t b).
198
Subtracting the last two inequalities from the inequality
_
t
0
L((s), (s)) ds+
c[0]t u
+
((t)) u
+
((0)) +, we nd
_
b
a
L((s), (s)) ds +c[0](b a) u
+
((b)) u
+
((a)) +. (*)
Moreover, since u

L +c[0], we have
u

((a)) u

((0))
_
a
0
L((s), (s)) ds +c[0]a.
Since, by the inequality (), this last quantity is not larger than
u
+
((a))u
+
((0))+, we obtain u

((a))u

((0)) u
+
((a))
u
+
((0)) + . The condition u

((0)) u
+
((0)) + gives then
u

((a)) u
+
((a)) +2, for each a [0, t]. We conclude that it
is enough to show the lemma with t = t
0
, taking smaller if neces-
sary. Let us argue by contradiction. We suppose that there exists
a sequence of extremal curves
n
: [0, t
0
] M and a sequence
n
,
such that the following conditions are satised
(1)
n
0;
(2) u

(
n
(0)) u
+
(
n
(0)) +
n
;
(3)
_
t
0
0
L(
n
(s),
n
(s)) ds+c[0]t
0
u
+
(
n
(t
0
))u
+
(
n
(0))+
n
;
(4) there exists s
n
[0, t
0
] such that the distance from (
n
(s
n
),
n
(s
n
))
with

1
(u

,u
+
)
is bigger than .
By conditions (1) and (3) above, there exists a constant C <
+ such that
_
t
0
0
L(
n
(s),
n
(s)) ds C, for each n 0. It
follows that there exists s

n
[0, t
0
] such that L(
n
(s

n
),
n
(s

n
))
C/t
0
. Therefore the (
n
(s

n
),
n
(s

n
)) are all in the compact subset
K = (x, v) [ L(x, v) C/t
0
TM. Since the
n
are extremal
curves, we have (
n
(t),
n
(t)) =
(ts

n
)
(
n
(s

n
),
n
(s

n
)), and thus
the point (
n
(t),
n
(t)) is in the compact subset

s[0,t
0
]

s
(K),
for each n 0 and each t [0, t
0
].
Extracting a subsequence if necessary, we can thus suppose
that the sequence of extremal curves
n
converges in the C
1
topol-
ogy to the extremal curve

: [0, t
0
] M. As we saw above, we
have u

(
n
(t)) u
+
(
n
(t)) + 2
n
, for each t [0, t
0
]. Going to
the limit and taking s

a value of adherence of the sequence s


n
,
we thus obtain
(1) for each t [0, t
0
], we have u

(t)) u
+
(

(t));
(2)
_
t
0
0
L(

(s),

(s)) ds +c[0]t
0
u
+
(

(t
0
)) u
+
(

(0));
199
(3) there exists a number s

[0, t
0
] such that the distance
from the point (

(s

),

(s

)) to the set

1
(u

,u
+
)
is at least .
However u

u
+
thus u

(t)) = u
+
(

(t)), for each t


[0, t
0
], which gives

(t) 1
(u

,u
+
)
. In the same way, the fact that
u
+
L+c[0], forces the equality in condition (2) above. This gives
_
t
0
0
L(

(s),

(s)) ds+c[0]t
0
= u
+
(

(t
0
))u
+
(

(0)). In par-
ticular, the derivative of u
+
at

(s), for s ]0, t


0
[ is the Legendre
transform of (

(s),

(s)). It follows that (

(s

),

(s

))

1
(u

,u
+
)
, which contradicts condition (3) above.
Corollary 5.4.2. Let u

and u
+
o
+
be a pair of conjugate
functions. If x, y M is such that h(x, y) = u
+
(y) u

(x), then
x, y 1
(u

,u
+
)
. Moreover, if (x, v
x
) and (y, v
y
) are the points of

1
(u

,u
+
)
above x and y, then for each > 0, we can nd a sequence
of points (x
i
, v
i
)

1
(u

,u
+
)
, i = 0, 1, . . . , k, with k 1, such that
(x
0
, v
0
) = (x, v
x
), (x
k
, v
k
) = (y, v
y
), that there exists t [1, 2] with
the distance from
t
(x
k1
, v
k1
) to (x
k
, v
k
) = (y, v
y
) is less than
and that, for i = 0, . . . , k 2, the distance in TM from
1
(x
i
, v
i
)
to (x
i+1
, v
i+1
) is also less than .
Proof. We know that h(x, y) u

(y) u
+
(x). Since u

u
+
, we
see that u

(x) = u
+
(x) and u

(y) = u
+
(y). By the properties of
h, there exist a sequence of extremals
n
: [0, t
n
] M, with t
n

, such that
n
(0) = x,
n
(t
n
) = y and
_
tn
0
L(
n
(s),
n
(s)) ds +
c[0]t
0
u
+
(
n
(t
n
)) u
+
(
n
(0)) +
n
with
n
0.
Let us x > 0. The compactness of

1
(u

,u
+
)
gives the exis-
tence of

such that if (a, v)



1
(u

,u
+
)
and (b, w) TM are at
distance less than

, then, for each s [0, 2], the distance from

s
(a, v) with
s
(b, w) is smaller than /2. By the previous propo-
sition, for n large enough, (
n
(s),
n
(s)), s [0, t
n
] is at distance
min(

, /2) from a point of



1
(u

,u
+
)
. Let us x such an inte-
ger n with t
n
1 and call k the greatest integer t. We thus
have k 1 and t = t
n
(k 1) [1, 2[. For i = 1, 2, . . . , k 2,
let us choose (x
i
, v
i
)

1
(u

,u
+
)
at distance min(

, /2) from
(
n
(i),
n
(i)) and let us set (x
0
, v
0
) = (x, v
x
), (x
k
, v
k
) = (y, v
y
).
For i = 0, . . . , k 2, by the choice of

, the distance between

1
(x
i
, v
i
) and
1
(
n
(i),
n
(i)) = (
n
(i + 1),
n
(i + 1)) is /2,
consequently the distance between
t
(x
i
, v
i
) and (x
i+1
, v
i+1
) is
less than . In the same way, as t [1, 2[ the distance between
200

t
(x
k1
, v
k1
) and
t
(
n
(k 1),
n
(k 1)) = (
n
(k),
n
(k))) =
(y, v
y
) is less than /2.
The following theorem is due to Ma ne, see [Mn97, Theorem V,
page 144]:
Theorem 5.4.3 (Ma ne). The Ma ne set

^
0
is chain transitive for
the ow
t
. In particular, it is connected.
Proof. We recall that

^
0
=


1
(u

,u
+
)
, where the union is taken on
all pairs of conjugate functions. Let us notice that

/
0


1
(u

,u
+
)
,
by denition of conjugate functions. If x, y /
0
, then by the
compactness of the set of dierences u

u
+
and the character-
ization of h, there exists a pair (u

, u
+
) of conjugate functions
such that h(x, y) = u

(y) u
+
(x), since x, y /
0
, we have also
h(x, y) = u
+
(y) u

(x). By the corollary above, we see that


the points (x, v
x
)

/
0
and (y, v
y
)

/
0
above x and y can,
for each > 0, be connected by an -chain of points, for
t
, in

1
(u

,u
+
)
. It follows that the set

/
0
is contained in a single chain
recurrent component of the union

^
0
=


1
(u

,u
+
)
. To nish
showing that

^
0
is chain transitive, it is enough to notice that if
(x, v) 1
(u

,u
+
)
, then the -limit and -limit sets, for
t
, of (x, v)
both contain points of

/
0
.
Chapter 6
A Closer Look at the
Lax-Oleinik semi-group
6.1 Semi-convex Functions
6.1.1 The Case of Open subsets of R
n
Proposition 6.1.1. Let U be an open convex subset of R
n
, and let
u : U R be a function. The following conditions are equivalent:
(i) There exists a C
2
function : U R with bounded second
derivative and such that u + is convex.
(ii) There exists a C

function : U R with bounded second


derivative and such that u + is convex.
(iii) There exists a nite constant K and for each x U there
exists a linear form
x
: R
n
R such that
y U, u(y) u(x)
x
(y x) K|y x|
2
.
Proof. Obviously (ii) implies (i). To prove that (i) implies (iii),
we denote by 2K an upper bound on U of the norm of the second
derivative of , using Taylors formula, we see that
x, y U, (y) (x) d
x
(y x) +K|y x|
2
.
We now use theorem 1.2.9 to obtain a supporting linear form
1
at x for the convex function u +. This gives
u(y) u(x)
1
(y x) (y) +(x).
201
202
Combining the two inequalities, we get
u(y) u(x)
1
(y x) d
x
(y x) K|y x|
2
,
but the map v
1
(v) d
x
(v) is linear.
It remains to prove that (iii) implies (ii). We consider (y) =
K|y|
2
, where K is the constant given by (iii). A simple compu-
tation gives
(y) (x) = K|y x|
2
+ 2Ky x, x).
Adding the inequality given by (iii) and the equality above, we
obtain
(u +)(y) (u +)(y)
x
(y x) + 2Ky x, x).
This shows that u+ admits the linear map v
x
(v) +2Kv, x)
as a supporting linear form at x. It follows from proposition 1.2.8
that u + is convex.
Denition 6.1.2 (Semi-convex). A function u : U R, dened
on the open convex subset U of R
n
, is said to be semi-convex if it
satises one of (and hence all) the three equivalent conditions of
proposition 6.1.1.
We will say that u is K-semi-convex if it satises condition (iii)
of 6.1.1 with K as a constant.
A function u is said to be semi-concave (resp. K-semi-concave)
is u semi-convex (resp. K-semi-convex).
We will say that a function u : V R, dened on an open
subset of R
n
is locally semi-convex (resp. semi-concave) if for each
point x V there exists an open convex neighborhood U
x
of x in V
such that the restriction u[U
x
is semi-convex (resp. semi-concave).
Here are some properties of locally semi-convex or semi-concave
functions:
Proposition 6.1.3. (1) A locally semi-convex (resp. semi-concave)
function is locally Lipschitz.
(2) If u is locally semi-convex (resp. semi-concave) then u is
dierentiable almost everywhere.
(3) If u : V R is locally semi-convex (resp. semi-concave)
and f : W V is a C
2
map then u f : W R is also is locally
semi-convex (resp. semi-concave).
203
Proof. It suces to prove these properties for a semi-convex func-
tion u : U R dened on the open convex subset U of R
n
. But
u is the sum of a C
2
function and a convex function. A C
2
func-
tion is obviously locally Lipschitz and dierentiable everywhere.
Moreover, a convex function on an open subset of R
n
is locally
Lipschitz by corollary 1.1.9 and dierentiable almost everywhere
by corollary 1.1.11 (or by Rademachers theorem 1.1.10). This
proves (1) and (2).
It suces to prove (3) for u : V R convex. We x some z
W, and we pick r > 0 such that the closed ball

B(f(z), 2r) V .
We rst show that there is a constant K
1
such that each supporting
linear form p of u at y

B(f(z), r) satises |p| K
1
. In fact,
for v such that |v| r, the point y +v is in V since it belongs to

B(f(z), 2r), hence we can write


u(y +v) u(y) p(v).
We set M = max[u(x)[ [ x

B(f(z), 2r). The constant M is
nite because

B(f(z), 2r) is compact and u is continuous by (1).
from the inequality above we obtain
v

B(0, r), p(v) 2M.
It is not dicult to conclude that |p| K
1
. We now pick an open
convex set O which contains x, such that its closure

O is compact
and contained in W and f(O) B(f(z), r). Since f is C
2
and O
is convex with compact closure, by Taylors formula, we can nd
a constant K
2
such that
z
1
, z
2
O, [f(z
2
) f(z
1
) Df(z
1
)[z
2
z
1
][ K
2
|z
2
z
1
|
2
. (*)
If z
1
, z
2
are both in O, then f(z
2
), f(z
1
) are both in B(f(z), r). If
we call p
1
a supporting linear form of u at f(z
1
), we have
|p
2
| K
1
and u f(z
2
) u f(z
1
) p
1
(f(z
2
) f(z
1
)).
Combining with (), we obtain
z
1
, z
2
O, uf(z
2
)uf(z
1
) p
1
Df(z
1
)[z
2
z
1
]K
1
K
2
|z
2
z
1
|
2
.
Thus u f is semi-convex on O.
204
Part (3) of the last proposition 6.1.3 implies that the notion
of locally semi-convex (or semi-concave) is well-dened on a dif-
ferentiable manifold (of class at least C
2
).
Denition 6.1.4 (Semi-convex). A function u : M R dened
on the C
2
dierentiable manifold M is locally semi-convex (resp.
semi-concave), if for each x M there is a C
2
coordinate chart
: U R
n
, with x U, such that u
1
: (U) R is
semi-convex (resp. semi-concave).
In that case for each C
2
coordinate chart : V R
n
, the map
u
1
: (U) R is semi-convex (resp. semi-concave).
Theorem 6.1.5. A function u : M R, dened on the C
2
dif-
ferentiable manifold M, is both locally semi-convex and locally
semi-concave if and only if it is C
1,1
.
Proof. Suppose that u is C
1,1
. Since the result is by nature local,
we can suppose that M is in fact the open ball subset

B(0, r) of R
n
and that the derivative du :

B(0, r) R
n
, x d
x
u is Lipschitz
with Lipschitz constant K. If x, y

B(0, r), we can write
u(y) u(x) =
_
1
0
d
ty+(1t)x
u(y x) dt.
Moreover, we have
|d
ty+(1t)x
u d
x
u| Kt|y x|.
Combining these two inequalities, we obtain
[u(y) u(x) d
x
u(y x)[ =

_
1
0
d
ty+(1t)x
u(y x) d
x
u(y x) dt

_
1
0
|d
ty+(1t)x
u d
x
u||y x| dt

_
1
0
K|ty + (1 t)x x||y x| dt
=
_
1
0
Kt|y x|
2
dt
=
K
2
|y x|
2
.
205
This implies that u is both semi-convex and semi-concave on the
convex set

B(0, r).
Suppose now that u is both locally semi-convex and locally
semi-concave. Again due to the local nature of the result, we can
assume M =

B(0, r) and that u is both K
1
-semi-convex and K
2
-
semi-concave. Given x

B(0, r), we can nd
1
x
,
2
x
R
n
such
that
y

B(0, r), u(y) u(x)
1
x
(y x) K
1
|y x|
2
,
and
y

B(0, r), u(y) u(x)
2
x
(y x) +K
2
|y x|
2
.
For a xed v R
n
, and for all small enough, the point x +v

B(0, r), thus combining the two inequalities above we obtain

1
x
(v) K
1
|v|
2

2
x
(v) +K
2
|v|
2
,
for all small enough. Dividing by and letting go to 0, we
obtain
v R
n
,
1
x
(v)
2
x
(v).
Changing v into v, and using the linearity of
1
x
,
2
x
R
n
, we
see that
1
x
=
2
x
. Thus we have
y

B(0, r), [u(y) u(x)
1
x
(y x)[ max(K
1
, K
2
)|y x|
2
.
It follows from 4.11.3 that u is C
1
on

B(0, r) with a Lipschitz
derivative.
6.2 The Lax-Oleinik Semi-group and Semi-
convex Functions
In this section we suppose that the compact manifold M is en-
dowed with a C
2
Lagrangian L : TM R which is superlin-
ear and C
2
strictly convex in the bers of the tangent bundle
: TM M. We can then dene the two Lax-Oleinik semi-
groups T

t
, T
+
t
: (
0
(M, R) (
0
(M, R). An immediate conse-
quence of proposition 4.11.1 is the following proposition:
206
Proposition 6.2.1. For each u (
0
(M, R) and each t > 0, the
function T

t
(u) (resp. T
+
t
(u)) is locally semi-concave (resp. semi-
convex)
Theorem 6.2.2. Suppose that T

t
(u) (resp. T
+
t
(u)) is C
1
, where
u (
0
(M, R) and t > 0, then we have u = T
+
t
T

t
(u) (resp.
u = T

t
T
+
t
(u).
In particular, the function u must be locally semi-convex. More-
over, for each t

]0, t[ the function T

(u) (resp. T
+
t
(u)) is C
1,1
.
Before embarking in the proof of this theorem, we will need the
following lemma, whose proof results easily from the denitions of
T

t
and T
+
t
:
Lemma 6.2.3. If u, v (
0
(M, R) and t 0, the following three
conditions are equivalent:
(1) v T

t
u;
(2) T
+
t
v u;
(3) for each C
1
curve : [0, t] M, we have
v((t)) u((0)) L() =
_
t
0
L((s), (s)) ds.
If anyone of these conditions is satised and : [0, t] M is
a C
1
curve with
v((t)) u((0)) = L() =
_
t
0
L((s), (s)) ds,
then v((t)) = T

t
u((t)) and u((0)) = T
+
t
v((0)).
Proof of theorem 6.2.2. We set v = T

t
u. For each x M, we
can nd a minimizing C
2
extremal curve
x
: [0, t] M with

x
(t) = x and
v(
x
(t)) =
_
t
0
L((s), (s)) ds +u(
x
(0)).
From lemma 6.2.3, we have u(
x
(0)) = T
+
t
v((0)). To show that
u = T
+
t
v, it then remains to see that the set
x
(0) [ x M
is the whole of M. Since v is C
1
, it follows from proposition
4.11.1 that d
x
v = L/v(x,
x
(t)) or
x
(t) = grad
L
v(x). Since
207

x
is an extremal we conclude that (
x
(s),
x
(s)) =
st
(x,
x
(t)),
where
t
is the Euler-Lagrange ow of L. In particular, the set

x
(0) [ x M is nothing but the image of the map f : M
M, x
t
(x, grad
L
v(x)). Since v = T

t
u is C
1
, the map f
is continuous and homotopic to the identity, a homotopy being
given by (x, s)
s
(x, grad
L
v(x)), with s [0, t]. Since M
is a compact manifold without boundary, it follows from degree
theory mod 2 that f is surjective. This nishes the proof of
u = T
+
t
T

t
(u).
Using t > 0, we obtain the local semi-convexity of u from
proposition 6.2.1.
If t

