Global Minimizers of
Autonomous Lagrangians
Gonzalo Contreras
Renato Iturriaga
cimat
mexico, gto.
c ( 2000
ii
ii
Contents
1 Introduction. 7
11 Lagrangian Dynamics. . . . . . . . . . . . . . . . . . . . . 7
12 The EulerLagrange equation. . . . . . . . . . . . . . . . . 9
13 The Energy function. . . . . . . . . . . . . . . . . . . . . . 12
14 Hamiltonian Systems. . . . . . . . . . . . . . . . . . . . . 14
15 Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Ma˜ n´e’s critical value. 23
21 The action potential and the critical value. . . . . . . . . 23
22 Continuity of the critical value. . . . . . . . . . . . . . . . 27
23 Holonomic measures. . . . . . . . . . . . . . . . . . . . . . 28
24 Invariance of minimizing measures. . . . . . . . . . . . . . 31
25 Ergodic characterization of the critical value. . . . . . . . 44
26 The AubryMather Theory. . . . . . . . . . . . . . . . . . 47
26.a Homology of measures. . . . . . . . . . . . . . . . . 47
26.b The asymptotic cycle. . . . . . . . . . . . . . . . . 47
26.c The alpha and beta functions. . . . . . . . . . . . 50
27 Coverings. . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3 Globally minimizing orbits. 55
31 Tonelli’s theorem. . . . . . . . . . . . . . . . . . . . . . . . 55
32 A priori compactness. . . . . . . . . . . . . . . . . . . . . 62
iii
iv CONTENTS
33 Energy of timefree minimizers. . . . . . . . . . . . . . . . 65
34 The ﬁnitetime potential. . . . . . . . . . . . . . . . . . . 67
35 Global Minimizers. . . . . . . . . . . . . . . . . . . . . . . 70
36 Characterization of minimizing measures. . . . . . . . . . 75
37 The Peierls barrier. . . . . . . . . . . . . . . . . . . . . . . 79
38 Graph Properties. . . . . . . . . . . . . . . . . . . . . . . 82
39 Coboundary Property. . . . . . . . . . . . . . . . . . . . . 86
310 Covering Properties. . . . . . . . . . . . . . . . . . . . . . 88
311 Recurrence Properties. . . . . . . . . . . . . . . . . . . . . 89
4 The Hamiltonian viewpoint. 97
41 The HamiltonJacobi equation. . . . . . . . . . . . . . . . 97
42 Dominated functions. . . . . . . . . . . . . . . . . . . . . . 99
43 Weak solutions of the HamiltonJacobi equation. . . . . . 103
44 Lagrangian graphs. . . . . . . . . . . . . . . . . . . . . . . 105
45 Contact ﬂows. . . . . . . . . . . . . . . . . . . . . . . . 111
46 Finsler metrics. . . . . . . . . . . . . . . . . . . . . . . . . 113
47 Anosov energy levels. . . . . . . . . . . . . . . . . . . . . . 118
48 The weak KAM Theory. . . . . . . . . . . . . . . . . . . . 121
49 Construction of weak KAM solutions . . . . . . . . . . . . 125
49.a Finite Peierls barrier. . . . . . . . . . . . . . . . . 126
49.b The compact case. . . . . . . . . . . . . . . . . . . 127
49.c Busemann weak KAM solutions. . . . . . . . . . . 130
410 Higher energy levels. . . . . . . . . . . . . . . . . . . . . . 134
411 The LaxOleinik semigroup. . . . . . . . . . . . . . . . . . 137
412 The extended static classes. . . . . . . . . . . . . . . . . . 145
5 Examples 155
51 Riemannian Lagrangians. . . . . . . . . . . . . . . . . . 155
52 Mechanic Lagrangians. . . . . . . . . . . . . . . . . . . . 155
53 Symmetric Lagrangians. . . . . . . . . . . . . . . . . . . 156
iv
CONTENTS v
54 Simple Pendulum. . . . . . . . . . . . . . . . . . . . . . . 156
55 The ﬂat Torus T
n
. . . . . . . . . . . . . . . . . . . . . . 157
56 Flat domain for the βfunction. . . . . . . . . . . . . . . 158
57 A Lagrangian with Peierls barrier h = +∞. . . . . . . 158
58 Horocycle ﬂow. . . . . . . . . . . . . . . . . . . . . . . . 160
6 Generic Lagrangians. 165
61 Generic Families of Lagrangians. . . . . . . . . . . . . . . 166
61.a Open. . . . . . . . . . . . . . . . . . . . . . . . . . 169
61.b Dense. . . . . . . . . . . . . . . . . . . . . . . . . . 170
7 Generic Lagrangians. 175
71 Generic Lagrangians. . . . . . . . . . . . . . . . . . . . . . 176
72 Homoclinic Orbits. . . . . . . . . . . . . . . . . . . . . . . 188
Appendix. 195
A Absolutely continuous functions. . . . . . . . . . . . . . . 195
B Measure Theory . . . . . . . . . . . . . . . . . . . . . . . 197
C Convex functions. . . . . . . . . . . . . . . . . . . . . . . . 198
D The Fenchel and Legendre Transforms. . . . . . . . . . . . 199
E Singular sets of convex funcions. . . . . . . . . . . . . . . 203
F Symplectic Linear Algebra. . . . . . . . . . . . . . . . . . 205
Bibliography. 207
Index. 213
v
vi CONTENTS
vi
Chapter 1
Introduction.
11 Lagrangian Dynamics.
Let M be a boundaryless ndimensional complete riemannian manifold.
An (autonomous) Lagrangian on M is a smooth function L : TM → R
satisfying the following conditions:
(a) Convexity: The Hessian
∂
2
L
∂v
i
∂v
j
(x, v), calculated in linear co
ordinates on the ﬁber T
x
M, is uniformly positive deﬁnite for all
(x, v) ∈ TM, i.e. there is A > 0 such that
w L
vv
(x, v) w ≥ A [w[
2
for all (x, v) ∈ TM and w ∈ T
x
M.
(b) Superlinearity:
lim
v→+∞
L(x, v)
[v[
= +∞, uniformly on x ∈ M,
equivalently, for all A ∈ R there is B ∈ R such that
L(x, v) ≥ A[v[ −B for all (x, v) ∈ TM.
7
8 1. introduction.
(c) Boundedness
1
: For all r ≥ 0,
(r) = sup
(x,v)∈TM,
v≤r
L(x, v) < +∞. (1.1)
g(r) = sup
w=1
(x,v)≤r
w L
vv
(x, v) w < +∞. (1.2)
The EulerLagrange equation associated to a lagrangian L is (in
local coordinates)
d
dt
∂L
∂v
x, ˙ x) =
∂L
∂x
(x, ˙ x). (EL)
The condition (c) implies that the EulerLagrange equation (EL) deﬁnes
a complete ﬂow ϕ
t
on TM (proposition 13.2), called the EulerLagrange
ﬂow, by setting ϕ
t
(x
0
, v
0
) =
x
v
(t), ˙ x
v
(t)
, where x
v
: R → M is the
solution of (EL) with x
v
(0) = x
0
and ˙ x
v
(0) = v
0
.
We shall be interested on coverings p : N → M of a compact man
ifold M and the lifted Lagrangian L = L ◦ dp : TN → R of a convex
superlinear lagrangian L on M. The lagrangian L then satisﬁes (a)–(c)
and its ﬂow ψ
t
is the lift of ϕ
t
.
Observe that when we add a closed 1form ω to the lagrangian L,
the new lagrangian L + ω also satisﬁes the hypothesis (a)(c) and has
the same EulerLagrange equation as L. This can also be seen using the
variational interpretation of the EulerLagrange equation (see 12.3).
1
The Boundedness condition (c) is equivalent to the condition that the associ
ated hamiltonian H is convex and superlinear, see remark 14.2. This condition is
immediate when the manifold M is compact.
8
12. the eulerlagrange equation. 9
12 The EulerLagrange equation.
The action of a diﬀerential curve γ : [0, T] →M is deﬁned by
A
L
(γ) =
T
0
L(γ(t), ˙ γ(t)) dt
One of the main problems of the calculus of variations is to ﬁnd and
to study the curves that minimize the action. Denote by C
k
(q
1
, q
2
; T)
the set of C
k
diﬀerentiable curves γ : [0, T] → M such that γ(0) = q
1
and γ(T) = q
2
.
12.1 Proposition. If a curve x(t) in the space C
k
(q
1
, q
2
; T) is a critical
point of the action functional on C
k
(q
1
, q
2
; T), then x satisﬁes the Euler
Lagrange equation
d
dt
L
v
(x(t), ˙ x(t)) = L
x
(x(t), ˙ x(t)) (EL)
in local coordinates. Consequently, this equation does not depend on the
coordinate system.
Proof: Choose a coordinate system (x
1
, . . . , x
n
) about x(t). Let h(t)
a diﬀerentiable curve such that h(0) = h(T) = 0. Then for every ε,
suﬃciently small the curve y
ε
= x+εh is on C
k
(q
1
, q
2
; T) and contained
in the coordinate system. Deﬁne
g(ε) = A
L
(y
ε
)
9
10 1. introduction.
Then g has a minimum in zero and
lim
ε→0
g(ε) −g(0)
ε
= lim
ε→0
T
0
L(x +εh, ˙ x +ε
˙
h) −L(x, ˙ x)
ε
dt
=
T
0
lim
ε→0
εL
x
h +εL
v
˙
h +o(ε)
ε
dt
=
T
0
L
x
h +L
v
˙
h dt
=
T
0
(L
x
−
d
dt
L
v
) h dt +L
v
h
T
0
=
T
0
(L
x
−
d
dt
L
v
) h dt.
Hence
0 =
T
0
¸
L
x
(x(t), ˙ x(t)) −
d
dt
L
v
(x(t), ˙ x(t))
h dt,
for any function h ∈ C
k
(0, 0; T). This implies that x(t) satisﬁes the
Euler Lagrange equation (EL).
The Euler Lagrange equation is a second order diﬀerential equation
on M, but the convexity hypothesis (L
vv
invertible) implies that this
equation can also be seen as a ﬁrst order diﬀerential equation on TM:
˙ x = v,
˙ v = (L
vv
)
−1
(L
x
−L
vx
v).
The associated vector ﬁeld X on TM is called the lagrangian vector
ﬁeld and its ﬂow ϕ
t
the lagrangian ﬂow. Observe that X is of the form
X(x, v) = (v, ).
12.2 Remark. It is possible to do the same thing in the space (
T
(p, q),
the set of absolutely continuous curves γ : [0, T] →M such that γ(0) = p
and γ(T) = q. A priori minimizers do not have to be diﬀerentiable and
there are examples where they are not, see Ball & Mizel [4]. However
10
12. the eulerlagrange equation. 11
when the lagrangian ﬂow is complete (cf. proposition 13.2), every ab
solutely continuous minimizers is C
2
and satisﬁes the EulerLagrange
equation. See Mather [46].
12.3 Remark. If we add a closed 1form ω to the lagrangian L, the
lagrangian L + ω also satisﬁes the hypothesis (a)(c). Moreover, the
action functional A
L+ω
on a neighbourbood of a curve γ ∈ C
k
(x
1
, x
2
, T)
satisﬁes
A
L+ω
(η) = A
L
(η) +
γ
ω,
because the curve η is homologous to γ. Therefore, the critical points
for A
L+ω
and for A
L
are the same. This implies that the EulerLagrange
equations for L and L + ω are the same. But since the values of A
L+ω
and A
L
are diﬀerent, minimizers of these two actions may be diﬀerent.
11
12 1. introduction.
13 The Energy function.
The energy function of the lagrangian L is E : TM →R, deﬁned by
E(x, v) =
∂L
∂v
(x, v) v −L(x, v). (1.3)
Observe that if x(t) is a solution of the EulerLagrange equation (EL),
then
d
dt
E(x, ˙ x) =
d
dt
L
v
−L
x
˙ x = 0.
Hence E : TM → R is an integral (i.e. invariant function
2
) for the
lagrangian ﬂow ϕ
t
and its level sets, called energy levels are invariant
under ϕ
t
. Moreover, the convexity implies that
d
ds
E(x, sv)[
s=1
= v L
vv
(x, v) v > 0.
Thus
min
v∈T
x
M
E(x, v) = E(x, 0) = −L(x, 0).
Write
e
0
:= max
x∈M
E(x, 0) = −min
x∈M
L(x, 0) > −∞, (1.4)
by the superlinearity e
0
> −∞, then
e
0
= min ¦ k ∈ R [ π : E
−1
¦k¦ →M is surjective ¦. (1.5)
By the uniform convexity, and the boundedness condition,
A := inf
(x,v)∈TM
w=1
w L
vv
(x, v) w > 0,
and then using (1.1) and (1.2),
E(x, v) = E(x, 0) +
v
0
d
ds
E
x, s
v
v
ds
≥ −(0) +
1
2
A [v[
2
. (1.6)
2
The energy is invariant only for autonomous (i.e. timeindependent) lagrangians.
12
13. the energy function. 13
Similarly, using (1.2),
E(x, v) ≤ e
0
+g([v[) [v[. (1.7)
Hence
13.1 Remark.
If k ∈ R and K ⊆ M is compact, then E
−1
¦k¦ ∩ T
K
M is compact.
13.2 Proposition. The EulerLagrange ﬂow is complete.
Proof: Suppose that ]α, β[ is the maximal interval of deﬁnition of t →
ϕ
t
(v), and −∞ < α or β < +∞. Let k = E(v). Since E(ϕ
t
(v)) ≡ k,
by (1.6), there is a > 0 such that 0 ≤ [ϕ
t
(v)[ ≤ a for α ≤ t ≤ β. Since
ϕ
t
(v) is of the form (γ(t), ˙ γ(t)), then ϕ
t
(v) remains in the interior of the
compact set
Q :=
¸
(y, w) ∈ TM
d(y, x) ≤ a
[β −α[ + 1
, [w[ ≤ a + 1
¸
,
where x = π(v). The EulerLagrange vector ﬁeld is uniformly Lipschitz
on Q. Then by the theory of ordinary diﬀerential equations, we can
extend the interval of deﬁnition ]α, β[ of t →ϕ
t
(v).
13
14 1. introduction.
14 Hamiltonian Systems.
Let T
∗
M be the cotangent bundle of M. Deﬁne the Liouville’s 1form
Θ on T
∗
M as
Θ
p
(ξ) = p (dπ ξ) for ξ ∈ T
p
(T
∗
M),
where π : T
∗
M →M is the projection. The canonical symplectic form
on T
∗
M is deﬁned as ω = dΘ.
A local chart x = (x
1
, . . . , x
n
) of M induces a local chart (x, p) =
(x
1
, . . . , x
n
; p
1
, . . . , p
n
) of T
∗
M writing p ∈ T
∗
M as p = Σ
i
p
i
dx
i
. In
these coordinates the forms Θ and ω are written
Θ = p dx =
¸
i
p
i
dx
i
,
ω = dp ∧ dx =
¸
i
dp
i
∧ dx
i
.
A hamiltonian is a smooth function H : T
∗
M →R. The hamiltonian
vector ﬁeld X
H
associated to H is deﬁned by
ω(X
H
, ) = dH. (1.8)
In local charts, the hamiltonian vector ﬁeld deﬁnes the diﬀerential equa
tion
˙ x = H
p
,
˙ p = − H
x
,
(1.9)
where H
x
and H
p
are the partial derivatives of H with respect to x and
p. Let ψ
t
be the hamiltonian ﬂow. Observe that it preserves H, because
d
dt
H = H
x
˙ x +H
p
˙ p = 0.
Moreover, it preserves the symplectic form ω, because
3
d
dt
(ψ
∗
t
ω) = L
X
H
ω = d i
X
H
ω +i
X
H
ω = d(dH) +i
X
H
(0) = 0.
3
L
X
is the Lie derivative, deﬁned on forms η by L
X
η = di
X
η + i
X
dη, where
i
X
η = η(X, ·) is the contraction by X. The Lie derivative satisﬁes L
X
η =
d
dt
ψ
∗
t
η
t=0
,
where ψ
t
is the ﬂow of X.
14
14. hamiltonian systems. 15
We shall be specially interested in hamiltonians obtained by the
Fenchel transform of a lagrangian:
H(x, p) = max
v∈T
x
M
p v −L(x, v).
Observe that H = E◦L
−1
, where E is the energy function (1.3) and
L(x, v) = (x, L
v
(x, v)) is the Legendre transform of L. Moreover
14.1 Proposition. The Legendre transform L : TM → T
∗
M,
L(x, v) = (x, L
v
(x, v)) is a conjugacy between the lagrangian ﬂow and
the hamiltonian ﬂow.
Proof: By corollary D.2, the convexity and superlinearity hypothesis
imply that L = L
∗∗
= H
∗
. So if p = L
v
(x, v) then v = H
p
(x, p). With
this notation:
H(x, p) = v L
v
(x, v) −L(x, v) = E ◦ L
−1
= p H
p
(x, p) −L(x, H
p
(x, p)).
Thus H
x
= −L
x
, and the EulerLagrange equation
˙ x =
d
dt
x = v = H
p
,
˙ p =
d
dt
L
v
= L
x
= −H
x
,
is the same as the hamiltonian equations.
14.2 Remark. Using that L
∗
= H and H
∗
= L, from proposition D.2
in the appendix we obtain that the boundedness condition is equivalent
to
(c) Boundedness: H = L
∗
is convex and superlinear.
We say that an energy level H
−1
(k) is regular, if k is a regular value
of H, i.e. dH(x, p) = 0 whenever H(x, p) = k.
15
16 1. introduction.
14.3 Proposition. Two hamiltonian ﬂows restricted to a same regular
energy level are reparametrizations of each other.
Proof: Suppose that H, G : T
∗
M → R are two hamiltonians with
H
−1
(k) = G
−1
() and k, are regular values for H and G respectively.
Then, if H(x, p) = k,
ker d
(x,p)
H = T
(x,p)
H
−1
(k) = T
(x,p)
H
−1
(k) = ker d
(x,p)
G.
Thus there exists λ(x, p) > 0 such that d
(x,p)
H = λ(x, p) d
(x,p)
G. Equa
tion (1.8) implies that X
H
= λ(x, p) X
G
when H(x, p) = k.
We shall need the following estimate on the norm of the partial
derivative L
v
(x, v).
14.4 Lemma. There is a function f : [0, ∞[→ R
+
such that
L
v
(x, v) ≤ f([v[) for all (x, v) ∈ TM.
Proof: The convexity condition implies that the maximum in
H(x, p) = max
w∈T
x
M
p w −L(x, w)
is attained at w = v
0
with p = L
v
(x, v
0
). Since H(x, L
v
(x, v)) = E(x, v),
L
v
(x, v) w ≤ E(x, v) +L(x, w), ∀v, w ∈ T
x
M, ∀x ∈ M.
Applying this inequality to −w, we get that
L
v
(x, v) w ≥ −E(x, v) −L(x, −w).
Thus using (1.6), (1.7) and (1.1), for [v[ ≤ r, we have
L
v
(x, v) ≤ [E(x, v)[ + max¦ [L(x, v)[, [L(x, −v)[ ¦
≤ max¦[(0)[ +
1
2
Ar
2
, e
0
+g(r)r¦ +[(r)[ =: f(r).
16
15. examples. 17
15 Examples.
We give here some basic examples of lagrangians.
Riemannian Lagrangians:
Given a riemannian metric g = ', `
x
on TM, the riemannian la
grangian on M is given by the kinetic energy
L(x, v) =
1
2
v
2
x
. (1.10)
Its EulerLagrange equation (EL) is the equation of the geodesics of g:
D
dt
˙ x ≡ 0, (1.11)
and its EulerLagrange ﬂow is the geodesic ﬂow. Its corresponding
hamiltonian is
H(x, p) =
1
2
p
2
x
.
Analogous to the riemannian lagrangian is the Finsler lagrangian,
given also by formula (1.10), but where 
x
is a Finsler metric, i.e. 
x
is a (non necessarily symmetric
4
) norm on T
x
M which varies smoothly
on x ∈ M. The EulerLagrange ﬂow of a Finsler lagrangian is called the
geodesic ﬂow of the Finsler metric 
x
.
Mechanic Lagrangians:
The mechanic lagrangian, also called natural lagrangian, is given
by the kinetic energy minus the potential energy U : M →R,
L(x, v) =
1
2
v
2
x
−U(x). (1.12)
Its EulerLagrange equation is
D
dt
˙ x = −∇U(x),
4
i.e. λv
x
= λ v
x
only for λ ≥ 0
17
18 1. introduction.
where
D
dt
is the covariant derivative and ∇U is the gradient of U with
respect to the riemannian metric g, i.e.
d
x
U(v) = '∇U(x), v`
x
for all (x, v) ∈ TM.
Its energy function and its hamiltonian are given by the kinetic energy
plus potential energy:
E(x, v) =
1
2
v
2
x
+U(x),
H(x, p) =
1
2
p
2
x
+U(x).
Symmetric Lagrangians.
The symmetric lagrangians is a class of lagrangian systems which
includes the riemannian and mechanic lagrangians. These are the la
grangians which satisfy
L(x, v) = L(x, −v) for all (x, v) ∈ TM. (1.13)
Their EulerLagrange ﬂow is reversible in the sense that ϕ
−t
(v) =
−ϕ
t
(−v).
Magnetic Lagrangians.
If one adds a closed 1form ω to a lagrangian, L(x, v) = L(x, v) +
ω
x
(v), the EulerLagrange ﬂow does not change. This can be seen by
ﬁrst observing that the solutions of the EulerLagrange equation are
the critical points of the action functional on curves on ((x, y, T) (with
ﬁxed time interval and ﬁxed endpoints). Since ω is closed, the action
functional of L and L on ((x, y, T) diﬀer by a constant and hence they
have the same critical points.
But adding a nonclosed 1form to a lagrangian does change the
EulerLagrange ﬂow. We call a magnetic lagrangian a lagrangian of the
form
L(x, v) =
1
2
v
x
+η
x
(v) −U(x), (1.14)
18
15. examples. 19
where 
x
is a riemannian metric, η is a 1form on M with dη = 0, and
U : M → R a smooth function. If Y : TM → TM is the bundle map
such that
dη(u, v) = 'Y (u), v`
then the EulerLagrange equation of (1.14) is
D
dt
˙ x = Y
x
( ˙ x) −∇U(x). (1.15)
This models the motion of a particle with unit mass and unit charge
under the eﬀect of a magnetic ﬁeld with Lorentz force Y and potential
energy U(x). The energy functional is the same as that of the mechanical
lagrangian but its hamiltonian changes because of the change in the
Legendre transform:
E(x, v) =
1
2
v
2
x
+U(x),
H(x, p) =
1
2
p −A(x)
2
x
+U(x),
where A : M →TM is the vector ﬁeld given by η
x
(v) = 'A(x), v`
x
.
Twisted geodesic ﬂows.
The twisted geodesic ﬂows correspond to the motion of a particle
under the eﬀect of a magnetic ﬁeld with no potential energy. This can be
modeled as the EulerLagrange ﬂow of a lagrangian of the form L(x, v) =
1
2
v
2
x
+η
x
(v), where dη ≡ 0. But the EulerLagrange equations depend
only on the riemannian metric and dη. A generalization of these ﬂows
can be made using a nonzero 2form Ω instead of dη and not requiring
Ω to be exact. This is better presented in the hamiltonian setting.
Fix a riemannian metric ' , ` and a 2form Ω on M. Let K : TTM →
TM be the connection map Kξ = ∇
˙ x
v, where ξ =
d
dt
x(t), v(t)
. Let
π : TM → M be the canonical projection. Let ω
0
be the symplectic
form in TM obtained by pulling back the canonical symplectic form via
the Legendre transform associated to the riemannian metric, i.e.
ω
0
(ξ, ζ) = 'dπ ξ , Kζ` −'dπ ζ , Kξ` .
19
20 1. introduction.
The coordinates T
θ
TM ÷ ξ ←→ (dπ ξ, Kξ) ∈ T
π(θ)
M ⊕ T
π
(θ)M =
H(θ) ⊕V (θ) are the standard way of writing the horizontal and vertical
components of a vector ξ ∈ T
θ
TM for a riemannian manifold M (see
Klingenberg [31]).
Deﬁne a new symplectic form ω
Ω
on TM by
ω
Ω
= ω
0
+π
∗
Ω.
This is called a twisted symplectic structure on TM. Let H : TM → R
be the hamiltonian
H(x, v) =
1
2
v
2
x
.
Consider the hamiltonian vector ﬁeld X
F
corresponding to (H, ω
Ω
), i.e.
ω
Ω
X
Ω
(θ),
= dH . (1.16)
Deﬁne Y : TM →TM as the bundle map such that
Ω
x
(u, v) = 'Y (u), v`
x
. (1.17)
The hamiltonian vector ﬁeld X
Ω
(θ) ∈ T
θ
TM is given by X
Ω
(θ) =
(θ, Y (θ)) ∈ H(θ) ⊕V (θ). Hence the hamiltonian equation is
D
dt
˙ x = Y
x
( ˙ x),
recovering equation (1.15) with U ≡ 0, but where Ω doesn’t need to be
exact.
If H
1
(M, R) = 0, both approaches coincide, and any twisted geodesic
ﬂow is the lagrangian ﬂow of a magnetic lagrangian of the form L(x, v) =
1
2
v
2
x
+ η
x
(v), with dη = Ω. For example if N is a compact manifold
Ω is a 2form in N and M is the abelian cover or the universal cover
of N; if Ω is not exact, then the corresponding twisted geodesic ﬂow is
a lagrangian ﬂow on M but not on N (where it is locally a lagrangian
ﬂow). This lagrangian ﬂow on M is actually the lift of the twisted
geodesic ﬂow on N.
20
15. examples. 21
Embedding ﬂows:
There is a way to embed the ﬂow of any bounded vector ﬁeld on a
lagrangian system. Given a smooth bounded vector ﬁeld F : M →TM,
let
L(x, v) =
1
2
v −F(x)
2
x
. (1.18)
Since F(x) is bounded, then the lagrangian L is convex, superlinear and
satisﬁes the boundedness condition. The lagrangian L on a ﬁber T
x
M
is minimized at (x, F(x)), hence the integral curves of the vector ﬁeld,
˙ x = F(x), are solutions to the EulerLagrange equation.
