Lagrangian Dynamics
Preliminary Version
Number 10
Albert FATHI
Lyon, Version 15 June 2008
ii
Contents
Preface vii
Introduction ix
0.1 The HamiltonJacobi 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 Quasiconvex 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 EulerLagrange 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 HamiltonJacobi 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 HamiltonJacobi 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 HamiltonJacobi 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 LaxOleinik Semigroup. . . . . . . . . . . . . 144
4.6 The LaxOleinik semigroup . . . . . . . . . . . . . 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 LaxOleinik semigroup 201
6.1 Semiconvex Functions . . . . . . . . . . . . . . . . 201
6.1.1 The Case of Open subsets of R
n
. . . . . . 201
6.2 The LaxOleinik Semigroup and Semiconvex Func
tions . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.3 Convergence of the LaxOleinik Semigroup . . . . 207
6.4 Invariant Lagrangian Graphs . . . . . . . . . . . . 210
7 Viscosity Solutions 213
7.1 The dierent forms of HamiltonJacobi 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 Quasiconvexity and viscosity subsolutions . . . . . 242
8.4 The viscosity semidistance . . . . . . . . . . . . . 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 Semistatic 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 dAubryMather, 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, 1522,
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 HamiltonJacobi
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 20022003 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 1620 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 semigroup 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), JeanChristophe Yoc
coz, Francis Clarke, Gabriel & Miguel Paternain, Gonzalo Con
treras, Renato Itturiaga, Guy Barles, JeanMichel Roquejore,
Ezequiel Maderna, Patrick Bernard, Italo CapuzzoDolcetta, 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 EulerLagrange 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 nondegenerate
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 EulerLagrange 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 EulerLagrange
ow using the theory of Dynamical Systems.
0.1 The HamiltonJacobi Method
A natural problem in dynamics is the search for subsets invari
ant by the ow
s
. Within the framework which concerns us the
HamiltonJacobi 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, dened by
L(x, v) = (x,
L
v
(x, v)),
is a dieomorphism of TM onto 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 (HamiltonJacobi). Let be a closed 1form 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 1form
0
, does there exist another
closed 1form 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
KolmogorovArnoldMoser 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 EulerLagrange 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 EulerLagrange
ow
t
. This is an instance of the socalled 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
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
) 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 nonempty 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 semicontinuous
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 semicontinuous 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 nonempty 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
 r/2. We
write y = tz + (1 t)y
= t(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
.
Denition 1.2.1 (Supporting Linear Form). We say that the
linear form p 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
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
= 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, nonempty, and convex in E R. Since (x
0
, f(x
0
)) is not
in O, by the HahnBanach 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
.
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) Kx + C(K), for all
x E.
When E is nitedimensional, 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
such that
p K, we have
p, v) L(v) p x Kx 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) = Kp, 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 nitedimensional linear space E, and H : E
such that
p, v) = H(p) +L(v).
15
Proof. We are assuming that E is nitedimensional, 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)[ pv, 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
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 nitedimensional 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
: 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 nitedimensional 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 nitedimensional 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) Kv
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 nitedimensional locally trivial vector
bundle : E X. We consider
: 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) pv (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
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
, Kv = sup
xA
K
DL(x)(v).
As L(v) DL(x)(v) +L(x) DL(x)(x) we see that
L(v) Kv + 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
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
, Kv = sup
xA
K
DL(x)(v).
As L(v) DL(x)(v) +L(x) DL(x)(x) we see that
L(v) Kv + 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 nitedimensional 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 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
) p
,
we rewrite it as
y R
n
DL(y)(v
) p
0
(v
) +p
.
It then remains to observe that p
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 quasiconvex.
2) Any monotonic function : I R, where I is an interval
in R, is quasiconvex.
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 nitedimensional 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
HahnBanach 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 nonempty,
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 nonempty, 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 quasiconvex
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 (KreinMilman). 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 KreinMilman 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 Rvector space E,
determines two open (resp. closed) halfspaces. If H is the set of
points where the ane function a : E R is 0, then the two
open (resp. closed) halfspaces 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 halfspaces 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
2y a
2
_
= y a +
d
2
2y 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
/[2y 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 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
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 halfspace 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 onedimensional 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
: T
M
M. A point of 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 (Nondegenerate Lagrangian). If L is a C
2
Lag
rangian on the manifold M, we say that L is nondegenerate if for
each (x, v) TM the second partial derivative
2
L/v
2
(x, v) is
nondegenerate 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 nondegenerate;