]0, t[, we also have v = T

t
u = T

tt

u. Applying the
rst part with T

u instead of u and t t

> 0 instead of t, we
see that T

u is locally semi-convex. It is also locally semi-concave


by 6.2.1, since t

> 0. It follows from theorem 6.1.5 that T

u is
C
1,1
.
6.3 Convergence of the Lax-Oleinik Semi-
group
Up to now in this book, all the statements given above do hold for
periodic time-dependent Lagrangians which satisfy the hypothesis
imposed by Mather in [Mat91, Pages 170172]. The results in
this section depend heavily on the invariance of the energy by the
Euler-Lagrange ow. Some of these results do not hold for the
time-dependent case, see [FM00].
The main goal of this section is to prove the following theorem:
Theorem 6.3.1 (Convergence of the Lax-Oleinik Semi-group).
Let L : TM R be a C
2
Lagrangian, dened on the the com-
pact manifold M, which is superlinear and C
2
strictly convex in
the bers of the tangent bundle : TM M. If T

t
, T
+
t
:
(
0
(M, R) (
0
(M, R) are the two Lax-Oleinik semi-groups asso-
ciated with L, then for each u (
0
(M, R), the limits, for t +,
of T

t
u + c[0]t and T
+
t
u c[0]t exist. The limit of T

t
u + c[0]t is
in o

, and the limit of T


+
t
u c[0]t is in o
+
.
Particular cases of the above theorem are due to Namah and
Roquejore, see [NR97b, NR97a, NR99] and [Roq98a]. The rst
208
proof of the general case was given in [Fat98b]. There now exists
other proofs of the general case due to Barles-Souganidis and to
Roquejore, see [BS00] and [Roq98b].
The main ingredient in the proof is the following lemma:
Lemma 6.3.2. Under the hypothesis of theorem 6.3.1 above, for
each > 0, there exists a t() > 0 such that for each u (
0
(M, R)
and each t t(), if T

t
u has a derivative at x M, then c[0]
H(x, d
x
T

t
u) c[0] + . Consequently lim
t
H
M
(T

t
u) c[0],
for each u (
0
(M, R).
Proof. By Carneiros theorem 4.12.5, the set
W

= (x, v) [ c[0] H

L(x, v) c[0] +
is a neighborhood of the Mather set

/
0
. We can now apply
lemma 5.3.4 with W

as neighborhood of

/
0
, to nd t() > 0 such
that for minimizing extremal curve : [0, t] M with t t(),
there exists a t

[0, t] with ((t

), (t

)) W

, which means that


H

L((t

), (t

)) is in [c[0] , c[0] +]. This implies that for such


a minimizing curve, we have
s [0, t], c[0] H

L((s), (s)) c[0] +.
In fact, since is a minimizing curve, its speed curve s ((s), (s))
is a piece of an orbit of the Euler-Lagrange ow. Since the en-
ergy H

L is invariant by the Euler-Lagrange ow, it follows that
H

L((s), (s)) does not depend on s [0, t], but for s = t

, we
know that this quantity is in [c[0] , c[0] +].
If we suppose that T

t
u is dierentiable at x M, and we pick
a curve : [0, t] M, with (t) = x, such that
T

t
u(x) = u((0)) +
_
t
0
L((s), (s)) ds,
then we know that is minimizing, and from proposition 4.11.1 we
also know that d
x
T

t
u = L(x, (t)). From what we saw above, we
indeed conclude that when t t() we must have H(x, d
x
T

t
u)
[c[0], c[0]+]. It follows that the Hamiltonian constant H
M
(T

t
u)
does converge to c[0] when t +.
209
Proof of Theorem 6.3.1. It is convenient to introduce

T

t
: (
0
(M, R)
(
0
(M, R) dened by

T

t
u = T

t
u +c[0]t. It is clear that

T

t
is it-
self a semi-group of non-expansive maps whose xed points are
precisely the weak KAM solutions in o

. We x some u
0

.
If u (
0
(M, R), since

T

t
is non-expansive, and u
0

is a xed
point of

T

t
, we obtain
|

T
t
u u
0

|
0
= |

T
t
u

T
t
u
0

|
0
|u u
0

|.
It follows that
|

T
t
u| |u u
0

| +|u
0

|.
The family of functions

T
t
u = T

t
u + ct, with t 1, is equi-
Lipschitzian by lemma ??, hence there exists a sequence t
n
+
such that

T
tn
u u

uniformly. Using the Lemma above 6.3.2 and


Theorem 4.2.5, we see that u

L+c[0], and hence u



T
t
u

,
for each t 0. Since

T
t
is order preserving, we conclude that
t

t 0, u



T
t
u



T
t
u

.
To show that u

is a xed point for



T
t
, it then remains to nd a
sequence s
n
+ such that

T
sn
u

.
Extracting if necessary, we can assume that t
n+1
t
n
+.
We will show that the choice s
n
= t
n+1
t
n
does work. We have

T
sn


T
tn
u =

T
t
n+1
u therefore
|

T
sn
u

|
0
|

T
sn
u



T
sn


T
tn
u|
0
+|

T
t
n+1
u u

|
0
|u



T
tn
u|
0
+|

Tu
t
n+1
u

|
0
,
where we used the inequality |

T
sn
u



T
sn


T
tn
u|
0
|u

T
tn
u|
0
, which is valid since the maps

T
t
are non-expansive. Since
|u



T
tn
u|
0
0, this indeed shows that

T
t
u

= u

.
We still have to see that

T
t
u u

, when t +. If t t
n
,
we can write
|

T
t
u u

|
0
= |

T
ttn


T
tn
u

T
ttn
u

|
0
|

T
tn
u u

|
0
.
This nishes the proof, since

T
tn
u u

.
A corollary is that the liminf in the denition of the Peierls
barrier is indeed a limit.
210
Corollary 6.3.3. For each x, y M, we have h(x, y) = lim
t+
h
t
(x, y)+
c[0]t. Moreover, the convergence is uniform on M M.
Proof. If we x y in M, we can dene continuous functions h
y
t
M
R by h
y
t
(x) = h
t
(x, y). It is not dicult to check that T

h
y
t
=
h
y
t

+t
. It follows from theorem 6.3.1 that the limit of h
t+1
(x, y) +
c[0]t = T

t
h
y
1
(x) + c[0]t exists, it must of course coincide with
liminf
t+
h
t+1
(x, y) +c[0]t. By the denition of the Peierls bar-
rier, this last quantity is h(x, y) c[0].
The fact that the limit is uniform on M follows from the fact
that the h
t
, t 1 are equi-Lipschitzian by part (6) of lemma 5.3.2.
6.4 Invariant Lagrangian Graphs
Theorem 6.4.1. Suppose that N T

M is a compact Lagran-
gian submanifold of T

M which is everywhere transverse to the


bers of the canonical projection

: T

M M. If the image

t
(N) is still transversal to the bers of the canonical projection

: T

M M, then the same is true for

s
(N), for any s [0, t].
Moreover, if there exists t
n
such that

tn
(N) is still transver-
sal to the bers of the canonical projection

: T

M M, for
each n, then N is in fact invariant by the whole ow

t
, t R.
Proof. We rst treat the case where N is the graph Graph(du) of
the derivative of a C
1
function u : M R. Using proposition
4.11.1, we see that the derivative of the Lipschitz function T

t
u
wherever it exists is contained in

t
(N). Since this last set is
a (continuous) graph over the base, the derivative of T

t
u can
be extended by continuity, hence T

t
u is also C
1
and

t
(N) =
Graph(dT

t
u). By theorem 6.2.2, it follows that T

s
u is C
1
and

t
s(N) = Graph(dT

s
u), for each s [0, t]. If moreover, there
exists a sequence t
n
with

tn
(Graph(du)) transversal to the
bers of

: T

M M, then T

tn
u is everywhere smooth. By
lemma 6.3.2, if > 0 is given, then the graph Graph(dT

tn
u) is
contained in H
1
([c[0] , c[0] +]), for n large enough. Since H
is invariant by the ow

t
and Graph(dT

tn
u) =

tn
(Graph(du)),
it follows that Graph(du) is contained in H
1
([c[0] , c[0] + ]),
211
for all > 0. Hence H is constant on Graph(du), with value c[0].
By lemma 2.5.10, the graph of du is invariant under

t
.
To treat the general case, we observe that

[N : N M is
a covering since N is compact and transversal to the bers of the
submersion . We will now assume that

[N is a dieomorphism.
Since N is Lagrangian, it is the graph of a closed form on M.
We choose some C

form which is cohomologous to . If we


introduce the Lagrangian L

: TM R, (x, v) L(x, v)
x
(v),
its Hamiltonian H

is (x, p) H(x, p +
x
). This means that
H

= H , where : T

M T

M, (x, p) (x, p+
x
). Since
is closed, the dieomorphism preserves the canonical symplec-
tic form on T

M, hence conjugates

t
, the Hamiltonian ow
associated to H

and with

t
, the Hamiltonian ow associated
to H. Moreover sends bers of

onto bers of

. It follows
that

tn
(
1
(N)) is transversal to the bers of
n
. Since N is the
graph of , its inverse
1
(N), is the graph of , which is exact
by the choice of . Hence we can apply what we already proved
to conclude that
1
(N) is invariant under

t
, from which we
obtain that N is invariant under

t
.
It remain to consider the case where the covering map

[N :
N M is not necessarily injective. To simplify notations, we
set p =

[N. If we consider the tangent map Tp : TN


TM, (x, v) (p(x), T
x
p(v), we obtain a covering, which reduces
to p on the 0-section identied to N. Since T
x
p is an isomorphism,
for each x N, we can also dene (Tp)

: T

N T

M, (x, p)
(p(x), p T
x
p
1
. We dene the Lagrangian

L = L Tp : TN R.
It is easy to see that

L is as dierentiable as L, is superlinear
in each ber of TN, and for each (x, v) TN
2

L/v
2
(x, v) is
positive denite. Moreover, the conjugate Hamiltonian of

L is

H = H (Tp)

. It is not dicult to check that the pullback of


the Liouville form
M
on T

M by (Tp)

is the Liouville form


N
on T

N. If we call

t
, the Hamiltonian ow associated with

H, it
follows that (Tp)

t
=

t
. We conclude that

N = [(Tp)

]
1
(N)
is a Lagrangian submanifold of T

N, which is transversal to the


bers of the projection

N
: T

N N, and

tn
(

N) is transversal
to the bers of

N
. Using the identity map of N, we see that we
can nd a section
0
of the covering map

N
:

N N. If we call
N
0
the image of
0
, we see that we can apply the previous case
212
to conclude that N
0
is invariant by

t
. Hence N = (Tp)

(N
0
) is
invariant by
t
.
Corollary 6.4.2. Suppose that N T

M is a compact Lagran-
gian submanifold of T

M which is everywhere transverse to the


bers of the canonical projection

: T

M M. If

t
0
(N) = N
for some t
0
,= 0 then N is in fact invariant by the whole ow

t
, t R.
Chapter 7
Viscosity Solutions
In this chapter, we will study the notion of viscosity solutions
which was introduced by Crandall and Lions, see [CL83]. There
are two excellent books on the subject by Guy Barles [Bar94]
and another one by Martino Bardi and Italo Capuzzo-Dolceta
[BCD97]. A rst introduction to viscosity solutions can be found
in Craig Evans book [Eva98]. Our treatment has been extremely
inuenced by the content of these three books. Besides introduc-
ing viscosity solutions, the main goal of this chapter is to show
that, at least, for the Hamiltonians introduced in the previous
chapters, the viscosity solutions and the weak KAM solutions are
the same.
7.1 The dierent forms of Hamilton-Jacobi
Equation
We will suppose that M is a xed manifold, and that H : T

M
R is a continuous function , which we will call the Hamiltonian.
Denition 7.1.1 (Stationary HJE). The Hamilton-Jacobi asso-
ciated to H is the equation
H(x, d
x
u) = c,
where c is some constant.
A classical solution of the Hamilton-Jacobi equation H(x, d
x
u) =
c (HJE in short) on the open subset U of M is a C
1
map u : U R
such that H(x, d
x
u) = c, for each x U.
213
214
We will deal usually only with the case H(x, d
x
u) = 0, since we
can reduce the general case to that case if we replace the Hamil-
tonian H by H
c
dened by H
c
(x, p) = H(x, p) c.
Denition 7.1.2 (Evolutionary HJE). The evolutionary Hamilton-
Jacobi associated equation to the Hamiltonian H is the equation
u
t
(t, x) +H
_
x,
u
x
(t, x)
_
= 0.
A classical solution to this evolutionary Hamilton-Jacobi equa-
tion on the open subset W of R T

M is a C
1
map u : W R
such that
u
t
(t, x) +H
_
x,
u
x
(t, x)
_
= 0, for each (t, x) W.
The evolutionary form can be reduced to the stationary form
by introducing the Hamiltonian

H : T

(R M) dened by

H(t, x, s, p) = s +H(x, p),


where (t, x) R M, and (s, p) T

t,x
(R M) = RT

x
M.
It is also possible to consider a time dependent Hamiltonian
dened on an open subset of M. Consider for example a Hamil-
tonian H : R TM

R, the evolutionary form of the HJE for


that Hamiltonian is
u
t
(t, x) +H
_
t, x,
u
x
(t, x)
_
= 0.
A classical solution of that equation on the open subset W of
R M is, of course, a C
1
map u : W R such that
u
t
(t, x) +
H
_
t, x,
u
x
(t, x)
_
= 0, for each (t, x) W. This form of the
Hamilton-Jacobi equation can also be reduced to the stationary
form by introducing the Hamiltonian

H : T

(RM) R dened
by

H(t, x, s, p) = s +H(t, x, p).


7.2 Viscosity Solutions
We will suppose in this section that M is a manifold and H :
T

M M is a Hamiltonian.
As we said in the introduction of this book, it is usually impos-
sible to nd global C
1
solutions of the Hamilton-Jacobi equation
H(x, d
x
u) = c. One has to admit more general functions. A rst
attempt is to consider Lipschitz functions.
215
Denition 7.2.1 (Very Weak Solution). We will say that u :
M R is a very weak solution of H(x, d
x
u) = c, if it is Lipschitz,
and H(x, d
x
u) = c almost everywhere (this makes sense since the
derivative of u exists almost everywhere by Rademachers theo-
rem).
This is too general because it gives too many solutions. A
notion of weak solution is useful if it gives a unique, or at least a
small number of solutions. This is not satised by this notion of
very weak solution as can be seen in the following example.
Example 7.2.2. We suppose M = R, so T

M = R R, and we
take H(x, p) = p
2
1. Then any continuous piecewise C
1
function
u with derivative taking only the values 1 is a very weak solution
of H(x, d
x
u) = 0. This is already too huge, but there are even
more very weak solutions. In fact, if A is any measurable subset
of R, then the function
f
A
(x) =
_
x
0
2
A
(t) 1 dt,
where
A
is the characteristic function of A, is Lipschitz with
derivative 1 almost everywhere.
Therefore we have to dene a more stringent notion of solu-
tions. Crandall and Lions have introduced the notion of viscosity
solutions, see [CL83] and [CEL84].
Denition 7.2.3 (Viscosity solution). A function u : V R is a
viscosity subsolution of H(x, d
x
u) = c on the open subset V M,
if for every C
1
function : V R and every point x
0
V such
that u has a maximum at x
0
, we have H(x
0
, d
x
0
) c.
A function u : V R is a viscosity supersolution of H(x, d
x
u) =
c on the open subset V M, if for every C
1
function : V R
and every point y
0
V such that u has a minimum at y
0
, we
have H(y
0
, d
y
0
) c.
A function u : V R is a viscosity solution of H(x, d
x
u) = c
on the open subset V M, if it is both a subsolution and a
supersolution.
This denition is reminiscent of the denition of distributions:
since we cannot restrict to dierentiable functions, we use test
216
functions (namely or ) which are smooth and on which we can
test the condition. We rst see that this is indeed a generalization
of classical solutions.
Theorem 7.2.4. A C
1
function u : V R is a viscosity solution
of H(x, d
x
u) = c on V if and only if it is a classical solution.
In fact, the C
1
function u is a viscosity subsolution (resp. su-
persolution) of H(x, d
x
u) = c on V if and only H(x, d
x
u) c
(resp. H(x, d
x
u) c), for each x V .
Proof. We will prove the statement about the subsolution case.
Suppose that the C
1
function u is a viscosity subsolution. Since u
is C
1
, we can use it as a test function. But u u = 0, therefore
every x V is a maximum, hence H(x, d
x
u) c for each x V .
Conversely, suppose H(x, d
x
u) c for each x V . If :
V R is C
1
and u has a maximum at x
0
, then the dieren-
tiable function u must have derivative 0 at the maximum x
0
.
Therefore d
x
0
= d
x
0
u, and H(x, d
x
0
) = H(x, d
x
0
u) c.
To get a feeling for these viscosity notions, it is better to re-
state slightly the denitions. We rst remark that the condition
imposed on the test functions ( or ) in the denition above is
on the derivative, therefore, to check the condition, we can change
our test function by a constant. Suppose now that (resp. )
is C
1
and u (resp. u ) has a maximum (resp. minimum)
at x
0
(resp. y
0
), this means that u(x
0
) (x
0
) u(x) (x)
(resp. u(y
0
) (y
0
) u(x) (x)). As we said, since we can add
to (resp. ) the constant u(x
0
) (x
0
) (resp. u(y
0
) (y
0
)),
these conditions can be replaced by u (resp. u) and
u(x
0
) = (x
0
) (resp. u(y
0
) = (y
0
)). Therefore we obtain an
equivalent denition for subsolution and supersolution.
Denition 7.2.5 (Viscosity Solution). A function u : V R is
a subsolution of H(x, d
x
u) = c on the open subset V M, if for
every C
1
function : V R, with u everywhere, at every
point x
0
V where u(x
0
) = (x
0
) we have H(x
0
, d
x
0
) c, see
gure 7.1.
A function u : V R is a supersolution of H(x, d
x
u) = c on
the open subset V M, if for every C
1
function : V R, with
u everywhere, at every point y
0
V where u(y
0
) = (y
0
) we
have H(y
0
, d
y
0
) c, see gure 7.2.
217
Graph(u)
(x
0
, u(x
0
))
Graph()
Figure 7.1: Subsolution: u, u(x
0
) = (x
0
) H(x
0
, d
x
0
) c
To see what the viscosity conditions mean we test them on the
example 7.2.2 given above.
Example 7.2.6. We suppose M = R, so T