21
22 1. introduction.
22
Chapter 2
Ma˜ n´e’s critical value.
21 The action potential and the critical value.
We shall be interested on action minimizing curves with free time in
terval. Unless otherwise stated, all the curves will be assumed to be
absolutely continuous. For x, y ∈ M, let
((x, y) = ¦ γ : [0, T] →M[ T > 0, γ(0) = x, γ(T) = y ¦.
For k ∈ R deﬁne the action potential Φ
k
: M M →R ∪ ¦−∞¦, by
Φ
k
(x, y) = inf
γ∈C(x,y)
A
L+k
(γ).
Observe that if there exists a closed curve γ on N with negative L + k
action, then Φ
k
(x, y) = −∞ for all x, y ∈ N, by going round γ many
times.
Deﬁne the critical level c = c(L) as
c(L) = sup¦ k ∈ R[ ∃ closed curve γ with A
L+k
(γ) < 0 ¦.
Observe that the function k →Φ
k
(x, y) is increasing. The superlinearity
implies that L is bounded below. Hence there is k ∈ R such that
23
24 2. ma˜ n´ e’s critical value.
L + k ≥ 0. Thus c(L) < +∞. Since k → A
L+k
(γ) is increasing for any
γ, we have that
c(L) = inf¦ k ∈ R[ A
L+k
(γ) ≥ 0 ∀ closed curve γ ¦.
21.1 Proposition.
1. (a) For k < c(L), Φ
k
(x, y) = −∞ for all x, y ∈ M.
(b) For k ≥ c(L), Φ
k
(x, y) ∈ R for all x, y ∈ M.
2. For k ≥ c(L), Φ
k
(x, z) ≤ Φ
k
(x, y) + Φ
k
(y, z), ∀ x, y, z ∈ M.
3. Φ
k
(x, x) = 0, ∀ x ∈ M.
4. Φ
k
(x, y) + Φ
k
(y, x) ≥ 0 ∀ x, y ∈ M.
For k > c(L), Φ
k
(x, y) + Φ
k
(y, x) > 0 if x = y.
5. For k ≥ c(L) the action potential Φ
k
is Lipschitz.
21.2 Remark. The action potential Φ
k
is not symmetric in general,
but items 2, 3, 4 imply that
d
k
(x, y) = Φ
k
(x, y) + Φ
k
(y, x)
is a metric for k > c(L) and a pseudometric for k = c(L) [i.e. perhaps
d
c
(x, y) = 0 for some x = y and c = c(L)].
Proof:
2. We ﬁrst prove 2 for all k ∈ R. Since Φ
k
(x, y) ∈ R ∪ ¦−∞¦, the
inequality in item 2 makes sense for all k ∈ R. If γ ∈ ((x, y),
η ∈ ((y, z), then γ ∗ η ∈ ((x, z) and hence
Φ
k
(x, z) ≤ A
L+k
(γ ∗ η) ≤ A
L+k
(γ) +A
L+k
(η).
Taking the inﬁma on γ ∈ ((x, y) and η ∈ ((y, z), we obtain 2.
24
21. the action potential and the critical value. 25
1. (a) If γ is a closed curve with A
L+k
(γ) < 0 and γ(0) = z, then
Φ
k
(z, z) ≤ lim
N→∞
A
L+k
(γ ∗
N
∗ γ) = lim
N
N A
L+k
(γ) = −∞.
For x, y ∈ M, item 2 implies that
Φ
k
(x, y) ≤ Φ
k
(x, z) + Φ
k
(z, z) + Φ
k
(z, y) = −∞.
Since the function k → Φ
k
(x, y) is increasing, then item 1(a) fol
lows.
(b) Conversely, if Φ
k
(x, y) = −∞ for some k ∈ R and x, y ∈ M,
then
Φ
k
(x, x) ≤ Φ
k
(x, y) + Φ
k
(y, x) = −∞.
Thus there is γ ∈ ((x, x) with A
L+k
(γ) < 0. Then k ≤ c(L). Ob
serve that the set ¦ k ∈ R[ A
L+k
(γ) < 0 for some closed curve γ ¦
is open. Hence Φ
k
(x, y) = −∞ actually implies that k < c(L).
This proves item 1(b).
3. Let k ∈ R by the boundedness condition there exists Q > 0 be
such that
[L(x, v) +k[ ≤ Q for [v[ ≤ 2. (2.1)
Now let γ : [0, ε] → M be a diﬀerentiable curve with [ ˙ γ[ ≡ 1 and
γ(0) = x. Then
Φ
k
(x, x) ≤ Φ
k
(x, γ(ε)) + Φ
k
(γ(ε), x)
≤ A
L+k
γ[
[0,ε]
+A
L+k
γ(t −ε)[
[0,ε]
≤ 2 Qε.
Letting ε →0 we get that Φ
k
(x, x) ≤ 0. But the deﬁnition of c(L)
and the monotonicity of k →Φ
k
(x, x) imply that Φ
k
(x, x) ≥ 0 for
all k ≥ c(L).
5. Let k ≥ c(L). Given x
1
, x
2
∈ M we have that
Φ
k
(x
1
, x
2
) ≤ A
L+k
(γ) ≤ Q d
M
(x
1
, x
2
),
25
26 2. ma˜ n´ e’s critical value.
where γ : [0, d(x
1
, x
2
)] → N is a unit speed minimizing geodesic
joining x
1
to x
2
and Q > 0 is from (2.1). If y
1
, y
2
∈ M, then the
triangle inequality implies that
Φ
k
(x
1
, y
1
) −Φ
k
(x
2
, y
2
) ≤ Φ
k
(x
1
, x
2
) + Φ
k
(y
2
, y
1
)
≤ Q [d
M
(x
1
, x
2
) +d
M
(y
1
, y
2
)].
Changing the roles of (x
1
, y
1
) and (x
2
, y
2
) we get item 5.
4. The ﬁrst part of item 4 follows from items 2 and 3. Now suppose
that k > c(L), x = y and d
k
(x, y) = 0. Let γ
n
: [0, T
n
] → M,
γ
n
∈ ((x, y) be such that Φ
k
(x, y) = lim
n
A
L+k
(γ
n
). We claim
that T
n
is bounded below.
Indeed, suppose that lim
n
T
n
= 0. Let A > 0, from the superlin
earity there is B > 0 such that L(x, v) ≥ A[v[ −B, ∀ (x, v) ∈ TM.
Then
Φ
k
(x, y) = lim
n
T
N
0
L(γ
n
, ˙ γ
n
) +k
≥ lim
n
A
[ ˙ γ[ + (k −B) T
n
= A d
M
(x, y)
Letting A →+∞ we get that Φ
k
(x, y) = +∞ which is false.
Now let η
n
: [0, S
n
] → M, η
n
∈ ((y, x) with lim
n
A
L+k
(η
n
) =
Φ
k
(y, x). Choose 0 < T < liminf
n
T
n
and 0 < S < liminf
n
S
n
.
Then for c = c(L) < k,
Φ
c
(x, x) ≤ lim
n
A
L+c
(γ
n
∗ η
n
)
≤ lim
n
A
L+k
(γ
n
) + (c −k)T +A
L+k
(η
n
) + (c −k)S
≤ lim
n
Φ
k
(x, y) + Φ
k
(y, x) + (c −k)(T +S)
≤ (c −k)(T +S) < 0,
which contradicts item 3.
26
22. continuity of the critical value. 27
22 Continuity of the critical value.
22.1 Lemma. The function C
∞
(M, R) ÷ ψ →c(L +ψ) is continuous
in the topology induced by the supremum norm.
Proof: Suppose that ψ
n
→ψ and let c
n
:= c(L+ψ
n
) and c := c(L+ψ).
We will prove that c
n
→c.
Fix ε > 0. Since c − ε < c, by the deﬁnition of critical value there
exists a closed curve γ : [0, T] → M such that A
L+ψ+c−ε
(γ) < 0, hence
for all n suﬃciently large
A
L+ψ
n
+c−ε
(γ) < 0.
Therefore for n suﬃciently large c −ε < c
n
, and thus c −ε ≤ liminf
n
c
n
.
Since ε was arbitrary we have that c ≤ liminf
n
c
n
.
We show now that limsup
n
c
n
≤ c. Suppose that c < limsup
n
c
n
.
Take ε such that
c < c +ε < limsup
n
c
n
. (2.2)
Since ψ
n
→ψ, there exists n
0
such that for all n ≥ n
0
,
−ε ≤ ψ −ψ
n
≤ ε. (2.3)
By (2.2), there exists m ≥ n
0
such that
c < c +ε < c
m
.
By the deﬁnition of critical value there exists a closed curve γ : [0, T] →
M such that
A
L+ψ
m
+c+ε
(γ) < 0,
and hence using (2.3) we have
A
L+ψ+c
(γ) ≤ A
L+ψ
m
+c+ε
(γ) < 0,
which yields a contradiction to the deﬁnition of the critical value c.
This proof also shows that L → c(L) is continuous if we endow the
set of lagrangians L with the topology induced by the supremum norm
on compact subsets of TM.
27
28 2. ma˜ n´ e’s critical value.
23 Holonomic measures.
Let C
0
be the set of continuous functions f : TM → R having linear
growth, i.e.
f
:= sup
(x,v)∈TM
[f(x, v)[
1 +v
< +∞.
Let ´
be the set of Borel probabilities µ on TM such that
TM
v dµ < +∞,
endowed with the topology such that lim
n
µ
n
= µ if and only if
lim
n
f dµ
n
=
f dµ (2.4)
for all f ∈ C
0
.
Let (C
0
)
the dual of C
0
. Then ´
is naturally embedded in (C
0
)
and its topology coincides with that induced by the weak* topology on
(C
0
)
.
We shall see that this topology is metrizable. Let ¦f
n
¦ be a sequence
of functions with compact support on C
0
which is dense on C
0
in the
topology of uniform convergence on compact sets of TM. Deﬁne a metric
d(, ) on ´
by
d(µ
1
, µ
2
) =
[v[ dµ
1
−
[v[ dµ
2
+
¸
n
1
2
n
1
c
n
f
n
dµ
1
−
f
n
dµ
2
(2.5)
where c
n
= sup
(x,v)
[f
n
(x, v)[.
23.1 Exercises:
1. Construct µ ∈ ´
such that
[v[
2
dµ = +∞.
2. Show that the ﬁrst term in (2.5) is necessary.
28
23. holonomic measures. 29
23.2 Proposition.
The metric d(, ) induces the weak* topology on ´
⊂ (C
0
)
.
Proof: We prove that d(, ) generates the weak* topology on ´
. Sup
pose that
f dµ
n
→
f dµ, ∀f ∈ C
0
.
Given ε > 0, choose M > 0 such that
¸
m≥M
1
2
m
2 < ε, and choose
N > 0 such that
f
m
dµ
n
−
f
m
dµ
< ε, for 0 ≤ m ≤ M, n ≥ N;
[v[ dµ
n
−
[v[ dµ
< ε, for n ≥ N.
Since
f
n
∞
c
n
= 1, then for n > N we have that
d(µ
n
, µ) ≤ ε +
M
¸
m=1
1
2
m
ε +
¸
m≥M+1
1
2
m
2
f
m
c
m
= 3 ε.
Thus d(µ
n
, µ) →0.
Now suppose that d(µ
n
, µ) →0. Let K
m
be compact sets such that
K
m
⊂ K
m+1
and that TM = ∪K
m
. Then
K
m
f dµ
n
−→
K
m
f dµ, ∀f ∈ C
0
, ∀m;
[v[ dµ
n
−→
[v[ dµ.
This implies that
lim
n→∞
TM−K
m
[v[ dµ
n
=
TM−K
m
[v[ dµ, ∀ m. (2.6)
Given ε > 0, choose m(ε) > 0 such that
TM−K
m(ε)
(1 +[v[) dµ <
ε
4
,
29
30 2. ma˜ n´ e’s critical value.
and N such that
TM−K
m(ε)
(1 +[v[) dµ
n
<
ε
2
, ∀n > N.
Fix f ∈ C
0
. Choose N > 0 such that
K
m(ε)
f dµ
n
−
K
m(ε)
f dµ
< ε, ∀ n > N.
Then
TM−K
m(ε)
[f[ dµ
n
≤ f
TM−K
m(ε)
(1+[v[) dµ
n
≤ f
ε
2
, ∀ n > N.
Using a similar estimate for µ we obtain that
f dµ
n
−
f dµ
≤ ε +f
(
ε
2
+
ε
4
).
If γ : [0, T] →M is a closed absolutely continuous curve, let µ
γ
∈ ´
be deﬁned by
f dµ
γ
=
1
T
T
0
f
γ(t), ˙ γ(t)
dt
for all f ∈ C
0
. Observe that µ
γ
∈ ´
because if γ is absolutely contin
uous then
[ ˙ γ(t)[ dt < +∞. Let ((M) be the set of such µ
γ
’s and let
((M) be its closure in ´
. Observe that the set ((M) is convex. We
call ((M) the set of holonomic measures on M.
30
24. invariance of minimizing measures. 31
24 Invariance of minimizing measures.
Given a Borel probability measure µ in TM deﬁne its action by
A
L
(µ) =
TM
L dµ.
Since by the superlinearity the lagrangian L is bounded below, this
action is well deﬁned. Observe that L / ∈ C
0
and that A
L
(µ) = +∞ for
some µ ∈ ´
(see exercise 23.1).
Let ´(L) be the set of ϕ
t
invariant probabilities on TM.
24.1 Theorem (Ma˜ n´e [38], prop. 1.1, 1.3, 1.2).
1. ´(L) ⊆ ((M) ⊆ ´
.
2. If µ ∈ ((M) satisﬁes
A
L
(µ) = min¦ A
L
(ν) [ ν ∈ ((M) ¦,
then µ ∈ ´(L).
3. If M is compact and a ∈ R, then the set ¦ µ ∈ ((M) [ A
L
(µ) ≤ a ¦
is compact.
Observe that item 3 implies the existence of a minimizer as in item 2.
The inclusion ´(L) ⊆ ((M) follows from Birkhoﬀ’s ergodic theorem
and the fact that ((M) is convex. Taking f = v in equation (2.4) we
see that ´
is closed, so that ((M) ⊆ ´
.
Proof of item 24.1.3:
Since ((M) is closed, it is enough to prove that the set
/(a) :=
¸
µ ∈ ´
[ A
L
(µ) ≤ a
¸
is compact in ´
. First we prove that /(a) is closed. Let k > 0 and
deﬁne L
k
:= min¦L, k¦. Let
B
k
:=
¸
µ ∈ ´
[
L
k
dµ ≤ a
¸
.
31
32 2. ma˜ n´ e’s critical value.
Since L
k
∈ (
0
, then B
k
is closed in ´
. Since /(a) = ∩
k>0
B
k
, then
/(a) is closed.
In order to prove the compactness, consider a sequence ¦µ
n
¦ ⊂ /(a).
Applying the Riesz’ theorem B.1, taking a subsequence we can assume
that there exists a measure µ on the Borel σalgebra of TM such that
f
i
dµ
n
−→
f
i
dµ, (2.7)
for every f
i
in the sequence used for the deﬁnition of d(, ). Approxi
mating the function 1 by the functions f
i
we see that µ is a probability.
Approximating L
k
by functions f
i
we have that
L
k
dµ = lim
n
L
k
dµ
n
≤ liminf
n
L dµ
n
≤ a.
Letting k ↑ +∞, by the monotone convergence theorem, we get that
A
L
(µ) ≤ a. (2.8)
Let B > 0 be such that [v[ < L(x, v) +B for all (x, v) ∈ TM. Then
[v[ dµ ≤ A
L
(µ) +B ≤ a +B < +∞. (2.9)
So that µ ∈ ´
.
We now prove that lim
n
[v[ dµ
n
−→
[v[ dµ. Let ε > 0. By
adding a constant we may assume that L > 0. Choose r > 0 such that
L(x, v) > a ε
−1
[v[ for all [v[ > r. Then
v>r
[v[ dµ
n
≤
ε
a
v>r
L dµ
n
≤
ε
a
L dµ
n
≤ ε.
Similarly, by (2.8),
v>r
[v[ dµ ≤ ε.
32
24. invariance of minimizing measures. 33
From (2.7) we obtain that there is N > 0 such that
v≤r
[v[ dµ −
v≤r
[v[ dµ
n
< ε, for n > N.
Adding these inequalities we get that
[v[ dµ
n
−
[v[ dµ
≤ 3 ε.
The proof of item 24.1.2 requires some preliminary results which we
present now. Item 24.1.2 is proved at the end of the section.
The following proposition is needed to show that the minimum of
the action in ((M) is the same as the minimum on ((M).
24.2 Proposition.
Given µ ∈ ((M), there are µ
η
n
∈ ((M) such that µ
η
n
→µ and
lim
n
L dµ
η
n
=
L dµ.
24.3 Remark. The statement of proposition 24.2 is not trivial. It is
easy to see that the function A
L
: ((M) → R is always lower semicon
tinuous (see the last argument of the proof of 24.2), but in general it
is not continuous. It is possible to give a sequence µ
γ
n
∈ ((M) such
that µ
γ
n
→ µ in ((M) but liminf
n
A
L
(µ
γ
n
) > A
L
(µ) for a quadratic
lagrangian L.
This can be made by calibrating the high speeds in γ
n
so that
[v>R]
[v[ dµ
γ
n
→ 0 but a := liminf
n
[v>R]
Ldµ
γ
n
> 0. Then the
limit measure µ will have support on [[v[ ≤ R] and “will not see” the
remnant a of the action.
33
34 2. ma˜ n´ e’s critical value.
Proof: Let A > 1 and let γ : [0, T] → M be a closed absolutely con
tinuous curve. We reparametrize γ to a curve η : [0, S] → M such that
˙ η = ˙ γ when [ ˙ γ[ < A and ˙ η =
˙ γ
 ˙ γ
A when [ ˙ γ[ > A. So that [ ˙ η[ ≤ A. Write
η(s(t)) = γ(t), w(s) = [ ˙ η(s)[ and v(t) = [ ˙ γ(t)[. We want
s(t)
0
w(s) ds =
t
0
v(t) dt,
so that
s
(t) =
v(t)
w(s(t))
=
1 when v(t) ≤ A.
v(t)
A
when v(t) ≥ A.
Then
S(T) =
[v(t)≤A]
dt +
[v(t)≥A]
v(t)
A
dt,
S(T)
T
= µ
γ
([[v[ ≤ A]) +
[v≥A]
[v[
A
dµ
γ
,
S(T)
T
−1
≤
[v>A]
dµ
γ
+
[v≥A]
[v[
A
dµ
γ
≤ 2
[v>A]
[v[ dµ
γ
. (2.10)
Suppose that f : TM → R is µ
η
integrable. Since
ds
dt
=
v(t)
A
when
v(t) > A then
[ ˙ γ(t(s))>A]
f(η(s),
˙ γ(t(s))
 ˙ γ(t(s))
A) ds =
[ ˙ γ(t)>A]
f
γ(t),
˙ γ(t)
 ˙ γ(t)
A
[ ˙ γ(t)[
A
dt.
34
24. invariance of minimizing measures. 35
Then
f dµ
η
=
1
S(T)
f(η(s), ˙ η(s)) ds
=
1
S(T)
¸
[ ˙ γ≤A]
f(γ(t), ˙ γ(t)) dt +
[ ˙ γ(t(s))>A]
f
η(s),
˙ γ(t(s))
 ˙ γ(t(s))
A
ds
¸
=
T
S(T)
¸
[v≤A]
f(v) dµ
γ
(v) +
[v>A]
f
v
v
A
[v[
A
dµ
γ
¸
For A > 1 big enough,
[v>A]
[v[ dµ
γ
< ε <
1
4
. (2.11)
Deﬁne
f
A
(v) :=
f(v) if [v[ ≤ A,
f
v
v
A
v
A
if [v[ > A.
Then
f dµ
η
=
T
S(T)
f
A
dµ
γ
. (2.12)
Observe that from (2.10) and (2.11), we have that
T
S(T)
−1
≤ 4 ε. (2.13)
Then
f dµ
η
−
f
A
dµ
γ
≤
T
S(T)
−1
[f
A
[ dµ
γ
≤ 4 ε
[f
A
[ dµ
γ
.
(2.14)
If f
∞
≤ 1, then
[f
A
[ dµ
γ
≤
[v≤A]
[f[ dµ
γ
+
[v>A]
[f −f
A
[ dµ
γ
≤ 1 +
[v>A]
[v[
A
dµ
γ
≤ 1 +ε.
35
36 2. ma˜ n´ e’s critical value.
f dµ
η
−
f dµ
γ
=
T
S(T)
f
A
dµ
γ
−
f dµ
γ
≤
T
S(T)
−1
[f
A
[ dµ
γ
+
[f −f
A
[ dµ
γ
≤ 4 ε (1 +ε) +ε ≤ 6 ε.
Also, using (2.14),
[v[ dµ
η
−
[v[ dµ
γ
=
T
S(T)
−1
[v[ dµ
γ
≤ 4 ε
[v[ dµ
γ
.
Hence
d
M
(µ
η
, µ
γ
) ≤ 6 ε
([v[ + 1) dµ
γ
. (2.15)
Now let µ ∈ ((M). Let
K := 1 +
([v[ + 1) dµ. (2.16)
For R > 0, deﬁne
L
R
(v) :=
L(v) if [v[ ≤ R.
L
v
v
R
v
R
if [v[ > R.
Claim: If E(v) > 0 for all [v[ ≥ R, then
L
R
(v) ≤ L(v) for all v ∈ TM.
Proof:
If [v[ ≤ R then L
R
(v) = L(v). Suppose that [v[ = R. For s ≥ 1 let
f(s) := L(sv) −L
R
(sv) = L(sv) −sL(v).
It is enough to prove that f(s) ≥ 0 for all s ≥ 1. We have
f
(s) = v L
v
(sv) −L(v)
f
(s) = v L
vv
(sv) v > 0.
36
24. invariance of minimizing measures. 37
We have that f(1) = 0, f
(1) = E(v) > 0, f
(s) > 0 for all s ≥ 1. This
implies that f(s) ≥ 0 for all s ≥ 1.
♦
Given N > 0, choose R = R(N) > 1 such that
E(v) > 0 if [v[ ≥ R and
v>R
[v[ dµ <
1
N
. (2.17)
Observe L
R(N)
has linear growth. Choose µ
γ
N
∈ ((M) such that
d
M
(µ
γ
N
, µ) <
1
N
, (2.18)
L
R(N)
dµ
γ
N
≤
L
R(N)
dµ +
1
N
,
and
[v≤R(N)]
[v[ dµ
γ
N
≥
[v≤R(N)]
[v[ dµ −
1
N
. (2.19)
Then
([v[ + 1) dµ
γ
N
≤ K from (2.16) and (2.18), (2.20)
[v>R(N)]
[v[ dµ
γ
N
<
3
N
from (2.17), (2.18) and (2.19). (2.21)
Construct η
N
as above for γ
N
and A = R(N). Then from (2.11), (2.15),
(2.20) and (2.21), d
M
(µ
η
N
, µ
γ
N
) <
18
N
K. From (2.18),
d
M
(µ
η
N
, µ) <
18
N
K +
1
N
.
Thus µ
η
N
N
−→µ in ((M). Moreover, from (2.12), (2.13), (2.11) and the
claim,
L dµ
η
N
=
T
N
S(T
N
)
L
R(N)
dµ
γ
N
≤
T
N
S(T
N
)
¸
L
R(N)
dµ +
1
N
≤
1 +
12
N
¸
L dµ +
1
N
.
37
38 2. ma˜ n´ e’s critical value.
Hence
limsup
N
L dµ
η
N
≤
L dµ.
Fix R > 0 such that E > 0 on [v[ > R. Then L
R
has linear growth
and by the claim L
R
≤ L. Therefore
liminf
N
L dµ
η
N
≥ lim
N
L
R
dµ
η
N
=
L
R
dµ.
Letting R ↑ +∞, by the dominated convergence theorem we get that
liminf
N
L dµ
η
N
≥
L dµ.
Given x, y ∈ M, deﬁne
S(x, y; T) := inf
γ∈C
T
(x,y)
A
L
(γ).
Observe that S(x, y; T) > −∞ because L is bounded below. If γ ∈
(
ac
([0, T], M), deﬁne
S
+
(γ) := A
L
(γ) −S(γ(0), γ(T); T).
The absolutely continuous curves γ with S
+
(γ) = 0 are called Tonelli
minimizers. Observe that a Tonelli minimizer is a solution of (EL).
Given γ
1
, γ
2
∈ (
ac
([0, T], M), the absolutely continuous distance
d
1
(γ
1
, γ
2
) is deﬁned by
d
1
(γ
1
, γ
2
) := sup
t∈[0,T]
d
γ
1
(t), γ
2
(t)
+
T
0
d
TM
[γ
1
(t), ˙ γ
1
(t)], [γ
2
(t), ˙ γ
2
(t)]
dt.
24.4 Proposition. Given a compact subset K ⊆ M and given C, ε > 0
there exist δ > 0 such that if γ : [0, T] →M is absolutely continuous and
satisﬁes
38
24. invariance of minimizing measures. 39
i. 1 ≤ T ≤ C.
ii. A
L
(γ) ≤ C.
iii. S
+
(γ) ≤ δ.
Then either γ([0, T]) ∩ K = or there exists a Tonelli minimizer γ
0
:
[0, T] →M such that d
1
(γ
0
, γ) ≤ ε.
Proof: If such δ does not exists then there is a sequence γ
n
∈
(
ac
([0, T
n
], M) such that γ
n
([0, T
n
]) ∩K = , 1 ≤ T
n
≤ C, S
+
(γ
n
) →0,
A
L
(γ
n
) ≤ C and d
1
(γ
n
, η) ≥ ε for any Tonelli minimizer η.