(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 nondegenerate, 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 nondegenerate 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
, b
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 (DuboisRaymond). 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
(t)A(t)(v
0
) dt and the function
(t)A(t)(v
0
) is continuous, nonnegative 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 EulerLagrange 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
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
1
(t) =
_
t
a
1
(s) C
0
(s) ds
is C
. ()
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 righthand 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
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
(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 EulerLagrange 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 EulerLagrange 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
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
, 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 <
, a
+] [b
, b
]) [
]. We can therefore
extend [[a
, b
] [
] to
[a, b] [
, 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 (EulerLagrange). 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
, b
, b
, b
] is also an extremal
curve. Since ([a
, b
, b
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
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 nondegenerate, 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 nondegeneracy 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
EulerLagrange 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 EulerLagrange 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 nondegenerate. 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 CauchyPeano 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 CauchyLipschitz theorem. Therefore, if
1
:
[[
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
EulerLagrange ow of the Lagrangian L.
54
2.5 Symplectic Aspects
Let M be dierentiable manifold of class C
, denote by
:
T
M
onto M and denote by T
: TT
M TM the derivative of
.
On T
M
is a linear map
(x,p)
: T
(x,p)
(T
: 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 1form 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
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
(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 1form dened on the open subset U of
M. This 1form 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 1form dened on the open
subset U of M. We have
= ,
where
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
is
n
i=1
i
(x)dx
i
= .
By taking, the exterior derivative of we can dene a sym
plectic structure on T
: 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
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
i=1
dp
i
dx
i
=
n
i=1
dx
i
dp
i
,
56
it is, then, easy to check the nondegeneracy 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 nondegenerate, 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
(F
) = p
E
= 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 1form on the manifold
M, then the graph Graph() of is a Lagrangian submanifold of
T
d =
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 (HamiltonJacobi). 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 CauchyPeano [Bou76], if N is
of class C
1
(or CauchyLipschitz, 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 EulerLagrange equa
tion, see 2.2.6.
Suppose that the restriction of the global Legendre transform
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
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 EulerLagrange
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 EulerLagrange vector eld X
L
is uniquely integrable on O, therefore the partial EulerLagrange
ow
L
t
is dened and C
1
on O. More generally, for every r 2,
the EulerLagrange 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 CauchyLipschitz, 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 nondegenerate C
r
Lagran
gian, with r 2, on the manifold M. Then for r = 2, the the
EulerLagrange 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 EulerLagrange 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 nondegenerate. Obviously, the 2form 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 EulerLagrange 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
, v
) = (k
, v
).
From (ii), it results that v
= v
. Since v
n
,= v
n
, we can set u
n
=
64
vnv
n
vnv
. This limit u
is also of norm 1. As v
n
i
and v
n
i
both converge to v
= 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
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, 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
_
(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 righthand side is obviously welldened for t = 0, and equal
to /t(k, 0, v) = v. Moreover, upon inspection of the righthand
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
(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 nondegenerate 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
) 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
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 HamiltonJacobi 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 2form 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 2form on the vector space
E is provided with the alternate bilinear 2form. 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 2form 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 2form 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 2form 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
) = dimF
1, since
Ker(a)
#
= Ker(a), where a
#
: E E
, a(, ). But a
#
(F
)
E
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 1form
H
on O R by
H
= Hdt,
where is the Liouville form on T
H
= p
1
(H p
1
) dt,
where p
1
: T
M R T
MR of the projection T
MR
R on the second factor. The exterior derivative
H
= d
H
denes a dierential 2form 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
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 1form.
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 CauchyLipschitz theo
rem. Instead as in the proof of 2.5.10, if (x, p, t) V , we apply the
CauchyPeano [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
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 1form 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 HamiltonJacobi equation
S
t
+H(x,
S
x
) = 0. (HJ)
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 HamiltonJacobi 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 HamiltonJacobi 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
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
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 HamiltonJacobi 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
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
oddLagrangian
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 1form
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 HamiltonJacobi 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
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
)
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
) 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
(s) ds
_
A
(s) ds.
We now x an integer big enough to satisfy