M = R R, and
we take H(x, p) = p
2
1. Any Lipschitz function u : R R
with Lipschitz constant 1 is in fact a viscosity subsolution of
H(x, d
x
u) = 0. To check this consider a C
1
function and x
0
R
such that (x
0
) = u(x
0
) and (x) u(x), for x R. We can write
(x) (x
0
) u(x) u(x
0
) [x x
0
[.
For x > x
0
, this gives
(x) (x
0
)
x x
0
1,
hence passing to the limit

(x
0
) 1. On the other hand, if
(x x
0
) < 0 we obtain
(x) (x
0
)
x x
0
1,
hence

(x
0
) 1.This yields [

(x
0
)[ 1, and therefore
H(x
0
,

(x
0
)) = [

(x
0
)[
2
1 0.
218
(x
0
, u(x
0
))
Graph(u)
Graph()
Figure 7.2: Supersolution: u, u(x
0
)=(x
0
)H(x
0
, d
x
0
)c
So in fact, any very weak subsolution (i.e. a Lipschitz function u
such that H(x, d
x
, u) 0 almost everywhere) is a viscosity subso-
lution. This is due to the fact that, in this example, the Hamilto-
nian is convex in p, see 8.3.4 below.
Of course, the two smooth functions x x, and x x are
the only two classical solutions in that example. It is easy to check
that the absolute value function x [x[, which is a subsolution
and even a solution on R 0 (where it is smooth and a classical
solution), is not a viscosity solution on the whole of R. In fact
the constant function equal to 0 is less than the absolute value
everywhere with equality at 0, but we have H(0, 0) = 1 < 0, and
this violates the supersolution condition.
The function x [x[ is a viscosity solution. It is smooth
and a classical solution on R0. It is a subsolution everywhere.
Moreover, any function with (0) = 0 and (x) [x[ every-
where cannot be dierentiable at 0. This is obvious on a picture
of the graphs, see gure 7.3. Formally it results from the fact that
both (x) x and (x) +x have a maximum at 0.
We now establish part of the relationship between viscosity
solutions and weak KAM solutions.
Proposition 7.2.7. Let L : TM R be a Tonelli Lagrangian on
the compact manifold M. If the function u : V R, dened on
the open subset V M, is dominated by L+c, then u is a viscosity
219
Figure 7.3: Graphs of (x) [x[ with (0) = 0.
220
subsolution of H(x, d
x
u) = c on V , where H is the Hamiltonian
associated to L, i.e. H(x, p) = sup
vTxM
p, v) L(x, v).
Moreover, any weak KAM solution u

is a viscosity
solution of H(x, d
x
u) = c[0].
Proof. To prove the rst part, let : V R be C
1
, and such
that u with equality at x
0
. This implies (x
0
) (x)
u(x
0
) u(x). Fix v T
x
0
M and choose :] , [ M, a C
1
path
with (0) = x
0
, and (0) = v. For t ] , 0[, we obtain
((0)) ((t)) u((0)) u((t))

_
0
t
L((s), (s)) ds ct.
Dividing by t > 0 yields
((t)) ((0))
t

1
t
_
0
t
L((s), (s)) ds +c.
If we let t 0, we obtain d
x
0
(v) L(x
0
, v) +c, hence
H(x
0
, d
x
0
) = sup
vTx
0
M
d
x
0
(v) L(x
0
, v) c.
This shows that u is a viscosity subsolution.
To prove that u

is a viscosity solution, it remains to


show that it is a supersolution of H(x, d
x
u

) = c[0]. Suppose that


: M R is C
1
, and that u

everywhere with u

(x
0
) =
(x
0
). We have (x
0
) (x) u

(x
0
) u

(x), for each x M.


We pick a C
1
path :] , 0] M, with (0) = x
0
, and such
that
t 0, u

((0)) u

((t)) =
_
0
t
L((s), (s)) ds c[0]t.
Therefore
((0)) ((t))
_
0
t
L((s), (s)) ds c[0]t.
If, for t < 0, we divide both sides by t > 0, we obtain
((t)) ((0))
t

1
t
_
0
t
L((s), (s)) ds +c[0].
If we let t tend to 0, this yields d
x
0
( (0)) L(x
0
, (0)) + c[0];
hence H(x
0
, d
x
0
) d
x
0
( (0)) L(x
0
, (0)) c[0].
221
The following theorem will be needed to prove the converse of
proposition 7.2.7.
Theorem 7.2.8. Suppose L : TM R is a Tonelli Lagran-
gian on the compact manifold M. Let T

t
be the associated Lax-
Oleinik semi-group. If u (
0
(M, R), then the continuous function
U : [0, +[M R dened by U(t, x) = T

t
u(x) is a viscosity
solution of
U
t
(t, x) +H(x,
U
x
(t, x)) = 0
on the open set ]0, +[M, where H : T

M R is the Hamil-
tonian associated to L, i.e. H(x, p) = sup
vTxM
p(v) L(x, v).
Proof. Suppose that : [a, b] M. Since T

b
u = T

ba
[T

a
(u)],
using the denition of T

ba
, we get
T

b
u((a)) = T

ba
[T

a
(u)]((a)
T

a
u((a)) +
_
b
a
L((s), (s)) ds,
therefore
U(b, (b)) U(a, (a))
_
b
a
L((s), (s)) ds, ()
We now show that U is a viscosity subsolution. Suppose
U, with of class C
1
and (t
0
, x
0
) = U(t
0
, x
0
), where t
0
> 0.
Fix v T
x
0
M, and pick a C
1
curve : [0, t
0
] M such that
((t
0
), (t
0
)) = (x, v).
If 0 t t
0
, we have by () and therefore
U(t
0
, (t
0
)) U(t, (t))
_
t
0
t
L((s), (s)) ds. ()
Since U, with equality at (t
0
, x
0
), noticing that (t
0
) = x
0
,
we obtain from ()
t ]0, t
0
[, (t
0
, (t
0
)) (t, (t))
_
t
0
t
L((s), (s)) ds.
Dividing by t
0
t > 0, and letting t t
0
, we get
v T
x
0
M,

t
(t
0
, x
0
) +

x
(t
0
, x
0
)(v) L(x
0
, v).
222
By Fenchels formula 1.3.1
H(x
0
,

x
(t
0
, x
0
)) = sup
vTx
0
M

x
(t
0
, x
0
)(v) L(x
0
, v),
therefore

t
(t
0
, x
0
) +H(x
0
,

x
(t
0
, x
0
)) 0.
To prove that U is a supersolution, we consider U, with of
class C
1
. Suppose U(t
0
, x
0
) = (t
0
, x
0
), with t
0
> 0.
We pick : [0, t
0
] M such that (t
0
) = x
0
and
U(t
0
, x
0
) = T

t
0
u(x
0
) = u((0)) +
_
t
0
0
L((s), (s)) ds.
Since U(0, (0)) = u((0)), this can be rewritten as
U(t
0
, x
0
) U(0, (0)) =
_
t
0
0
L((s), (s)) ds. ( )
Applying (*) above twice, we obtain
U(t
0
, x
0
) U(t, (t))
_
t
0
0
L((s), (s)) ds
U(t, (t)) U(0, (0))
_
t
0
0
L((s), (s)) ds.
Adding this two inequalities we get in fact by (**) an equality,
hence we must have
t [0, t
0
], U(t
0
, (t
0
)) U(t, (t)) =
_
t
0
t
L((s), (s)) ds.
Since U, with equality at (t
0
, x
0
), we obtain
(t
0
, (t
0
)) (t, (t))
_
t
0
t
L((s), (s)) ds.
Dividing by t
0
t > 0, and letting t t
0
, we get

t
(t
0
, x
0
) +

x
(t
0
, x
0
)( (t
0
)) L(x
0
, (t
0
)).
223
By Fenchels formula 1.3.1
H(x
0
,

x
(t
0
, x
0
))

x
(t
0
, x
0
)( (t
0
)) L(x
0
, (t
0
)),
Therefore

t
(t
0
, x
0
) +H(x
0
,

x
(t
0
, x
0
)) 0.
7.3 Lower and upper dierentials
We need to introduce the notion of lower and upper dierentials.
Denition 7.3.1. If u : M R is a map dened on the manifold
M, we say that the linear form p T

x
0
M is a lower (resp. upper)
dierential of u at x
0
M, if we can nd a neighborhood V of x
0
and a function : V R, dierentiable at x
0
, with (x
0
) = u(x
0
)
and d
x
0
= p, and such that (x) u(x) (resp. (x) u(x)), for
every x V .
We denote by D

u(x
0
) (resp. D
+
u(x
0
)) the set of lower (resp.
upper) dierentials of u at x
0
.
Exercise 7.3.2. Consider the function u : R R, x [x[,
for each x R, nd D

u(x), and D
+
u(x). Same question with
u(x) = [x[.
Denition 7.3.1 is not the one usually given for M an open
set of an Euclidean space, see [Bar94], [BCD97] or [Cla90]. It is
nevertheless equivalent to the usual denition as we now show.
Proposition 7.3.3. Let u : U R be a function dened on the
open subset U of R
n
, then the linear form p is in D

u(x
0
) if and
only if
liminf
xx
0
u(x) u(x
0
) p(x x
0
)
|x x
0
|
0.
In the same way p D
+
u(x
0
) if and only if
limsup
xx
0
u(x) u(x
0
) p(x x
0
)
|x x
0
|
0.
224
Proof. Suppose p D

u(x
0
), we can nd a neighborhood V of x
0
and a function : V R, dierentiable at x
0
, with (x
0
) = u(x
0
)
and d
x
0
= p, and such that (x) u(x), for every x V .
Therefore, for x V , we can write
(x) (x
0
) p(x x
0
)
|x x
0
|

u(x) u(x
0
) p(x x
0
)
|x x
0
|
.
Since p = d
x
0
the left hand side tends to 0, therefore
liminf
xx
0
u(x) u(x
0
) p(x x
0
)
|x x
0
|
0.
Suppose conversely, that p R
n
satises
liminf
xx
0
(u(x) u(x
0
) p(x x
0
))
|x x
0
|
0.
We pick r > 0 such that the ball

B(x
0
, r) U, and for h R
n
such that 0 < |h| < r, we set
(h) = min(0,
u(x
0
+h) u(x
0
) p(h)
|h|
).
It is easy to see that lim
h0
(h) = 0. We can therefore set (0) =
0. The function :

B(x
0
, r) R, dened by (x) = u(x
0
)+p(x
x
0
) + |x x
0
|(x x
0
), is dierentiable at x
0
, with derivative
p, it is equal to u at x
0
and satises (x) u(x), for every x

B(x
0
, r).
Proposition 7.3.4. Let u : M R be a function dened on the
manifold M.
(i) For each x in M, we have D
+
u(x) = D

(u)(x) = p [
p D

(u)(x) and D

u(x) = D
+
(u)(x).
(ii) For each x in M, both sets D
+
u(x), D

u(x) are closed con-


vex subsets of T

x
M.
(iii) If u is dierentiable at x, then D
+
u(x) = D

u(x) = d
x
u.
(iv) If both sets D
+
u(x), D

u(x) are non-empty then u is dier-


entiable at x.
225
(v) if v : M R is a function with v u and v(x) = u(x), then
D

v(x) D

u(x) and D
+
v(x) D
+
u(x).
(vi) If U is an open convex subset of an Euclidean space and
u : U R is convex then D

u(x) is the set of supporting


linear forms of u at x U. In particular D
+
u(x) ,= if and
only if u is dierentiable at x.
(vii) Suppose M has a distance d obtained from the Riemannian
metric g. If u : M R is Lipschitz for d with Lipschitz
constant Lip(u), then for any p D

u(x) we have |p|


x

Lip(u).
In particular, if M is compact then the sets D

u = (x, p) [
p D

u(x), x M are compact.


Proof. Part (i) and and the convexity claim in part (ii) are obvious
from the denition 7.3.1.
To prove the fact that D
+
u(x
0
) is closed for a given for x
0
M,
we can assume that M is an open subset of R
k
. We will apply
proposition 7.3.3. If p
n
D
+
u(x
0
) converges to p R
k
, we can
write
u(x) u(x
0
) p(x x
0
)
|x x
0
|

u(x) u(x
0
) p
n
(x x
0
)
|x x
0
|
+|p
n
p|.
Fixing n, and letting x x
0
, we obtain
limsup
xx
0
u(x) u(x
0
) p(x x
0
)
|x x
0
|
|p
n
p|.
If we let n , we see that p D
+
u(x
0
).
We now prove (iii) and (iv) together. If u is dierentiable at
x
0
M then obviously d
x
0
u D
+
u(x
0
) D

u(x
0
). Suppose
now that both D
+
u(x
0
) and D

u(x
0
) are both not empty, pick
p
+
D
+
u(x
0
) and p

u(x
0
). For h small, we have
p

(h) +|h|

(h) u(x
0
+h) u(x
0
) p
+
(h) +|h|
+
(h), (*)
where both

(h) and
+
(h) tend to 0, a h 0. If v R
n
, for
t > 0 small enough, we can replace h by tv in the inequalities (*)
above. Forgetting the middle term and dividing by t, we obtain
p

(v) +|v|

(tv) p
+
(v) +|v|
+
(tv),
226
letting t tend to 0, we see that p

(v)+ p
+
(v), for every v R
n
.
Replacing v by v gives the reverse inequality p
+
(v)+ p

(v),
therefore p

= p
+
. This implies that both D
+
u(x
0
) and D

u(x
0
)
are reduced to the same singleton p. The inequality (*) above
now gives
p(h) +|h|

(h) u(x
0
+h) u(x
0
) p(h) +|h|
+
(h),
this clearly implies that p is the derivative of u at x
0
.
Part (v) follows routinely from the denition.
To prove (vi), we remark that by convexity u(x
0
+th) (1
t)u(x
0
) +tu(x
0
+h), therefore
u(x
0
+h) u(x
0
)
u(x
0
+th) u(x
0
)
t
.
If p is a linear form we obtain
u(x
0
+h) u(x
0
) p(h)
|h|

u(x
0
+th) u(x
0
) p(h)
|th|
.
If p D

u(x
0
), then the liminf as t 0 of the right hand side is
0, therefore u(x
0
+h) u(x
0
) p(h) 0, which shows that p is
a supporting linear form. Conversely, a supporting linear form is
clearly a lower dierential.
It remains to prove (vii). Suppose, for example that : V R
is dened on some neighborhood V of a given x
0
M, that it is
dierentiable at x
0
, and that u on V , with equality at x
0
.
If v T
x
0
M is given, we pick a C
1
path : [0, ] V , with
> 0, (0) = x
0
, and (0) = v.We have
t [0, ], [u((t)) u(x
0
)[ Lip(u)d((t), x
0
)
Lip(u)
_
t
0
| (s)| ds.
Therefore u((t)) u(x
0
) Lip(u)
_
t
0
| (s)| ds. Since u on
V , with equality at x
0
, it follows that
((t)) (x
0
) Lip(u)
_
t
0
| (s)| ds.
227
Dividing by t > 0, and letting t 0, we get
d
x
0
(v) Lip(u)|v|.
Since v T
x
0
M is arbitrary, we can change v into v in the
inequality above to conclude that we also have
d
x
0
(v) Lip(u)|v|.
It then follows that |d
x
0
| Lip(u).
Lemma 7.3.5. If u : M R is continuous and p D
+
u(x
0
)
(resp. p D

u(x
0
)), there exists a C
1
function : M R,
such that (x
0
) = u(x
0
), d
x
0
phi = p, and (x) > u(x) (resp.
(x) < u(x)) for x ,= x
0
.
Moreover, if W is any neighborhood of x
0
and C > 0, we can
choose such that (x) u(x) + C, for x / W (resp. (x)
u(x) C).
Proof. Assume rst M = R
k
. To simplify notations, we can as-
sume x
0
= 0. Moreover, subtracting from u the ane function
x u(0) + p(x). We can assume u(0) = 0 and p = 0. The fact
that 0 D
+
u(0) gives
limsup
x0
u(x)
|x|
0.
If we take the non-negative part u
+
(x) = max(u(x), 0) of u, this
gives
lim
x0
u
+
(x)
|x|
= 0. ()
If we set
c
n
= supu
+
(x) [ 2
(n+1)
|x| 2
n

then c
n
is nite and 0, because u
+
0 is continuous. Moreover
using that 2
n
u
+
(x) u
+
(x)/|x|, for |x| 2
n
, and the limit in
() above, we obtain
lim
n
[ sup
mn
2
m
c
m
] = 0. ()
228
We now consider : R
k
R a C

bump function with = 1 on


the set x R
k
[ 1/2 |x| 1, and whose support is contained
in x R
k
[ 1/4 |x| 2. We dene the function : R
k
R
by
(x) =

nZ
(c
n
+ 2
2n
)(2
n
x).
This function is well dened at 0 because every term is then 0. For
x ,= 0, we have (2
n
x) ,= 0 only if 1/4 < |2
n
x| < 2. Taking the
logarithm in base 2, this can happen only if 2 log
2
|x| < n <
1 log
2
|x|, therefore this can happen for at most 3 consecutive
integers n, hence the sum is also well dened for x ,= 0. Moreover,
if x ,= 0, the set V
x
= y ,= 0 [ 1 log
2
|x| < log
2
|y| <
1 log
2
|x| is a neighborhood of x and
y V
y
, (y) =