Adding a constant we can assume that L > 0. Let B > 0 be such
that L(x, v) > [v[ −B for all (x, v) ∈ TM. Choose s
0
∈ [0, T
n
] such that
γ
n
(s
0
) ∈ K. Then
d(K, γ
n
(t)) ≤ d(γ
n
(s
0
), γ
n
(t)) ≤
[s
0
,t]
[ ˙ γ
n
[
≤
[s
0
,t]
L(γ
n
, ˙ γ
n
) +B
≤ C +BC.
Let Q := ¦ y ∈ M[ d(y, K) ≤ C+BC ¦. Then we have that γ
n
([0, T
n
]) ⊆
Q.
We can assume that T
n
→ T, γ
n
(0) → x ∈ Q and γ
n
(T
n
) → y ∈ Q.
Moreover, we can assume that T
n
≡ T, γ
n
(0) ≡ x and γ
n
(T) ≡ y. By
theorem 31.2, the set /(b) =
¸
γ ∈ (
T
(x, y) [ A
L
(γ) ≤ b
¸
is compact in
the C
0
topology. Then we can assume that there is γ
0
∈ (
T
(x, y) such
that γ
n
→ γ
0
in the C
0
topology. Since the action functional is lower
semicontinuous, then A
L
(γ
0
) ≤ liminf
n
A
L
(γ
n
) = S(x, y; T), because
S
+
(γ
n
) → 0. Thus γ
0
is a Tonelli minimizer. Moreover, we have that
A
L
(γ
n
) →A
L
(γ
0
). By proposition 31.3, γ
n
→γ
0
in the d
1
topology.
Let
H :=
¸
h : TM →R
f
∞
≤ 1, [h]
Lip
≤ 1, h with compact support
¸
,
39
40 2. ma˜ n´ e’s critical value.
where
[h]
Lip
= sup
(x,v)=(y,w)
[h(x, v) −h(y, w)[
d
TM
(x, v), (y, w)
is the smallest Lipschitz constant for h.
24.5 Corollary.
Given h ∈ H and C > 0 there exist δ = δ(C, h) > 0 such that if
γ : [0, T] →M satisﬁes conditions 24.4.i, 24.4.ii, 24.4.iii then
γ
h −
γ
h ◦ ϕ
1
≤ 5. (2.22)
Proof: Let K = π
supp(h) ∪ ϕ
−1
(supp(h))
. Given C > 0 and ε > 0
let δ = δ(C, ε) > 0 and A > 0 be given by proposition 24.4 then if
γ : [0, T] → M satisﬁes conditions 24.4.i, 24.4.ii, 24.4.iii we have
that either γ([0, T]) ∩ K = , or we can take γ
0
minimizing such that
d
1
(γ
0
, γ) ≤ ε.
Observe that if γ([0, T])∩K = , then h(γ, ˙ γ) ≡ 0 and h◦ϕ
1
(γ, ˙ γ) ≡
0. This implies (2.22). Suppose then that d
1
(γ
0
, γ) ≤ ε.
We have that
γ
h −
γ
0
h
≤ [h]
Lip
d
1
(γ, γ
0
) ≤ 1 1 ε,
where [h]
Lip
is the smallest Lipschitz constant of h. Let Q(h) :=
ϕ
−1
(supp(h)), then
γ
h ◦ ϕ
1
−
γ
0
h ◦ ϕ
1
≤ [h]
Lip
[ϕ
1
[
Q(h)
]
Lip
d
1
(γ, γ
0
) ≤ 1 [ϕ
1
[
Q(h)
]
Lip
ε.
Since γ
0
is a solution of (EL), we have that
γ
0
h −
γ
0
h ◦ ϕ
1
=
T
0
h
γ
0
(t), ˙ γ
0
(t)
−h
γ
0
(t + 1), ˙ γ
0
(t + 1)
dt
≤
1
0
[h(γ
0
, ˙ γ
0
)[ dt +
T+1
T
[h(γ
0
, ˙ γ
0
)[ dt ≤ 2.
40
24. invariance of minimizing measures. 41
Hence
γ
h −
γ
h ◦ ϕ
1
≤ ε
1 + [ϕ
1
[
Q(h)
]
Lip
+ 2.
Proof of item 24.1.2:
Observe that to prove that µ is invariant it is enough to prove that
h dµ =
h d(ϕ
∗
1
µ) for all h ∈ H. (2.23)
By proposition 24.2, there exists a sequence µ
γ
n
∈ ((M) such that
µ
γ
n
→µ and
lim
n
A
L
(µ
γ
n
) = A
L
(µ) = min
¸
A
L
(ν) [ ν ∈ ((M)
¸
=: k. (2.24)
Let T
n
be a period of the curve γ
n
: R → M. Take an integer N > 0.
By joining a constant curve if necessary, we can assume that every T
n
is a multiple of N and that lim
n→∞
T
n
= +∞. Given C > 0 let
B
n
(C) :=
¸
j ∈ N[ 1 ≤ j ≤
T
n
N
, A
L
(γ
n,j
) ≥ C
¸
,
where
γ
n,j
:= γ
n
[
[jN,(j+1)N]
.
By the superlinearity L is bounded below, adding a constant we can
assume that L > 0. Then we can assume that
A
L
(µ
γ
n
) =
1
T
n
T
n
0
L(γ
n
, ˙ γ
n
) dt ≤ 2 k ∀n.
Hence
2 k T
n
≥
¸
j∈B
n
(C)
A
L
(γ
n,j
) ≥ C #B
n
(C).
Thus
#B
n
(C)
T
n
≤
2k
C
. (2.25)
41
42 2. ma˜ n´ e’s critical value.
Given δ > 0, let
B
n
(δ) :=
¸
j ∈ N[ 1 ≤ j ≤
T
n
N
−1, S
+
(γ
n,j
) > δ
¸
.
Then
S
+
(γ
n
) ≥
(T
n
/N)−1
¸
j=1
S
+
(γ
n,j
) ≥ δ #B
n
(δ).
Moreover,
k ≤
1
T
n
S(γ
n
(0), γ(T
n
); T
n
) = A
L
(µ
γ
n
) −
1
T
n
S
+
(γ
n
).
Hence
S
+
(γ
n
) ≤ T
n
A
L
(µ
γ
n
) −k
.
Therefore
#B
n
(δ)
T
n
≤
1
δ
A
L
(µ
γ
n
) −k
. (2.26)
Now ﬁx h ∈ H. Then
h dµ
γ
n
−
h d(ϕ
∗
1
µ
γ
n
)
≤
1
T
n
(T
n
/N)−1
¸
j=0
γ
n,j
h −
γ
n,j
h ◦ ϕ
1
.
Denote B
n
:= B
n
(C) ∪ B
n
(δ). Since sup [h[ ≤ 1, then
h dµ
γ
n
−
h d(ϕ
∗
1
µ
γ
n
)
≤
1
T
n
¸
j / ∈B
n
γ
n,j
h −
γ
n,j
h ◦ ϕ
1
+
1
T
n
2N #B
n
.
Now choose C ≥ N
2
and δ = δ(C, h) > 0 from corollary 24.5. Using
equations (2.25), (2.26) and corollary 24.5 we obtain that
h dµ
γ
n
−
h d(ϕ
∗
1
µ
γ
n
)
≤
5
T
n
T
n
N
−#B
n
+
1
T
n
2N #B
n
≤
5
N
+ 2N
2k
C
+
1
δ
A
L
(µ
γ
n
) −k
.
42
24. invariance of minimizing measures. 43
Now let n →∞. Using equation (2.24) and that C ≥ N
2
, µ
γ
n
→µ and
h, h ◦ ϕ
1
∈ C
0
(because they have compact support), we obtain that
h dµ −
h d(ϕ
∗
1
µ)
≤
5
N
+
4k
N
.
Since N is arbitrary, this diﬀerence is zero and we get (2.23).
43
44 2. ma˜ n´ e’s critical value.
25 Ergodic characterization of the critical
value.
Given a Borel probability measure µ in TM deﬁne its action by
A
L
(µ) =
TM
L dµ.
Since by the superlinearity the lagrangian L is bounded below, this
action is well deﬁned.
Let ´(L) be the set of ϕ
t
invariant probabilities on TM.
25.1 Theorem (Ma˜ n´e [35]). If M is compact, then
c(L) = −min¦ A
L
(µ) [ µ ∈ ´(L) ¦.
We will obtain theorem 25.1 from theorem 25.2 below, which also
applies to the noncompact case. If M is noncompact, theorem 25.1
may not hold, as seen in example 57.
Recall that if γ : [0, T] →M is a closed absolutely continuous curve,
the measure µ
γ
∈ ´
is deﬁned by
f dµ
γ
=
1
T
T
0
f
γ(t), ˙ γ(t)
dt
for all f ∈ C
0
, and that ((M) is the closure of the set of such µ
γ
’s in
´
.
25.2 Theorem.
c(L) = −inf¦ A
L
(µ) [ µ ∈ ((M) ¦
= −inf¦ A
L
(µ) [ µ ∈ ((M) ¦.
(2.27)
25.3 Deﬁnition. We say that a holonomic measure µ ∈ ((M) is (glob
ally) minimizing if A
L
(µ) = −c(L).
44
25. ergodic characterization of the critical value. 45
25.4 Remarks.
1. Recall that by theorem 24.1, if a minimizing measure exists, then
it is invariant under the lagrangian ﬂow.
2. The equality between the two inﬁma in theorem 25.2 is nontrivial
and follows from proposition 24.2.
3. Theorems 25.2 and 24.1 imply theorem 25.1.
4. If p : N → M is a covering, M is compact and L = L ◦ dp is the
lifted lagrangian, then theorems 25.2 and 24.1 imply that
c(L) = −min¦ A
L
(µ) [ µ ∈ ´(L) ∩ dp
∗
((N) ¦, (2.28)
by noticing that A
L
(dp
∗
ν) = A
L
(ν) for ν ∈ ((N). Here dp
∗
((N) is
the set of probabilities µ
γ
on TM where γ is a curve on M whose
lifts to N are closed. The compactness property on theorem 2
4.1(3) allows to obtain a minimum on (2.28) instead of the inﬁmum
on (2.27) which may not be attained in the noncompact case.
5. The statement for coverings in equation (2.28) allows to obtain
minimizing measures which don’t appear in the Mather’s theory.
For example if c
u
is the critical value of the universal cover
M of
M and c
0
is the critical value of the abelian cover
´
M of M; the
minimizing measures on a ﬁxed homology class (corresponding to
Mather’s theory) all have action A(µ) ≥ −c
0
, (see equation (2.31)
and proposition 27.3), while the minimizing measures for
M have
action c
u
< c
0
.
The measures for
M correspond to “minimizing in the zero homo
topy class” while the measures for
´
M are minimizing in the zero
homology class.
The drawback of this approach is that we obtain honest minimizing
invariant measures on TM which may not lift to ﬁnite measures
on the covering TN.
45
46 2. ma˜ n´ e’s critical value.
Proof of theorem 25.2:
If µ
γ
∈ ((M), then A
L+c(L)
(µ
γ
) ≥ 0. Hence A
L
(µ
γ
) ≥ −c(L). Thus
−c(L) ≤ inf¦ A
L
(µ) [ µ ∈ ((M) ¦ = inf¦ A
L
(µ) [ µ ∈ ((M) ¦,
where the last equality follows from proposition 24.2.
If k < c(L) then there is a closed absolutely continuous curve γ on
M such that A
L+k
(γ) < 0. Thus µ
γ
∈ ((M) and
−k > A
L
(µ
γ
) ≥ inf¦ A
L
(µ) [ µ ∈ ((M)¦.
Now let k ↑ c(L).
46
26. the aubrymather theory. 47
26 The AubryMather Theory.
Through this section we shall assume that M is compact.
26.a Homology of measures.
Observe that since M is compact, any 1form is in C
0
. By deﬁnition,
an holonomic probability µ ∈ ((M) satisﬁes
TM
[v[ dµ < +∞ and
TM
df dµ = 0 for all f ∈ C
∞
(M, R).
Then we can deﬁne its homology class as ρ(µ) ∈ H
1
(M, R) ≈ H
1
(M, R)
∗
by
ρ(µ) , [ω]
=
TM
ω dµ, (2.29)
for any closed 1form ω on M, where [ω] ∈ H
1
(M, R) is the coho
mology class of ω. Here we have used the identiﬁcation
1
H
1
(M, R) ≈
H
1
(M, R)
∗
and equation (2.29) shows how the homology class ρ(µ) acts
on H
1
(M, R). Since µ is holonomic, the integral in (2.29) depends only
on the cohomology class of ω. The class ρ(µ) is called the homology of
µ or the rotation of µ by analogy to the twist map theory.
Using a ﬁnite basis ¦ [ω
1
], . . . , [ω
k
] ¦ for H
1
(M, R) and the topology
of ((M), we have that
26.1 Lemma. The map ρ : ((M) →H
1
(M, R) is continuous.
26.b The asymptotic cycle.
Given a diﬀerentiable ﬂow ϕ
t
on a compact manifold N and a ϕ
t

invariant probability µ, the Schwartzman’s [65] asymptotic cycle of an
1
In fact, H
1
(M, R) ≈ hom
`
H
1
(M, R), R
´
= H
1
(M, R)
∗
by the universal coeﬃcient
theorem. Since M is compact, then H
1
(M, R) is a ﬁnite dimensional vector space and
hence it is naturally isomorphic to its double dual H
1
(M, R)
∗
.
47
48 2. ma˜ n´ e’s critical value.
invariant probability µ is deﬁned to be the homology class A(µ) ∈
H
1
(N, R) ≈ H
1
(N, R)
∗
such that
A(µ) , [ω]
=
N
ω(X) dµ,
for any closed 1form ω, where [ω] ∈ H
1
(N, R) is the cohomology class
of ω and X is the vector ﬁeld of ϕ
t
. This integral depends only on
the cohomology class of ω because the integral of a coboundary by an
invariant measure is zero: in fact, if df is an exact 1form, then deﬁne
F(y) := lim
T→+∞
1
T
T
0
df
X(ϕ
t
y)
dt = lim
T→+∞
1
T
f(ϕ
t
y) −f(y)
= 0,
by Birkhoﬀ’s theorem,
N
df(X) dµ =
N
F dµ = 0.
If µ is ergodic and x ∈ N is a generic point
2
for µ, then
A(µ) , [ω]
= lim
T→+∞
1
T
T
0
ω
X(ϕ
t
x)
dt .
Applying this to a basis ¦ω
1
, . . . , ω
k
¦ for H
1
(N, R), we get that
A(µ) = lim
T→+∞
1
T
[ γ
T
∗ δ
T
] ∈ H
1
(N, R),
where γ
T
(t) = ϕ
t
(x), t ∈ [0, T], the curve δ
T
is a unit speed geodesic
from ϕ
T
(x) to x, and the limit is on the ﬁnite dimensional vector space
H
1
(N, R).
In the case of a lagrangian ﬂow, the phase space N = TM is not com
pact, but it has the same homotopy type as the conﬁguration space M
because M is a deformation retract of TM (contracting TM along the
2
i.e. lim
T→+∞
1
T
R
T
0
f(ϕ
t
x) dt =
R
f dµ for all f ∈ C
0
(N, R).
48
26. the aubrymather theory. 49
ﬁbers to the zero section M 0). Moreover, the ergodic components
of an invariant measure of a lagrangian ﬂow are contained in a unique
energy level, which is a compact submanifold of TM by remark 13.1.
We see that the homology of an invariant probability and its asymp
totic cycle coincide under the identiﬁcation H
1
(TM, R)
π
∗
≈ H
1
(M, R).
26.2 Proposition.
π
∗
A(µ)
= ρ(µ) for all µ ∈ ´(L),
where π
∗
: H
1
(TM, R) →H
1
(M, R) is the map induced by the projection
TM
π
→M.
Proof: If ω is a closed 1form on M, then
(π
∗
ω)(X(x, v)) = ω
dπ
X(x, v)
= ω
x
(v) ,
because the lagrangian vector ﬁeld X has the form X(x, v) = (v, ∗).
Then
π
∗
A(µ) , [ω]
=
A(µ) , π
∗
[ω]
=
TM
(π
∗
ω)(X) dµ
=
TM
ω dµ =
ρ(µ) , [ω]
.
26.3 Lemma. The map ρ : ´(L) →H
1
(M, R) is surjective.
Proof: Let h ∈ H
1
(M, Z) be an integer homology class. Let η : [0, 1] →
M be a closed curve with homology class h. Let γ be a minimizer of the
action of L among the set of absolutely continuous curves [0, 1] → M
with the same homotopy class as η. Then by remark 12.2, γ is a periodic
orbit for the lagrangian ﬂow with period 1. The invariant measure µ
γ
satisﬁes ρ(µ
γ
) = h.
The map ρ is aﬃne and ´(L) is convex; hence ρ
´(L)
is con
vex and, in particular, it contains the convex hull of H
1
(M, Z). Thus,
H
1
(M, R) ⊆ ρ
´(L)
.
49
50 2. ma˜ n´ e’s critical value.
26.c The alpha and beta functions.
The action functional A
L
: ´(L) →R is lower semicontinuous
3
and the
sets
´(h) := ¦ µ ∈ ´(L) [ ρ(µ) = h¦
are closed. Hence we can deﬁne the Mather’s beta function β :
H
1
(M, R) →R, as
β(h) := min
µ∈M(h)
A
L
(µ).
We shall prove below that the βfunction is convex. The Mather’s
alpha function α = β
∗
: H
1
(M, R) → R is the convex dual of the
βfunction:
α([ω]) = max
h∈H
1
(M,R)
¸
'[ω] , h` −β(h)
¸
= − min
µ∈M(L)
¸
A
L
(µ) −'[ω] , ρ(µ)`
¸
using 26.3,
= − min
µ∈M(L)
A
L−ω
(µ)
= c(L −ω) , by 25.1. (2.30)
Observe that since L − ω is also a convex superlinear lagrangian, then
α([ω]) is ﬁnite.
26.4 Theorem. The α and β functions are convex and superlinear.
Proof: We ﬁrst prove that β is convex. Let h
1
, h
2
∈ H
1
(M, R) and
0 ≤ λ ≤ 1. Let µ
1
, µ
2
∈ ´(L) be such that ρ(µ
i
) = h
i
and A
L
(µ
i
) =
β(h
i
) for i = 1, 2. The probability ν = λµ
1
+ (1 −λ) µ
2
satisﬁes ρ(ν) =
λh
1
+ (1 −λ) h
2
. Hence
β
λh
1
+(1 −λ) h
2
≤ A
L
λµ
1
+(1 −λ) µ
2
= λβ(h
1
) + (1 −λ) β(h
2
).
By proposition 26.3, ρ is surjective, and hence β is ﬁnite. By propo
sition D.1 on the appendix, α is superlinear. By D.1, α and β are
3
A
L
is lower semicontinuous iﬀ liminf
n
A
L
(ν
n
) ≥ A
L
(µ) when ν
n
→ µ.
50
26. the aubrymather theory. 51
convex. Formula (2.30), implies that α is ﬁnite and then by D.1, β is
superlinear.
For h ∈ H
1
(M, R) and ω ∈ H
1
(M, R), write
´
h
(L) := ¦ µ ∈ ´(L) [ ρ(µ) = h, A
L
(µ) = β(h) ¦ ,
´
ω
(L) := ¦ µ ∈ ´(L) [ A
L−ω
(µ) = −c(L −ω) ¦.
Since the βfunction has a supporting hyperplane at each homology
class h, if ω ∈ ∂β(h), then ´
h
(L) ⊆ ´
ω
(L). Conversely, since by
corollary D.2 α
∗
= β, then ´
ω
(L) ⊆ ´
h
(L) if h ∈ ∂α(ω). Thus
¸
h∈H
1
(M,R)
´
h
(L) =
¸
ω∈H
1
(M,R)
´
ω
(L).
We call these measures Mather minimizing measures and the set
´:= ´
0
(L) = ¦µ ∈ ´(L) [ A
L
(µ) = −c(L) ¦,
the Mather set.
Deﬁne the strict critical value as
c
0
(L) : = min
ω∈H
1
(M,R)
c(L −ω) = min
ω∈H
1
(M,R)
α(ω)
= −β(0).
(2.31)
By corollary 36.3 the strict critical value is the lowest energy level which
supports Mather minimizing measures and since c
0
(L) = −β(0), these
minimal energy Mather minimizing measures have trivial homology.
51
52 2. ma˜ n´ e’s critical value.
27 Coverings.
We shall deal mainly with compact manifolds M, but there are some
important noncompact cases, for example the coverings of M. Partic
ularly interesting are the abelian cover
´
M, the universal cover
M and
the ﬁnite coverings.
The abelian cover
´
M of M is the covering whose fundamental group
is the kernel of the Hurewicz homomorphism π
1
(M) → H
1
(M, Z). Its
deck transformation group is H
1
(M, Z) and H
1
(
M, Z) = ¦0¦. When
π
1
(M) is abelian,
´
M =
M. A closed curve in
´
M projects to a closed
curve in M with trivial homology.
If M
1
p
−→ M is a covering, denote by L
1
:= L ◦ dp : TM
1
→ R the
lifted lagrangian to TM
1
.
27.1 Lemma. If M
1
p
−→M is a covering, then c(L
1
) ≤ c(L).
Proof: The lemma follows form the fact that closed curves on N project
to closed curves on M.
27.2 Proposition. If M
1
is a ﬁnite covering of M
2
then c(L
1
) = c(L
2
).
Proof: We know that c(L
1
) ≤ c(L
2
). Suppose that the strict inequality
holds and let k be such that c(L
1
) < k < c(L
2
). Hence there exists
a closed curve γ in M
2
with negative (L
2
+ k)action. Since M
1
is a
ﬁnite covering of M
2
some iterate of γ lifts to a closed curve in M
1
with
negative (L
1
+k)action which contradicts c(L
1
) < k.
27.3 Proposition. [60]
c
0
(L) = c
a
(L) = critical value of the abelian cover.
Then we have
c
u
(L) ≤ c
a
(L) = c
0
(L) ≤ c(L −ω) ∀[ω] ∈ H
1
(M, R),
52
27. coverings. 53
where c
u
is the critical value of the lift of the lagrangian to the universal
cover. When c
u
(L) < c
0
(L), the method in equation (2.28) gives some
minimizing measures which are not Mather minimizing. For symmetric
lagrangians c(L) = e
0
= c
0
(L) = c
u
(L). Ma˜ n´e [39] gives an example in
which e
0
< c
a
(L) = c
0
(L) < c(L). G. Paternain and M. Paternain [60]
give an example in which c
u
(L) < c
a
(L).
Proof: Let ω be a closed form in M. Since H
1
(M, R) = ¦0¦, the lift ´ ω
of ω to
´
M is exact, then
c
a
(L) := c(
´
L) = c(
´
L − ´ ω) ≤ c(L −ω).
Hence
c
a
(L) ≤ min
ω∈H
1
(M,R)
c(L −ω) = c
0
(L).
Moreover,
−c
a
(L) = inf
A
b
L
(µ) [ µ ∈ ((
´
M)
¸
= inf¦ A
b
L
(µ
γ
) [ µ
γ
∈ ((
´
M) ¦ by proposition 24.2,
= inf¦ A
L
(µ
γ
) [ µ
γ
∈ ((M), ρ(µ) = 0 ¦,
because a closed curve γ on M has homology [γ] = 0 if and only if it
has a closed lift to
´
M. Then
−c
a
(L) = inf¦ A
L
(µ) [ µ ∈ ((M), ρ(µ) = 0 ¦, by 24.2 and 26.1
≤ min¦ A
L
(µ) [ µ ∈ ´(L), ρ(µ) = 0 ¦ because ´(L) ⊂ ((M)
= −β(0) = c
0
(L).
The real abelian cover is the covering
ˇ
M of M with h : π(M) →
H
1
(M, R) is the Hurewicz homomorphism. It is an intermediate covering
M →
ˇ
M → M and the deck transformations of
´
M →
ˇ
M are given by
the torsion
4
of H
1
(M, Z). Hence
´
M →
ˇ
M is a ﬁnite cover so that they
have the same critical value c
a
(L) = c
0
(L).
4
i.e. the elements of ﬁnite order Z
n
1
⊕· · · ⊕Z
n
k
⊆ H
1
(M, Z).
53
54 2. ma˜ n´ e’s critical value.
54
Chapter 3
Globally minimizing orbits.
31 Tonelli’s theorem.
Given x, y ∈ M and T > 0, let
(
T
(x, y) :=
¸
γ ∈ (
ac
([0, T], M) [ γ(0) = x, γ(T) = y
¸
.
We say that γ ∈ (
T
(x, y) is a Tonelli minimizer if
A
L
(γ) = min
η∈C
T
(x,y)
A
L
(η) .
In this section we shall prove
31.1 Tonelli’s Theorem.
For all x, y ∈ M and T > 0 there exists a Tonelli minimizer on (
T
(x, y).
The only diﬀerence in the proof of this theorem when M is noncom
pact is corollary 31.8. An independent proof of this corollary is given
in remark 31.9.
The idea of Tonelli’s theorem is to prove that the sets
/(c) :=
¸
γ ∈ (
T
(x, y) [ A
L
(γ) ≤ c ¦ (3.1)
55
56 3. globally minimizing orbits.
are compact in the C
0
topology. Then a Tonelli minimizer will be a
curve in
¸
c≥α
/(c) = ,
where α = inf
η∈C
T
(x,y)
A
L
(η) ≥ inf L > −∞.
An addendum to Tonelli’s theorem due to Mather [46] states that
these sets are compact in the topology of absolutely continuous curves.
Given γ
1
, γ
2
∈ (
ac
([0, T], M) deﬁne their absolutely continuous distance
by
d
1
(γ
1
, γ
2
) := sup
t∈[0,T]
d
M
γ
1
(t), γ
2
(t)
+
T
0
d
TM
[γ
1
(t), ˙ γ
1
(t)], [γ
2
(t), ˙ γ
2
(t)]
dt.