_
U
(s) ds
_
A
(s)ds ,
and m(U
\A
n
(s) ds
_
U
\A

n
(s)ds .
Since 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 wellknown 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 Lipschitzequivalent 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
, 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
]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
, b
], we see that
n, L(
n
[[a
, b
]) L(
n
) C
0
[(b a) (b
)].
It follows
[a
, b
, 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
Aw + (C
2
C
1
) +C
1
= Aw +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
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) Av +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
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
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 semicontinuous 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) +Av
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).
If > 0 is given, we choose A such that [CC(0)(ba)]/A /2.
It follows that for
K,C
and t, 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
, 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 dballs 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 nondegenerate 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 EulerLagrange 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 welldened 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 EulerLagrange 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 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 nondegenerate, 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 EulerLagrange 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
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
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 HamiltonJacobi and Minimizers
Theorem 3.5.1 (Lagrangian Gradient and HamiltonJacobi). Let
L : TM R be a C
r
Tonelli Lagrangian. If S :]a, b[U R is a
C
1
solution of the HamiltonJacobi equation
S
t
+H(x,
S
x
) = 0,
102
then, for every absolutely continuous curve : [, ] U, with
[, ] ]a, b[, we have
L() =
_
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 HamiltonJacobi 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(, ())
_
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
(()). The
two extremal curves and
1
go through the same point with the
same speed at t = . Since the EulerLagrange 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 HamiltonJacobi 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
CauchyLipschitz 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
() =
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
(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 HamiltonJacobi 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 EulerLagrange 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 [
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 EulerLagrange ow is welldened 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 HamiltonJacobi 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) =
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 HamiltonJacobi 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
HamiltonJacobi inequality.
Proposition 4.1.2. Suppose that L is a Tonelli Lagrangian on
the manifold M. Call H : T
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
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
.
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 HamiltonJacobi 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
HamiltonJacobi 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
EulerLagrange ow. If U = M and M is compact then the graph
Graph(grad
L
u) is invariant by the EulerLagrange 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 EulerLagrange ow. Given x U, since the vector eld
grad
L
u is continuous, we can apply the CauchyPeano 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
] I, with t t
. If :
[t, t
] U is a (continuous) piecewise C
1
, from u L + c, it
follows that
u((t
t).
Moreover, since is (u, L, c)calibrated, we have equality when
= [[t, t
]. If (t
) = (t
]) +c(t
t) = u((t
t),
hence L([[t, t
(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 pointwise 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 pointwise 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) Kv
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 pointwise
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 equiLipschitz 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; , ) =
_
: [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(()
_
I 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
. 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
[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
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
(y) u(y)
+
(y), with equality at 0, the C
1
func
tion
+
+
= d
x
.
By the denition of the derivative, using
(x) = u(x) =
+
(x),
we can write
(y x), with
lim
h0
(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 realvalued
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
nonincreasing 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 welldened 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
] 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
] 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
timeindependent. 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 semigroup, 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
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
=
st
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
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
, 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
: [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
,y
,y
: [0, t]
M by
x
,y
= on [, t ],
x
,y
=
i
0
x,
on [0, ], and
,y
=
j
1
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 
,y
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
, 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
, 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
, y
), param
eterized proportionally arclength and connecting respectively x to
y and x
to y
= 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 LaxOleinik Semigroup.
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 LaxOleinik semigroup
We introduce a semigroup of nonlinear operators (T
t
)
t0
from
T(M, [, +]) into itself. This semigroup is wellknown in PDE
and in Calculus of Variations, it is called the LaxOleinik semi
group. Again it has been rediscovered many times in dierent
145
forms; For the author, it came as a natural byproduct of the
proof of the hamiltonJacobi Theorem for C
1
functions, see the
proof of Theorem 4.1.1 and the discussion in section 1.
Denition 4.6.1 (LaxOleinik semigroup). 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 LaxOleinik semigroup.
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
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 LaxOleinik
semigroup.
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 semigroup T
t
, we recall
the following wellknown denition
Denition 4.6.4 (Nonexpansive Map). A map : X Y ,
between the metric spaces X and Y , is said to be nonexpansive
if it is Lipschitzian with a Lipschitz constant 1.
Proposition 4.6.5 (Nonexpansiveness of the Lax Oleinik semi
group). The maps T
t
: are nonexpansive 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 semigroup 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 semigroup 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 equiLipschitzian 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) Kv
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 semigroup 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 equiLipschitzianity, 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 semigroup 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 semigroup 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 nitevalued, 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
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]
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
(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
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
equiLipschitzian. 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
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
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 pointwise 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 nonexpansiveness of
the Lax Oleinik semigroup 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
nonexpansive. 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 nonexpansive
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 nonexpansive.
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
SchauderTykhonov 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 nonexpansive maps. Since,
for each t > 0, the image of T
t
is an equiLipschitzian 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
_
TM
Ld,
where varies among Borel probability measures on TM invariant
by the EulerLagrange 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
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 2u
t
u
= u
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 EulerLagrange 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 EulerLagrange ow, which satises
c[0] =
_
TM
Ld.
Exercise 4.8.4. Show that each Borel probability measure , in
variant by the EulerLagrange 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
L +c[0] and
thus
H(x, d
x
u
) = sup
vTxM
d
x
u
(v
0
) = L(x, v
0
) +c[0].
For that, we pick extremal curves
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( ) = 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 LaxOleinik
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 EulerLagrange 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 EulerLagrange 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
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
: 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
. 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
k
i
= k
_
_
k
i
= k
_
i
,
from which we obtain the equality
k = max
_
k
1
, . . . ,
_
k
1
, . . . ,
_
k
_
,
where is a C
(
0
(M, R) such that
u
+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 1form, the EulerLagrange
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
_
TM
(L )d, where
varies among the Borel probability measures on TM invariant
under the EulerLagrange 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])+Kyx
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])Kyx
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 ALipschitz. We then
pick > 0 such that for each ball
B(y, A), for y (
B(0, 1)) is
contained in (
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
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
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])+Kv
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
:] , 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 258259], 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)[ Ky 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
) d
x
ud
x
u 6Kxx
.
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
). If x = x
 > 0.
If h is such that h = xx
+h
are in
B. This allows us to write
[u(x +h) u(x)
x
(h)[ Kh
2
[u(x) u(x
)
x
(x x
) Kx x