3log
2
x<n<2log
2
x
(c
n
+ 2
2n
)(2
n
y). (*)
This sum is nite with at most 5 terms, therefore is C

on
R
k
0.
We now check that is continuous at 0. Using equation (*),
and the limit () we see that
0 (x)

3log
2
x<n<2log
2
x
(c
n
+ 2
2n
)
5 sup
n>3log
2
x
(c
n
+ 2
2n
) 0 as x 0.
To show that is C
1
on the whole of R
k
with derivative 0 at 0,
it suces to show that d
x
tends to 0 as |x| 0. Dierentiating
equation (*) we see that
d
x
=

3log
2
x<n<2log
2
x
(c
n
+ 2
2n
)2
n
d
2
n
x
.
Since has compact support K = sup
xR
n|d
x
| is nite. The
equality above and the limit in () give
|d
x
| 5K sup2
n
c
n
+ 2
n
[ n 2 log
2
|x|,
but the right hand side goes to 0 when |x| 0.
229
We now show (x) > u(x), for x ,= 0. There is an integer n
0
such that |x| [2
n
0
+1
, 2
n
0
], hence (2
n
0
x) = 1 and (x)
(2
n
0
x)(c
n
0
+2
2n
0
) c
n
0
+2
2n
0
, since c
n
0
= supu
+
(y) [ |y|
[2
(n
0
+1)
, 2
n
0
], we obtain c
n
0
u
+
(x) and therefore (x) >
u
+
(x) u(x).
It remains to show that we can get rid of the assumption
M = R
k
, and to show how to obtain the desired inequality on the
complement of W. We pick a small open neighborhood U W
of x
0
which is dieomorphic to an Euclidean space. By what we
have done, we can nd a C
1
function : U R with (x
0
) =
u(x
0
), d
x
0
= p, and (x) > u(x), for x U x
0
. We then
take a C

bump function : M [0, 1] which is equal to 1


on a neighborhood of x
0
and has compact support contained in
U W. We can nd a C

function

: M R such that

u + C. It is easy to check that the function : M R


dened by (x) = (1 (x))

(x) + (x)(x) has the required
property.
The following simple lemma is very useful.
Lemma 7.3.6. Suppose : M R is C
r
, with r 0. If x
0

M, C 0, and W is a neighborhood of x
0
, there exist two C
r
functions
+
,

: M R, such that
+
(x
0
) =

(x
0
) = (x
0
),
and
+
(x) > (x) >

(x), for x ,= x
0
. Moreover
+
(x)
C > (x) >

(x) + C, for x / W. If r 1, then necessarily


d
x
0

+
= d
x
0

= d
x
0

Proof. The last fact is clear since


+
(resp.

) achieves
a minimum (resp. maximum) at x
0
.
Using the same arguments as in the end of the proof in the
previous lemma to obtain the general case, it suces to assume
C = 0 and M = R
n
. In that case, we can take

(x) = (x)
|x x
0
|
2
.
Proposition 7.3.7. Suppose M is a compact manifold. Let L :
TM R be a C
2
Tonelli Lagrangian. Consider the associ-
ated Lax-Oleinik semi-groups T

t
, T
+
t
. Suppose that t > 0, and
that : [0, t] M is a C
1
curve with (t) = x (resp. (0) =
x), and such that T

t
u(x) = u((0)) +
_
t
0
L((s), (s)) ds (resp.
T
+
t
u(x) = u((t))
_
t
0
L((s), (s)) ds), then L/v((t), (t))
230
D
+
[T

t
u](x), and L/v((0), (0)) D

u((0)) (resp. L/v((0), (0))


D

[T

t
u](x) and L/v((t), (t)) D
+
[u]((t))).
Proof. A Faire!!!!!
7.4 Criteria for viscosity solutions
We x in this section a continuous function H : T

M R.
Theorem 7.4.1. Let u : M R be a continuous function.
(i) u is a viscosity subsolution of H(x, d
x
u) = 0 if and only if
for each x M and each p D
+
u(x) we have H(x, p) 0.
(ii) u is a viscosity supersolution of H(x, d
x
u) = 0 if and only if
for each x M and each p D

u(x) we have H(x, p) 0.


Proof. Suppose that u is a viscosity subsolution. If p D
+
u(x),
since u is continuous, it follows from 7.3.5 that there exists a C
1
function : M R, with u on M, u(x
0
) = (0) and d
x
= p.
By the viscosity subsolution condition H(x, p) = H(x, d
x
) 0, .
Suppose conversely that for each x M and each p D
+
u(x
0
)
we have H(x, p) 0. If : M R is C
1
with u , then at each
point x where u(x) = (x), we have d
x
D
+
u(x) and therefore
H(x, d
x
) 0.
Since D

u(x) depends only on the values of u in a neighbor-


hood of x, the following corollary is now obvious. It shows the
local nature of the viscosity conditions.
Corollary 7.4.2. Let u : M R be a continuous function.
If u is a viscosity subsolution (resp. supersolution, solution) of
H(x, d
x
u) = 0 on M, then any restriction u
|U
to an open sub-
set U M is itself a viscosity subsolution (resp. supersolution,
solution) of H(x, d
x
u) = 0 on U.
Conversely, if there exists an open cover (U
i
)
iI
of M such that
every restriction u
|U
i
is a viscosity subsolution (resp. supersolu-
tion, solution) of H(x, d
x
u) = 0 on U
i
, then u itself is a viscosity
subsolution (resp. supersolution, solution) of H(x, d
x
u) = 0 on M.
Here is another straightforward consequence of theorem 7.4.1.
231
Corollary 7.4.3. Let u : M R be a locally Lipschitz func-
tion. If u is a viscosity subsolution (resp. supersolution solu-
tion) of H(x, d
x
u) = 0, then H(x, d
x
u) 0 (resp. H(x, d
x
u)
0, H(x, d
x
u) = 0) for almost every x M.
In particular, a locally Lipschitz viscosity solution is always a
very weak solution.
We end this section with one more characterization of viscosity
solutions.
Proposition 7.4.4 (Criterion for viscosity solution). Suppose
that u : M R is continuous. To check that u is a viscosity
subsolution (resp. supersolution) of H(x, d
x
u) = 0, it suces to
show that for each C

function : M R such that u has


a unique strict global maximum (resp. minimum), attained at x
0
,
we have H(x
0
, d
x
0
) 0 (resp. H(x
0
, d
x
0
) 0).
Proof. We treat the subsolution case. We rst show that if :
M R is a C

function such that u achieves a (not necessarily


strict) maximum at x
0
, then we have H(x
0
, d
x
0
) 0. In fact
applying 7.3.6, we can nd a C

function
+
: M R such that

+
(x
0
) = (x
0
), d
x
0

+
= d
x
0
,
+
(x) > (x), for x ,= x
0
. The
function u
+
has a unique strict global maximum achieved at
x
0
, therefore H(x
0
, d
x
0

+
) 0. Since d
x
0

+
= d
x
0
, this nishes
our claim.
Suppose now that : M R is C
1
and that u has a global
maximum at x
0
, we must show that H(x
0
, d
x
0
) 0. We x a
relatively compact open neighborhood W of x
0
, by 7.3.6, applied to
the continuous function , there exists a C
1
function
+
: M R
such that
+
(x
0
) = (x
0
), d
x
0

+
= d
x
0
,
+
(x) > (x), for x ,=
x
0
, and even
+
(x) > (x) + 3, for x / W. It is easy to see that
u
+
has a strict global maximum at x
0
, and that u(x)
+
(x) <
u(x
0
)
+
(x
0
) 3, for x / W. By smooth approximations, we
can nd a sequence of C

functions
n
: M R such that
n
converges to
+
in the C
1
topology uniformly on compact subsets,
and sup
xM
[
n
(x)
+
(x)[ < 1. This last condition together with
u(x)
+
(x) < u(x
0
)
+
(x
0
)3, for x / W, gives u(x)
n
(x) <
u(x
0
)
n
(x
0
) 1, for x / W. This implies that the maximum of
u
n
on the compact set

W is a global maximum of u
n
. Choose
y
n


W where u
n
attains its global maximum. Since
n
is C

,
232
from the beginning of the proof we must have H(y
n
, d
yn

n
) 0.
Extracting a subsequence, if necessary, we can assume that y
n
converges to y



W. Since
n
converges to
+
uniformly on the
compact set

W, necessarily u
+
achieves its maximum on

W at
y

. This implies that y

= x
0
, because the strict global maximum
of u

is precisely attained at x
0
W. The convergence of
n
to
+
is in the C
1
topology, therefore (y
n
, d
yn

n
) (x
0
, d
x
0

+
),
and hence H(y
n
, d
yn

n
) H(x
0
, d
x
0

+
), by continuity of H. But
H(y
n
, d
yn

n
) 0 and d
x
0
= d
x
0

+
, hence H(x
0
, d
x
0
) 0.
7.5 Coercive Hamiltonians
Denition 7.5.1 (Coercive). A continuous function H : T

M
R is said to be coercive above every compact subset, if for each
compact subset K M and each c R the set (x, p) T

M [
x K, H(x, p) c is compact.
Choosing any Riemannian metric on M, it is not dicult to
see that H is coercive, if and only if for each compact subset
K M, we have lim
p
H(x, p) = + the limit being uniform
in x K.
Theorem 7.5.2. Suppose that H : T

M R is coercive above
every compact subset, and c R then a viscosity subsolution
of H(x, d
x
u) = c is necessarily locally Lipschitz, and therefore
satises H(x, d
x
u) c almost everywhere.
Proof. Since this is a local result we can assume M = R
k
, and
prove only that u is Lipschitz on a neighborhood of the origin 0.
We will consider the usual distance d given by d(x, y) = |y x|,
where we have chosen the usual Euclidean norm on R
k
. We set

0
= sup|p| [ p R
k
, x R
k
, |x| 3, H(x, p) c.
Suppose u : R
k
R is a subsolution of H(x, d
x
u) = c. Choose

0
+ 1 such that
2 > sup[u(y) u(x)[ [ x, y R
k
, |x| 3, |y| 3.
Fix x, with |x| 1, and dene : R
k
R by (y) = |yx|.
Pick y
0


B(x, 2) where the function y u(y) (y) attains its
233
maximum for y

B(x, 2). We rst observe that y
0
is not on the
boundary of

B(x, 2). In fact, if |y x| = 2, we have u(y)(y) =
u(y)2 < u(x) = u(x)(x). In particular y
0
is a local maximum
of u . If y
0
is not equal to x, then d
y
0
exists, with d
y
0
(v) =
y
0
x, v)/|y
0
x|, and we obtain |d
y
0
| = . On the other
hand, since u(y) u(y
0
) (y
0
) + (y), for y in a neighborhood
of y
0
, we get d
y
0
D
+
u(y
0
), and therefore have H(y
0
, d
y
0
) c.
By the choice of
0
, this gives |d
y
0
|
0
<
0
+ 1 . This
contradiction shows that y
0
= x, hence u(y) |y x| u(x),
for every x of norm 1, and every y

B(x, 2). This implies that
u has Lipschitz constant on the unit ball of R
k
.
7.6 Viscosity and weak KAM
In this section we nish showing that weak KAM solutions and
viscosity solutions are the same.
Theorem 7.6.1. Let L : TM R be a Tonelli Lagrangian on
the compact manifold M. Denote by H : T

M R its associated
Hamiltonian. A continuous function u : U R is a viscosity
subsolution of H(x, d
x
u) = c on the open subset U if and only if
u L +c.
Proof. By proposition 7.2.7, it remains to prove that a viscosity
subsolution of H(x, d
x
u) = c is dominated by L + c on U. Since
H is superlinear, we can apply theorem 7.5.2 to conclude that
u is locally Lipschitz, and hence, by corollary 7.4.3, we obtain
H(x, d
x
u) c almost everywhere on U. From lemma 4.2.3, we
infer u L +c on U.
Theorem 7.6.2. Let L : TM R be a Tonelli Lagrangian on
the compact manifold M. Denote by H : T

M R its associated
Hamiltonian. A continuous function u : M R is a viscosity
solution of H(x, d
x
u) = c if and only if it is Lipschitz and satises
u = T

t
u+ct, for each t 0. (In particular, we must have c=c[0].)
Proof. If u satises u = T

t
u +ct, for each t 0, then necessarily
c = c[0], and therefore, by proposition 7.2.7, the function u is a
viscosity solution of H(x, d
x
u) = c = c[0].
234
Suppose now that u is a viscosity solution. From proposition
7.6.1, we know that u L + c, and that u is Lipschitz. We then
dene u(t, x) = T

t
u(x). We must show that u(t, x) = u(x) ct.
Since we know that u is locally Lipschitz on ]0, +[M, it suf-
ces to show that
t
u(t, x) = c at each (t, x) where u admits
a derivative. In fact, since u is locally Lipschitz, it is dieren-
tiable almost everywhere, therefore for almost every x the deriva-
tive d
(t,x)
u exists for almost every t. If we x such an x, it follows
that
t
u(t, x) = c, for almost every t. But, since the t u(t, x)
is locally Lipschitz, it is the integral of its derivative, therefore
u(t, x) u(0, x) = ct. This is valid for almost every x M, and
by continuity for every x M.
It remains to show that at a point (t, x) where u is dieren-
tiable, we have
t
u(t, x) = c. From proposition 7.2.8, we know
that u is a viscosity solution of
t
u + H(x,
x
u) = 0. Hence we
have to show that H(x,
x
u(t, x)) = c. In fact, we already have
that H(x,
x
u(t, x)) c, because u(t, ) = T

t
u is dominated by
L +c, like u. It remains to prove that H(x,
x
u(t, x)) c. To do
this, we identify the derivative of
x
u(t, x). We choose : [0, t]
M with (t) = x and T

t
u(x) = u((0)) +
_
t
0
L((s), (s)) ds.
The curve is a minimizer of the action. In particular, the
curve s ((s), (s)) is a solution of the Euler-Lagrange equa-
tion, it follows that the energy H((s),
L
v
((s)), (s))) is con-
stant on [0, t]. By proposition 7.3.7, we have L/v((t)), (t))
D
+
(T
+
t
u)(x), therefore
x
u(t, x) =
L
v
((t)), (t)). Using the fact
that H((s),
L
v
((s)), (s))) is constant, we are reduced to see
that H((0),
L
v
((0)), (0))) c. But the same proposition 7.3.7
yields also L/v((0), (0)) D

u((0)). We can therefore


conclude using theorem 7.4.1, since u is a viscosity solution of
H(x, d
x
u) = c.
Chapter 8
More on Viscosity
Solutions
We further develop the theory of viscosity solutions. Although
many things are standard, whatever is not comes from joint work
with Antonio Siconol, see [FS04] and [FS05].
8.1 Stability
Theorem 8.1.1 (Stability). Suppose that the sequence of contin-
uous functions H
n
: T

M R converges uniformly on compact


subsets to H : T

M R. Suppose also that u


n
: M R is a se-
quence of continuous functions converging uniformly on compact
subsets to u : M R. If, for each n, the function u
n
is a viscos-
ity subsolution (resp. supersolution, solution) of H
n
(x, d
x
u
n
) = 0,
then u is a viscosity subsolution (resp. supersolution, solution) of
H(x, d
x
u) = 0.
Proof. We show the subsolution case. We use the criterion 7.4.4.
Suppose that : M R is a C

function such that u has


a unique strict global maximum, achieved at x
0
, we have to show
H(x
0
, d
x
0
) 0. We pick a relatively compact open neighbor-
hood W of x
0
. For each n, choose y
n


W where u
n
attains its
maximum on the compact subset

W. Extracting a subsequence,
if necessary, we can assume that y
n
converges to y



W. Since
u
n
converges to u uniformly on the compact set

W, necessarily
235
236
u achieves its maximum on

W at y

. But u has a strict


global maximum at x
0
W therefore y

= x
0
. By continuity
of the derivative of , we obtain (y
n
, d
yn
) (x
0
, d
x
0
). Since
W is an open neighborhood of x
0
, dropping the rst terms if nec-
essary, we can assume y
n
W, this implies that y
n
is a local
maximum of u
n
, therefore d
yn
D
+
u
n
(y). Since u
n
is a vis-
cosity subsolution of H
n
(x, d
x
u
n
) = 0, we get H
n
(y
n
, d
yn
) 0.
The uniform convergence of H
n
on compact subsets now implies
H(x
0
, d
x
0
) = lim
n
H
n
(y
n
, d
yn
) 0.
8.2 Construction of viscosity solutions
Proposition 8.2.1. Let H : T