31.2 Theorem (Mather [46]). For any x, y ∈ M, T > 0, b ∈ R, the
set
/(b) :=
¸
γ ∈ (
T
(x, y) [ A
L
(γ) ≤ b
¸
is compact in the C
0
topology.
This theorem is proved in 31.12. We quote here the following propo
sition (addendum on page 175 of Mather [46]).
31.3 Proposition. (Mather [46])
If N ⊆ M is a compact subset and γ
1
, γ
2
, . . . is a sequence in
(
ac
([a, b], N) which converges C
0
to γ and A
L
(γ
i
) converges to A
L
(γ),
then γ
1
, γ
2
, . . . converges in the d
1
topology to γ.
In fact, the set /(b) in proposition 31.2 is not compact in the d
1

topology as the following example shows. Since the action functional is
only lower semicontinuous on (
T
(x, y) a priori it is not possible to ensure
the convergence of the action in a C
0
convergent sequence obtained from
theorem 31.2. Unless, for example, if the action is converging to the
minimal action (e.g. proposition 24.4).
56
31. tonelli’s theorem. 57
31.4 Example. /(b) is not compact in the d
1
topology.
Let L =
1
2
[v[
2
be the riemannian lagrangian on R
2
with the ﬂat
metric. Let η(t) = (t, 0) and γ
n
(t) = (t,
1
n
sin(2πnt)), t ∈ [0, 1]. The
action A
L
(γ
n
) = 1 + 2π
2
is bounded. γ
n
→η in the C
0
topology.
The length of γ
n
is bounded below by a polygonal curve join
ing the maxima and minimums of its second component. Hence
(γ
n
) ≥ 2n
4/n
2
+ 1/(4n
2
) > 4. Therefore (γ
n
) → (η), and thus
γ
n
→ η in the d
1
topology. Moreover, since a reparametrization pre
serves length, there is no reparametrization of the γ
n
’s which converges
in the d
1
topology to η.
We shall split the proof of Tonelli’s theorem in several parts:
31.5 Deﬁnition.
A family T ⊆ C
0
([a, b], M) is absolutely equicontinuous if ∀ ε > 0
∃ δ > 0 such that
N
¸
i=1
[t
i
−s
i
[ < δ =⇒
N
¸
i=1
d(x
s
i
, x
t
i
) < ε,
whenever ]s
1
, t
1
[, . . . , ]s
N
, t
N
[ are disjoint intervals in [a, b].
31.6 Remark.
(i) An absolutely equicontinuous family is equicontinuous.
(ii) A uniform limit of absolutely equicontinuous functions is abso
lutely continuous.
31.7 Lemma. For all c ∈ R and T > 0, the family
T(c) :=
¸
γ ∈ (
ac
([0, T], M) [ A
L
(γ) ≤ c
¸
is absolutely equicontinuous.
57
58 3. globally minimizing orbits.
Proof: Since by the superlinearity, the lagrangian L is bounded below;
by adding a constant we may assume that L ≥ 0. For a > 0 let
K(a) := inf
L(x, v)
[v[
(x, v) ∈ TM, [v[ ≥ a
¸
. (3.2)
The superlinearity implies that lim
a→+∞
K(a) = +∞. Given ε > 0 let
a > 0 be such that
c
K(a)
<
ε
2
.
Let 0 ≤ s
1
< t
1
≤ ≤ s
N
< t
N
≤ T, J := ∪
N
i=1
[s
i
, t
i
] and E :=
J ∩
[ ˙ x[ > a
, then L(x
s
, ˙ x
s
) ≥ K(a) [ ˙ x
s
[ for s ∈ E. We have that
K(a)
N
¸
i=1
d(x
s
, x
t
) ≤ K(a)
E
[ ˙ x[ +K(a)
J\E
[ ˙ x[
≤
E
L(x, ˙ x) +a K(a) m(J)
≤ c +a K(a) m(J), (because L ≥ 0),
where m is the Lebesgue measure on [0, T]
N
¸
i=1
d(x
s
i
, x
t
i
) ≤
J
[ ˙ x
s
[ ≤
c
K(a)
+a m(J) ≤
ε
2
+a m(J). (3.3)
This implies the absolute equicontinuity of T(c).
In order to apply the ArzelaAscoli theorem we need a compact
range, for this we have:
31.8 Corollary. For all c ∈ R and T > 0 there is R > 0 such that for
all x, y ∈ M,
/(c) ⊆ (
ac
[0, T], B(x, R)
,
where B(x, R) := ¦ z ∈ M [ d
M
(x, z) ≤ R¦.
58
31. tonelli’s theorem. 59
Proof: Inequality (3.3) for N = 1 and J = [s, t] is d(x
s
, x
t
) ≤
ε
2
+a [t−s[.
It is enough to take R =
ε
2
+a T.
31.9 Remark.
Corollary 31.8 is the only diﬀerence for the proof of Tonelli’s theorem when
M is noncompact. Another proof for corollary 31.8 is the following:
Adding a constant we may assume that L ≥ 0. There is B > 0 such that
L(x, v) ≥ [v[ −B for all (x, v) ∈ TM. Then for 0 ≤ s ≤ t ≤ T, we have that
d(x
s
, x
t
) ≤
t
s
[ ˙ x[ ≤ BT +
t
s
L(x, ˙ x) ≤ BT +c.
Recall that
T(c) :=
¸
γ ∈ (
ac
([0, T], M) [ A
L
(γ) ≤ c
¸
31.10 Theorem.
If γ
n
∈ T(c) and γ
n
→γ in the uniform topology, then γ ∈ T(c).
We shall need the following lemma. We may assume that M = R
n
,
31.11 Lemma. Given K compact, a > 0 and ε > 0, there exists δ > 0
such that if x ∈ K, [x −y[ ≤ δ, [v[ ≤ a and w ∈ R
n
, then
L(y, w) ≥ L(x, v) +L
v
(x, v) (w −v) −ε. (3.4)
Proof of theorem 31.10:
It is immediate from the deﬁnition 31.5 that a uniform limit of abso
lutely equicontinuous curves is absolutely continuous. We may assume
that γ
n
([0, T]) is contained in a compact neighbourhood K of γ([0, T]).
By the superlinearity we may assume that L ≥ 0. Let ε > 0 and
E =
[ ˙ γ[ ≤ a
, then by lemma 31.11, for n large,
E
L( ˙ γ) +L
v
( ˙ γ)( ˙ γ
n
− ˙ γ) −ε
≤
E
L( ˙ γ
n
) ≤ c (since L ≥ 0). (3.5)
59
60 3. globally minimizing orbits.
Claim:
lim
n
E
L
v
( ˙ γ) ( ˙ γ
n
− ˙ γ) = 0.
Letting n →+∞ on (3.5), we have that
E
L( ˙ γ) −ε T ≤ c.
Since E ↑ [0, T] when a →+∞ and L ≥ 0, then
T
0
L( ˙ γ) = lim
a→+∞
E
L( ˙ γ) ≤ c +ε T.
Now let ε →0.
We now prove the claim. We have to prove that
lim
n
ψ ˙ z
n
= 0, (3.6)
where ψ = L
v
( ˙ γ)1
E
is bounded, z
n
= γ
n
−γ and 1
E
is the characteristic
function of E. Since z
n

∞
→0, if A = [a, b] is an interval then
lim
n
A
˙ z
n
= lim
n
z
n
b
a
= 0. (3.7)
Since γ is absolutely continuous, ˙ γ ∈ L
1
. From ˙ γ ∈ L
1
and inequality
(3.3), given ε > 0 there is δ > 0 such that
m(D) < δ =⇒
D
[ ˙ z
n
[ < ε, ∀n ∈ N. (3.8)
Equations (3.7) and (3.8) imply that (3.7) holds for any Borel set A,
approximating A by a ﬁnite union of intervals. Hence (3.6) holds if ψ is
a simple function
¸
k
a
k
1
A
k
.
Let f be a simple function such that f
∞
≤ 2 ψ
∞
and
[ψ −f[ <
ε
2
. Let B = [[ψ −f[ > ε], then m(B) < ε.
(ψ −f) ˙ z
n
≤ ε
B
c
[ ˙ z
n
[ + 3 ψ
∞
B
[ ˙ z
n
[ ≤ h(ε).
60
31. tonelli’s theorem. 61
By (3.8) one can take h(ε) with lim
ε
h(ε) = 0. Then
lim
n
ψ ˙ z
n
≤ lim
n
f ˙ z
n
+ limsup
n
(ψ −f) ˙ z
n
≤ 0 +h(ε)
ε
−→0.
31.12. Proof of Tonelli’s theorem:
By lemma 31.7, the family /(c) in (3.1) is equicontinuous, and by
corollary 31.8, the curves in /(c) have a uniform compact range. By
Arzel´aAscoli’s theorem and theorem 31.10, /(c) is compact. Then
γ ∈
¸
c≥inf
C
T
(x,y)
A
L
/(c)
is a Tonelli’s minimizer on (
T
(x, y).
Proof of lemma 31.11:
Let
C
1
:= sup
¸
[L
v
(x, v)[ [ x ∈ K, [v[ ≤ a
¸
C
2
:= sup
¸
L(x, v) −L
v
(x, v) v [ x ∈ K, [v[ ≤ a
¸
.
Let b > 0 be such that
K(b) r ≥ C
2
+C
1
r for all r ≥ b,
where K(b) is from (3.2). Then if y ∈ M and [w[ > b,
L(y, w) ≥ K(b) [w[
≥ C
2
+C
1
[w[
≥ C
2
+L
v
(x, v) w
≥ L(x, v) +L
v
(x, v) (w −v) for [w[ ≥ b.
This gives (3.4) when [w[ ≥ b. Since L is convex,
L(x, w) ≥ L(x, v) +L
v
(x, v) (w −v) ∀ w ∈ R
n
.
Then there is δ > 0 such that for [x − y[ ≤ δ, [v[ ≤ a and [w[ ≤ b
inequality (3.4) holds.
61
62 3. globally minimizing orbits.
32 A priori compactness.
The following lemma, due to Mather [46] for Tonelli minimizers in the
nonautonomous case, will be very useful. In the autonomous case its
proof is very simple.
32.1 Lemma.
For C > 0 there exists A = A(C) > 0 such that if x, y ∈ M and γ ∈
(
T
(x, y) is a solution of the EulerLagrange equation with A
L
(γ) ≤ C T,
then [ ˙ γ(t)[ < A for all t ∈ [0, T].
Proof: By the superlinearity there is D > 0 such that L(x, v) ≥ [v[ −D
for all (x, v) ∈ TM. Since A
L
(γ) ≤ C T, the mean value theorem implies
that there is t
0
∈]0, T[ such that
[ ˙ γ(t
0
)[ ≤ D +C.
The conservation of the energy and the uniform bounds (1.7) and (1.6)
imply that there is A = A(C) > 0 such that [ ˙ γ[ ≤ A.
For k ≥ c(L) and x, y ∈ M, deﬁne
Φ
k
(x, y; T) := inf
γ∈C
T
(x,y)
A
L+k
(γ).
32.2 Corollary. Given ε > 0, there are constants A(ε), B(ε), C(ε) > 0
such that if T ≥ ε, d(x, y) < R and k ∈ R, then
(i) Φ
k
(x, y; T) ≤ [C(ε, R) +k] T.
(ii) If γ ∈ (
T
(x, y) is a solution of the EulerLagrange equation such
that A
L+k
(γ) ≤ [C(ε, R) + k] T + 1, then [ ˙ γ[ ≤ A(ε, R) and
E(γ, ˙ γ) ≤ B(ε, R).
62
32. a priori compactness. 63
Proof: Comparing with the action of a geodesic on (
T
(x, y), we get (i)
with
C(ε, R) = sup
¸
L(v)
[v[ ≤
d(x,y)
ε
≤
R
ε
¸
.
Using (i) and lemma 32.1,we obtain A(ε, R). Using A(ε, R) and in
equality (1.7) we obtain B(ε, R).
32.3 Lemma.
There exists A > 0 such that if x, y ∈ M and γ ∈ (
T
(x, y) is a
solution of the EulerLagrange equation with
A
L+c
(γ) ≤ Φ
c
(x, y) +d
M
(x, y),
then
(a) T >
1
A
d
M
(x, y).
(b) [ ˙ γ(t)[ < A for all t ∈ [0, T].
Proof: Let η : [0, d(x, y)] → M be a minimal geodesic with [ ˙ η[ ≡ 1.
Let (r) be from (1.1) and D = (1) + c + 2. From the superlinearity
condition there is B > 0 such that
L(x, v) +c > D[v[ −B, ∀(x, v) ∈ TM.
Then
[(1) +c] d(x, y) ≥ A
L+c
(η) ≥ Φ
c
(x, y) (3.9)
≥ A
L+c
(γ) −d(x, y) (3.10)
≥
T
0
D [ ˙ γ[ −B
dt −d(x, y)
≥ D d(x, y) −BT −d(x, y).
Hence
T ≥
D−−c−1
B
d(x, y) ≥
1
B
d(x, y).
63
64 3. globally minimizing orbits.
From (3.9) and (3.10), we get that
A
L
(γ) ≤
(1) +c + 1
d(x, y) −c T,
≤
¸
B[ (1) +c + 1 ] −c
¸
T.
Then lemma 32.1 completes the proof.
64
33. energy of timefree minimizers. 65
33 Energy of timefree minimizers.
A curve γ ∈ ((x, y) is a global minimizer or time free minimizer for
L +k if k ≥ c(L) and A
L+k
(γ) = Φ
k
(x, y).
33.1 Proposition. A timefree minimizer for L+k has energy E ≡ k.
We need the following
33.2 Lemma. Let : x : [0, T] → M be an absolutely continuous curve
and k ∈ R. For λ > 0, let x
λ
(t) := x(λt) and /(λ) := A
L+k
(x
λ
). Then
/
(1) =
T
0
E(x, ˙ x) −k
dt.
Proof: Since ˙ x
λ
(t) = λ ˙ x(λt), then
/(λ) =
T
λ
0
L(x(λt), λ ˙ x(λt)) +k
dt.
Diﬀerentiating /(λ) and evaluating at λ = 1, we have that
/
(1) = −T
L(x(T), ˙ x(T)) +k
+
T
0
L
x
t ˙ x +L
v
( ˙ x +t ¨ x)
dt.
Integrating by parts the term (L
x
˙ x +L
v
¨ x) t =
d
dt
L
t, we have that
/
(1) = −T
L(x(T), ˙ x(T)) +k
+Lt [
T
0
+
T
0
L
v
˙ x −L
dt
= −T k +
T
0
E(x, ˙ x) dt =
T
0
E(x, ˙ x) −k
dt.
Proof of proposition 33.1.
Since γ is a solution of the EulerLagrange equation its energy
E(γ, ˙ γ) is constant. Since it minimizes with free time, the derivative
in lemma 33.2 must be zero. So that E(γ, ˙ γ) ≡ k.
65
66 3. globally minimizing orbits.
33.3 Corollary.
Let x ∈ (
ac
([0, 1], M) and k > 0. For T > 0, write y
T
(t) = x(
t
T
) :
[0, T] →M and B(T) = A
L+k
(y
T
). Then
B
(T) = −
1
T
T
0
E(y
T
, ˙ y
T
) −k
dt.
Proof: Using λ =
T
S
on lemma 33.2, we have that
d
dS
= −
T
S
2
d
dλ
. Thus
d
dS
S=T
B = −
1
T
d
dλ
λ=1
/ = −
1
T
T
0
E(y
T
, ˙ y
T
) −k
dt.
66
34. the finitetime potential. 67
34 The ﬁnitetime potential.
Recall that if k ≥ c(L), x, y ∈ M and T > 0,
Φ
k
(x, y; T) := inf
γ∈C
T
(x,y)
A
L+k
(γ).
Here we shall prove the following proposition. See also corollary 411.8.
34.1 Proposition.
1. For k ∈ R, a compact subset R ⊆ M and ε > 0, the function
(x, y, t) → Φ
k
(x, y; t) is Lipschitz on R R [ε, +∞[ for any
ε > 0.
2. Φ
k
(x, y; T) = Φ
c
(x, y; T) + (k −c) T, for k ≥ c(L), x, y ∈ M.
3. lim
ε→0
+ Φ
k
(x, y; ε) = +∞, for k ≥ c(L), x = y.
4. lim
T→+∞
Φ
k
(x, y; T) = +∞, for k > c(L), x, y ∈ M.
5. lim
T→+∞
Φ
k
(x, y; T) = −∞, for k < c(L), x, y ∈ M.
6. When M is compact, the limits in items 4 and 5 are uniform in
(x, y).
Proof: Item 2 follows from the fact that both action potentials satify
an equivalent variational principle. We now compute the limits.
3. Given A > 0, let B > 0 be such that L(x, v) > A[v[ −B. Then
Φ
k
(x, y; ε) = inf
γ∈C
ε
(x,y)
A
L+k
(γ) ≥ inf
γ
ε
0
A[ ˙ γ[ −B +k
≥ Ad(x, y) + (k −B) ε.
Thus, liminf
ε→0
+ Φ
k
(x, y; ε) ≥ Ad(x, y). Now let A →+∞.
67
68 3. globally minimizing orbits.
4. Observe that if k > c(L),
lim
T→+∞
Φ
k
(x, y; T) ≥ lim
T→+∞
Φ
c
(x, y) + (k −c) T
= +∞.
When M is compact, this limit is uniform because
Φ
k
(x, y; T) ≥ inf
x,y∈M
Φ
c
(x, y) + (k −c) T.
5. For k < c(L) there exists a closed curve γ ∈ (
S
(z, z), S > 0,
z ∈ M such that A
L+c
(γ) < 0. Then for T = nS +τ, with S ≤ τ < 2S,
we have that
Φ
k
(x, y; T) ≤ Φ
k
(x, z;
τ
2
) +nA
L+c
(γ) + Φ
k
(z, y;
τ
2
).
This implies that lim
T→+∞
Φ
k
(x, y; T) = −∞. The uniformity of this
limit for (x, y) on compact subsets follows from the Lipschitz condition
of item 1 that we now prove.
Fix ε > 0. We prove now that the function ]ε, +∞[÷ t →Φ
k
(x, y; t)
is uniformly Lipschitz. If T > ε and γ ∈ (
T
(x, y) is a Tonelli minimizer,
from corollary 32.2 and inequality (1.6) there exists D = D(ε, R) > 0
such that [E(γ, ˙ γ) − k[ ≤ D(ε, R) + [k[. Denote h(s) := Φ
k
(x, y; s).
If γ
s
(t) := γ(
T
s
t), t ∈ [0, s], then h(s) ≤ A
L+k
(γ
s
) =: B(s). Using
corollary 33.3 we have that
f(T) : = limsup
δ→0
h(T +δ) −h(T)
δ
≤ B
(T) =
1
T
T
0
k −E(γ, ˙ γ)
dt
≤ D(ε, R) +[k[ .
If S, T > ε we have that
Φ
k
(x, y; S) ≤ Φ
k
(x, y; T) +
S
T
f(t) dt
≤ Φ
k
(x, y; T) +
D(ε, R) +[k[
[T −S[.
68
34. the finitetime potential. 69
Since we can reverse the roles of S and T, this implies the uniform
Lipschitz condition for T →Φ
k
(x, y; T).
Now we prove the Lipschitz condition on a neighbourhood of y
0
for
y →Φ
k
(x
0
, y; T
0
). The proof for a neighbourhood of x
0
is similar. Since
R is compact, this is enough to show that the function is Lipschitz on
R R]ε, +∞[.
Fix 0 < δ <
ε
2
small. Let λ ∈ (
T
(x
0
, y
0
) be a Tonelli minimizer.
Observe that λ realizes Φ
k
(x
0
, y
0
, T). Let γ : [0, 1] → M be a length
minimizing geodesic joining y
0
to y, with [ ˙ γ[ ≡ 1. Let η(s, t), s ∈ [0, 1],
t ∈ [T −δ, T] be a variation by solutions of (EL) such that η(s, T −ε) =
λ(T −δ), η(s, T) = γ(s) and η(0, t) = λ(t). Let
c(s) := A
L+k
λ[
[0,T−δ]
∗ η(s, )[
[T−δ,T]
.
Then
c
(s) =
d
ds
T
T−δ
L
∂η
∂t
(s, t)
dt =
T
T−δ
L
x
∂η
∂s
+L
v
∂
2
η
∂s ∂t
= L
v
∂η
∂s
T
T−δ
+
T
T−δ
L
x
−
d
dt
L
v
∂η
∂s
= L
v
γ(s),
∂η
∂t
(s, T)
˙ γ(s).
By Weierstrass theorem ??, if δ and d(y, y
0
) are small enough, the curves
t → η(s, t) are Tonelli minimizers — and thus they realize Φ
k
(y
0
, y, δ).
Then by corollary 32.2, there exists A(δ, R) > 0 such that [
∂η
∂t
(s, T)[ <
A(δ, R). By lemma 14.4, L
v
(x,
∂η
∂t
(s, T)) ≤ f(A(δ, R)). Thus
Φ
k
(x
0
, y; T) ≤ Φ
k
(x
0
, y
0
; T) +
1
0
c
(s) ds
≤ Φ
k
(x
0
, y
0
; T) +
1
0
f(A(δ, R)) [ ˙ γ(s)[ ds
= Φ
k
(x
0
, y
0
; T) +f(A(δ, R)) d(y, y
0
).
The value od δ can be taken locally constant on a neighbourhood of
y
0
. Changing the roles of y and y
0
we obtain that f(A(δ, R)) is a local
Lipschitz constant for y →Φ
k
(x, y; T), x, y ∈ R.
69
70 3. globally minimizing orbits.
35 Global Minimizers.
Here we construct curves that realize the action potential.
For k < c(L), Φ
k
≡ −∞, so there are no minimizers.
35.1 Proposition.
If k > c(L) and x, y ∈ M, x = y, then there is γ ∈ ((x, y) such that
A
L+k
(γ) = Φ
k
(x, y).
Moreover, the energy of γ is E(γ, ˙ γ) ≡ k.
Proof: Let f(t) := Φ
k
(x, y; t). By proposition 34.1, f(t) is continuous
and f(t) →+∞when t →0
+
or t →+∞. Hence it attains its minimum
at some T > 0. Moreover, Φ
k
(x, y) = inf
t>0
Φ
k
(x, y; t) = Φ
k
(x, y; T).
Now take a Tonelli minimizer γ on (
T
(x, y). From lemma 33.2, the
energy of γ is k.
We now study minimizers at k = c(L). Observe that for c = c(L)
and any absolutely continuous curve γ ∈ ((x, y), we have that
A
L+c
(γ) ≥ Φ
c
(x, y) ≥ −Φ
c
(y, x). (3.11)
35.2 Deﬁnition. Set c = c(L).
An absolutely continuous curve γ ∈ ((x, y) is said semistatic if
A
L+c
(γ) = Φ
c
(x, y).
An absolutely continuous curve γ ∈ ((x, y) is said static if
A
L+c
(γ) = −Φ
c
(y, x).
These names are justiﬁed by the following remark: For mechanic
lagrangians L =
1
2
[v[
2
x
− U(x), static orbits are the ﬁxed points (x
0
, 0)
of the lagrangian ﬂow where U(x) is maximal; and semistatic orbits lie
in the stable or unstable manifolds of those ﬁxed points.
70
35. global minimizers. 71
By the triangle inequality for Φ
c
the deﬁnition of semistatic curve
x : [a, b] →M is equivalent to
A
L+c
x[
[s,t]
= Φ
c
x(s), x(t)
, ∀ a ≤ s ≤ t ≤ b. (3.12)
Inequality (3.11) implies that static curves are semistatic.
Moreover, a curve γ ∈ ((x, y) is static if
(a) γ is semistatic, and
(b) d
c
(x, y) = Φ
c
(x, y) + Φ
c
(y, x) = 0.
From proposition 33.1 we get
35.3 Corollary. Semistatic curves have energy E ≡ c(L).
35.4 Deﬁnition.
´= ∪
¸
supp(µ)
µ ∈ ´(L), A
L
(µ) = −c(L)
¸
¯
^ = Σ(L) :=
¸
w ∈ TM
x
w
: R →M is semistatic
¸
/ =
´
Σ(L) :=
¸
w ∈ TM
x
w
: R →M is static
¸
Σ
−
(L) :=
¸
w ∈ TM
x
w
:] −∞, 0] →M is semistatic
¸
Σ
+
(L) :=
¸
w ∈ TM
x
w
: [0, +∞[→M is semistatic
¸
We call ´ the Mather set,
¯
^ the Ma˜ n´e set, { = π(
´
Σ(L)) the Peierls
set
1
and / =
´
Σ(L) the Aubry set.
Using the characterization of minimizing measures 36.1 and corol
lary 35.3 we have that
2
´⊆ / ⊆
¯
^ ⊆
´
c , (3.13)
where ´ is the Mather set, / is the Aubry set,
¯
^ is the Ma˜ n´e set and
´
c is the energy level
´
c = [E ≡ c(L)]. All these inclusions can be made
1
The name is justiﬁed by proposition 37.1.5.
2
The typographical relationship was observed by Albert Fathi.
71
72 3. globally minimizing orbits.
proper constructing examples of embedded ﬂows as in equation (1.18)
and adding a properly chosen potential φ(x).
Denote by α(v) and ω(v) the α and ωlimits of v under the
EulerLagrange ﬂow.
35.5 Proposition.
A local static is a global static, i.e. if x
v
[
[a,b]
is static then v ∈
´
Σ(L)
(i.e. the whole orbit is static).
Proof: Let η(t) = π ϕ
t
(v) and let γ
n
∈ (
T
n
(η(b), η(a)) be solutions
of (EL) with
A
L+c
(γ
n
) < Φ
c
(η(b), η(a)) +
1
n
.
By the a priori bounds 32.3 [ ˙ γ
n
[ < A. We can assume that ˙ γ
n
(0) →w.
Let ξ(s) = π ϕ
s
(w). If w = ˙ η(b) then the curve η[
[b−ε,b]
∗ ξ[
[0,ε]
is not C
1
,
and hence by remark 12.2, it can not be a Tonelli minimizer. Thus
Φ
c
(η(b −ε), ξ(ε)) < A
L+c
(η[
[b−ε,b]
) +A
L+c
(ξ[
[0,ε]
).