2
[u(x +h) u(x
)
x
(x x
+h)[ Kx x
+h
2
.
As h = x x
) u(x +h) +
x
(x x
+h)[ 4Kx 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)[ 6Kx x

2
,
hence

x
x
 = sup
h=xx
[
x
(h)
x
(h)[
x x

6Kx 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(()) =
_
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
. It is
not then dicult to deduce that /
,u
is closed. It also results from
() that
T
+
u(x) u(x) +c
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(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)), ,
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) +Ky x
2
T
+
u(
i
(y) T
+
u(
i
( x))
L
v
(x, (0)) D
i
( x)(y x) Ky x
2
,
Using the inequalities () we get
[u(
i
(y)) u(
i
( x))
L
v
(x, (0)) D
i
( x)(y x)[ Ky 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))
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 CauchyPeano
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
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 EulerLagrange 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
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
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
(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 EulerLagrange 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
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
(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
=
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
(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
is an extremal
curve. Let us suppose that 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)). We then
conclude that H(x, 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
at x, we nd
that H
_
(t),
L
v
(
x
(t),
x
(t))
)
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
(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
)
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
) we necessarily have d
x
u
=
L
v
(x, v) = p. As
Graph(du
) 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
(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
(x). If u
) 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
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
t
,
for t 0, if (x,
_
supV V (x)) 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
(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
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
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 equiLipschitzian,
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
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 equiLipschitzian 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
. We of course
do have u = (u
1
+ u
2
, 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
+u
2
)/2) = H(x, d
x
u
1
) = H(x, d
x
u
2
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 biLipschitzian homeomorphism
1
(u
,u
+
)
on 1
(u
,u
+
)
. The set
1
(u
,u
+
)
is compact and invariant by
the EulerLagrange 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 EulerLagrange 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 biLipschitz
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 equiLipschitzian 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
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
.
Therefore d
x
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
such that
t t
0
, x, y M, 2u
h
t
(x, y)+c[0]t 2u
+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
2u
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 equiLipschitzian.
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
, 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, 2u

0
h
t
(x, y) +c[0]t 2u

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
+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
: ] , 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)) 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
(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) 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),
(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
,u
+
)
and (b, w) TM are at
distance less than
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(
,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
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
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
LaxOleinik semigroup
6.1 Semiconvex 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
_
1
0
d
ty+(1t)x
u(y x) d
x
u(y x) dt
_
1
0
d
ty+(1t)x
u d
x
uy x dt
_
1
0
Kty + (1 t)x xy x dt
=
_
1
0
Kty x
2
dt
=
K
2
y x
2
.
205
This implies that u is both semiconvex and semiconcave on the
convex set
B(0, r).
Suppose now that u is both locally semiconvex and locally
semiconcave. Again due to the local nature of the result, we can
assume M =
B(0, r) and that u is both K
1
semiconvex and K
2