M R be a continuous func-
tion. Suppose (u
i
)
iI
is a family of continuous functions u
i
:
M R such that each u
i
is a subsolution (resp. supersolution) of
H(x, d
x
u) = 0. If sup
iI
u
i
(resp. inf
iu
u
i
) is nite and continu-
ous everywhere, then it is also a subsolution (resp. supersolution)
of H(x, d
x
u) = 0 .
Proof. Set u = sup
iI
u
i
. Suppose : M R is C
1
, with (x
0
) =
u(x
0
) and (x) > u(x), for every x M x
0
. We have to show
H(x
0
, d
x
0
) 0. Fix some distance d on M. By continuity of the
derivative of , it suces to show that for each > 0 small enough
there exists x

B(x
0
, ), with H(x, d
x
) 0.
For > 0 small enough, the closed ball

B(x
0
, ) is compact.
Fix such an > 0. There is a > 0 such that (y) u(y) =
sup
iI
u
i
(y), for each y B(x
0
, ).
Since (x
0
) = u(x
0
), we can nd i

I such that (x
0
) <
u
i
(x
0
). It follows that the maximum of the continuous function
u
i
on the compact set

B(x
0
, ) is not attained on the boundary,
therefore u
i
has a local maximum at some x



B(x
0
, ). Since
the function u
i
is a viscosity subsolution of H(x, d
x
u) = 0, we
have H(x

, d
x
) 0.
Theorem 8.2.2 (Perron Method). Suppose the Hamiltonian H :
TM R is coercive above every compact subset. If M is con-
nected and there exists a viscosity subsolution u : M R of
H(x, d
x
u) = 0, then for every x
0
M, the function S
x
0
: M R
dened by S
x
0
(x) = sup
v
v(x), where the supremum is taken over
237
all viscosity subsolutions v satisfying v(x
0
) = 0, has indeed nite
values and is a viscosity subsolution of H(x, d
x
u) = 0 on M.
Moreover, it is a viscosity solution of H(x, d
x
u) = 0 on M
x
0
.
Proof. Call oo
x
0
the family of viscosity subsolutions v : M R
of H(x, d
x
v) = 0 satisfying v(x
0
) = 0.
Since H is coercive above every compact subset of M, by
theorem 7.5.2, we know that each element of this family is lo-
cally Lipschitz. Moreover, since for each compact set K, the set
(x, p) [ x K, H(x, p) 0 is compact, it follows that the fam-
ily of restrictions v
|K
, v oo
x
0
is equi-Lipschitzian. We now
show, that S
x
0
is nite everywhere. Since M is connected, given
x M, there exists a compact connected set K
x,x
0
containing
both x and x
0
. By the equicontinuity of the family of restric-
tions v
|Kx,x
0
[ v oo
x
0
, we can nd > 0, such that for each
y, z K
x,x
0
with d(y, z) , we have [v(y) v(z)[ 1, for each
v oo
x
0
.
By the set K
x,x
0
is connected , we can nd a sequence x
0
, x
1
, , x
n
=
x in K
x,x
0
with d(x
i
, x
i+1
) . It follows that [v(x)[ = [v(x)
v(x
0
)[

n1
i=0
[v(x
i+1
) v(x
i
)[ n, for each v oo
x
0
. Therefore
sup
vSSx
0
v(x) is nite everywhere. Moreover, as a nite-valued
supremum of a family of locally equicontinuous functions, it is
continuous.
By the previous proposition 8.2.1, the function S
x
0
is a vis-
cosity subsolution on M itself. It remains to show that it is a
viscosity solution of H(x, d
x
u) on M x
0
.
Suppose : M R is C
1
with (x
1
) = S
x
0
(x
1
), where
x
1
,= x
0
, and (x) < S
x
0
(x) for every x ,= x
1
. We want to show
that necessarily H(x
1
, d
x
1
) 0. If this were false, by continu-
ity of the derivative of , endowing M with a distance dening
its topology, we could nd > 0 such that H(y, d
y
) < 0, for
each y

B(x
1
, ). Taking > 0 small enough, we assume that

B(x
1
, ) is compact and x
0
/

B(x
1
, ). Since < S
x
0
on the
boundary B(x
1
, ) of

B(x
1
, ), we can pick > 0, such that
(y) + S
x
0
(y), for every y B(x
1
, ). We dene

S
x
0
on

B(x
1
, ) by

S
x
0
(x) = max((x) + /2, S
x
0
(x)). The function

S
x
0
is a viscosity subsolution of H(x, d, u) on

B(x
1
, ) as the maxi-
mum of the two viscosity subsolutions +/2 and S
x
0
. Moreover,
238
this function

S
x
0
coincides with S
x
0
outside K = x

B(x
1
, ) [
(x)+/2 S
x
0
(x)) which is a compact subset of

B(x
1
, ), there-
fore we can extend it to M itself by

S
x
0
= S
x
0
on M K. It is a
viscosity subsolution of H(x, d
x
u) on M itself, since its restrictions
to both open subsets MK and

B(x
1
, ) are viscosity subsolutions
and M =

B(x
1
, ) (M K).
But

S
x
0
(x
0
) = S
x
0
(x
0
) = 0 because x
0
/

B(x
1
, ). Moreover

S
x
0
(x
1
) = max((x
1
)+/2, S
x
0
(x
1
)) = max(S
x
0
(x
1
)+/2, S
x
0
(x
1
)) =
S
x
0
(x
1
)+/2 > S
x
0
(x
1
). This contradicts the denition of S
x
0
.
The next argument is inspired by the construction of Buse-
mann functions in Riemannain Geometry, see [BGS85].
Corollary 8.2.3. Suppose that H : T

M R is a continuous
Hamiltonian coercive above every compact subset of the connected
non-compact manifold M. If there exists a viscosity subsolution
of H(x, d
x
u) = 0 on M, then there exists a viscosity solution on
M.
Proof. Fix x M, and pick a sequence x
n
(this means such
that each compact subset of M contains only a nite number of
points in the sequence).
By arguments analogous to the ones used in the previous proof,
the sequence S
xn
is locally equicontinuous and moreover, for each
x M, the sequence S
xn
(x) S
xn
( x) is bounded. Therefore, by
Ascolis theorem, extracting a subsequence if necessary, we can
assume that S
xn
S
xn
( x) converges uniformly to a continuous
function u : M R. It now suces to show that the restriction
of u to an arbitrary open relatively compact subset V of M is a
viscosity solution of H(x, d
x
u) = 0 on V . Since n [ x
n


V is
nite, for n large enough, the restriction of S
xn
S
xn
( x) to V is
a viscosity solution; therefore by the stability theorem 8.1.1, the
restriction of the limit u to V is also a viscosity solution.
The situation is dierent for compact manifolds as can be seen
from the following theorem:
Theorem 8.2.4. Suppose H : T

M R is a coercive Hamil-
tonian on the compact manifold M. If there exists a viscosity
subsolution of H(x, d
x
u) = c
1
and a viscosity supersolution of
H(x, d
x
u) = c
2
, then necessarily c
2
c
1
.
239
In particular, there exists a most one c for which the Hamilton-
Jacobi equation H(x, d
x
u) = c has a global viscosity solution u :
M R. This only possible value is the smallest c for which
H(x, d
x
u) = c admits a global viscosity subsolution u : M R.
In order to prove this theorem, we will need a lemma (the last
part of the lemma will be used later).
Lemma 8.2.5. Suppose M compact, and u : M R is a Lips-
chitz viscosity subsolution (resp. supersolution) of H(x, d
x
u) = c.
For every > 0, there exists a locally semi-convex (resp. semi-
concave) function u

: M R such that |u

u|

< , and u

is
a viscosity subsolution (resp. supersolution) of H(x, d
x
u) = c +
(resp. H(x, d
x
u) = c ).
Moreover, if K U are respectively a compact and an open
subsets of M such that the restriction u
|U
is a viscosity subsolu-
tion (resp. supersolution) of H(x, d
x
u) = c

on U, for some c

< c
(resp. c

> c), we can also impose that restriction u


|U
is a viscos-
ity subsolution (resp. supersolution) of H(x, d
x
u) = c

+ (resp.
H(x, d
x
u) = c

) on a neighborhood of K.
Proof. Suppose > 0 is given. Fix a C

Riemannian metric
g on M. We call T

t
and T
+
t
the two Lax-Oleinik semi-groups
associated to the Lagrangian L(x, v) =
1
2
g
x
(v, v) =
1
2
|v|
2
x
. If
> 0 and u : M R is a Lipschitz viscosity subsolution of
H(x, d
x
u) = c, we consider the locally semi-convex function T
+
t
u.
By the analogous of part (7) of corollary ??, the map t T
+
t
u
is continuous as a map with values in (
0
(M, R) endowed with the
sup norm, therefore |T
+
t
uu|

< , for each t > 0 small enough.


Assume that t > 0. Since T
+
t
u is locally semi-convex, at each point
x, the set D

T
+
t
u(x) is not empty, therefore the points x where
D
+
T
+
t
u(x) ,= are the points where to T
+
t
u is dierentiable.
Hence to check that it is a subsolution of H(x, du) = c + , it
suces to show that if d
x
T
+
t
u exists then H(x, d
x
T
+
t
u) c +.
Suppose that d
x
T
+
t
u exists. Choose a geodesic : [0, t] M
with (0) = x, and T
+
t
u(x) =
_
t
0
1
2
| (t)|
2
(t)
dt u((t)). We
have d
x
T
+
t
u(x) = (0)

, where for v T
y
M, the linear form
v

y
M is given by v

(w) = g
y
(v, w), for every w T
y
M.
Moreover, we also have (t)

D
+
u((t)). Therefore d
x
T
+
t
u

g
t
(Graph(D
+
u)), where
g
t
: T

M T

M is the Hamiltonian
240
form of the geodesic ow
g
t
: TM TM of g, and Graph(D
+
u) =
(x, p) [ x M, p D
+
u(x).
Since u is a subsolution of H(x, d
x
u) = c, we have H(x, p)
c, for each p D
+
u(x). The compactness of P = p [ p
T

x
M, |p|
x
Lip(u), which contains Graph(D
+
u), and the con-
tinuity of both the ow
g
t
and the Hamiltonian H imply that
there exists t
0
> 0 such that
g
t
D
+
u H
1
(] , c + ]), for
every t [0, t
0
].
To prove the last part, we choose an open subset V with K
V

V U. We observe that Graph(D
+
u) is contained in the
compact set

P = (x, p) P [ x / U, H(x, p) c (x, p) P [
x

U, H(x, p) c

. Again by compactness, we can nd a t

0
such
that the intersection of
g
t
D
+
u with T

V = (x, p) [ x V is
contained in H
1
(] , c

+]), for every t [0, t

0
].
Proof of theorem 8.2.4. Suppose c
1
< c
2
, and choose > 0, with
c
1
+ < c
2
, by the previous lemma 8.2.5, we can nd a locally
semi-convex function u
1
: M R which is a viscosity subsolution
of H(x, d
x
u) = c
1
+.
We now show that for every x M, there exists p D

u
1
(x)
with H(x, p) c
1
+ . Since a locally semi-convex function is
Lipschitz, by Rademachers theorem, if x M, we can nd a se-
quence of points x
n
M converging to x such that the derivative
d
xn
u
1
exists. We have H(x
n
, d
xn
u
1
) c
1
+ . Since u
1
is Lip-
schitz, the points (x
n
, d
xn
u
1
) are contained in a compact subset
of T

M. Extracting a sequence if necessary, we can assume that


(x
n
, d
xn
u
1
) (x, p), of course H(x, p) c
1
+, and p D

u
1
(x),
because u
1
is locally semi-convex.
We x u
2
: M R a viscosity supersolution of H(x, d
x
u) = c
2
.
Call x
0
a point where the continuous function u
2
u
1
on the com-
pact manifold M achieves its minimum. We have u
2
u
2
(x
0
)
u
1
(x
0
) +u
1
with equality at x
0
, therefore D

u
1
(x
0
) D

u
2
(x
0
),
this is impossible since D

u
2
(x
0
) H
1
([c
2
, +[), D

u
1
(x
0
)
H
1
(] , c
1
+]) ,= and c
1
+ < c
2
.
Exercise 8.2.6. Prove a non-compact version of lemma 8.2.5:
Suppose H : T

M R is a continuous Hamiltonian on the


(not necessarily compact) manifold M. If u : M R is a locally
Lipschitz viscosity subsolution (resp. supersolution) of H(x, d
x
u) =
241
c, show that for every > 0, and every relatively compact open
subset U, there exists a locally semi-convex (resp. semi-concave)
function u

: U R such that |u

u
|U
|

< , and u

is a vis-
cosity subsolution (resp. supersolution) of H(x, d
x
u) = c + (resp.
H(x, d
x
u) = c ) on U.
Denition 8.2.7 (strict subsolution). We say that a viscosity
subsolution u : M R of H(x, d
x
u) = c is strict at x
0
M if
there exists an open neighborhood V
x
0
of x
0
, and c
x
0
< c such
that u[V
x
0
is a viscosity subsolution of H(x, d
x
u) = c
x
0
on V
x
0
.
Here is a way to construct viscosity subsolutions which are
strict at some point.
Proposition 8.2.8. Suppose that u : M R is a viscosity sub-
solution of H(y, d
y
u) = c on M, that is also a viscosity solution
on M x. If u is not a viscosity solution of H(y, d
y
u) = c on M
itself then there exists a viscosity subsolution of H(y, d
y
u) = c on
M which is strict at x.
Proof. If u is not a viscosity solution, since it is a subsolution on
M, it is the supersolution condition that is violated. Moreover,
since u is a supersolution on M x, the only possibility is that
there exists : M R of class C
1
such that (x) = u(x), (y) <
u(y), for y ,= x, and H(x, d
x
) < c. By continuity of the derivative
of , we can nd a compact ball

B(x, r), with r > 0, and a c
x
< c
such that H(y, d
y
) < c
x
, for every y

B(0, r). In particular, the
C
1
function is a subsolution of H(z, d
z
v) = c
x
on

B(x, r), and
therefore also of H(z, d
z
v) = c on the same set since c
x
< c.
We choose > 0 such that for every y B(x, r) we have
u(y) > (y) + . This is possible since B(x, r) is a compact
subset of M x where we have the strict inequality < u.
If we dene u : M R by u(y) = u(y) if y /

B(x, r) and
u(y) = max(u(y), (y) + ), we obtain the desired viscosity sub-
solution of H(y, d
y
u) c which is strict at x. In fact, by the
choice of > 0, the subset K = y

B(x, r) [ (y) + u(y)
is compact and contained in the open ball

B(x, r). Therefore M
is covered by the two open subsets M K and

B(x, r). On the
rst open subset u is equal to u, it is therefore a subsolution of
H(y, d
y
u) = c on that subset. On the second open subset

B(x, r),
242
the function u is the maximum of u and + which are both
subsolutions of H(y, d
y
u) = c on

B(x, r), by proposition 8.2.1, it
is therefore a subsolution of H(y, d
y
u) = c on that second open
subset. Since u(x) = (x); we have u(x) = (x) + > u(x),
therefore by continuity u = + on a neighborhood N

B(x, r)
of x. On that neighborhood H(y, d
y
) < c
x
, hence u is strict at
x.
8.3 Quasi-convexity and viscosity subsolu-
tions
In this section we will be mainly interested in Hamiltonians H :
T

M R quasi-convex in the bers, i.e. for each x M, the


function p H(x, p) is quasi-convex on the vector space T

x
M,
see denition 1.5.1
Our rst goal in this section is to prove the following theorem:
Theorem 8.3.1. Suppose H : T

M R is quasi-convex in the
bers. If u : M R is locally Lipschitz and H(x, d
x
u) c almost
everywhere, for some xed c R, then u is a viscosity subsolution
of H(x, d
x
u) = c.
Before giving the proof of the theorem we need some prelimi-
nary material.
Let us rst recall from denition 4.2.4 that the Hamiltonian
constant H
U
(u) of a locally Lipschitz function u : U R, where
U is an open subset of M is the essential supremum on U of
H(x, d
x
u).
We will use some classical facts about convolution. Let (

)
>0
be a family of functions

: R
k
[0, [ of class C

, with

(x) =
0, if |x| , and
_
R
k

(x) dx = 1. Suppose that V, U are open


subsets of R
k
, with

V compact and contained in U. Call 2
0
the
Euclidean distance of the compact set

V to the boundary of U,
we have
0
> 0, therefore the closed
0
-neighborhood

0
(

V ) = y R
k
[ x

V , |y x|
0

of

V is compact and contained in U.
243
If u : U R is a continuous function, then for <
0
, the
convolution
u

(x) =

u(x) =
_
R
k

(y)u(x y) dy.
makes sense and is of class C

on a neighborhood of

V . Moreover,
the family u

converges uniformly on

V to u, as t 0.
Lemma 8.3.2. Under the hypothesis above, suppose that u :
U R is a locally Lipschitz function. Given any Hamiltonian
H : T

U R quasi-convex in the bers and any > 0, for


every > 0 small enough, we have sup
xV
[u

(x) u(x)[ and


H
V
(u

) H
U
(u) +.
Proof. Because u is locally Lipschitz the derivative d
z
u exists for
almost every z U. We rst show that, for <
0
, we must have
x V, d
x
u

=
_
R
k

(y)d
xy
udy. (*)
In fact, since u

is C

, it suces to check that


lim
t0
u

(x +th) u

(x)
t
=
_
R
k

(y)d
xy
u(h) dy, (**)
for x V, <
0
, and h R
k
. Writing
u

(x +th) u

(x)
t
=
_
R
k

(y)
u(x +th y) u(x y)
t
dy,
We see that we can obtain (**) from Lebesgues dominated con-
vergence theorem, since

has a compact support contained in


y R
k
[ |y| < , and for y R
k
, t R such that |y| <
, |th| <
0
, the two points x + th y, x y are contained
in the compact set