Φ
c
(η(a), η(a)) ≤ Φ
c
(η(a), η(b −ε)) + Φ
c
(η(b −ε), ξ(ε)) + Φ
c
(ξ(ε), η(a))
< A
L+c
(η
[a,b−ε]
) +A
L+c
(η[
[b−ε,b]
) +A
L+c
(ξ[
[0,ε]
) + liminf
n
A
L+c
(γ
n
[
[ε,T
n
]
)
≤ A
L+c
(η[
[a,b]
) + lim
n
γ
n
[
[0,ε]
∗ γ
n
[
[ε,T
n
]
≤ −Φ
c
(η(b), η(a)) + Φ
c
(η(b), η(a)) = 0,
which contradicts proposition 21.1(3). Thus w = ˙ η(b) and similarly
lim
n
˙ γ
n
(T
n
) = ˙ η(a).
If limsup T
n
< +∞, we can assume that τ = lim
n
T
n
> 0 exists. In
this case η is a (semistatic) periodic orbit of period τ + b − a and then
it is static.
72
35. global minimizers. 73
Now suppose that lim
n
T
n
= +∞. If s > 0, we have that
A
L+c
(η[
[a−s,b+s]
) + Φ
c
(η(b +s), η(a −s)) ≤
≤ lim
n
¸
A
L+c
(γ
n
[
[T
n
−s,T
n
]
) +A
L+c
(η) +A
L+c
(γ
n
[
[0,s]
)
¸
+ Φ
c
(η(b +s), η(a −s))
≤ Φ
c
(η(a), η(b))
+ lim
n
¸
A
L+c
(γ
n
[
[0,s]
) +A
L+c
(γ
n
[
[s,T
n
−s]
) +A
L+c
(γ
n
[
[T
n
−s,T
n
]
)
¸
≤ Φ
c
(η(a), η(b)) + Φ
c
(η(b), η(a)) = 0.
Thus η
[a−s,b+s]
is static.
35.6 Deﬁnition. Set c = c(L).
An absolutely continuous curve γ : [0, +∞[→M (resp. η :]−∞, 0] →M)
is a ray if γ[
[0,t]
(resp. η[
[−t,0]
) is a Tonelli minimizer for all t > 0; i.e.
A
L+c
(γ[
[0,t]
) = Φ
c
(γ(0), γ(t); t) for all t > 0.
Clearly a semiinﬁnite semistatic curve is a ray. We shall see in
corollary 411.9 that rays are semistatic.
35.7 Proposition.
If v ∈ Σ is semistatic, then α(v) ⊂
´
Σ(L) and ω(v) ⊂
´
Σ(L). Moreover
α(v) and ω(v) are each contained in a static class.
Proof: We prove only that ω(v) ⊂
´
Σ. Let γ(t) = π ϕ
t
(v). Suppose that
t
n
→ +∞ and ˙ γ(t
n
) → w ∈ TM. Let η(t) = π ϕ
t
(w). Since γ and η
are solutions the EulerLagrange equation, then γ[
[t
n
−s,t
n
+s]
−→
C
1
η[
[−s,s]
.
73
74 3. globally minimizing orbits.
Then
A
L+c
η[
[−s,s]
+ Φ
c
(η(s), η(−s)) =
= lim
n
¸
A
L+c
(γ[
[t
n
−s,t
n
+s]
) + lim
m
A
L+c
(γ[
[t
n
+s,t
m
−s]
)
¸
= lim
n
lim
m
A
L+c
(γ[
[t
n
−s,t
m
−s]
)
= lim
n
lim
m
Φ
c
(γ(t
n
−s), γ(t
m
−s))
= Φ
c
(η(−s), η(−s)) = 0.
Thus w ∈
´
Σ(L). Let u ∈ ω(v). We may assume that ˙ γ(s
n
) → u with
t
n
< s
n
< t
n+1
. Then
d
c
(πw, πu) = Φ
c
(πw, πu) + Φ
c
(πu, πw)
= lim
n
A
L+c
(γ[
[t
n
,s
n
]
) +A
L+c
(γ[
[s
n
,t
n+1
]
)
= lim
n
A
L+c
(γ
n
[
[t
n
,t
n+1
]
) = Φ
c
(πw, πw) = 0.
Thus w and u are in the same static class.
74
36. characterization of minimizing measures. 75
36 Characterization of minimizing measures.
Recall that minimizing measures µ satisfy A
L
(µ) = −c(L) and that by
theorem 24.1 they are invariant.
36.1 Theorem (Ma˜ n´e [39]).
µ ∈ ´(L) is a minimizing measure if and only if supp(µ) ⊆
´
Σ(L).
Proof: Since
´
Σ is closed, it is enough to prove the theorem for ergodic
measures. Suppose that µ ∈ ´(L) is ergodic and supp(µ) ⊆ Σ(L).
Since µ is ﬁnite, by Birkhoﬀ’s theorem there is a set of total µmeasure
A such that if θ ∈ A then liminf
T→+∞
d
TM
(θ, ϕ
T
θ) = 0 and
M
L +c dµ = lim
T→+∞
1
T
T
0
L(ϕ
t
θ) +c
dt
≤ liminf
T→+∞
1
T
Φ
c
π(ϕ
T
θ), π(θ)
= 0.
Now suppose that µ ∈ ´(L) is minimizing. Then
L + c dµ = 0.
Applying corollary 36.5 to F = L + c and X = TM, we get that there
is a set A ⊂ TM of total µmeasure such that if θ ∈ A then there is a
sequence T
n
→+∞ such that d(θ, ϕ
T
n
θ) →0 and
lim
n
T
n
0
L(ϕ
t
θ) +c
dt = 0.
Then
0 ≤ d
c
(π θ, πϕ
1
θ) = lim
n
Φ
c
(π θ, π ϕ
1
θ) + Φ
c
(π ϕ
1
θ, π ϕ
T
n
θ)
≤ lim
n
T
n
0
L(ϕ
t
θ) +c
dt = 0.
This implies that θ is static. Since
´
Σ(L) is closed and A is dense in
supp(µ) then supp(µ) ⊆
´
Σ(L).
75
76 3. globally minimizing orbits.
It follows from theorem 25.1 that
36.2 Corollary.
If M is compact then
´
Σ(L) = .
Combining theorem 36.1 with corollary 35.3, we get
36.3 Corollary (Dias Carneiro [9]). If µ is a minimizing measure then
it is supported in the energy level E(supp(µ)) = c(L).
36.4 Lemma. Let (X, B, ν) be a probability space, f an ergodic mea
sure preserving map and F : X → R an integrable function. Given
A ∈ B with ν(A) > 0 denote by
ˆ
A the set of point p ∈ A such that for
all ε > 0 there exists an integer N > 0 such that f
N
(p) ∈ A and
N−1
¸
j=0
F
f
j
(p)
−N
F dν
< ε .
Then ν(
ˆ
A) = ν(A).
Proof: Without loss of generality we can assume that
F dν = 0. For
p ∈ X denote
S
N
F(p) :=
N−1
¸
n=0
F
f
n
(p)
.
Let
A(ε) :=
¸
p ∈ A[ ∃N > 0 such that f
N
(p) ∈ A and [S
N
F(p)[ < ε
¸
.
It is enough to prove that ν(A(ε)) = ν(A), because
ˆ
A =
¸
n
A(
1
n
). Let
ˆ
X be the set of points for which the Birkhoﬀ’s theorem holds for F and
the characteristic functions of A and of A(ε). Take x ∈ A ∩
ˆ
X and let
N
1
< N
2
< be the integers for which f
N
i
(x) ∈ A. Deﬁne δ(k) by
N
k
δ(k) = [S
N
k
F(x)[ .
76
36. characterization of minimizing measures. 77
Since x ∈
ˆ
X we have that lim
k→+∞
δ(k) = 0. Set
c
k
:= S
N
k
F(x) ,
o :=
¸
k ∈ N
∀ > k, [c
−c
k
[ ≥ ε
¸
,
S(k) :=
¸
1 ≤ j ≤ k −1
j ∈ o
¸
.
Observe that if j / ∈ S(k), then there is an > j such that [c
−c
j
[ ≤ ε,
so that
S
N
−N
j
F
f
N
j
(x)
= [c
−c
j
[ ≤ ε .
Hence
j / ∈ S(k) =⇒f
N
j
(x) ∈ A(ε) .
This implies that
ν
A−A(ε)
= lim
k→+∞
1
N
k
#
¸
0 ≤ j < N
k
[ f
j
(x) ∈ A−A(ε)
¸
≤
1
N
k
#S(k) . (3.14)
If o is ﬁnite, from inequality (3.14) we get that ν
A − A(ε)
= 0,
concluding the proof of the lemma.
Assume that o is inﬁnite. This implies that the set ¦ c
k
[ k ∈ o ¦ is
unbounded. Choose an inﬁnite sequence / in o such that for all k ∈ /
we have,
[c
k
[ = max
j∈S(k)
[c
j
[.
Then, for k ∈ /,
1
2
(2ε) #S(k) ≤ [c
k
[ = δ(k) N
k
.
From (3.14), we get that
ν
A−A(ε)
≤ lim
k∈K
δ(k)
ε
= 0.
77
78 3. globally minimizing orbits.
36.5 Corollary. If besides the hypothesis of lemma 36.4, X is a com
plete separable metric space, and B is its Borel σalgebra, then for a.e.
x ∈ X the following property holds: for all ε > 0 there exists N > 0
such that d(f
N
(x), x) < ε and
N−1
¸
j=0
F
f
j
(x)
−N
F dν
< ε
Proof: Given ε > 0 let ¦V
n
(ε)¦ be a countable basis of neighbourhoods
with diameter < ε and let
ˆ
V
n
be associated to V
n
as in lemma 36.4. Then
the full measure subset ∩
m
∪
n
ˆ
V
n
(
1
m
) satisﬁes the required property.
78
37. the peierls barrier. 79
37 The Peierls barrier.
For T > 0 and x, y ∈ M deﬁne
h
T
(x, y) = Φ
c
(x, y; T) := inf
γ∈C
T
(x,y)
A
L+c
(γ).
So that the curves which realize h
T
(x, y) are the Tonelli minimizers on
(
T
(x, y). Deﬁne the Peierls barrier as
h(x, y) := liminf
T→+∞
h
T
(x, y).
The diﬀerence between the action potential and the Peierls barrier is
that in the Peierls barrier the curves must be deﬁned on large time
intervals. Clearly
h(x, y) ≥ Φ
c
(x, y).
37.1 Proposition.
If h : M M →R is ﬁnite, then
1. h is Lipschitz.
2. ∀ x, y ∈ M, h(x, x) ≥ Φ
c
(x, y), in particular h(x, x) ≥ 0, ∀ x ∈ M.
3. h(x, z) ≤ h(x, y) +h(y, z), ∀ x, y, z ∈ M.
4. h(x, y) ≤ Φ
c
(x, p) +h(p, q) + Φ
c
(q, y), ∀ x, y, p, q ∈ M.
5. h(x, x) = 0 ⇐⇒x ∈ π(
´
Σ) = {.
6. If
´
Σ = , h(x, y) ≤ inf
p∈π(
b
Σ)
Φ
c
(x, p) + Φ
c
(p, y).
Proof: Item 2 is trivial. Observe that for all S, T > 0 and y ∈ M,
h
T+S
(x, z) ≤ h
T
(x, y) +h
S
(y, z).
Taking liminf
T→+∞
we get that
h(x, z) ≤ h(x, y) +h
S
(y, z), for all S > 0.
79
80 3. globally minimizing orbits.
Taking liminf
S→+∞
, we obtain item 3.
1. Taking the inﬁmum on S > 0, we get that
h(x, z) ≤ h(x, y) + Φ
c
(y, z) ∀ x, y, z ∈ M.
≤ h(x, y) +A d
M
(y, z),
where A is a Lipschitz constant for Φ
c
. Changing the roles of x, y, z,
we obtain that h is Lipschitz.
4. Observe that
inf
S>T
h
S
(x, y) ≤ Φ
c
(x, p) +h
T
(p, q) + Φ
c
(q, x).
Taking liminf
T→+∞
we get item 4.
5. We ﬁrst prove that if p ∈ { = π(
´
Σ), then h(p, p) = 0. Take
v ∈
´
Σ such that π(v) = p and y ∈ π(ωlimit(v)). Let γ(t) := π ϕ
t
(v)
and choose t
n
↑ +∞ such that γ(t
n
) →y. Then
0 ≤ h(p, p) ≤ h(p, y) + Φ
c
(y, p)
≤ lim
n
A
L+c
(γ[
[0,t
n
]
) + Φ
c
(y, p)
≤ lim
n
−Φ
c
(γ(t
n
), p) + Φ
c
(y, p) = 0.
Conversely, if h(x, x) = 0, then there exists a sequence of Tonelli
minimizers γ
n
∈ ((x, x; T
n
) with T
n
→ +∞ and A
L+c
(γ
n
)
n
→ 0. By
lemma 32.3, [ ˙ γ[ is uniformly bounded. Let v be an accumulation point
of ˙ γ
n
(0) and η(t) := π ϕ
t
(v). Then if ˙ γ
n
k
(0)
k
→v, for any s > 0 we have
that
0 ≤ Φ
c
(x, π ϕ
s
v) + Φ
c
(π ϕ
s
v, x)
≤ A
L+c
η[
[0,s]
+ Φ
c
(π ϕ
s
v, x)
≤ lim
k
A
L+c
γ
n
k
[
[0,s]
+A
L+c
γ
n
k
[
[s,T
n
]
= 0.
Thus v ∈
´
Σ.
80
37. the peierls barrier. 81
6. Using items 4 and 5, we get that
h(x, y) ≤ inf
p∈π(
b
Σ)
Φ
c
(x, p) + 0 + Φ
c
(p, y)
.
37.2 Proposition. If M is compact, then
h(x, y) = inf
p∈π(
b
Σ)
Φ
c
(x, p) + Φ
c
(p, y)
.
Proof:
From proposition 37.1.6 we have that
h(x, y) ≤ inf
p∈π(
b
Σ)
Φ
c
(x, p) + Φ
c
(p, y)
.
In particular h(x, y) < +∞ for all x, y ∈ M. Now let γ
n
∈ (
T
n
(x, y)
with T
n
→+∞ and A
L+c
(γ
n
) →h(x, y) < +∞. Then
1
T
A
L+c
(γ
n
) →0.
Let µ be a weak limit of a subsequence of the measures µ
γ
n
. Then µ
is minimizing. Let q ∈ π
supp(µ)
and q
n
∈ γ
n
([0, T
n
]) be such that
lim
n
q
n
= q. Then,
Φ
c
(x, q) + Φ
c
(q, y) ≤ Φ
c
(x, q
n
) + Φ
c
(q
n
, y) + 2Ad(q
n
, q)
≤ A
L+c
(γ
n
) + 2Ad(q
n
, q).
Letting n →∞, we get that
Φ
c
(x, q) + Φ
c
(q, y) ≤ h(x, y).
81
82 3. globally minimizing orbits.
38 Graph Properties.
In this section we shall see that the projection π :
´
Σ → M is injective.
We shall call { := π(
´
Σ) the Peierls set.
3
Thus the projection π[
b
Σ
gives
an identiﬁcation { ≈
´
Σ.
For v ∈ TM, write x
v
(t) = π ϕ
t
(v). Given ε > 0, let
Σ
ε
:= ¦w ∈ TM[ x
w
: [0, ε) →M or x
w
: (−ε, 0] →M is semistatic ¦.
38.1 Theorem. (Ma˜ n´e) [39]
For all p ∈ π(
´
Σ) there exists a unique ξ(p) ∈ T
p
M such that
(p, ξ(p)) ∈ Σ
ε
, in particular (p, ξ(p)) ∈
´
Σ and
´
Σ = graph(ξ).
Moreover, the map ξ : π(
´
Σ) →Σ is Lipschitz.
The proofs of the injectivity of π in this book only need that the
solutions of the EulerLagrange equation are diﬀerentiable
4
. The reader
may provide those proofs as exercises. The proof of the Lipschitz con
dition need the following lemma, due to Mather. For the proof see [46]
or Ma˜ n´e [36].
38.2 Mather’s Crossing lemma. [46]
Given A > 0 there exists K > 0 ε
1
> 0 and δ > 0 with the following
property: if [v
i
[ < A, (p
i
, v
i
) ∈ TM, i = 1, 2 satisfy d(p
1
, p
2
) < δ and
d((p
1
, v
1
), (p
2
, v
2
)) ≥ K
−1
d(p
1
, p
2
) then, if a ∈ R and x
i
: R → M,
i = 1, 2, are the solutions of L with x
i
(a) = p
i
, ˙ x
i
(a) = v
i
, there exist
solutions γ
i
: [a −ε, a +ε] →M of L with 0 < ε < ε
1
, satisfying
γ
1
(a −ε) =x
1
(a −ε) , γ
1
(a +ε) = x
2
(a +ε) ,
γ
2
(a −ε) =x
2
(a −ε) , γ
2
(a +ε) = x
1
(a +ε) ,
A
L
(x
1
[
[a−ε,a+ε]
) +A
L
(x
2
[
[a−ε,a+ε]
) > A
L
(γ
1
) +A
L
(γ
2
)
3
This name is justiﬁed by proposition 37.1(5).
4
and hence a nondiﬀerentiable curve can not be a Tonelli minimizer.
82
38. graph properties. 83
Proof of theorem 38.1:
We prove that if (p, v) ∈
´
Σ, (q, w) ∈ Σ
ε
, and d(p, q) < δ, then
d
TM
(p, v), (q, w)
< K d
M
(p, q) .
Observe that this implies the theorem. For simplicity, we only prove
the case in which x
v
[
[−ε,0]
is semistatic. Suppose it is false. Then by
lemma 38.2 there exist α, β : [−ε, ε] →M such that
α(−ε) = x
w
(−ε) =: q
−ε
, α(0) = p ,
β(−ε) = x
v
(−ε) =: p
ε
, β(0) = q ,
and
A
L
(α) +A
L
(β) < A
L
(x
w
[
[−ε,0]
) +A
L
(x
v
[
[−ε,0]
).
So
Φ
c
(q
−ε
, p) + Φ
c
(p
−ε
, q) < Φ
c
(q
−ε
, q) + Φ
c
(p
−ε
, p)
= Φ
c
(q
−ε
, q) −Φ
c
(p, p
−ε
)
Thus
fig. 1: graph property.
Φ
c
(q
−ε
, q) ≤ Φ
c
(q
−ε
, p) + Φ
c
(p, p
−ε
) + Φ
c
(p
−ε
, q) < Φ
c
(q
−ε
, q)
which is a contradiction.
83
84 3. globally minimizing orbits.
Using the graph property 38.1 we can deﬁne an equivalence relation on
´
Σ by
u, v ∈
´
Σ, u ≡ v ⇐⇒ d
c
(π(u), π(v)) = 0.
The equivalence classes are called static classes. The continuity of the
pseudometric d
c
implies that a static class is closed, and it is invariant
by proposition 35.5.
For v ∈ TM denote by ω(v) its ωlimit. Let Γ be a static class, the
set
Γ
+
=
¸
v ∈ Σ
+
(L) [ ω(v) ⊆ Γ
¸
is called the (forward) basin of Γ. Clearly Γ
+
is forward invariant. Let
Γ
+
0
=
¸
t>0
ϕ
t
(Γ
+
)
=
¸
ε>0
¸
v ∈ TM [ x
v
[
]−ε,+∞[
is semistatic and ω(v) ⊆ Γ
¸
.
The set π(Γ
+
` Γ
+
0
) is called the cut locus of Γ
+
.
38.3 Theorem. (Ma˜ n´e)[39]
For every static class Γ, the projection π : Γ
+
0
→M is injective with
Lipschitz inverse.
The projection π : Γ
+
→ M may not be surjective. But when M is
compact for generic lagrangians π(Γ
+
) = π(Γ
−
) = M because there is
only one static class (cf. theorem 70.1.(B)). But π may not be injective
on Γ
+
` Γ
+
0
even for generic lagrangians.
Proof: We prove that for K as in lemma 38.2, if v , w ∈ Γ
+
0
then
d
TM
(v, w) ≤ K d
M
(π(v), π(w)). (3.15)
Suppose it is false. Then there are v, w ∈ Γ
+
0
such that inequality (3.15)
does not hold. Let ε > 0 be such that x
v
[
[−ε,+∞[
and x
w
[
[−ε,+∞[
are
semistatic. By lemma 38.2, there exist α ∈ (
2ε
(x
v
(−ε), x
w
(ε)) and
β ∈ (
2ε
(x
w
(−ε), x
v
(ε)) such that
A
L+c
(α) +A
L+c
(β) +δ < A
L+c
(x
v
[
[−ε,ε]
) +A
L+c
(x
w
[
[−ε,ε]
),
84
38. graph properties. 85
for some δ > 0. Let p, q ∈ π(Γ) and s
n
t
n
→+∞be such that x
v
(s
n
) →
+∞ and x
v
(s
n
) →p, x
w
(t
n
) →q. Then
Φ
c
(x
v
(−ε), x
w
(t
n
)) + Φ
c
(x
w
(−ε), x
v
(s
n
)) +δ
≤ A
L+c
(α ∗ x
w
[
[ε,t
n
]
) +A
L+c
(β ∗ x
v
[
[ε,s
n
]
) +δ
< A
L+c
(x
v
[
[−ε,s
n
]
) +A
L+c
(x
w
[
[−ε,t
n
]
)
= Φ
c
(x
v
(−ε), x
v
(s
n
)) + Φ
c
(x
w
(−ε), x
w
(t
n
)).
Letting n →∞ and adding d
c
(p, q) = 0, we have that
Φ
c
(x
v
(−ε), p) + Φ
c
(x
w
(−ε), q)
≤ Φ
c
(x
v
(−ε), q) + Φ
c
(x
w
(−ε), p) + Φ
c
(q, p) + Φ
c
(p, q)
< Φ
c
(x
v
(−ε), p) + Φ
c
(x
w
(−ε), q).
85
86 3. globally minimizing orbits.
39 Coboundary Property.
The coboundary property was ﬁrst presented by R. Ma˜ n´e in [38] and
further developed in [39] and by A. Fathi.
39.1 Theorem. (Ma˜ n´e) [39]
If c = c(L), then (L+c)
b
Σ
is a Lipschitz coboundary. More precisely,
taking any p ∈ M and deﬁning G :
´
Σ →R by
G(w) = Φ
c
p, π(w)
,
then
(L +c)
b
Σ
=
dG
df
,
where
dG
dϕ
(w) := lim
h→0
1
h
G(ϕ
h
(w) −G(w)
.
Proof: Let w ∈
´
Σ and deﬁne F
w
(v) := Φ
c
(π(w), π(v)). We have that
dF
w
dϕ
w
= lim
h→0
1
h
F
w
(ϕ
h
w) −F
w
(w)
= lim
h→0
1
h
Φ
c
(πw, πϕ
h
w) −Φ
c
(πw, πw)
= lim
h→0
1
h
S
L+c
x
w
[
[0,h]
= lim
h→0
1
h
h
0
L
x
w
(s), ˙ x
w
(s)
+c
ds
= L(w) +c.
We claim that for any p ∈ M and any w ∈
´
Σ, h ∈ R,
G(ϕ
h
w) = Φ
c
p, π(ϕ
h
w)
= Φ
c
p, π(w)
+ Φ
c
π(w), π(ϕ
h
w)
G(ϕ
h
w) = Φ
c
p, π(w)
+F
w
ϕ
h
(w)
. (3.16)
86
39. coboundary property. 87
This is enough to prove the theorem because then
dG
dϕ
w
=
d
dh
F
h
(ϕ
h
w)
h=0
=
F
w
dϕ
w
= L(w) +c,
and G is Lipschitz by proposition 21.1.
We now prove (3.16). Let q := π(w), x := π(ϕ
h
w). We have to
prove that
Φ
c
(p, x) = Φ
c
(p, q) + Φ
c
(q, x). (3.17)
Since the points q and x can be joined by the static curve x
w
[
[0,h]
, then
Φ
c
(x, q) = −Φ
c
(q, x).
Using twice the triangle inequality for Φ
c
we get that
Φ
c
(p, q) ≤ Φ
c
(p, x) + Φ
c
(x, q) = Φ
c
(p, x) −Φ
c
(q, x) ≤ Φ
c
(p, q).
This implies (3.17).
87
88 3. globally minimizing orbits.
310 Covering Properties.
310.1 Theorem. π(Σ
+
(L)) = M.
But in general π : Σ
+
(L) →M is not injective.
Proof: First suppose that
´
Σ(L) = . Take p ∈ π(
´
Σ). Given x ∈
M ` π(
´
Σ), take a Tonelli minimizer γ
n
∈ (
T
n
(x, p) such that
A
L+c
(γ
n
) < Φ
c
(x, p) +
1
n
.
By the a priori bounds 32.3, [ ˙ γ
n
[ < A and T
n
>
1
A
d(x, p). Let v =
lim
k
˙ γ
n
k
(0) be an accumulation point of ' ˙ γ
n
(0)`. Let η(t) := π ϕ
t
(v).
Then, if 0 < s < liminf
k
T
n
k
, we have that
A
L+c
η[
[0,s]
= lim
k
A
L+c
γ
n
k
[
[0,s]
≤ lim
k
Φ
c
γ
n
k
(0), γ
n
k
(s)
+
1
n
k
= Φ
c
η(0), η(s)
.
Then η is semistatic on [0, S], where S = liminf
k
T
n
k
. If S < +∞ then
η(S) = lim
k
γ
n
k
(T
k
) = p. Since x / ∈ π(
´
Σ), this contradicts the graph
property 38.1; hence S = +∞. Thus η[
[0,+∞[
is semistatic and v ∈ Σ
+
.