semiconcave. 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
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 LaxOleinik Semigroup 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 LaxOleinik 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 semiconcave (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 semiconvex. More
over, for each 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 EulerLagrange 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 semiconvexity of u from
proposition 6.2.1.
If 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
C
1,1
.
6.3 Convergence of the LaxOleinik Semi
group
Up to now in this book, all the statements given above do hold for
periodic timedependent 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
EulerLagrange ow. Some of these results do not hold for the
timedependent case, see [FM00].
The main goal of this section is to prove the following theorem:
Theorem 6.3.1 (Convergence of the LaxOleinik Semigroup).
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 LaxOleinik semigroups 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
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
), (t
)) W
), (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 semigroup of nonexpansive 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 nonexpansive, 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
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
.
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 nonexpansive. 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 equiLipschitzian 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
: T
M M. If the image
t
(N) is still transversal to the bers of the canonical projection
: T
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
: 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 =
: 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)
M by (Tp)
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
: 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
: 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 CapuzzoDolceta
[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 HamiltonJacobi
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 HamiltonJacobi asso
ciated to H is the equation
H(x, d
x
u) = c,
where c is some constant.
A classical solution of the HamiltonJacobi 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 HamiltonJacobi 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
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
(RM) R dened
by
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 HamiltonJacobi 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
everywhere with u
(x
0
) =
(x
0
). We have (x
0
) (x) u
(x
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 semigroup. 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
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
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, p) [
p 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 nonnegative 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
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
nd
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
function
: M R such that
: M R, such that
+
(x
0
) =
(x
0
) = (x
0
),
and
+
(x) > (x) >
(x), for x ,= x
0
. Moreover
+
(x)
C > (x) >
+
= d
x
0
= d
x
0
) 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 LaxOleinik semigroups 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
[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
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
= 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
= supp [ 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 EulerLagrange 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
W. Since
u
n
converges to u uniformly on the compact set
W, necessarily
235
236
u achieves its maximum on
W at 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 equiLipschitzian. 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 nitevalued
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
noncompact 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 semiconvex (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
+ (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 LaxOleinik semigroups
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 semiconvex 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
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
0
such
that the intersection of
g
t
D
+
u with T
V = (x, p) [ x V is
contained in H
1
(] , c
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
semiconvex 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 semiconvex 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
u
1
(x),
because u
1
is locally semiconvex.
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 noncompact version of lemma 8.2.5:
Suppose H : T
: 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 Quasiconvexity and viscosity subsolu
tions
In this section we will be mainly interested in Hamiltonians H :
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 quasiconvex 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
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
) 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
(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
0
(
) = H(x,
_
R
k
(y)d
xy
udy). (***)
Since
N
0
(
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
, p)
H(z, p)[ . If
, since
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
) H
U
(u) +, for
. The
inequality sup
xV
[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
M R is
continuous and quasiconvex 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 quasiconvexity 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 quasiconvex 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 quasiconvex 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
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
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 quasiconvex 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 quasiconvexity 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
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
H(y, d
y
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 semidistance
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
equiLipschitzian, 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 semidistance,
S
c
(x, y) =
S
c
(x, y)+S
c
(y, x) is a distance which is locally Lipschitzequivalent
to any distance coming from a Riemannian metric.
Proof. The fact that S
c
is a semidistance 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 supp [ 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), 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
on
B(x
0
,
1
) x
0
, for every
x
0
N
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
y
) c.
Since S
c[0]
(x
0
, ) is a viscosity subsolution of H(x, d
x
u) = c[0],
and
y
is C
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 realvalued function u
2
dened on a
253
neighborhood of
B(x, /2) and semiconvex 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 semiconvex 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
equiLipschitzian. 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
) < 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
u
uniformly as 0. It then suces to choose small enough and
to apply 8.3.6 to v
.
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 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
, 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 equiLipschitzian 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 arclength, 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 semidistance 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 Semistatic 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 (Semistatic curve). A curve : [a, b] M is
called semistatic, 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 semistatic 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
) + 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
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
DuboisRaymond, 42
Equation
EulerLagrange, 42
Erdmann, 44
Euler, 42, 49, 52, 53
EulerLagrange Equation, 42
EulerLagrange Flow, 53
EulerLagrange Theorem, 49
EulerLagrange 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
quasiconvex, 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
HamiltonJacobi, 57, 110
HamiltonJacobi 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
nondegenerate, 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
DuboisRaymond, 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
Nondegenerate Lagrangian, 38
NonExpansive Map, 149
Odd Lagrangian Submanifold, 73
Odd Lagrangian Subspace, 71
Peierls Barrier, 190
Point
Exposed, 31
Extremal, 30
proper map, 23
quasiconvex, 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
EulerLagrange, 49
Fenchel, 13
KreinMilman, 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, EulerLagrange, 52,
53
very weak solution, 215
Viscosity
solution, 215, 216
subsolution, 215, 216
Weak KAM, 118
Weak KAM Theorem, 153, 162
Viel mehr als nur Dokumente.
Entdecken, was Scribd alles zu bieten hat, inklusive Bücher und Hörbücher von großen Verlagen.
Jederzeit kündbar.