N

0
(

V ) on which u is Lipschitz. Equation (*)


yields
H(x, d
x
u

) = H(x,
_
R
k

(y)d
xy
udy). (***)
Since

N

0
(

V ) is compact and contained in U, and u is locally


Lipschitz, we can nd K < such that |d
z
u| K, for each
z

N

0
(

V ) for which d
z
u exists. Since H is continuous, by a
244
compactness argument, we can nd

]0,
0
[, such that for z, z

0
(

V ), with |z z

, and |p| K, we have [H(z

, p)
H(z, p)[ . If

, since

(y) = 0, if |y| , we deduce that


for all x in V and almost every y with |y| , we have
H(x, d
xy
u) H(x y, d
xy
u) + H
U
(u) +.
Since H is quasi-convex in the bers, and

dy is a probability
measure whose support is contained in y [ |y| , we can now
apply proposition 1.5.6 to obtain

, H(x,
_
R
k

(y)d
xy
udy)H
U
(u) +.
It from inequality (***) above that H(x, d
x
u

) H
U
(u) + , for

and x V . This gives H


V
(u

) H
U
(u) +, for

. The
inequality sup
xV
[u

(x) u(x)[ < also holds for every small


enough, since u

converges uniformly on

V to u, as t 0.
Proof of theorem 8.3.1. We have to prove that for each x
0
M,
there exists an open neighborhood V of x
0
such that u
|V
is a
viscosity subsolution of H(x, d
x
u) on V . In fact, if we take V
any open neighborhood such that

V is contained in a domain of a
coordinate chart, we can apply lemma 8.3.2 to obtain a sequence
u
n
: V R, n 1, of C

functions such that u


n
converges
uniformly to u
|V
on V and H(x, d
x
u
n
) c + 1/n. If we dene
H
n
(x, p) = H(x, p) c 1/n, we see that u
n
is a smooth classical,
and hence viscosity, subsolution of H
n
(x, d
x
w) = 0 on V . Since H
n
converges uniformly to H c, the stability theorem 8.1.1 implies
that u
|V
is a viscosity subsolution of H(x, d
x
u) c = 0 on V .
Corollary 8.3.3. Suppose that the Hamiltonian H : T

M R is
continuous and quasi-convex in the bers. For every c R, the set
of Lipschitz functions u : M R which are viscosity subsolutions
of H(x, d
x
u) = c is convex.
Proof. If u
1
, . . . , u
n
are such viscosity subsolutions. By 7.4.3, we
know that at every x where d
x
u
j
exists we must have H(x, d
x
u
j
)
c. If we call A the set of points x where d
x
u
j
exists for each j =
1, . . . , n, then A has full Lebesgue measure in M. If a
1
, . . . , a
n
0,
and a
1
+ +a
n
= 1, then u = a
1
u
1
+ +a
n
u
n
is dierentiable
245
at each point of x A with d
x
u = a
1
d
x
u
1
+ +a
n
d
x
u
n
. There-
fore by the quasi-convexity of H(x, p) in the variable p, for every
x A, we obtain H(x, d
x
u) = H(x, a
1
d
x
u
1
+ + a
n
d
x
u
n
)
max
n
i=1
H(x, d
x
u
i
) c. Since A is of full measure, by theo-
rem 8.3.1, we conclude that u is also a viscosity subsolution of
H(x, d
x
u) = c.
The next corollary shows that the viscosity subsolutions are
the same as the very weak subsolutions, at least in the geometric
cases we have in mind. This corollary is clearly a consequence of
theorems 7.5.2 and 8.3.1.
Corollary 8.3.4. Suppose that the Hamiltonian H : T

M R is
continuous, coercive, and quasi-convex in the bers. A continuous
function u : M R is a viscosity subsolution of H(x, d
x
u) = c, for
some c R if and only if u is locally Lipschitz and H(x, d
x
u) c,
for almost every x M.
We now give a global version of lemma 8.3.2.
Theorem 8.3.5. Suppose that H : T

M R is a Hamiltonian,
which is quasi-convex in the bers. Let u : M R be a locally
Lipschitz viscosity subsolution of H(x, d
x
u) = c on M. For every
couple of continuous functions , : M ]0, +[, we can nd
a C

function v : M R such that [u(x) v(x)[ (x) and


H(x, d
x
v) c +(x), for each x M.
Proof. We endow M with an auxiliary Riemannian metric. We
pick up a locally nite countable open cover (V
i
)
iN
of M such
that each closure

V
i
is compact and contained in the domain U
i
of
a chart which has a compact closure

U
i
in M. The local niteness
of the cover (V
i
)
iN
and the compactness of

V
i
imply that the set
J(i) = j N [ V
i
V
j
,= is nite. Therefore, denoting by #A
for the number of elements in a set A, we obtain
j(i) =#J(i) = #j N [ V
i
V
j
,= < +,

j(i) = max
J(i)
j() < +.
We dene R
i
= sup
x

U
i
|d
x
u|
x
< +, where the sup is in fact
taken over the subset of full measure of x U
i
where the locally
246
Lipschitz function u has a derivative. It is nite because

U
i
is
compact. Since J(i) is nite, the following quantity

R
i
is also
nite

R
i
= max
J(i)
R

< +.
We now choose (
i
)
iN
a C

partition of unity subordinated to


the open cover (V
i
)
iN
. We also dene
K
i
= sup
xM
|d
x

i
|
x
< +,
which is nite since
i
is C

with support in V
i
which is relatively
compact.
Again by compactness, continuity, and niteness routine argu-
ments the following numbers are > 0

i
= inf
x

V
i
(x) > 0,

i
= min
J(i)

> 0

i
= inf
x

V
i
(x) > 0,
i
= min
J(i)

> 0.
Since

V
i
is compact, the subset (x, p) T

M [ x

V
i
, |p|
x

R
i
+1 is also compact, therefore by continuity of H, we can nd

i
> 0 such that
x

V
i
, p, p

x
M,|p|
x


R
i
+ 1, |p

|
x

i
, H(x, p) c +

i
2
H(x, p +p

) c +
i
.
We can now choose
i
> 0 such that

j(i)K
i

i
< min
J(i)

.
Noting that H(x, p) and |p|
x
are both quasi-convex in p, and
that

V
i
is compact and contained in the domain U
i
of a chart, by
lemma 8.3.2, for each i N, we can nd a C

function u
i
: V
i
R
such that
x V
i
, [u(x) u
i
(x)[ min(

i
,
i
),
H(x, d
x
u
i
) sup
zV
i
H(z, d
z
u) +

i
2
c +

i
2
|d
x
u
i
|
x
sup
zV
i
|d
z
u|
z
+ 1 = R
i
+ 1,
where the sup in the last two lines is taken over the set of points
z V
i
where d
z
u exists.
247
We now dene v =

iN

i
u
i
, it is obvious that v is C

. We
x x M, and choose i
0
N such that x V
i
0
. If
i
(x) ,= 0
then necessarily V
i
V
i
0
,= and therefore i J(i
0
). Hence

iJ(i
0
)

i
(x) = 1, and v(x) =

iJ(i
0
)

i
(x)u
i
(x).We can now
write
[u(x) v(x)[

iJ(i
0
)

i
(x)[u(x) u
i
(x)[

iJ(i
0
)

i
(x)

iJ(i
0
)

i
(x)
i
0
=
i
0
(x).
We now estimate H(x, d
x
u). First we observe that

iJ(i
0
)

i
(y) =
1, and v(x) =

iJ(i
0
)

i
(y)u
i
(y), for every y V
i
0
. Since V
i
0
is a
neighborhood of x, we can dierentiate to obtain

iJ(i
0
)
d
x

i
=
0, and
d
x
v =

iJ(i
0
)

i
(x)d
x
u
i
. .
p(x)
+

iJ(i
0
)
u
i
(x)d
x

i
. .
p

(x)
.
Using the quasi-convexity of H in p, we get
H(x, p(x)) max
iJ(i
0
)
H(x, d
x
u
i
) max
iJ(i
0
)
c +

i
2
c +

i
0
2
, (*)
where for the last inequality we have used that i J(i
0
) means
V
i
V
i
0
,= , and therefore i
0
J(i), which implies
i

i
0
, by the
denition of
i
.
In the same way, we have
|p(x)|
x
max
iJ(i
0
)
)|d
x
u
i
|
x
max
iJ(i
0
)
R
i
+ 1

R
i
0
+ 1. (**)
We now estimate |p

(x)|
x
. Using

iJ(i
0
)
d
x

i
= 0, we get
p

(x) =

iJ(i
0
)
u
i
(x)d
x

i
=

iJ(i
0
)
(u
i
(x) u(x))d
x

i
.
Therefore
|p

(x)|
x
= |

iJ(i
0
)
(u
i
(x) u(x))d
x

i
|
x
=

iJ(i
0
)
[u
i
(x) u(x)[|d
x

i
|
x

iJ(i
0
)

i
K
i
. (***)
248
From the denition of
i
, we get K
i

i


i
0
j(i
0
)
, for all i J(i
0
).
Hence |p

(x)|
x


iJ(i
0
)

i
0
j(i
0
)
=
i
0
. The denition of
i
0
, to-
gether with the inequalities (*), (**) and (***), above implies
H(x, d
x
v) = H(x, p(x) +p

(x)) c +
i
0
c +(x).
Theorem 8.3.6. Suppose H : T

M R is a Hamiltonian quasi-
convex in the bers. Let u : M R be a locally Lipschitz
viscosity subsolution of H(x, d
x
u) = c which is strict at every
point of an open subset U M. For every continuous function
: U ]0, +[, we can nd a viscosity subsolution u

: M R of
H(x, d
x
u) = c such that u = u

on MU, [u(x)u

(x)[ (x), for


every x M, and the restriction u
|U
is a C

with H(x, d
x
u) < c
for each x U.
Proof. We dene : M R by (x) = min((x), d(x, M U)
2
),
for x U, and (x) = 0, for x / U. It is clear that is continuous
on M and > 0 on U.
For each x U, we can nd c
x
< c, and V
x
V an open
neighborhood of x such that H(y, d
y
u) c
x
, for almost every
y V
x
. The family (V
x
)
xU
is an open cover of U, therefore we
can nd a locally nite partition of unity (
x
)
xU
on U submitted
to the open cover (V
x
)
xU
. We dene : U ]0, +[ by (g) =

xU

x
(y)(c c
x
), for y U. It is not dicult to check that
H(y, d
y
u) c (y) for almost every y U.
We can apply theorem 8.3.5 to the Hamiltonian

H : T

U
R dened by

H(y, p) = H(y, p) + (y) and u[U which satises

H(y, d
y
u) c for almost every y U, we can therefore nd
a C

function u

: U R, with [u

(y) u(y)[ (y), and

H(y, d
y
u

) c + (y)/2, for each y U. Therefore, we obtain


[u

(y) u(y)[ (y), and H(y, d


y
u

) c (y)/2 < c, for each


y U. Moreover, since (y) d(y, MU)
2
, it is clear that we can
extend continuously u

by u on M U. This extension satises


[u

(x) u(x)[ d(x, M U)


2
, for every x M. We must verify
that u

is a viscosity subsolution of H(x, d


x
u

) = c. This is clear
on U, since u

is C

on U, and H(y, d
y
u

) < c, for y U. It
remains to check that if : M R is such that u

with
equality at x
0
/ U then H(x
0
, d
x
0
) c. For this, we note that
u

(x
0
) = u(x
0
), and u(x) u

(x) d(x, M U)
2
d(x, x
0
)
2
.
Hence u(x) (x) + d(x, x
0
)
2
, with equality at x
0
. The function
249
x (x) + d(x, x
0
)
2
has a derivative at x
0
equal to d
x
0
, there-
fore H(x
0
, d
x
0
) c, since u is a viscosity solution of H(x, d
x
u)
c.
8.4 The viscosity semi-distance
We will suppose that H : T

M R is a continuous Hamiltonian
coercive above every compact subset of the connected manifold
M.
We dene c[0] as the inmum of all c R, such that H(x, d
x
u) =
c admits a global subsolution u : M R. This denition is
coherent with the one we gave in earlier chapters for particular
Hamiltonians.
As before we denote by oo
c
the set of viscosity subsolutions
of H(x, d
x
u) = c, and by oo
c
x
oo
c
the subset of subsolutions
vanishing at a given x M. Of course, since we can always add a
constant to a viscosity subsolution and still obtain a subsolution,
we have oo
c
x
,= if and only if oo
c
,= , and in that case oo
c
=
R+oo
c
x
.
Proposition 8.4.1. Under the above hypothesis, the constant c[0]
is nite and there exists a global u : M R viscosity subsolution
of H(x, d
x
u) = c[0].
Proof. Fix a point x M. Subtracting u( x) if necessary, we
will assume that all the viscosity subsolutions of H(x, d
u
) = c we
consider vanish at x. Since H is coercive above every compact
subset of M, for each c the family of functions in oo
c
x
is locally
equi-Lipschitzian, therefore
x M, sup
vSS
c
x
[v(x)[ < +,
since M is connected, and every v oo
c
x
vanish at x. We pick a
sequence c
n
c[0], and a sequence u
n
oo
cn
x
. Since, by Ascolis
theorem, the family oo
c
x
is relatively compact in the topology
of uniform convergence on each compact subset, extracting a se-
quence if necessary, we can assume that u
n
converges uniformly to
u on each compact subset of M. By the stability theorem 8.1.1,
the function u is a viscosity subsolution of H(x, d
x
u) = c
n
, for each
250
n, it is therefore locally Lipschitz. Pick a point where d
x
0
u exists,
we have H(x
0
, d
x
0
u) c
n
, for each n, therefore c[0] H(x
0
, d
x
0
u)
has to be nite. Again by the stability theorem 8.1.1, the function
u is a viscosity subsolution of H(x, d
x
u) = c[0].
For c c[0], we dene
S
c
(x, y) = sup
uSS
c
u(y) u(x) = sup
uSS
c
x
u(y).
It follows from the 8.2.2, that for each x M the function S
c
(x, .)
is a viscosity subsolution of H(y, d
y
u) = c on M itself, and a
viscosity solution on M x.
Theorem 8.4.2. For each c c[0], the function S
c
is a semi-
distance, i.e. it satises
(i) for each x M, S
c
(x, x) = 0,
(ii) for each x, y, z M, S
c
(x, z) S
c
(x, y) +S
c
(y, z)
Moreover, for c > c[0], the symmetric semi-distance,

S
c
(x, y) =
S
c
(x, y)+S
c
(y, x) is a distance which is locally Lipschitz-equivalent
to any distance coming from a Riemannian metric.
Proof. The fact that S
c
is a semi-distance follows easily from the
denition
S
c
(x, y) = sup
uSS
c
u(y) u(x).
Fix a Riemannian metric on the connected manifold M whose
associated norm is denoted by ||, and associated distance is d.
Given a compact subset K M, the constant sup|p| [ x
K, p T
x
M, H(x, p) c, is nite since H is coercive above com-
pact subsets of M. It follows from this that for each compact
subset K M, there exists a constant L
K
< such that.
x, y K, S
c
(x, y) L
K
d(x, y).
It remains to show a reverse inequality for c > c[0]. Fix such a c,
and a compact set K M. Choose > 0, such that

N

(K) =
x M [ d(x, K) is also compact. By the compactness of
the set
(x, p) [ x

N

(K), H(x, p) c[0],


251
and the continuity of H, we can nd > 0 such that
x

N

(K), p, p

T
x
M,H(x, p) c[0] and |p

|
H(x, p +p

) c.
(*)
We can nd
1
> 0, such that the radius of injectivity of the
exponential map, associated to the Riemannian metric, is at least

1
at every point x in the compact subset

N

(K). In particular, the


distance function x d(x, x
0
) is C

on

B(x
0
,
1
) x
0
, for every
x
0


N

(K). The derivative of x d(x, x


0
) at each point where
it exists has norm 1, since this map has (local) Lipschitz constant
equal to 1. We can assume
1
< . We now pick : R R a
C

function, with support in ]1/2, 2[, and such that (1) = 1. If


x
0
K and 0 < d(y, x
0
)
1
/2, the function

y
(x) = (
d(x, x
0
)
d(y, x
0
)
)
is C

. In fact, if d(x, x
0
)
1
, then
y
is zero in a neighborhood of
x, since d(x, x
0
)/d(y, x
0
)
1
/(
1
/2) = 2; if 0 < d(x, x
0
) <
1
< ,
then it is C

on a neighborhood of x; nally
y
(x) = 0 for x such
that d(x, x
0
) d(y, x
0
)/2. In particular, we obtained that d
x

y
=
0, unless 0 < d(x, x
0
) < , but at each such x, the derivative of
z d(z, x
0
) exists and has norm 1. It is then not dicult to see
that sup
xM
|d
x

y
| A/d(y, x
0
), where A = sup
tR
[

(t)[.
Therefore if we set = d(y, x
0
)/A, we see that |d
x

y
| ,
for x M. Since is 0 outside the ball B(x
0
,
1
) N

1
(K), it
follows from the property (*) characterizing that we have
(x, p) T

M, H(x, p) c[0] H(x, p +d


x

y
) c.
Since S
c[0]
(x
0
, ) is a viscosity subsolution of H(x, d
x
u) = c[0],
and
y
is C