If
´
Σ = , then by corollary 36.2, M is noncompact. Let x ∈ M
and 'y
n
` ⊆ M such that d
M
(x, y
n
) → +∞. Let γ
n
∈ (
T
n
(x, y
n
) be a
Tonelli minimizer such that
A
L+c
(γ
n
) < Φ
c
(x, y
n
) +
1
n
.
Then by lemma 32.3, [ ˙ γ
n
[ < A, and hence T
n
→ +∞. The rest of the
proof is similar to the case above.
By corollary 35.3 E(Σ
+
) = c(L), using (1.5) we get that c(L) ≥ e
0
and then
e
0
≤ c
u
≤ c
a
≤ c
0
≤ c(L). (3.18)
88
311. recurrence properties. 89
311 Recurrence Properties.
Let Λ be the set of static classes. Deﬁne a reﬂexive partial order in
Λ by
(a) is reﬂexive.
(b) is transitive.
(c) If there is v ∈ Σ with the αlimit set α(v) ⊆ Λ
i
and ωlimit set ω(v) ⊆ Λ
j
, then Λ
i
Λ
j
.
311.1 Theorem.
Suppose that M is compact and the number of static classes is ﬁnite.
Then given Λ
i
and Λ
j
in Λ, we have that Λ
i
Λ
j
.
fig. 2: connecting orbits between static classes.
The three closed curves represent the static classes and the
other curves represent semistatic orbits connecting them.
Theorem 311.1 could be restated by saying that if the cardinality
of Λ is ﬁnite, then given two static classes Λ
i
and Λ
j
there exist classes
Λ
i
= Λ
1
, . . . , Λ
n
= Λ
j
and semistatic vectors v
1
, . . . , v
n−1
∈ Σ such that
for all 1 ≤ k ≤ n − 1 we have that α(v
k
) ⊆ Λ
k
and ω(v
k
) ⊆ Λ
k+1
.
In other words, between two static classes there exists a chain of static
classes connected by heteroclinic semistatic orbits (cf. ﬁgure 2).
89
90 3. globally minimizing orbits.
A proof of the following theorem can be found in [12]
311.2 Theorem. If M is compact, then
1. Σ(L) is chain transitive.
2.
´
Σ(L) is chain recurrent.
Now we proceed to prove theorem 311.1. Assume for the rest of this
section that M is compact.
311.3 Proposition.
If v ∈ Σ is semistatic, then α(v) ⊂
´
Σ(L) and ω(v) ⊂
´
Σ(L). Moreover
α(v) and ω(v) are each contained in a static class.
Proof: We prove only that ω(v) ⊂
´
Σ. Let γ(t) = π ϕ
t
(v). Suppose that
t
n
→ +∞ and ˙ γ(t
n
) → w ∈ TM. Let η(t) = π ϕ
t
(w). Since γ and η
are solutions the EulerLagrange equation, then γ[
[t
n
−s,t
n
+s]
−→
C
1
η[
[−s,s]
.
Then
A
L+c
η[
[−s,s]
+ Φ
c
(η(s), η(−s)) =
= lim
n
¸
A
L+c
(γ[
[t
n
−s,t
n
+s]
) + lim
m
A
L+c
(γ[
[t
n
+s,t
m
−s]
)
¸
= lim
n
lim
m
A
L+c
(γ[
[t
n
−s,t
m
−s]
)
= lim
n
lim
m
Φ
c
(γ(t
n
−s), γ(t
m
−s))
= Φ
c
(η(−s), η(−s)) = 0.
Thus w ∈
´
Σ(L). Let u ∈ ω(v). We may assume that ˙ γ(s
n
) → u with
t
n
< s
n
< t
n+1
. Then
d
c
(πw, πu) = Φ
c
(πw, πu) + Φ
c
(πu, πw)
= lim
n
A
L+c
(γ[
[t
n
,s
n
]
) +A
L+c
(γ[
[s
n
,t
n+1
]
)
= lim
n
A
L+c
(γ
n
[
[t
n
,t
n+1
]
) = Φ
c
(πw, πw) = 0.
Thus w and u are in the same static class.
90
311. recurrence properties. 91
311.4 Proposition. Every static class is connected.
Proof: Let Λ be a static class and suppose that it is not connected. Let
U
1
, U
2
be disjoint open sets such that Λ ⊆ U
1
∪ U
2
and Λ ∩ U
i
= ,
i = 1, 2. Let p
i
∈ π(U
i
∩ Λ), i = 1, 2. Since U
1
and U
2
are disjoint sets
we can take a solution x
v
n
: [a
n
, b
n
] → M, a
n
< 0 < b
n
of (EL) such
that x
v
n
(0) / ∈ π(U
1
∪ U
2
), x
v
n
(a
n
) = p
1
, x
v
n
(b
n
) = p
2
and
A
L+c
(x
v
n
) ≤ Φ
c
(p
1
, p
2
) +
1
n
. (3.19)
Let u be a limit point of v
n
, then x
u
: R → M is semistatic (see the
proof of claim 2 item (a)). Then, for a
n
≤ s ≤ t ≤ b
n
,
d
c
(p
1
, p
2
) ≤ Φ
c
(p
1
, x
v
n
(s))+Φ
c
(x
v
n
(s), x
v
n
(t))+Φ
c
(x
v
n
(t), p
2
)+Φ
c
(p
2
, p
1
),
therefore
d
c
(p
1
, p
2
) ≤ Φ
c
(p
2
, p
1
)
+ liminf
n
[Φ
c
(p
1
, x
v
n
(s)) + Φ
c
(x
v
n
(s), x
v
n
(t)) + Φ
c
(x
v
n
(t), p
2
)]
≤ Φ
c
(p
2
, p
1
) + liminf
n
A
L+c
(x
v
n
)
≤ d
c
(p
1
, p
2
) = 0,
where in the last inequality we used (3.19). Hence
Φ
c
(p
1
, x
u
(s)) + Φ
c
(x
u
(s), x
u
(t)) + Φ
c
(x
u
(t), p
2
) + Φ
c
(p
2
, p
1
) = 0.
Combining the last equation with the triangle inequality we obtain
d
c
(x
u
(s), x
u
(t)) ≤
≤ Φ
c
(x
u
(s), x
u
(t)) + [Φ
c
(x
u
(t), p
2
) + Φ
c
(p
2
, p
1
) + Φ
c
(p
1
, x
u
(s))] = 0.
So that u ∈
´
Σ. Moreover, for s = 0, t = 1:
d
c
(x
u
(0), p
1
) ≤
≤ Φ
c
(p
1
, x
u
(0)) + [Φ
c
(x
u
(0), x
u
(1)) + Φ
c
(x
u
(1), p
2
) + Φ
c
(p
2
, p
1
)] = 0.
Hence x
u
(0) ∈ π(Λ). On the other hand x
u
(0) / ∈ π(U
1
∪ U
2
). This
contradicts the fact that Λ ⊆ U
1
∪ U
2
.
91
92 3. globally minimizing orbits.
Proof of theorem 311.1.
Given v ∈ TM denote by α(v) and ω(v) its α and ωlimits respec
tively. By proposition 311.4 the static classes are connected. Hence
if we assume that there are only ﬁnitely many of them, the connected
components of
´
Σ are ﬁnite and must coincide with the static classes. For
ε > 0, let
´
Σ(ε) be the εneighborhood of
´
Σ, i.e.
´
Σ(ε) := ¦ v ∈ TM[ d
TM
(v,
´
Σ) < ε ¦.
Fix ε > 0 small enough such that the connected components of
´
Σ(ε) are
the εneighborhoods of the static classes. So that for 0 < δ < ε,
´
Σ(δ) =
¸
N(ε)
i=1
Λ
i
(δ), where Λ
i
(δ) are disjoint open sets containing exactly one
static class and the number of components N(ε) is ﬁxed for all 0 < δ < ε.
Now suppose that the theorem is false. Let Λ
i
, Λ
k
∈ Λ be such that
Λ
i
Λ
k
. Let
A :=
¸
{ Λ
j
∈Λ Λ
i
Λ
j
}
Λ
j
, B :=
¸
{ Λ
j
∈Λ Λ
i
Λ
j
}
Λ
j
.
Given v ∈ Σ with α(v) ⊆ A and 0 < δ < ε, deﬁne inductively s
k
(v),
t
k
(v), T
k
(v) as follows. Let
s
1
(v) := inf¦ s ∈ R[ f
s
(v) / ∈ A(ε) ¦ ∈ R ∪ ¦+∞¦.
If s
k
(v) < +∞, k ≥ 1, deﬁne
t
k
(v) := sup¦ t < s
k
(v) [ f
t
(v) ∈ A(δ) ¦,
T
k
(v) := inf¦ t > s
k
(v) [ f
t
(v) ∈ A(δ) ¦.
Observe that s
k
(v) < +∞ implies that T
k
(v) < +∞ because by the
deﬁnition of B and the transitivity of we have that ω(v) ⊆ A. Deﬁne
A
k
= A
k
(δ) := sup¦ [ T
k
(v) −t
k
(v) [ : v ∈ Σ, α(v) ⊆ A, s
k
(v) < +∞¦,
92
311. recurrence properties. 93
if s
k
(v) = +∞for all v ∈ Σ with α(v) ⊆ A, write A
(δ) ≡ 0 for all ≥ k.
Now set:
s
k+1
(v) := inf¦ s > T
k
(v) [ f
t
(v) / ∈ A(ε) ¦.
Observe that s
k
(v), t
k
(v) and T
k
(v) are invariant under f
t
.
We split the rest of the proof of theorem 311.1 into the following
claims:
Claim 1. A
k
(δ) < +∞ for all k = 1, 2, . . . and all 0 < δ < ε.
Deﬁne
M:= ¦ v [ v ∈ Σ, α(v) ⊆ A¦.
Claim 2.
(a) M∩ B = .
(b) limsup
k
A
k
(δ) = sup
k
A
k
(δ) = +∞.
Claim 3. There exist sequences v
n
∈ Σ, 0 < s
n
< t
n
such that v
n
→
u
1
∈ A, f
s
n
(v
n
) →u
2
/ ∈ A(ε), f
t
n
(v
n
) →u
3
∈ A and d
c
(πu
1
, πu
3
) = 0.
We now use claim 3 to complete the proof of theorem 311.1. If u
1
∈
Λ
j
⊆ A, we shall prove that u
2
∈ Λ
j
`A(ε), obtaining a contradiction and
thus proving theorem 311.1. It is enough to show that d
c
(πu
1
, πu
2
) = 0.
Indeed
d
c
(πu
1
, πu
2
) = Φ
c
(πu
1
, πu
2
) + Φ
c
(πu
2
, πu
1
)
≤ Φ
c
(πu
1
, πu
2
) + Φ
c
(πu
2
, πu
3
) + Φ
c
(πu
3
, πu
1
)
≤ lim
n
Φ
c
(πv
n
, πf
s
n
(v
n
)) +Φ
c
(πf
s
n
(v
n
), πf
t
n
(v
n
))
+ Φ
c
(πu
3
, πu
1
)
= lim
n
Φ
c
(πv
n
, πf
t
n
(v
n
)) + Φ
c
(πu
3
, πu
1
)
= d
c
(πu
1
, πu
3
) = 0,
where the fourth equation holds because v
n
is a semistatic vector.
93
94 3. globally minimizing orbits.
Proof of claim 1:
Suppose that A
i
< +∞ for i = 1, . . . , k − 1 and A
k
= +∞. The
case k = 1 is similar. Then there exists v
n
∈ Σ, with α(v
n
) ⊂ A and
T
k
(v
n
) − t
k
(v
n
) → +∞. We can assume that t
k
(v
n
) = 0 and that v
n
converges (Σ is compact). Let u = lim
n
v
n
∈ ∂A(δ). Then for all n, we
have that
m¦ t < 0 [ f
t
(v
n
) / ∈ A(ε) ¦ ≤
k−1
¸
i=1
A
i
,
where m is the Lebesgue measure on R. This implies that
m¦ t < 0 [ f
t
(u) / ∈ A(ε) ¦ ≤
k−1
¸
i=1
A
i
and hence α(u) ∩ A(ε) = . By proposition 311.3, α(u) ⊂ A. Since
f
t
(v
n
) / ∈ A(ε) for 0 < t < T
k
(v
n
) and T
k
(v
n
) → +∞, then f
t
(u) / ∈ A(ε)
for all t > 0 and hence ω(u) ⊆ B. But then the orbit of u contradicts
the deﬁnition of B.
Proof of claim 2:
(a) Let p ∈ πA, q ∈ πB. For n > 0, let x
v
n
: [a
n
, b
n
] → M be a
solution of (EL) such that x
v
n
(a
n
) = p, x
v
n
(b
n
) = q and
A
L+c
(x
v
n
) ≤ Φ
c
(p, q) +
1
n
.
This implies that
A
L+c
x
v
n
[
[s,t]
≤ Φ
c
x
v
n
(s), x
v
n
(t)
+
1
n
(3.20)
for all a
n
≤ s ≤ t ≤ b
n
. We can assume that
inf¦ s > a
n
[ x
v
n
(s) ∈ B(δ) ¦ = 0,
and that the sequence v
n
converges (cf. lemma 32.3). Let u = lim
n
v
n
∈
π
−1
(∂ π B(δ)). Taking limits in (3.20) we obtain that x
u
[
[s,t]
is semistatic
for all liminf
n
a
n
≤ s ≤ t ≤ limsup
n
b
n
.
94
311. recurrence properties. 95
Any limit point w of ˙ x
v
n
(a
n
) = f
a
n
(v
n
) satisﬁes π(w) = p ∈ πA,
and by the graph property (theorem 38.1), w ∈ A. Similarly, any limit
point of f
b
n
(v
n
) is in B. Since A ∪ B is invariant and u / ∈ A ∪ B, then
lim
n
a
n
= −∞, lim
n
b
n
= +∞. Hence u ∈ Σ. Since f
t
(v
n
) / ∈ B(δ) for
all a
n
≤ t < 0 and a
n
→ −∞, then f
t
(u) / ∈ B(δ) for all t < 0. Hence
α(u) ⊆ A and thus u ∈ M. Since u ∈ π
−1
(∂ π B(δ)) there exists z ∈ B
such that d
M
(π(u), π(z)) ≤ δ. Since z ∈
´
Σ and u ∈ Σ, by theorem 38.1
we have that
d
TM
(π(u), u), (π(z), z)
≤ K δ.
Thus u ∈ M∩ B(Kδ). Letting δ →0, we obtain that M∩ B = .
(b) By claim 1 it is enough to show that sup
k
A
k
(δ) = +∞. If
sup
k
A
k
(δ) < T, then M⊆ M(δ, T), where
M(δ, T) = ¦ v ∈ Σ[ f
[−T,T]
(v) ∩ A(δ) = ¦.
Then M∩B ⊆ M(δ, T)∩B = , because B is invariant and B∩A(δ) = .
This contradicts item (a).
Proof of claim 3:
Given 0 < δ < ε, by claim 2(b) there exists k > N(ε) such that
A
k
(δ) > 0. Hence there is v = v
δ
∈ Σ with α(v) ⊂ A, such that the orbit
of v leaves A(ε) and returns to A(δ) at least k times. Since k > N(ε)
there is one component Λ
j
(δ) ⊆ A(δ) with two of these returns, i.e.
there exist τ
1
(δ) < s(δ) < τ
2
(δ) with f
τ
1
(v) ∈ Λ
j
(δ), f
s
(v) / ∈ A(ε) and
f
τ
2
(v) ∈ Λ
j
(δ). We can choose v
δ
so that τ
1
(δ) = 0. Now, there exists a
sequence such that the repeated component Λ
j
⊂ Λ
j
(δ
n
) is always the
same. Let s
n
:= s(δ
n
), t
n
:= τ
2
(δ
n
) and choose a subsequence such that
v
n
, f
s
n
(v
n
) and f
t
n
(v
n
) converge. Let u
1
= lim
n
v
n
∈ ∩
n
Λ
j
(δ
n
) = Λ
j
,
u
3
= lim
n
f
t
n
(v
n
) ∈ Λ
j
and u
2
= lim
n
f
s
n
(v
n
) / ∈ A(ε). Since u
1
, u
3
∈ Λ
j
,
then d
c
(πu
1
, πu
3
) = 0.
95
96 3. globally minimizing orbits.
96
Chapter 4
The Hamiltonian
viewpoint.
41 The HamiltonJacobi equation.
Let ω be the canonical symplectic form on T
∗
M. A subspace λ of
T
p
T
∗
M is called isotropic if ω(X, Y ) = 0 for all X, Y on λ. Since ω is
nondegenerate, the isotropic subspaces have dimension ≤ n, half of the
dimension of T
∗
M. Isotropic spaces of dimension n are called lagrangian
subspaces We say that a submanifold W ⊂ T
∗
M is lagrangian if at
each point θ ∈ W, its tangent space T
θ
W is a lagrangian subspace of
T
θ
T
∗
M. In particular, dimW = dimM = n.
41.1 Theorem (HamiltonJacobi).
If the hamiltonian H is constant on a lagrangian submanifold N,
then N is invariant under the hamiltonian ﬂow.
Proof: We only have to show that the hamiltonian vector ﬁeld X is
tangent to N. Since H is constant on N, then dH[
TN
≡ 0. Since
ω(X, ) = dH, then ω(X, ξ) = 0 for all ξ ∈ TN. Since the tangent spaces
to N are lagrangian, they are maximal isotropic subspaces, therefore
X ∈ TN.
97
98 4. the hamiltonian viewpoint.
Some distinguished ndimensional manifolds on T
∗
M are the graph
submanifolds, which are of the form
G
η
= ¦ (x, η
x
) [ x ∈ M ¦ ⊂ T
∗
M, (4.1)
where η
x
is a 1form on M. A lagrangian graph is a lagrangian graph
submanifold. In fact,
41.2 Lemma. G
η
is a lagrangian graph if and only if the form η is
closed:
G
η
is lagrangian ⇐⇒ dη ≡ 0
Proof: Choose local coordinates q
1
, . . . , q
n
of M. Then η(q) =
¸
k
p
k
(q) dq
k
. A basis of the tangent space to the graph G
η
is given
by E
i
=
∂
∂q
i
,
¸
k
∂p
k
∂q
i
∂
∂p
k
. Applying ω = dp ∧ dq,
ω(E
i
, E
j
) =
∂p
i
∂q
j
−
∂p
j
∂q
i
.
Since
dη =
¸
i<j
∂p
i
∂q
j
−
∂p
j
∂q
i
dq
j
∧ dq
i
,
then ω[
TG
η
≡ 0 ⇐⇒dη ≡ 0.
Thus, we can associate a cohomology class [η] ∈ H
1
(M, R) to each
lagrangian graph G
η
. Lagrangian graphs with zero cohomology class are
the graphs of the exact 1forms: G
df
, with η = df and f : M → R a
smooth function. These are called exact lagrangian graphs.
The HamiltonJacobi equation for autonomous hamiltonians is
H(x, d
x
u) = k, u : M →R. (HJ)
Thus a smooth solution of the HamiltonJacobi equation corresponds to
an exact invariant lagrangian graph.
98
42. dominated functions. 99
42 Dominated functions.
We say that a function u is dominated by L+k, and write u ≺ L+k if
u(y) −u(x) ≤ Φ
k
(x, y) for all x, y ∈ M.
The triangle inequality implies that the functions u(x) = Φ
k
(y, x)
and v(x) = −Φ
k
(x, y) are dominated, for any y ∈ M.
42.1 Lemma.
1. If u ≺ L +k, then u is Lipschitz with the same Lipschitz constant
as Φ
c
. In particular, a family of dominated functions is equicon
tinuous.
2. If u ≺ L+k then H(x, d
x
u) ≤ k at any diﬀerentiability point x of
u.
Proof:
1. We have that u(y) −u(x) ≤ Φ
c
(x, y) ≤ A d
M
(x, y), where A is a
Lipschitz constant for Φ
c
. Changing the roles of x and y, we get that u
is Lipschitz.
2. We have that
u(y) −u(x) ≤
γ
L(γ, ˙ γ) +k
for all curves γ ∈ ((x, y). This implies that
d
x
u v ≤ L(x, v) +k
for all v ∈ T
x
M when u is diﬀerentiable at x ∈ M. Since
H(x, d
x
u) = sup¦ d
x
u v −L(x, v) [ v ∈ T
x
M¦,
then H(x, d
x
u) ≤ k.
99
100 4. the hamiltonian viewpoint.
42.2 Exercises:
1. If u : M →R is diﬀerentiable, then
H(x, d
x
u) ≤ k ⇐⇒u ≺ L +k.
2. Fix x
0
∈ M. For k > c(L) choose f
k
∈ C
∞
(M, R) such that f(x
0
) = 0
and H(df
k
) < k (cf. 44.4). Let u(x) := limsup
k→c(L)
f
k
(x).
Then u < +∞, u is Lipschitz and H(x, d
x
u) ≤ c(L) for a.e. x ∈ M.
42.3 Deﬁnition. Given a dominated function u ≺ L +k, we say that
an absolutely continuous curve γ : [a, b] →M realizes u, if
u(γ(t)) −u(γ(s)) = A
L+k
(γ[
[s,t]
), for all a ≤ s ≤ t ≤ b. (4.2)
Observe that such realizing curves must be global minimizers. In
particular for k = c(L), they are semistatic.
The following proposition shows that we actually get a solution
of (HJ) if there are (semistatic) curves which realize a dominated func
tion u.
42.4 Proposition. Suppose that u ≺ L +k.
1. If γ :] −ε, ε[→M realizes u, then u is diﬀerentiable at γ(0).
2. If γ :]−ε, 0] →M or γ : [0, ε[→M realizes u and u is diﬀerentiable
at x = γ(0), then d
x
u = L
v
(x, ˙ γ(0)) and H(x, d
x
u) = k.
42.5 Remarks.
1. Equation d
x
u = L
v
(x, ˙ γ(0)) means that the tangent vector
(x, ˙ γ(0)) of any a.c. curve γ realizing u is sent by the Legendre
transform to d
x
u.
2. In particular, since the functions u(x) = Φ
k
(p, x) (resp. v(x) =
−Φ
c
(x, p)) are dominated, then they are diﬀerentiable at any point
which is not at the (backward) (resp. forward) (L + k)cut locus
of p.
100
42. dominated functions. 101
3. Observe that the energy E(x, ˙ γ(0)) = H(x, d
x
u). In proposition 4
9.7, we show that any semistatic orbit realizes some dominated
function. Thus we obtain another proof for Σ ⊂ E
−1
¦c¦, i.e. that
the semistatic orbits have energy c(L).
Proof: 1. Let w ∈ T
x
M and let η(s, t) be a variation of γ ﬁxing the
endpoints γ(−ε), γ(ε) such that η(0, t) = γ(t) and
∂
∂s
η(0, 0) = w. Deﬁne
/(s) :=
0
−ε
L(
∂
∂t
η(s, t)) +k dt.
Then, integrating by parts and using the EulerLagrange equation (EL),
/
(0) = L
v
ξ[
0
−ε
+
0
−ε
L
x
−
d
dt
L
v
ξ dt = L
v
(x, ˙ γ(0)) w,
where ξ(t) :=
∂
∂s
η(0, t). Also
1
s
u(η(s, 0)) −u(x)
=
1
s
u(η(s, 0)) −u(γ(−ε)) +u(γ(−ε)) −u(γ(0))
≤
1
s
[ /(s) −/(0) ] ,
where we used that u ≺ L +k and (4.2). Hence
lim
s→0
sup
1
s
u(η(s, 0)) −u(x)
≤ /
(0). (4.3)
Similarly, if B(s) := A
L+k
η(s, )[
[0,ε]
, then
u(γ(ε)) −u(η(s, 0)) −u(γ(ε)) +u(x) ≤ B(s) −B(0),
limsup
s→0
1
s
u(x) −u(η(s, 0))
≤ B
(0) = −L
v
(x, ˙ γ(0)) w.
Hence
liminf
s→0
1
s
u(η(s, 0)) −u(x)
≥ L
v
(x, ˙ γ(0)) w. (4.4)
From (4.3) and (4.4) we get that u is diﬀerentiable at x = γ(0). and
d
x
u = L
v
(x, ˙ γ(0)).
101
102 4. the hamiltonian viewpoint.
2. Now assume that γ :] −ε, 0] →M realizes u and that u is diﬀer
entiable at x = γ(0). The same argument as in 4.3 shows that
d
x
u w ≤ L
v
(x, ˙ γ(0)) w for all w ∈ T
x
M.
Applying this inequality to −w and combining both inequalities we get
that d
x
u = L
v
(x, ˙ γ(0)).
Now, since u ≺ L +k, by lemma 42.1, H(x, d
x
u) ≤ k. Since
u(γ(0)) −u(γ(t)) = A
L+k
(γ[
[t,0]
) =
0
t
L(γ(s), ˙ γ(s)) +k
ds,
then
d
x
u ˙ γ(0) = L(γ(0), ˙ γ(0)) +k.
Hence
H(x, d
x
u) = sup
v∈T
x
M
¦ d
x
u v −L(x, v) ¦ ≥ k.
102
43. weak solutions of the hamiltonjacobi equation. 103
43 Weak solutions of the HamiltonJacobi
equation.
We shall see in corollary 44.7 that there are no weakly diﬀerentiable
subsolutions of (HJ) for k < c(L).
In the next proposition we show that when k ≥ c(L) there are always
Lipschitz solutions of (HJ). On the other hand, in theorem 48.4 we
show that when M is compact the only energy level that supports a dif
ferentiable solution is k = c(L). When M is noncompact there may
be diﬀerentiable solutions on k > c(L), as example ?? shows.
43.1 Proposition. If k ≥ c(L), then for any y ∈ M, the function
u(x) = Φ
k
(y, x) satisﬁes H(x, d
x
u) = k for a.e. x ∈ M.
Proof: Since u is Lipschitz, by by Rademacher’s theorem [19], it is
diﬀerentiable at Lebesguealmost every point. Since u is dominated by
proposition 42.4.2, it is enough to see that u is onesided realized at
every point.