, we conclude that the function u(.) = S


c[0]
(x
0
, .) +

y
(.) is a viscosity subsolution of H(x, d
x
u) = c. But the value
of u at x
0
is 0, and its value at y is S
c[0]
(x
0
, y) +
y
(y) =
S
c[0]
(x
0
, y) + d(y, x
0
)/A, since
y
(y) = (1) = 1. Therefore
S
c
(x
0
, y) S
c[0]
(x
0
, y) +d(y, x
0
)/A. Hence we obtained
x, y K, d(x, y)
1
/2 S
c
(x, y) S
c[0]
(x, y) +A
1
d(x, y).
252
Adding up and using S
c[0]
(x, y) + S
c[0]
(y, x) S
c[0]
(x, x) = 0, we
get
x, y K, d(x, y)
1
/2 S
c
(x, y) +S
c
(y, x)
2
A
d(x, y).
8.5 The projected Aubry set
Theorem 8.5.1. Assume that H : T

M R is a Hamiltonian
coercive above every compact subset of the connected manifold M.
For each c c[0], and each x M, the following two conditions
are equivalent:
(i) The function S
c
(x, ) is a viscosity solution of H(z, d
z
u) = c.
(ii) There is no viscosity subsolution of H(z, d
z
u) = c which is
strict at x.
Proof. The implication (ii)(i) follows from proposition 8.2.8.
To prove (i)(ii), x x M such that S
x
is a viscosity solution
on the whole of M, and suppose that u : M R is a viscosity
subsolution of H(y, d
y
u) = c which is strict at x. Therefore we
can nd an open neighborhood V
x
of x, and a c
x
< c such that
u
|Vx
is a viscosity subsolution of H(y, d
y
u) = c
x
on V
x
. We can
assume without loss of generality that V
x
is an open subset of R
n
and u(x) = 0. We have
u(y) S(x, y)
and u(x) = S(x, x) = 0. On V
x
R
n
, we can dene u
1
(y) =
u(y)
1
2
|x y|
2
. Dene () > 0 by
() = max
xy
H(y, p +p

) c
x
[ H(y, p) c
x
|p

| .
Since H is continuous and coercive above compact subsets we have
() 0, when 0. Since the derivative at y
0
of y
1
2
|y x|
2
is y
0
x, ), we see that u
1
|

B(x,)
is a viscosity sub solution of
H(y, d
y
u
1
) = c
x
+ (). We x > 0 such that c
x
+ 2() <
c. By 8.2.5, we can nd a real-valued function u
2
dened on a
253
neighborhood of

B(x, /2) and semi-convex such that u
2
is as close
as we want to u
1
on

B(x/2), and u
2
is a viscosity subsolution of
H(x, d
x
u
2
) = c
x
+2(). In a neighborhood of

B(0, /2). We have
u
1
(x) = S(x, x) = 0 and u
1
(y) u(y)
1
2
|xy|
2
u(y) S(x, y)
hence S(x, y) u
1
(y)
1
2
on the boundary B(x, /2). We can
therefore choose u
2
close enough to u
1
so that S(x, )u
2
() attains
its minimum on

B(x, /2) at a point y
0


B(x, /2). Therefore
S(x, y) S(x, y
0
) u
2
(y
0
) + u
2
(y) in a neighborhood of y
0
, and
therefore D

u
2
(y
0
) D

S
x
(y
0
). Since u
2
is semi-convex and is a
viscosity subsolution of H(y, dyu
2
) = c
x
+2() on a neighborhood
of

B(x, /2), by an argument analogous to the proof of theorem
8.2.4, we can nd p
0
D

u
2
(x
0
) with H(y
0
, p
0
) c
x
+ 2().
Since D

u
2
(x
0
) D

S
x
(y
0
), and S
x
is a viscosity solution of
H(y, d
y
S
x
) = c on M, we must have H(y
0
, p
0
) c. This is a
contradiction since c > c
x
+ 2().
Denition 8.5.2 (Projected Aubry set). If H : T

M R is a
continuous Hamiltonian, coercive above every compact subset of
the connected manifold M. We dene the projected Aubry set as
the set of x M such that that S
c[0]
(x, ) is a viscosity solution of
H(z, d
z
u) = c[0].
Proposition 8.5.3. Assume that H : T

M R s a continuous
Hamiltonian, convex is the bers, and coercive above every com-
pact subset of the connected manifold M. There exists a viscosity
subsolution v : M R of H(x, d
x
v) = c[0], which is strict at every
x M /.
Proof. We x some base point x M. For each x / /, we can nd
u
x
: M R, an open subset V
x
containing x, and c
x
< c[0], such
that u
x
is a viscosity subsolution of H(y, d
y
u
x
) = c[0] on M, and
u
x
[V
x
is a viscosity subsolution of H(y, d
y
u
x
) c
x
, on V
x
. Sub-
tracting u
x
( x) if necessary, we will assume that u
x
( x) = 0. Since
U = M / is covered by the family of open sets V
x
, x / /, we
can extract a countable subfamily (V
x
i
)
iN
covering U. Since H is
coercive above every compact set the sequence (u
x
i
)
iN
is locally
equi-Lipschitzian. Therefore, since M is connected, and all the u
x
i
vanish at x, the sequence (u
x
i
)
iN
is uniformly bounded on every
compact subset of M. It follows that the sum V =

iN
1
2
i+1
u
x
i
is uniformly convergent on each compact subset. If we set u
n
=
254
(1 2
(n+1)
)
1

0in
1
2
i+1
u
x
i
, then u
n
is a viscosity subsolution
of H(x, d
x
u
n
) = c[0] as a convex combination of viscosity subso-
lutions, see proposition 8.3.3. Since u
n
converges uniformly on
compact subsets to u, the stability theorem 8.1.1 implies that v is
also a viscosity subsolution of H(x, d
x
v) = c[0].
On the set V
xn
0
, we have H(x, d
x
u
xn
0
) c
xn
0
, for almost every
x V
xn
0
. Therefore, if we x n n
0
, we see that for almost every
x V
xn
0
we have
H(x, d
x
u
n
) (1 2
(n+1)
)
1
n

i=0
1
2
i+1
H(x, d
x
u
x
i
)
(1 2
(n+1)
)
1
_
n

i=0
1
2
i+1
c[0] +
(c
xn
0
c[0])
2
n
0
+1
_
.
Therefore u
n
[V
xn
0
is a viscosity subsolution of H(x, d
x
u
n
) c[0] +
(c
xn
0
c[0])/2
n
0
+1
.
By the stability theorem, this is also true for v[V
xn
0
. Since
c
xn
0
c[0] < 0, we conclude that u[V
xn
0
is a strict subsolution
of H(x, d
x
v) = c[0], for each x V
xn
0
, and therefore at each
x U
nN
V
xn
.
Theorem 8.5.4. Assume that H : T

M R is a Hamiltonian
convex in the bers and coercive, where M is a compact connected
manifold. Its projected Aubry set / is not empty.
If two viscosity solutions of H(x, d
x
u) = c[0] coincide on /,
they coincide on M.
Theorem 8.5.5. Suppose u
1
, u
2
: M R are respectively a
viscosity subsolution and a viscosity supersolution of H(x, d
x
u) =
c[0]. If u
1
u
2
on the projected Aubry set /, then u
1
u
2
everywhere on M.
Proposition 8.5.6. Assume that H : T

M R is a Hamiltonian
convex in the bers and coercive, where M is a compact connected
manifold. If M is compact and connected, for each viscosity sub-
solution u : M R of H(x, d
x
u) = c[0], and each > 0, we can
nd a viscosity subsolution u

: M Rof H(x, d
x
u

) = c[0] such
that |uu

< , and u

is C

on M/, with H(x, d


x
u

) < c[0],
for each x M /.
255
Proof. Call v the strict subsolution given by the previous propo-
sition 8.5.3. By a similar argument to the one used in the proof
of that proposition v

= (1 )u +v is a viscosity subsolution of
H(x, d
x
v

) = c[0] which is strict a each point of M/, and v

u
uniformly as 0. It then suces to choose small enough and
to apply 8.3.6 to v

to obtain the function u

.
Proof of theorem 8.5.5. Assume m = inf(u
2
u
1
) < 0. Choose
> 0 such that m + 2 < 0. If we apply proposition 8.5.6, we
obtain u
1
: M R, with | u
1
u
1
|

< , and u
1
of class C

on
M /, with H(x, d
x
u
1
) < c[0], for every x / /. We have u
2
(x)
u
1
(x) u
2
(x)u
1
(x)+ u
1
(x)u
1
(x) u
2
(x)u(x), therefore
u
2
(x) u
1
(x) , for x /. Moreover, inf(u
2
u
1
) inf(u
2

u
1
)+|u
1
u
1
|

m+. Since m+ < , on the compact space


M, the inmum of (u
2
u
1
) is attained at a point x
0
/ /. Since
u
2
(x) [u
2
(x
0
) u
1
(x
0
)] + u
1
(x), with equality at x
0
, the function
u
1
is dierentiable on M / x
0
, and u
2
is a supersolution
of H(x, d
x
u
2
) = c[0], we must have H(x, d
x
u
1
) c[0]. This is
impossible by the choice of u
1
.
8.6 The representation formula
We still assume that M is compact, and that H : T

M R is a
coercive Hamiltonian convex in the bers.
Theorem 8.6.1. Any viscosity solution u : M Rfor H(x, d
x
u) =
c[0] satises
x M, u(x) = inf
x
0
A
u(x
0
) +S
c[0]
(x
0
, x)
.
This theorem follows easily from the uniqueness theorem 8.5.4
and the following one:
Theorem 8.6.2. For any function v : / R bounded below, the
function
v(x) = inf
x
0
A
v(x
0
) +S
c[0]
(x
0
, x)
is a viscosity solution of H(x, d
x
v) = c[0]. Moreover, we have
v
|A
= v, if and only if
x, y /, v(y) v(x) S(x, y).
256
We start with a lemma.
Lemma 8.6.3. Suppose H : T

M R is a continuous Hamilto-
nian convex in the bers, and coercive above each compact subset
of the connected manifold M. Let u
i
: M R, i I be a fam-
ily of viscosity subsolutions of H(x, d
x
u) = c. If inf
iI
u
i
(x
0
), is
nite for some x
0
M, then inf
iI
u
i
is nite everywhere. In
that case, the function u = inf
iI
u
i
is a viscosity subsolution of
H(x, d
x
u) c.
Proof. We choose an auxiliary Riemannian metric on M, and use
the associated distance.
By the coercivity condition, the family (u
i
)
iI
is locally equi-
Lipschitzian, therefore for if K compact connected subset of M,
there exists a constant C(K) such that
x, y K, i I, [u
i
(x) u
i
(y)[ C(K).
If x M is given, we can nd a compact connected subset K
x
containing x
0
and x, it follows that
inf
iI
u
i
(x
0
) inf
iI
u
i
(x) +C(K
x
)
therefore inf
iI
u
i
is nite everywhere. It now suces to show that
for a given x M, we can nd an open neighborhood V of x such
that inf
iI
u
i
[V is a viscosity subsolution of H(x, d
x
u) = c on V .
We choose an open neighborhood V of x such that its closure

V is
compact. Since (
0
(

V , R) is metric and separable in the topology


of uniform convergence, we can nd a countable subset I
0
I
such that u
i
|

V
, i I
0
is dense in u
i
|

V
[ i I, for the topology of
uniform convergence. Therefore inf
iI
u
i
= inf
iI
u
i
= inf
iI
0
u
i
on

V . Since I
0
is countable, we have reduced to the case I
0
=
0, , N, or I
0
= N.
Let us start with the rst case. Since u
0
, , u
N
, and u =
inf
N
i=0
u
i
are all Lipschitzian on V , we can nd E V of full
Lebesgue measure such that d
x
u, d
x
u
0
, , d
x
u
N
exists, for each
x E. At each such x E, we necessarily have d
x
u d
x
u
0
, . . . , d
x
u
N
.
In fact, if n is such that u(x) = u
n
(x), since u u
n
with equality
at x and both derivative at x exists, they must be equal. Since
each u
i
is a viscosity subsolution of H(x, d
x
v) = c, we obtain
257
H(x, d
x
u) c, for every x in the subset E of full measure in V .
The convexity of H in the bers imply that u is a viscosity sub-
solution of H(x, d
x
u) = c in V . It remains to consider the case
I
0
= N. Dene u
N
(x) = inf
0iN
u
i
(x), by the previous case, u
N
is a viscosity subsolution of H(x, d
x
u
N
) = c on V .
Now u
N
(x) inf
iI
0
u
i
(x), for each x

V , the convergence
is in fact, uniform on

V since (u
i
)
iI
0
is equi-Lipschitzian on the
compact set

V . It remains to apply the stability theorem 8.1.1.
258
Chapter 9
Ma nes Point of View
Ricardo Ma nes last paper [Mn97] contained a version of the weak
KAM theorem. The point of view is probably the closest to the
theory of optimal contral. His ideas after his untimely death were
carried out much further by G. Contreras, J. Delgado, R. Iturriaga,
Gabriel and Miguel Paternain [CDI97, CIPP98].
There is an excellent reference on Ma nes point of view and th
subsequent developments [CI99].
9.1 Ma nes potential
As in denition 5.3.1, we set
h
t
(x, y) = inf

_
t
0
L((s), (s)) ds
Where the inmum is taken over all continuous piecewise C
1
curves
: [0, t] M with (0) = x, (t) = y.
Denition 9.1.1 (Ma nes potential). Fix c R for each x, y,
M, we set
m
c
(x, y) = inf
t>0
h
t
(x, y) +ct
Here are some properties of m
c
:
Proposition 9.1.2. For each c R, the Ma ne potential m
c
has
values in R , and satises the following properties:
259
260
(i) If x, u M, and c, c

R, with c c

, we have m
c
(x, y)
m
c
(x, y).
(ii) For all c R, y, z M, we have
m
c
(x, z) m
c
(x, y) +m
c
(x, z).
(iii) If A = supL(x, v) [ (x, v) TM, |v|
x
1 we have
m
c
(x, y) (A +c)d(x, y).
(iv) For a given c R either m
c
is equal identically to or
m
c
is nite everywhere.
(v) For every c R, either m
c

or m
c
(x, x) = 0 for every
x M.
(vi) If m
c
is nite then it is Lipschitz.
(vii) For u : M R, we have u L +c if and only if
x, y M, u(y) u(x) m
c
(x, y).
(viii) If m
c
is nite, then for each x M, the function m
c,x
: M
R, y m
c
(x, y) (resp. m
x
c
: M R, y m
c
(y, x)) is
dominated by L +c.
(ix) The Ma ne critical value c[0] is equal to the inmum of the
set of c R such that m
c
is nite. Moreover, the critical
Ma ne potential m
0
= m
c[0]
is nite everywhere.
Proof. Property (i) is obvious. Property 3 (ii) results from
h
t
(x, z) h
t
(x, y) = h
t
(y, z).
For property (iii), if we use a geodesic
x,y
: [0, d(x, y)] M
from x to y parametrized by arc-length, we see that m
c
(x, y)
h
d(x,y)
(x, y) + cd(x, y) L(
x,y
) + cd(x, y) (A + c)d(x, y). For
property (iv), we remark that m
c
(x

, y

) m
c
(x

, x) +m
c
(x, y) +
m
c
(y, y

) m
c
(x, y)+(A+c)[d(x

, x)+d(y

, y)] hence if m
c
(x, y) =
for some (x, y) M M then m
c
(x

, y

) = for every
(x

, y

) M M.
261
For property (v), using constant paths, we rst remark that
m
c
(x, x) (L(x, 0) + c)t, for every t > 0, therefore m
c
(x, x) 0.
Moreover, by (ii)
m
c
(x, x) m
c
(x, x) +m
c
(x, x) nm
c
(x, x).
Hence m
c
(x, x) < 0 implies m
c
(x, x) = .
Property (vi) follows from the proof of (iv), since we obtained
there
m
c
(x

, y

) m
c
(x, y) + (A+c)[d(x, x

) +d(y, y

)]
which gives by symmetry
[m
c
(x

, y

) m
c
(x, y)[ [A +c][d(x, x

) +d(y, y

)].
Property (vii) is obvious since u L +c if and only if
t > 0, u(y) u(x) h
t
(x, y) +ct.
For (viii), the inequality obtained in (ii)
m
c
(x, z)

S
c
(x, y) +m
c
(y, z)
gives, when m
c
is nite
m
c
(x, z) m
c
(x, y) m
c
(y, z).
But this can be rewritten as
m
c,x
(z) m
c,x
(y) m
c
(y, z),
therefore m
c,x
L +c by (vii).
For (ix), if c c[0], there exists u : M R with u L + c
therefore by (vii), we have m
c
nite.
Conversely if m
c
is nite m
c,x
L +c therefore c c[0].
Corollary 9.1.3. For each c c[0], the Ma ne potential m
c
is
equal to the viscosity semi-distance S
c
, and therefore
x, y M, m
c
(x, y) = supu(y) u(x) [ u L +c.
Proof. The function m
c,x
(resp. S
c
x
) is a viscosity subsolution of
H(x, d
u
) = c (resp. is dominated by L + c), therefore m
c
(x, y) =
m
c,x
(y) m
c,x
(x) S
c
(x, y) (resp. S
c
(x, y) = S
c
x
(y) S
c
x
(x)
m
c
(x, y)).
Denition 9.1.4 (Ma nes critical potential). We will call m
0
=
m
c[0]
the Ma ne critical potential.
262
9.2 Semi-static and static curves
Proposition 9.2.1. Given c R, a curve : [a, b] M is an
absolute (L +c)-minimizer if and only if
m
c
((a), (b)) =
_
b
a
L((s), (s)) ds +c(b a).
Proof. Suppose that : [a, b] M is an absolute (L+c)-minimizer,
then for any curve : [0, t] M, with t > 0, (0) = (a), and
(t) = (b), we have
_
t
0
L((s),