If k > c(L), by proposition 35.1, for all x ∈ M, x = y, there exists a
ﬁnitetime global minimizer γ ∈ (
T
(y, x) with A
L+k
(γ) = Φ
k
(y, x). By
the triangle inequality, the function δ(t) = A
L+k
(γ[
[0,t]
) −Φ
k
(y, γ(t)) is
increasing. Also δ(t) ≥ 0 and δ(T) = 0. So that δ(t) ≡ 0 and hence γ
backwardrealizes u at x.
If k = c(L) then u may be realized by an inﬁnite semistatic orbit as
follows. Let γ
n
∈ ((y, x), γ
n
: [−T
n
, 0] →M be a Tonelli minimizer such
that
Φ
c
(y, x) ≤ A
L+c
(γ
n
) ≤ Φ
c
(y, x) +
1
n
.
This implies that
Φ
c
(y, γ(s)) ≤ A
L+c
(γ
n
[
[−T
n
,s]
) ≤ Φ
c
(y, γ(s)) +
1
n
,
for all −T
n
≤ s ≤ 0. Thus
u(x) −u(γ
n
(s)) −A
L+c
(γ
n
[
[s,0]
)
≤
1
n
. (4.5)
103
104 4. the hamiltonian viewpoint.
By proposition 32.3, [γ
n
[ < A for all n. Hence, if x = y, liminf
n
T
n
> 0.
We can assume that ˙ γ
n
(0) → v ∈ T
x
M and T
n
> ε > 0. Let λ(t) =
π ϕ
t
(v). Then γ
n
[
[−ε,0]
→ λ[
[−ε,0]
in the C
1
topology. Letting n → ∞
on (4.5), we get that λ realizes u.
104
44. lagrangian graphs. 105
44 Lagrangian graphs.
We say that a function u : M → R is a subsolution of the Hamilton
Jacobi equation if
H(x, d
x
u) ≤ k,
We shall prove that for k > c(L) there is always a C
∞
subsolution of
the HamiltonJacobi equation and for k < c(L) there are no (weakly)
diﬀerentiable subsolutions. Hence
44.1 Theorem. If M is any covering of a closed manifold, then
c(L) = inf
f∈C
∞
(M,R)
sup
x∈M
H(x, d
x
f)
= inf¦k ∈ R : there exists f ∈ C
∞
(M, R) such that H(df) < k¦.
where H is the hamiltonian associated with L.
Theorem 44.1 could be restated by saying that c(L) is the inﬁmum of
the values of k ∈ R for which H
−1
(−∞, k) contains an exact lagrangian
graph. This is a very geometric way of describing the critical value.
In exercise 42.2. there is an elementary construction of a weak
subsolution for k = c(L). Theorem 44.1 is an immediate consequence
of lemma 44.2 and proposition 44.4 below.
44.2 Lemma. If there exists a C
1
function f : M → R such that
H(df) ≤ k, then k ≥ c(L).
Proof: Recall that
H(x, p) = max
v∈T
x
M
¦p(v) −L(x, v)¦.
Since H(df) ≤ k it follows that for all (x, v) ∈ TM,
d
x
f(v) −L(x, v) ≤ k.
105
106 4. the hamiltonian viewpoint.
Therefore, if γ : [0, T] → M is any absolutely continuous closed curve
with T > 0, we have
T
0
(L(γ, ˙ γ) +k) dt =
T
0
L(γ, ˙ γ) +k −d
γ
f( ˙ γ)
dt ≥ 0,
and thus k ≥ c(L).
44.3 Remark. The utility of a diﬀerentiable subsolution of the
HamiltonJacobi equation can be seen in lemma 44.2. If H(du) ≤ c(L),
then the lagrangian L can be replaced by the lagrangian
L(x, v) := L(x, v) −d
x
u(v) +c(L) ≥ 0
The new lagrangian L is positive, has the same minimizing measures as
L, its α and β functions are translates of those for L. The static set
´
Σ(L) is contained in the level set L = 0.
44.4 Proposition.
For any k > c(L) there exists f ∈ C
∞
(M, R) such that H(df) < k.
Proof: Set c = c(L). Fix q ∈ M and let u(x) := Φ
c
(q, x). By the
triangle inequality, u ≺ L + c. By lemma 42.1.2, H(d
x
u) ≤ c at any
point x ∈ M where u(x) is diﬀerentiable.
We proceed to regularize u. Since u is Lipschitz, by Rademacher’s
theorem (cf. [19]) it is diﬀerentiable at Lebesgue almost every point.
Moreover it is weakly diﬀerentiable (cf. [19, Section 4.2.3]), that is, for
any C
∞
function ϕ : U →R with compact support, equation (4.6) holds.
The next lemma completes the proof.
44.5 Lemma. Let M be a riemannian covering of a compact manifold
and suppose that sup
v≤k
∂L
∂x
(x, v)
< +∞. If u : M → R is weakly
diﬀerentiable and
H(x, d
x
u) ≤ k for a.e. x ∈ M,
106
44. lagrangian graphs. 107
then for all δ > 0 there exists f ∈ C
∞
(M, R) such that H(x, d
x
f) < k+δ
for all x ∈ M.
44.6 Remark. If M is a covering M
ρ
→ N of a compact manifold N
and the lagrangian L on M is lifted from a lagrangian on N: L = ◦dρ,
then the condition on
∂L
∂x
follows. In general it may not be true as the
lagrangian L(x, v) =
1
2
[v[
2
+ sin(x
2
) on R shows.
The condition on M can be replaced by some bounds in the rieman
nian metric of M, see appendix ??.
Proof: We shall explain ﬁrst how to prove the proposition in the case
in which M is compact and then we will lift the construction to an
arbitrary covering M.
We can assume that M ⊂ R
N
. Let U be a tubular neighbourhood
of M in R
N
, and ρ : U → M a C
∞
projection along the normal bun
dle. Extend u(x) to U by u(z) = u
ρ(z)
. Then u(z) is also weakly
diﬀerentiable.
Extend the lagrangian to U by
L(z, v) = L
ρ(z), d
z
ρ(v)) +
1
2
[v −d
z
ρ(v)[
2
.
Then the corresponding hamiltonian satisﬁes H(z, p ◦ d
z
ρ) = H
ρ(z), p
for p ∈ T
∗
ρ(z)
M. At any point of diﬀerentiability of u, we have that
d
z
u = d
ρ(z)
u ◦ d
z
ρ, and H(d
z
u) = H(d
ρ(z)
u) ≤ k.
Let ε > 0 be such that
(a) The 3εneighbourhood of M in R
N
is contained in U.
(b) If x ∈ M, (y, p) ∈ T
∗
R
N
= R
2N
, H(y, p) ≤ k and d
R
N(x, y) < ε,
then H(x, p) < k +δ.
Let ψ : R → R be a C
∞
function such that ψ(x) ≥ 0, support(ψ) ⊂
(−ε, ε) and
R
N
ψ([x[) dx = 1. Let K : R
N
R
N
→ R be K(x, y) =
ψ([x − y[). Let N
ε
be the εneighbourhood of M in R
N
. Deﬁne f :
107
108 4. the hamiltonian viewpoint.
N
ε
→R by
f(x) =
R
N
u(y) K(x, y) dy .
Then f is C
∞
on N
ε
.
Observe that ∂
x
K(x, y) = −∂
y
K(x, y). Since u(y) is weakly diﬀer
entiable, for any C
∞
function ϕ : U →R with compact support
R
N
(ϕ du +u dϕ) dx = 0 . (4.6)
Hence
−
R
N
u(y) ∂
y
K(x, y) dy =
R
N
K(x, y) d
y
u dy.
Now, since
d
x
f =
R
N
u(y) ∂
x
K(x, y) dy,
we obtain
d
x
f =
R
N
K(x, y) d
y
u dy.
From the choice of ε > 0 we have that H(x, d
y
u) < k +δ for almost
every y ∈ suppK(x, ) and x ∈ M. Since K(x, y) dy is a probability
measure, by Jensen’s inequality
H(d
x
f) ≤ H(d
x
f) ≤
R
N
H(x, d
y
u) K(x, y) dy < k +δ .
for all x ∈ M.
Now, suppose that M is a covering of a compact manifold N with
covering projection p. Assume that N ⊆ R
N
. We can regularize our
function u similarly as we shall now explain. For ¯ x ∈ M let x be the
projection of ¯ x to N and let µ
x
be the Borel probability measure on N
deﬁned by
N
ϕdµ
x
=
R
N
(ϕ ◦ ρ)(y) K(x, y) dy,
108
44. lagrangian graphs. 109
for any continuous function ϕ : N →R. Then the support of µ
x
satisﬁes
supp(µ
x
) ⊂ ¦y ∈ N : d
N
(x, y) < ε¦.
Let ´ µ
e x
be the Borel probability measure on M uniquely deﬁned by the
conditions: supp(´ µ
e x
) ⊂ ¦´ y ∈ M : d
M
(¯ x, ´ y) < ε¦ and p
∗
´ µ
e x
= µ
x
. Then
we have
d
d¯ x
M
ϕd´ µ
e x
=
M
d
b y
ϕd´ µ
e x
(´ y),
for any weakly diﬀerentiable function ϕ : M → R. The condition (b)
above is now granted by the bound on
∂L
∂x
. Then the same arguments
as above show that
f(¯ x) =
M
u(´ y) d´ µ
e x
(´ y),
satisﬁes H(d
e x
f) < k.
Combining lemma 44.5 with lemma 44.2, we obtain:
44.7 Corollary.
There are no weakly diﬀerentiable subsolutions of (HJ) for k < c(L).
If we consider lagrangian graphs with other cohomology classes, we
obtain Mather’s alpha function:
44.8 Corollary. If M is compact,
α(κ) = inf
[ω]=κ
sup
x∈M
H(x, ω(x)).
In particular the critical value of the abelian cover c
0
= min
κ
α(κ)
is the inﬁmum of the energy levels which contain a lagrangian graph of
any cohomology class in is interior.
Proof: Let us ﬁx a closed one form ω
0
such that [ω
0
] = κ. By equality
(2.30) we have that α(κ) = c(L −ω
0
). Hence, it suﬃces to show that
c(L −ω
0
) = inf
[ω]=κ
sup
x∈M
H(x, ω(x)). (4.7)
109
110 4. the hamiltonian viewpoint.
It is straightforward to check that the hamiltonian associated with L−ω
0
is H(x, p+ω
0
(x)). Since all the closed one forms in the class κ are given
by ω
0
+df where f ranges among all smooth functions, equality (4.7) is
now an immediate consequence of theorem 44.1.
110
45. contact ﬂows. 111
45 Contact ﬂows.
Let N be a 2n+1 smooth manifold and let α be a nondegenerate
1form, i.e. α ∧ (dα)
n
is a volume form for N.
dimker dα ≡ 1, where
ker d
x
α := ¦ v ∈ T
x
N [ d
x
α(v, w) = 0, ∀w ∈ T
x
N ¦, x ∈ N.
Deﬁne a vector ﬁeld Y on N by Y (x) ∈ ker d
x
α,
etc
45.1 Proposition. The hamiltonian ﬂow on the energy level [H ≡ k]
is a reparametrization of a Θpreserving ﬂow if and only if k > c(L).
¿Tambi´en se tiene desigualdad estricta cuando M no es com
pacta?
Proof: Let Σ = [H ≡ k]. Observe that
d
dt
ψ
∗
t
Θ = L
X
Θ[
Σ
= i
X
dΘ +d i
X
Θ = dH[
Σ
+dΘ(X) = dΘ(X)[
Σ
.
Thus a reparametrization of ψ
t
preserves Θ if and only if its vector ﬁeld
is a constant multiple of Y =
1
Θ(X)
X. Such reparametrization exists if
and only if Θ(X) = 0 on Σ.
Since H is convex, the sets Σ
x
:= T
∗
x
M ∩ Σ have compact convex
interiors. The outward normal vector to Σ
x
at (x, p) is H
p
(x, p). Observe
that Θ(X) = p H
p
. Then if the point (x, 0) lies on the exterior of Σ
x
and (x, p) is the tangency point of a tangent line to Σ
x
passing through
(x, 0), we have that Θ(X(x, p)) = 0. Thus Θ(X) = 0 on Σ implies that
111
112 4. the hamiltonian viewpoint.
the zero section of T
∗
M lies in H < k. Since the zero section is the
derivative of a constant function, by theorem 44.1, k ≥ c(L).
Moreover, if the zero section lies inside H < k, then Θ(X) > 0.
Now suppose that k > c(L). By theorem 44.1, there exists
f ∈ C
∞
(M, R) with H(df) < k. Deﬁne the new convex hamiltonian
H(x, p) := H(x, p + d
x
f). Then the energy level [H ≡ k] contains the
zero section. Thus, if X is its hamiltonian vector ﬁeld, then Θ(X) > 0.
Let L := L −df. Let H be the hamiltonian of L and X its hamiltonian
ﬂow. Then
112
46. finsler metrics. 113
46 Finsler metrics.
In this section we prove that if k > c(L) then the EulerLagrange ﬂow
on the energy level E ≡ k is a reparametrization of the geodesic ﬂow
on the unit tangent bundle of a Finsler metric. This allows to borrow
theorems from Finsler geometry.
We give ﬁrst a lagrangian proof and afterwards a hamiltonian proof,
with a ﬂavor of symplectic geometry.
When k > c(L), the subsolution H(df) < k obtained in proposi
tion 44.4 can be used to replace the lagrangian L by the lagrangian
L = L − df, then L + k > 0. These lagrangians have the same energy
function and equivalent variational principles. Hence they have the same
lagrangian ﬂow, minimizing orbits, and the same action functional on
closed curves and invariant measures. Their action potentials are related
by
P
k
(x, y) = Φ
k
(x, y) +f(y) −f(x).
Given a Finsler metric
√
F and an absolutely continuous curve γ,
deﬁne its Finsler length as
l
F
(γ) =
F( ˙ γ).
Observe that since the Finsler metric is homogeneous of degree one, the
deﬁnition does not depend on the parametrization of the curve. Deﬁne
the Finsler distance as
D
F
(x, y) = inf
γ∈C(x,y)
l
F
(γ).
46.1 Theorem. [14, 30] If k > c(L) then the lagrangian ﬂow on the
energy level E ≡ k is a reparametrization of the geodesic ﬂow of a Finsler
metric on its unit tangent bundle.
Moreover, if f ∈ C
∞
(M, R) is such that H(df) < k, the Finsler
lagrangian F can be taken to be
F(x, v) = L(x, v) +k −d
x
f(v)
113
114 4. the hamiltonian viewpoint.
on E(x, v) = k, and then
Φ
k
(x, y) = D
F
(x, y) +f(y) −f(x), for all x, y ∈ M. (4.8)
If k > −inf L then f can chosen f ≡ 0.
Proof: By theorem 44.1, if k > c(L) then there exists f ∈ C
∞
(M, R)
such that H(df) < k. Observe that
H(x, 0) = max
v∈T
x
M
0 v −L(x, v)
= − inf
v∈T
x
M
L(x, v).
So that if k > −inf L, we can choose f ≡ 0.
Observe that if H(df) < k then k > c(L) > e
0
. So that the zero
section M 0 ⊂ E
−1
¦k¦ ⊂ TM. Also, L + k − df > 0. Then we can
deﬁne a Finsler metric on TM by
√
F = L+k−df on E
−1
¦k¦ and extend
it by homogeneity. Since k > c(L) = c(L−df), by proposition 35.1, for
any x = y there exists a global minimizer on ((x, y) for L+k −df which
has energy k. By the homogeneity of
√
F, we can restrict the curves in
the deﬁnition of D
F
to those with energy k. Thus Φ
k
−∆f = D
F
.
To show that the lagrangian ﬂow on E ≡ k is a reparametrization
of the geodesic ﬂow on the unit tangent bundle of
√
F, we only need to
prove that suﬃciently small EulerLagrange solutions with energy k are
geodesics of
√
F. Let L = L − df. The equality (4.8) implies that any
(L +k)global minimizer is a geodesic for
√
F. So it is enough to prove
that suﬃciently small orbits with energy k are global minimizers.
Fix x ∈ M and a small neighbourhood ^(x) of x such that for all
y ∈ ^(x) there exists a unique EulerLagrange solution contained in
^(x), with energy k and joining x to y. Let P
k
be the action potential
for L and let
ε = inf¦ P
k
(x, z) [ z / ∈ ^(x) ¦ > 0,
´(x) = ¦ y ∈ M[ P
k
(x, y) <
ε
2
¦.
Then ´(x) is a neighborhood of ^(x) with ´(x) ⊂ ^(x). By the
triangle inequality, any (L+k)global minimizer joining x to a point y ∈
114
46. finsler metrics. 115
´(x) must be contained in ^(x). By proposition 35.1 such minimizer
exists. Hence all the small solutions contained in ^(x) joining x to
points y ∈ ´(x) are global minimizers.
Now we shall give a hamiltonian proof of theorem 46.1.
First we need a hamiltonian characterization of Finsler lagrangians.
We say that a function f : R
n
→R is a Finsler energy if f(x) > 0 when
x = 0 and f is positively homogeneous of order 2. Observe that L is a
Finsler lagrangian if and only if it is a Finsler energy on each tangent
space T
x
M
46.2 Lemma. Let f ∈ C
2
(M, R) be strictly convex and superlinear.
Then f is a Finsler energy if and only if its convex dual f
∗
is a Finsler
energy.
Proof: If f is homogeneous, then, writing v = λw, λ > 0, we have
f
∗
(λp) = max
v
λp v −f(v)
= max
w
λ
2
p w −λ
2
f(w)
= λ
2
f
∗
(p).
Let L
f
be the Legendre transform of f. Observe that
f
∗
◦L
f
(p) = f
(p) p−f(p) =
d
dt
f(tp)
t=1
−f(p) = 2 f(p) −f(p) = f(p).
So that f
∗
> 0.
Since f is strictly convex and superlinear, then f
∗∗
= f. Thus the
argument above shows that if f
∗
is homogeneous then so is f.
46.3 Lemma. If two convex hamiltonians have a common regular level
set Σ, then their hamiltonian vector ﬁelds are parallel on Σ.
Proof: Suppose that H
−1
¦k¦ = G
−1
¦¦ = Σ. If p ∈ Σ, then d
p
H =
ker d
p
G = T
p
Σ. Since Σ is a regular energy level, these derivatives are
nonzero. Moreover, since H and G are convex, they are positive on
vectors pointing outwards Σ. Thus d
p
H = λ(p) d
p
G for some λ(p) > 0.
Also,
ω(X
H
, ) = d
p
H = λ(p) d
p
G = λ(p) ω(X
G
, )
So that X
H
= λ(p) X
G
.
115
116 4. the hamiltonian viewpoint.
46.4 Lemma. If F is a Finsler lagrangian, then the orbits of its la
grangian ﬂow are reparametrizations of the unit speed geodesics of F.
Proof: We prove it for a Finsler hamiltonian G. The result follows
because the EulerLagrange ﬂow of F is the hamiltonian ﬂow of its
energy function on TM with respect to the symplectic form L
∗
F
(ω),
where ω is the canonical symplectic form on T
∗
M.
Let G be a Finsler hamiltonian. Since G(x, λp) = λG(x, p), then
d
λp
G(w) =
d
ds
G(x, λp +sw)
s=1
=
d
ds
λ
2
G(x, p +s
w
λ
)
s=1
= λd
p
G(w).
If X is the hamiltonian ﬂow of G, then ω(X(p), ) = d
p
G. Therefore
X(λp) = λX(p).
Hamiltonian proof of theorem 46.1:
If k > c(L), by theorem 44.1, there exists f ∈ C
∞
(M, R) with
H(df) < k. Let H(x, p)
def
= H(x, p +d
x
f). Then H
−1
(] −∞, k[) contains
the zero section of T
∗
M. Deﬁne a new hamiltonian G : T
∗
M−M0 →R
by G ≡
1
4
on H
−1
¦k¦ and G(x, λp) = λ
2
G(x, p) for all λ ≥ 0. By
lemma 46.2, the convex dual G
∗
of G is a Finsler metric on TM.
Since by deﬁnition G
−1
¦
1
4
¦ = H
−1
¦k¦, it follows from lemma 4
6.3 that the hamiltonian ﬂows of G and H
−1
¦k¦ coincide up to
reparametrization on the energy level G
−1
¦
1
4
¦ = H
−1
¦k¦.
The Legendre transforms L
G
(x, p) = (x, G
p
) and L
H
(x, p) = (x, H
p
)
on H
−1
¦k¦ satisfy
G
p
w = 0 = H
p
p, for all w ∈ T
p
(T
∗
X
M ∩ H
−1
¦k¦);
G
p
p = 2 G(p) =
1
2
> 0, H
p
p = Θ(X) > 0, (by 45.1).
(4.9)
So that L
G
(x, p) = λL
H
(x, p) for some λ(x, p) > 0. Also L
H
conjugates
the hamiltonian ﬂow on H
−1
¦k¦ to the lagrangian ﬂow of L on E
−1
¦k¦.
By lemma 46.4 the orbits of the EulerLagrange ﬂow of L on E
−1
¦k¦
are reparametrization of unit speed geodesic of G
∗
.
116
46. finsler metrics. 117
We now compute G
∗
. From (4.9), H
p
= 2 Θ(X) G
p
, thus
G
∗
(x, H
p
) = G
∗
(x, 2 Θ(X) G
p
) = 4 Θ(X)
2
G
∗
(x, G
p
)
= 4 Θ(X)
2
G(p) = Θ(X)
2
= (p H
p
)
2
= (L
v
v)
2
= (L +k)
2
.
Let h : T
∗
M → T
∗
M be the map h(x, p) = (x, p + d
x
f). Then the
hamiltonian ﬂows φ
t
of H and ψ
t
of H
df
(x, p) = H(x, p+d
x
f) are related
by h ◦ φ
t
= ψ
t
◦ h. Thus, the hamiltonian ﬂow of H is conjugate to a
Finsler hamiltonian ﬂow.
A hamiltonian level set can be made a Finsler level set if and only
if it contains the zero section. On the other hand, a lagrangian energy
level [E = k] with k > e
0
always contains the zero section.
117
118 4. the hamiltonian viewpoint.
47 Anosov energy levels.
An Anosov energy level is a regular energy level on which the ﬂow φ
t
is
an Anosov ﬂow.
47.1 Theorem. If the energy level E
−1
(k) is Anosov, then
k > c
u
(λ).
Proof: Suppose that the energy level k is Anosov and set Σ
def
= H
−1
(k).
Let π : T
∗
N →N denote the canonical projection. G.P. Paternain and
M. Paternain proved in [57] that Σ must project onto the whole manifold
N and that the weak stable foliation J
s
of φ
∗
t
is transverse to the ﬁbers
of the ﬁbration by (n −1)spheres given by
π[
Σ
: Σ →N.
Let
¯
N be the universal covering of N. Let
¯
Σ denote the energy level k
of the lifted hamiltonian H. We also have a ﬁbration by (n−1)spheres
¯ π[
e
Σ
:
¯
Σ →
¯
N.
Let
J
s
be the lifted foliation which is in turn a weak stable foliation
for the hamiltonian ﬂow of H restricted to
¯
Σ. The foliation
J
s
is also
transverse to the ﬁbration ¯ π[
e
Σ
:
¯
Σ →
¯
N. Since the ﬁbers are compact a
result of Ehresmann (cf. [7]) implies that for every (x, p) ∈
¯
Σ the map
¯ π[
f
W
s
(x,p)
:
J
s
(x, p) →
¯
N,
is a covering map. Since
¯
N is simply connected, ¯ π[
f
W
s
(x,p)
is in fact a dif
feomorphism and
J
s
(x, p) is simply connected. Consequently,
J
s
(x, p)
intersects each ﬁber of the ﬁbration ¯ π[
e
Σ
:
¯
Σ →
¯
N at just one point.
In other words, each leaf
J
s
(x, p) is the graph of a one form. On the
other hand it is well known that the weak stable leaves of an Anosov
118
47. anosov energy levels. 119
energy level are lagrangian submanifolds. Since any closed one form in
the universal covering must be exact, it follows that each leaf
J
s
(x, p) is
an exact lagrangian graph. The theorem now follows from lemma 44.2
and the fact that by structural stability there exists ε > 0 such that for
all k
∈ (k −ε, k +ε) the energy level k
is Anosov.
For e ∈ R, let /
e
be the set of φ ∈ C
∞
(M) such that the ﬂow of
H + φ is Anosov in (H + φ)
−1
(e) and let B
e
be the set of φ ∈ C
∞
(M)
such that (H +φ)
−1
(e) contains no conjugate points. As is well known
/
e
is open in C
k
topology and B
e
is closed. On the other hand G. and
M. Paternain [54] have shown that /
e
is contained in B
e
. It is proved
in [15] the following
47.2 Theorem. The interior of B
e
in the C
2
topology is /
e
.
This theorem is an extension to the Hamiltonian setting of a result
of R. O. Ruggiero for the geodesic ﬂow [64]. Theorems 47.2 and 47.1
have as corollary:
47.3 Corollary. Given a convex superlinear lagrangian L, k < c
u
(L)
and ε > 0 there exists a smooth function ψ : N →R with [ψ[
C
2 < ε and
such that the energy level k of L +ψ possesses conjugate points.
Proof: Suppose now that there exists > 0 such that for every ψ with
[ψ[
C
2 < , the energy level k of λ + ψ has no conjugate points. The
main result in [15] says that in this case the energy level k of λ must be
Anosov thus contradicting theorem 47.1.
47.4 Proposition. If k is a regular value of the energy such that k < e,
then the energy level k has conjugate points.
Proof: If an orbit does not have conjugate points then there exist along
it two subbundles called the Green subbundles. They have the following
properties: they are invariant, lagrangian and they have dimension n =
dimN. Moreover, they are contained in the same energy level as the
orbit and they do not intersect the vertical subbundle (cf. [13]). If k is
119
120 4. the hamiltonian viewpoint.
a regular value of the energy with k < e, then π(E
−1
(k)) is a manifold
with boundary and at the boundary the vertical subspace is completely
contained in the energy level. Therefore the orbits that begin at the
boundary must have conjugate points, because at the boundary two n
dimensional subspaces contained in the energy level (which is (2n −1)
dimensional) must intersect.