(s)) ds +ct
_
b
a
L((s), (s))ds +c(b a)
therefore h
t
((a), (b)) +ct
_
a
L((s), (s)) ds +c(b a).
On the other hand reparametrizing linearly by [0, b a], we
see that h
ba
((b), (a)) +c(b a)
_
b
a
L((s), (s)) ds+c(b a).
It follows that
m
c
((b), (a)) = h
ba
((b), (a)) +c(b a)
=
_
b
a
L((s), (s))ds +c(b a).
Conversely, since m
c
((a), (a)) = 0, the equality
m
c
((a), (b)) =
_
b
a
L((s), (s)) +ds +c(b a),
can be rewritten as
m
c,(a)
((b)) m
c,(a)
((a)) =
_
b
a
L((s), (s))ds +c(b a).
This means that is (m
c,(a)
, L, c)-calibrated.
Denition 9.2.2 (Semi-static curve). A curve : [a, b] M is
called semi-static, if a < b and m
0
((a), (b)) =
_
b
a
L((s), (s))ds+
c[0](b a). (Recall that m
0
= m
c[0]
is the Ma ne potential).
We therefore have the following proposition;
263
Proposition 9.2.3. A curve : [a, b] M semi-static if and
only if it is absolutely minimizing, if and only if it is (u, L, c[0])-
calibrated for some u : M R dominated by L +c[0].
Ma ne has also dened a notion of static curve.
Denition 9.2.4 (Static curve). A curve : [a, b] M is static,
if a < b and
_
b
a
L((s), (s)) ds +c[0](b a) = m
0
((b), (a)).
Proposition 9.2.5. A curve is static if and only if it is a part of
a projected Aubry curve
Proof. We have
0 =
_
b
a
L((s), (s)) ds +c[0](b a) +m
0
((b), (a)) = 0.
For every > 0, we can nd a curve

: [b, b

] M with

(b) =
(b),

(b

) = (a), and
_
b
b
L(

(s),

(s))ds +c[0](b

b) m
0
((b), (a)) +.
Therefore, if we consider the concatenated closed curve

,
we nd a curve

that is a loop at (a), is parametrized by an


interval of length

b a > 0 and satises L(

) + c[0]


. Going n times through the loop

/n
, we nd a loop

at
(a), parametrized by an interval of length n

/n
n(b a), with
L(

n,
) + c[0]n

/n
. Since b a > 0, we have n(b a) +
as n +. It follows that h((a), (a)) , for every > 0,
where h is the Peierls barrier. Therefore (a) /. Since the loop

goes through every point of ([a, b]) a similar argument


shows ([a, b]) /.
It remains to show that is (u, L, c[0])-calibrated for every
u : M R which is dominated by L +c[0].
In fact, if we add up the two inequalities
u((b)) u((a))
_
b
a
L((s)), (s) ds +c[0](b a)
264
u((a)) u((b)) m
0
((b), (a)),
we obtain the equality 0 = 0 therefore both inequalities above
must be equalities.
Bibliography
[Bar94] Guy Barles. Solutions de viscosite des equations de
Hamilton-Jacobi. Springer-Verlag, Paris, 1994.
[BCD97] Martino Bardi and Italo Capuzzo-Dolcetta. Opti-
mal control and viscosity solutions of Hamilton-Jacobi-
Bellman equations. Systems & Control: Foundations
& Applications. Birkhauser Boston Inc., Boston, MA,
1997. With appendices by Maurizio Falcone and Pier-
paolo Soravia.
[BGH98] Giuseppe Buttazo, Mariano Giaquinta, and Stefan
Hildebrandt. One-dimensional variational problems, An
introduction. Oxford University Press, Oxford, 1998.
Oxford Lecture Series in Mathematics and its Applica-
tions, Vol. 15.
[BGS85] Werner Ballmann, Mikhael Gromov, and Viktor
Schroeder. Manifolds of nonpositive curvature, vol-
ume 61 of Progress in Mathematics. Birkhauser Boston
Inc., Boston, MA, 1985.
[Bos86] Jean-Benot Bost. Tores invariants des syst`emes dy-
namiques hamiltoniens (dapr`es Kolmogorov, Arnold,
Moser, R ussmann, Zehnder, Herman, P oschel, . . .).
Asterisque, (133-134):113157, 1986. Seminar Bour-
baki, Vol. 1984/85.
[Bou76] Nicolas Bourbaki.

Elements de mathematique. Her-
mann, Paris, 1976. Fonctions dune variable reelle,
Theorie elementaire, Nouvelle edition.
265
266
[Bou81] Nicolas Bourbaki. Espaces vectoriels topologiques.
Chapitres 1 ` a 5. Masson, Paris, new edition, 1981.

Elements de mathematique. [Elements of mathematics].


[BS00] Guy Barles and P.E. Souganidis. On the long time be-
havior of Hamilton-Jacobi equations. J. Math. Analysis,
2000. to appear.
[Car95] M. J. Dias Carneiro. On minimizing measures of
the action of autonomous Lagrangians. Nonlinearity,
8(6):10771085, 1995.
[CC95] Luis A. Caarelli and Xavier Cabre. Fully nonlinear el-
liptic equations. American Mathematical Society, Prov-
idence, RI, 1995.
[CDI97] Gonzalo Contreras, Jorge Delgado, and Renato Itur-
riaga. Lagrangian ows: the dynamics of globally
minimizing orbits. II. Bol. Soc. Brasil. Mat. (N.S.),
28(2):155196, 1997.
[CEL84] M. G. Crandall, L. C. Evans, and P.-L. Lions. Some
properties of viscosity solutions of Hamilton-Jacobi
equations. Trans. Amer. Math. Soc., 282(2):487502,
1984.
[CHL91] Dean A. Carlsson, Alain B. Haurie, and Arie
Leizarowitz. Innite horizon optimal control: deter-
mistic and stochastic systems. Springer-Verlag, New
York Berlin Heidelberg, 1991. 2nd, rev. and enl. ed.
[CI99] Gonzalo Contreras and Renato Iturriaga. Global min-
imizers of autonomous Lagrangians. Instituto de
Matematica Pura e Aplicada (IMPA), Rio de Janeiro,
1999.
[CIPP98] G. Contreras, R. Iturriaga, G. P. Paternain, and M. Pa-
ternain. Lagrangian graphs, minimizing measures and
Ma nes critical values. Geom. Funct. Anal., 8(5):788
809, 1998.
267
[CL83] Michael G. Crandall and Pierre-Louis Lions. Viscosity
solutions of Hamilton-Jacobi equations. Trans. Amer.
Math. Soc., 277(1):142, 1983.
[Cla90] F. H. Clarke. Optimization and nonsmooth analysis. So-
ciety for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, second edition, 1990.
[DG82] James Dugundji and Andrzej Granas. Fixed point the-
ory. I. Pa nstwowe Wydawnictwo Naukowe (PWN),
Warsaw, 1982.
[Dug66] James Dugundji. Topology. Allyn and Bacon Inc.,
Boston, Mass., 1966.
[EG92] Lawrence C. Evans and Ronald F. Gariepy. Measure
theory and ne properties of functions. CRC Press,
Boca Raton, FL, 1992.
[Eva98] Lawrence C. Evans. Partial dierential equations, vol-
ume 19 of Graduate Studies in Mathematics. American
Mathematical Society, Providence, RI, 1998.
[Fat97a] Albert Fathi. Solutions KAM faibles conjuguees et
barri`eres de Peierls. C. R. Acad. Sci. Paris Ser. I Math.,
325(6):649652, 1997.
[Fat97b] Albert Fathi. Theor`eme KAM faible et theorie de
Mather sur les syst`emes lagrangiens. C. R. Acad. Sci.
Paris Ser. I Math., 324(9):10431046, 1997.
[Fat98a] Albert Fathi. Orbites heteroclines et ensemble de
Peierls. C. R. Acad. Sci. Paris Ser. I Math.,
326(10):12131216, 1998.
[Fat98b] Albert Fathi. Sur la convergence du semi-groupe de
Lax-Oleinik. C. R. Acad. Sci. Paris Ser. I Math.,
327(3):267270, 1998.
[Fat03] Albert Fathi. Regularity of C
1
solutions of the
Hamilton-Jacobi equation. Ann. Fac. Sci. Toulouse
Math. (6), 12(4):479516, 2003.
268
[Fle69] Wendell H. Fleming. The Cauchy problem for a nonlin-
ear rst order partial dierential equation. J. Dieren-
tial Equations, 5:515530, 1969.
[FM00] Albert Fathi and John N. Mather. Failure of the conver-
gence of the Lax-Oleinik semi-group in the time periodic
case. Bull. Soc. Math. France, 128, 2000.
[FS04] Albert Fathi and Antonio Siconol. Existence of C
1
critical subsolutions of the Hamilton-Jacobi equation.
Invent. Math., 155(2):363388, 2004.
[FS05] Albert Fathi and Antonio Siconol. PDE aspects of
Aubry-Mather theory for quasiconvex Hamiltonians.
Calc. Var. Partial Dierential Equations, 22(2):185
228, 2005.
[GK90] Kazimierz Goebel and W. A. Kirk. Topics in metric
xed point theory. Cambridge University Press, Cam-
bridge, 1990.
[Gro90] M. Gromov. Convex sets and K ahler manifolds. In
Advances in dierential geometry and topology, pages
138. World Sci. Publishing, Teaneck, NJ, 1990.
[Her89] Michael-R. Herman. Inegalites a priori pour des
tores lagrangiens invariants par des dieomorphismes
symplectiques. Inst. Hautes

Etudes Sci. Publ. Math.,
(70):47101 (1990), 1989.
[Kis92] Christer O. Kiselman. Regularity classes for operations
in convexity theory. Kodai Math. J., 15(3):354374,
1992.
[Kni86] Gerhard Knieper. Mannigfaltigkeiten ohne konjugierte
Punkte. Universit at Bonn Mathematisches Insti-
tut, Bonn, 1986. Dissertation, Rheinische Friedrich-
Wilhelms-Universit at, Bonn, 1985.
[Lio82] Pierre-Louis Lions. Generalized solutions of Hamilton-
Jacobi equations. Pitman (Advanced Publishing Pro-
gram), Boston, Mass., 1982.
269
[LPV87] Pierre-Louis Lions, George Papanicolaou, and S.R.S.
Varadhan. Homogenization of Hamilton-Jacobi equa-
tion. unpublished preprint, 1987.
[Mat91] John N. Mather. Action minimizing invariant mea-
sures for positive denite Lagrangian systems. Math.
Z., 207(2):169207, 1991.
[Mat93] John N. Mather. Variational construction of connecting
orbits. Ann. Inst. Fourier (Grenoble), 43(5):13491386,
1993.
[Min61] George J. Minty. On the maximal domain of a mono-
tone function. Michigan Math. J., 8:135137, 1961.
[Min64] George J. Minty. On the monotonicity of the gradient of
a convex function. Pacic J. Math., 14:243247, 1964.
[Mn] Ricardo Ma ne. Global Variational methods in conser-
vative dynamics. IMPA, Rio de Janeiro. Lecture notes,
18 Col oquio Brasileiro de Matematica.
[Mn97] Ricardo Ma ne. Lagrangian ows: the dynamics of glob-
ally minimizing orbits. Bol. Soc. Brasil. Mat. (N.S.),
28(2):141153, 1997.
[NR97a] Gawtum Namah and Jean-Michel Roquejore. Com-
portement asymptotique des solutions dune classe
dequations paraboliques et de Hamilton-Jacobi. C. R.
Acad. Sci. Paris Ser. I Math., 324(12):13671370, 1997.
[NR97b] Gawtum Namah and Jean-Michel Roquejore. Conver-
gence to periodic fronts in a class of semilinear parabolic
equations. NoDEA Nonlinear Dierential Equations
Appl., 4(4):521536, 1997.
[NR99] Gawtum Namah and Jean-Michel Roquejore. Remarks
on the long time behaviour of the solutions of Hamilton-
Jacobi equations. Comm. Partial Dierential Equa-
tions, 24(5-6):883893, 1999.
270
[Roq98a] Jean-Michel Roquejore. Comportement asymptotique
des solutions dequations de Hamilton-Jacobi monodi-
mensionnelles. C. R. Acad. Sci. Paris Ser. I Math.,
326(12):185189, 1998.
[Roq98b] Jean-Michel Roquejore. Convergence to steady states
of periodic solutions in a class of Hamilton-Jacobi equa-
tions. preprint, 1998.
[Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill
Book Co., New York, third edition, 1987.
[Rud91] Walter Rudin. Functional analysis. McGraw-Hill Inc.,
New York, second edition, 1991.
[RV73] A. Wayne Roberts and Dale E. Varberg. Convex func-
tions. Academic Press [A subsidiary of Harcourt Brace
Jovanovich, Publishers], New York-London, 1973. Pure
and Applied Mathematics, Vol. 57.
[Smi83] Kennan T. Smith. Primer of modern analysis. Springer-
Verlag, New York, second edition, 1983.
[WZ77] Richard L. Wheeden and Antoni Zygmund. Measure
and Integral. Marcel Dekker, Inc., New York and Basel,
1977. Monographs and textbooks in Pure and Applied
Mathematics, Vol. 43.
Index
D
+
u, 223
D

u, 223
H
U
(u), 124
Lip, 118
(L
1
, L

), 83

/
0
, 188

1
(u

,u
+
)
, 187

/
0
, 175

^
0
, 188
h
t
, 135
/
0
, 188
(
ac
([a, b], M), 88
1
(u

,u
+
)
, 186
/
0
, 175
o
+
, 173
o

, 118, 173
C
2
-strictly convex function, 2
classical solution, 213
Absolutely continuous Curve, 81,
88
Action, 38
Minimal, 135
Aubry, 253
projected set, 253
set, 253
Aubry Set, 188
Bounded Below, 89
Calibrated Curve, 115
Carneiro, 177
Conjugate Functions, 185
convex function, 1
Convex in the Fibers, 18
Curve
extremal, 41
Curve, Calibrated, 115
Defect of a Curve, 126
dierential
lower, 223
upper, 223
Dominated Function, 111
Dubois-Raymond, 42
Equation
Euler-Lagrange, 42
Erdmann, 44
Euler, 42, 49, 52, 53
Euler-Lagrange Equation, 42
Euler-Lagrange Flow, 53
Euler-Lagrange Theorem, 49
Euler-Lagrange Vector Field, 52,
53
Exposed Point, 31
Extremal Curve, 49
Extremal curve, 41
Extremal Point, 30
Fenchel, 12, 13
formula, 12
inequality, 12
271
272
transform, 12
Fenchel, Theorem, 13
First Variation Formula, 47, 50
Flow
Hamiltonian, 56
formula, Fenchel, 12
Function
Dominated, 111
function
quasi-convex, 28
function, convex, 1, 2
function, strictly convex, 1
function, C
2
-strictly convex, 2
Functions, Conjugate, 185
Global Legendre transform, 39
Gradient, Lagrangian, 101
Hamilton, 110
Hamilton-Jacobi, 57, 110
Hamilton-Jacobi equation, 213,
214
Hamiltonian, 56, 58
Hamiltonian Constant, 124
Hamiltonian Flow, 56
inequality, Fenchel, 12
Jacobi, 110
Lagrange, 42, 49, 52, 53
Lagrangian, 37, 98
non-degenerate, 38
Symmetrical, 161
Tonelli, 98
Lagrangian Gradient, 101
Lagrangian Submanifold, 57
Lagrangian Subspace, 57
Lagrangian, Reversible, 181
Legendre, 20, 39
Legendre Transform, 20
Legendre transform
global, 39
Lemma
Dubois-Raymond, 42
Erdmann, 44
Linear Supporting Form, 7
Lipschitz, 118
locally, 118
Lipschitz Constant, 118
Locally Lipschitz, 118
locally Lipschitz, 5
locally, Lipschitz, 5
Ma ne, 200
Ma ne Set, 188
Mather, 165, 175, 176
Mather Set, 175
Mathers Function, 165
Method of Characteristics, 76
Minimal Action, 135
Minimizer, 38, 98
Minimizing Measure, 160
Minty, 26
Non-degenerate Lagrangian, 38
Non-Expansive Map, 149
Odd Lagrangian Submanifold, 73
Odd Lagrangian Subspace, 71
Peierls Barrier, 190
Point
Exposed, 31
Extremal, 30
proper map, 23
quasi-convex, 28
Rademacher, 6
273
Reversible Lagrangian, 181
Solution
Weak KAM, 118
Straszewicz, 31
strict subsolution, 241
strictly convex function, 1
subsolution
strict, 241
Superlinear, 16
superlinear, 12
Supporting Form, Linear, 7
Symmetrical Lagrangian, 161
Symplectic Structure, 55
Theorem
Carneiro, 177
Euler-Lagrange, 49
Fenchel, 13
Krein-Milman, 30
Ma ne, 200
Mather, 165, 176
Minty, 26
Rademacher, 6
Straszewicz, 31
Tonelli, 95, 97, 98
Weak KAM, 153, 162
Tonelli, 98
Tonellis Theorem, 95, 97, 98
transform, Fenchel, 12
Transform, Legendre, 20
Variation of a curve, 40
Vector Field, Euler-Lagrange, 52,
53
very weak solution, 215
Viscosity
solution, 215, 216
subsolution, 215, 216
Weak KAM, 118
Weak KAM Theorem, 153, 162