120
48. the weak kam theory. 121
48 The weak KAM Theory.
In the rest of this chapter we develop the theory of global weak KAM
solutions, discovered by Albert Fathi. Recall that for an autonomous
hamiltonian H : T
∗
M →R, the HamiltonJacobi equation is
H(x, d
x
u) = k, (HJ)
where u : U ⊆ M → R. Here we are interested on global solutions
of (HJ), i.e. u : M →R satisfying (HJ).
It may not be possible to obtain a smooth global solution of (HJ).
Instead, for certain values of k, we shall ﬁnd weak solutions of (HJ),
which are Lipschitz. By Rademacher’s theorem [19], a Lipschitz function
is Lebesgue almost everywhere diﬀerentiable so that (HJ) makes sense
in a.e. point.
In fact, we have seen in corollary 44.7 that there are no weakly dif
ferentiable
1
global solutions for k ≤ c(L). In theorem 48.4 we shall see
that when M is compact, there are no C
1+Lip
solutions unless k = c(L).
In proposition 49.7 we show that there are always Lipschitz solutions
for k = c(L), and in proposition 410.2 we show that when M is non
compact there are Lipschitz solutions for k > c(L).
Given a dominated function u ≺ L +c deﬁne the sets
Γ
+
0
(u) := ¦ v ∈ Σ
+
[ u(x
v
(t)) −u(x
v
(0)) = Φ
c
(x
v
(0), x
v
(t)), ∀t > 0 ¦,
Γ
−
0
(u) := ¦ v ∈ Σ
−
[ u(x
v
(0)) −u(x
v
(t)) = Φ
c
(x
v
(t), x
v
(0)), ∀t < 0 ¦,
Γ
+
(u) :=
¸
t>0
ϕ
t
Γ
+
(u)
, Γ
−
(u) :=
¸
t<0
ϕ
t
Γ
+
(u)
,
where x
v
(t) = π ϕ
t
(v). We call Γ
+
(u) (resp. Γ
+
(u)) the basin of u and
π
Γ
+
0
(u) ` Γ
+
(u)
(resp. π
Γ
+
0
(u) ` Γ
+
(u)
) the cut locus of u.
48.1 Deﬁnition.
A function u
−
: M →R is a backward weak KAM solution of (HJ) if
1
In particular, by Rademacher’s theorem, there are no Lipschitz global solutions.
121
122 4. the hamiltonian viewpoint.
1. u
−
≺ L +c.
2. π
Γ
−
0
(u
−
)
= M.
A function u
+
: M →R is a forward weak KAM solution of (HJ) if
1. u
+
≺ L +c.
2. π
Γ
+
0
(u
+
)
= M.
48.2 Remark.
From the domination condition it follows that u is Lipschitz and that
the curve γ is semistatic. From proposition 42.4, at an interior point
x of such a curve γ, the function u is diﬀerentiable and H(x, d
x
u) = c.
Moreover, item 2 in proposition 42.4 shows that if u is diﬀerentiable
at an endpoint of a curve γ, then H(x, d
x
u) = c. By Rademacher’s
theorem [19], u is diﬀerentiable at (Lebesgue) almost every point in M.
So that u is indeed a weak solution of the HamiltonJacobi equation for
k = c(L).
48.3 Theorem.
If u ∈ S
+
(resp. u ∈ S
−
) is a weak KAM solution, then
1. u is Lipschitz and hence diﬀerentiable (Lebesgue)almost every
where.
2. u ≺ L +c.
3. H(x, d
x
u) = c(L) at any diﬀerentiability point x of u.
4. Covering Property: π(Γ
+
0
(u)) = M.
5. Graph Property: π : Γ
+
(u) →M is injective and its inverse is
Lipschitz, with Lipschitz constant depending only on L.
6. Smoothness Property: u is diﬀerentiable on Γ
+
(u) and its
derivative d
x
u is the image of (π[
Γ
+
(u)
)
−1
(x) under the Legendre
transform L of L. In particular, the energy of Γ
+
0
(u) is c(L).
122
48. the weak kam theory. 123
Proof: Items 2 and 4 are the deﬁnition of u ∈ S
+
. Item 1 follows from
proposition 42.1.1 and Rademacher’s theorem [19]. Item 3 follows from
proposition 42.4.2 and the fact that by item 4, u is onesided realized
at every point. Item 6 follows from proposition 42.4 and remark 42.5.
We prove item 5. Let (z
1
, v
1
), (z
2
, v
2
) ∈ Γ
+
(u) and suppose that
d
TM
(v
1
, v
2
) > K d
M
(z
1
, z
2
), where K is from lemma 38.2 and the A
that we input on lemma 38.2 is from lemma 32.3. Let 0 < ε < ε
1
, (with
ε
1
from lemma 38.2) be such that ϕ
−ε
(z
i
, v
i
) ∈ Γ
+
(u). Let y
i
= x
v
i
(ε),
i = 1, 2, then u(y
i
) = u(x
i
) + Φ
c
(x
i
, y
i
), i = 1, 2. Then lemma 38.2
implies that
Φ
c
(x
1
, y
2
) + Φ
c
(x
2
, y
1
) < Φ
c
(x
1
, y
1
) + Φ
c
(x
2
, y
2
).
Adding u(y
1
) +u(y
2
) and using that u ≺ L +c, we get that
u(x
1
) +u(x
2
) ≤ Φ
c
(x
1
, y
2
) +u(y
2
) + Φ
c
(x
2
, y
1
) +u(y
1
)
< Φ
c
(x
1
, y
1
) +u(y
1
) + Φ
c
(x
2
, y
2
) +u(y
2
)
= u(x
1
) +u(x
2
),
which is a contradiction. This proves item 5.
fig. 1: graph property.
123
124 4. the hamiltonian viewpoint.
48.4 Theorem (Fathi [23]).
If u ∈ C
1+Lip
(M, R), M is compact and
H(x, d
x
u) = k, ∀x ∈ M. (4.10)
Then k = c(L) and u is a weak KAM solution u ∈ S
−
∩ S
+
.
Conversely, if u ∈ S
+
∩ S
−
, then u ∈ C
1+Lip
.
Proof: From (4.10) we get that
max
v∈T
x
M
d
x
u v −L(x, v) = k. (4.11)
The strict convexity of L on T
x
M implies that the maximum is attained
at a unique vector ξ(x) ∈ T
x
M. The implicit function theorem implies
that the vector ﬁeld x →ξ(x) is Lipschitz. Thus it can be integrated to
obtain a ﬂow ψ
t
on M.
From (4.11), u ≺ L +k and the ﬂow lines of ψ
t
realize L +k, i.e.
u
ψ
t
(x)
= u
ψ
s
(x)) +
t
s
L(ψ
τ
(x),
d
dτ
ψ
τ
(x)) +k
dτ, ∀s < t, ∀x ∈ M.
(4.12)
From (4.11), L +k −du ≥ 0. This implies that
c(L −du) = c(L) ≥ k.
Let µ be and invariant measure for the ﬂow ψ
t
. Observe that the measure
ν := ξ
∗
(µ) is holonomic and L +k = du on supp(ν). Then
(L +k) dν =
(du) dν = 0.
Hence k ≤ c(L) and thus k = c(L). Then by (4.12), u ∈ S
−
∩ S
+
.
Conversely, if u ∈ S
−
∩ S
+
, by 48.3.6, u is diﬀerentiable and
L
−1
(d
x
u) = ξ(x). Moreover, Γ
∓
(u) = M and by 48.3.5, ξ is Lipschitz.
124
49. construction of weak kam solutions 125
49 Construction of weak KAM solutions
In this section we present three ways to construct weak KAM solutions:
when the Peierls set is nonempty (in remark 49.3.4), when the Peierls
barrier is ﬁnite (in proposition 49.2), and the general case (in proposi
tion 49.7). In the horocycle ﬂow (example 58), the Peierls barrier is
ﬁnite, but the Peierls set is empty. In example 57, h
c
= +∞ and also
{ = .
When M is compact we characterize all weak KAM solutions in terms
of their values on each static class.
We begin by observing that
49.1 Lemma.
1. If  ⊆ S
−
is such that v(x) := inf
u∈U
u(x) > −∞, for all x ∈ M;
then v ∈ S
−
.
2. If  ⊆ S
+
is such that v(x) := sup
u∈U
u(x) < +∞, for all x ∈ M;
then v ∈ S
+
.
Proof: We only prove item 1. Since u ≺ L +c for all u ∈ , then
v(y) = inf
u∈U
u(y) ≤ inf
u∈U
u(x) + Φ
c
(x, y) = v(x) + Φ
c
(x, y). (4.13)
Thus v ≺ L +c.
Let x ∈ M and choose u
n
∈  such that u
n
(x) → v(x). Choose
w
n
∈ Γ
−
(u
n
) ∩ T
x
M. Since by lemma 32.3 [w
n
[ < A, we can assume
that w
n
→ w ∈ T
x
M. By lemma 42.1.1, all the functions u ∈  have
the same Lipschitz constant K as Φ
c
. For t < 0, we have that
v(x
w
(t)) ≤ liminf
n
u
n
(x
w
n
(t)) +K d
M
(x
w
(t), x
w
n
(t))
= liminf
n
u
n
(x) −Φ
c
(x
w
(t), x) +K d
M
(x
w
(t), x
w
n
(t))
= v(x) −Φ
c
(x
w
(t), x) ≤ v(x
w
(t)), because v ≺ L +c.
Hence w ∈ Γ
−
(v).
125
126 4. the hamiltonian viewpoint.
49.a Finite Peierls barrier.
49.2 Proposition. If h
c
< +∞ and f : M → R is a continuous
function. Suppose that
v
−
(x) := inf
z∈M
f(z) +h
c
(z, x) > −∞,
v
+
(x) := sup
z∈M
f(z) −h
c
(x, z) > −∞.
Then v
−
∈ S
−
and v
+
∈ S
+
.
Proof: We only prove that v
−
∈ S
−
. By lemma 49.1 it is enough to
prove that the functions u(x) →h
c
(z, x) are in S
−
for all z ∈ M.
By proposition 37.1.4, u ≺ L + c. Now ﬁx x ∈ M. Choose Tonelli
minimizers γ
n
: [T
n
, 0] →M such that γ
n
∈ ((z, x), T
n
< −n and
A
L+c
γ
n
[
[T
n
,0]
≤ h
c
(z, x) +
1
n
.
By lemma 32.3, [ ˙ γ
n
(0)[ < A for all n. We can assume that ˙ γ
n
(0)
n
→
w ∈ T
x
M. If −n ≤ s ≤ 0, then s > T
n
and
A
L+c
γ
n
[
[T
n
,s]
+ Φ
c
(γ
n
(s), x) ≤
≤ A
L+c
γ
n
[
[T
n
,s]
+A
L+c
γ
n
[
[s,0]
≤ h
c
(z, x) +
1
n
≤ h
c
(z, γ
n
(s)) + Φ
c
(γ
n
(s), x) +
1
n
, for −n ≤ s < 0.
Taking liminf
n→∞
, we get that
h
c
(z, x
w
(s)) +A
L+c
x
w
[
[s,0]
= h
c
(z, x).
Hence w ∈ Γ
−
(u).
49.3 Remarks.
126
49. construction of weak kam solutions 127
1. Observe that, since Φ
c
(x, x) = 0,
u ≺ L +c ⇐⇒u(x) = inf
z∈M
u(z) + Φ
c
(z, x).
2. Item 49.3.1 implies that the function h
c
in proposition 49.2 can
not be replaced by Φ
c
. In fact, the function u
z
(x) = Φ
c
(z, x)
satisﬁes u
z
≺ L + c, but in general u / ∈ S
−
, if z is not properly
chosen.
3. For any z ∈ M the function u
z
(x) = h
c
(z, x) ∈ S
−
and v
z
(x) :=
−h
c
(x, z) ∈ S
+
.
4. If p ∈ { then u
p
(x) := Φ
c
(p, x) ∈ S
−
, because
Φ
c
(p, x) ≤ h
c
(p, x) ≤ h
c
(p, p) + Φ
c
(p, x) ≤ Φ
c
(p, x).
Similarly, v
p
(x) := −Φ
c
(x, p) ∈ S
+
.
49.b The compact case.
In the next theorem we characterize all weak KAM solutions when M
is compact. They are determined by their values on one point of each
static class.
Let Γ = {/d
c
be the set of static classes of L. For each γ ∈ Γ choose
p
Γ
∈ Γ and let P = ¦ p
Γ
[ Γ ∈ Γ¦. We say that a function f : P → R is
dominated (f ≺ L +c) if f(p) ≤ f(q) + Φ
c
(q, p), for all p, q ∈ P.
49.4 Theorem.
If M is compact, the maps ¦ f : P →R[ f ≺ L +c ¦ →S
−
,
f −→u
f
(x) := inf
p∈P
f(p) + Φ
c
(p, x),
and ¦ f : P →R[ f ≺ L +c ¦ →S
+
,
f −→v
f
(x) := sup
p∈P
f(p) −Φ
c
(x, p),
are bijective isometries in the sup norm.
127
128 4. the hamiltonian viewpoint.
Proof: We only prove it for f →u
f
.
The domination condition f ≺ L + c implies that u
f
> −∞. Then
remark 49.3.4 and lemma 49.1.1 imply that u
f
∈ S
−
.
The injectivity follows from
u
f
(p) = min
q∈P
f(q) + Φ
c
(q, p) = f(p) ∀p ∈ P,
because f is dominated.
To prove the surjectivity, let u ∈ S
−
and let f = u[
P
. Given x ∈ M
choose w ∈ Γ
−
0
(u) ∩ T
x
M and let γ(t) = π ϕ
t
(w). So that
u(x) −u(γ(t)) = A
L+c
(γ[
[t,0]
) = Φ
c
(γ(t), x) for t < 0. (4.14)
Choose q ∈ π[αlim(v)] ⊂ { (by proposition 311.3), and t
n
→−∞such
that γ(t
n
)
n
→ q. Using t = t
n
on equation (4.14), in the limit we have
that
u(x) = u(q) + Φ
c
(q, x). (4.15)
Now take p ∈ P such that d
c
(p, q) = 0. Since u ≺ L +c, then
u(q) ≤ u(p) + Φ
c
(p, q) ≤ u(q) + Φ
c
(q, p) + Φ
c
(p, q) = u(q).
So that
u(q) = u(p) + Φ
c
(p, q). (4.16)
By the triangle inequality
Φ
c
(p, x) ≤ Φ
c
(p, q) + Φ
c
(q, x)
≤ Φ
c
(p, q) + Φ
c
(q, p) + Φ
c
(p, x) = Φ
c
(p, x).
So that
Φ
c
(p, x) = Φ
c
(p, q) + Φ
c
(q, x). (4.17)
Combining equalities (4.15), (4.17) and (4.16), we have that
u(x) = u(p) + Φ
c
(p, x),
128
49. construction of weak kam solutions 129
with p ∈ P. So that u
f
≤ u. But since u ≺ L + c and f = u[
P
, using
remark 49.3.1, we have that u ≤ u
f
.
Now we see that f →u
f
is an isometry in the supremum norm  
0
.
Given x ∈ M, choose p
n
∈ P such that u
g
(x) = lim
n
f(p
n
) + Φ
c
(p
n
, x).
Since u
f
≺ L +c, we have that
u
f
(x) −u
g
(x) ≤ liminf
n
f(p
n
) + Φ
c
(p
n
, x) −g(p
n
) −Φ
c
(p
n
, x)
≤ f −g
0
.
Changing the roles of f and g we get that u
f
−u
g

0
≤ f −g
0
. Since
u
f
[
P
= f and u
g
[
P
= g, then u
f
−u
g

0
≥ f −g
0
.
49.5 Corollary. There is only one static class if and only if S
−
(resp.
S
+
) is unitary modulo an additive constant.
This characterization of weak KAM solutions allows us to recover
the following theorem: We say that two weak KAM solutions u
−
∈ S
−
and u
+
∈ S
+
are conjugate if u
−
= u
+
on { and denote it by u
−
∼ u
+
.
49.6 Corollary. (Fathi [20]) If M is compact, then
h(x, y) = sup
u
∓
∈S
∓
u
−
∼u
+
¦ u
−
(y) −u
+
(x) ¦.
Proof: If u
+
∼ u
−
and p ∈ {, from the domination we get
u
+
(p) ≤ u
=
(x) + Φ
c
(x, p),
u
−
(y) ≤ u
−
(p) + Φ
c
(p, y).
Adding these equations and using that u
=
(p) = u
−
(p), we get
u
−
(y) −u
+
(x) ≤ Φ
c
(x, p) + Φ
c
(p, y).
129
130 4. the hamiltonian viewpoint.
Taking inf
p∈P
and then sup
u
=
∼u
−
we obtain
sup
u
+
∼u
−
¸
u
−
(y) −u
+
(x)
¸
≤ h(x, y).
On the other hand, let u
=
(z) := −h(z, y) and
u
−
(z) : = min
q∈P
¸
u
=
(q) + Φ
c
(q, z) ¦ (4.18)
= min
q∈P
¸
−h(q, y) + Φ
c
(q, z)
¸
= min
q∈P
¸
−φ
c
(q, y) + Φ
c
(q, z) ¦ (4.19)
From remark 49.3.3 and corollary 49.4, u
±
∈ S
±
. Since u
+
is domi
nated, from (4.18) we get that u
+
∼ u
−
. From (4.19), u
−
(y) = 0 and
hence u
−
(y) −u
+
(x) = h(x, y).
49.c Busemann weak KAM solutions.
When h
c
= +∞, we use another method to obtain weak KAM solutions,
resembling the constructions of Busemann functions in riemannian ge
ometry. By proposition 310.1, Σ
+
= and Σ
−
= even when M
is noncompact. We call the functions of proposition 49.7 weak KAM
Busemann functions.
49.7 Proposition.
1. If w ∈ Σ
−
(L) and γ(t) = x
w
(t), then
u
w
(x) = inf
t<0
[Φ
c
(γ(t), x) −Φ
c
(γ(t), γ(0))]
= lim
t→−∞
[Φ
c
(γ(t), x) −Φ
c
(γ(t), γ(0))]
is in S
−
.
130
49. construction of weak kam solutions 131
2. If w ∈ Σ
+
(L) and γ(t) = x
w
(t), then
u
w
(x) = sup
t>0
[Φ
c
(γ(0), γ(t)) −Φ
c
(x, γ(t))]
= lim
t→+∞
[Φ
c
(γ(0), γ(t)) −Φ
c
(x, γ(t))]
is in S
+
.
Proof: We only prove item 1. We start by showing that the function
δ(t) = Φ
c
(γ(t), x) −Φ
c
(γ(t), 0)
is increasing. If s < t, then
δ(t) −δ(s) = Φ
c
(γ(t), x) −Φ
c
(γ(s), x) +
Φ
c
(γ(s), γ(0)) −Φ
c
(γ(t), γ(0))
= Φ
c
(γ(t), x) −Φ
c
(γ(s), x) + Φ
c
(γ(s), γ(t))
≥ 0,
where the last inequality follows from the triangle inequality applied to
the triple (γ(s), γ(t), x). By the triangle inequality, δ(t) ≤ Φ
c
(γ(0), x),
hence lim
t↓−∞
δ(t) = inf
t<0
δ(t) and this limit is ﬁnite.
Since
u(y) = inf
t<0
Φ
c
(γ(t), y) −Φ
c
(γ(t), γ(0))
≤ inf
t<0
Φ
c
(γ(t), x) + Φ
c
(x, y) −Φ
c
(γ(t), γ(0))
= u(x) + Φ
c
(x, y),
then u ≺ L +c.
Suppose that x ∈ { = . Let (x, v) ∈
´
Σ and t < 0. Let p = x
v
(t)
and y ∈ M. Since d
c
(x, p) = 0, then
Φ
c
(y, x) = Φ(y, x) + Φ
c
(x, p) + Φ
c
(p, x)
≥ Φ
c
(y, p) + Φ
c
(p, x) ≥ Φ
c
(y, x).
131
132 4. the hamiltonian viewpoint.
Hence Φ
c
(y, x) = Φ
c
(y, p) + Φ
c
(p, x). For y = γ(s) (and p = x
v
(t)), we
have that
u(x) −u(x
v
(t)) = lim
s→+∞
[Φ
c
(γ(s), x) −Φ
c
(γ(s), x
v
(t))] = Φ
c
(x
v
(t), x)
= A
L+c
x
v
[
[t,0]
.
Now let x ∈ M ` { and choose y
n
: [T
n
, 0] →M a Tonelli minimizer
such that y
n
(T
n
) = γ(n), y
n
(0) = x and
A
L+c
(y
n
[
[T
n
,0]
) ≤ Φ
c
(γ(n), x) +
1
n
.
This implies that
A
L+c
(y
n
[
[s,t]
) ≤ Φ
c
(y
n
(s), y
n
(t)) +
1
n
, for T
n
≤ s < t ≤ 0. (4.20)
By lemma 32.3, [ ˙ y
n
[ < A. We can assume that ˙ y
n
(0) → v ∈ T
x
M.
Then
A
L+c
(x
v
[
[t,0]
) = Φ
c
(γ(t), x) for liminf
n
T
n
≤ t ≤ 0. (4.21)
We prove below that lim
n
T
n
= −∞. Then v ∈ Σ
−
(L). Observe that
for T
n
≤ s ≤ 0 we have that
Φ
c
(γ(n), x) ≤ Φ
c
(γ(n), y
n
(s)) + Φ
c
(y
n
(s), x)
≤ A
L+c
(y
n
[
[T
n
,0]
) ≤ Φ
c
(γ(n), x) +
1
n
.
Since y →Φ
c
(z, y) is uniformly Lipschitz, we obtain that
u(x) = lim
n
Φ
c
(γ(n), x) −Φ
c
(γ(n), γ(0))
= lim
n
Φ
c
(γ(n), x
v
(s)) + Φ
c
(x
v
(s), x) −Φ
c
(γ(n), γ(0))
= u(x
v
(s)) + Φ
c
(x
v
(s), x) for all s < 0.
= u(x
v
(s)) +A
L+c
(x
v
[
[s,0]
) because v ∈ Σ
−
.
Now we prove that lim
n
T
n
= −∞. Suppose, for simplicity, that
lim
n
T
n
= T
0
> −∞. Since ˙ y
n
(0) → v, then y
n
[
[T
n
,0]
C
1
−→ x
v
[
[T
0
,0]
and
132
49. construction of weak kam solutions 133
hence γ(n) = y
n
(T
n
) → x
v
(T
0
) =: p. Since by lemma 32.3 [ ˙ γ[ is
bounded, we can assume that lim
n
˙ γ(n) = (p, w
1
). By lemma 311.3,
w
1
∈ αlim( ˙ γ) ⊆
´
Σ. From (4.21), ˙ x
v
(T
0
) ∈ Σ
ε
. Since π(w
1
) = x
v
(T
0
) =
p, then lemma 38.1 implies that ˙ x
v
(T
0
) ∈
´
Σ. Since
´
Σ is invariant, then
v ∈
´
Σ and hence x = π(v) ∈ π(
´
Σ) = {. This contradicts the hypothesis
x ∈ M ` {.
49.8 Corollary. S
−
= and S
+
= .
133
134 4. the hamiltonian viewpoint.
410 Higher energy levels.
The method in proposition 49.7 allows us to obtain analogous weak
KAM solutions on energy levels k > c(L) when M is noncompact.
Let
Σ
−
(k) :=
¸
v ∈ TM[ A
L+k
x
v
[
[t,0]
= Φ
k
x
v
(t), x
v
(0)
, ∀t ≤ 0
¸
.
For u ≺ L +k deﬁne
Γ
−
(u, k) =
¸
v ∈ TM[ A
L+k
x
v
[
[t,0]
= u
x
v
(0)
−u
x
v
(t)
, ∀t ≤ 0
¸
.
(4.22)
Let
S
−
(k) =
¸
u : M →R[ u ≺ L +k and π
Γ
−
(u, k)
= M
¸
.
Deﬁne similarly Σ
+
(k), Γ
+
(u, k) and S
+
(k). Then the functions u ∈ S
±
satisfy Lipschitz, covering, graph and smoothness properties analogous
to theorem 48.3.
Observe that we are requiring that the global minimizers in (4.22)
are deﬁned in the whole ray ] −∞, 0]. In the following lemma we show
that there are not such weak KAM solutions when M is compact and
k > c(L). If we look for weak solutions of (HJ) with realizing curves
deﬁned only on ﬁnite intervals then the action potential u(x) = Φ
k
(p, x),
v(x) = −Φ
k
(p, x) give such examples.
410.1 Proposition.
If M is compact, then Σ
±
(k) = and hence S
±
(k) = for all k > c(L).
If M is noncompact then Σ
±
(k) = for all k > c(L).
Moreover, if v ∈ Σ
±
(k), then ωlim(v) = (resp. αlim(v) = ).
Proof: Suppose that v ∈ Σ
+
(k) = and that w ∈ ωlim(v) = .
Observe that the orbit of w can not be a ﬁxed point (i.e. w = 0) because
by (3.18), k > c(L) ≥ e
0
. Let p = π(w) and choose s > 0 such that
134
410. higher energy levels. 135
q = π(ϕ
s
(w)) = p. Then the same argument as in proposition 311.3
shows that
d
k
(p, q) = Φ
k
(p, q) + Φ
l
(q, p) = 0,
which contradicts proposition 21.1.4.
This proves that Σ
±
(k) = when M is compact. Observe that if
u
+
∈ S
+
(k) then = Γ
+
(u, k) ⊆ Σ
+
(k). Hence S
±
(k) = when M is
compact.
Assume now that M is noncompact. We show that Σ
+
(k) = 0. Let
x ∈ M and 'y
n
` ⊆ M such that d
M
(x, y
n
) → +∞. Let γ
n
∈ (
T
n
(x, y
n
)
be a Tonelli minimizer such that
A
L+c
(γ
n
) < Φ
k
(x, y
n
) +
1
n
. (4.23)
Then by lemma 32.3, [ ˙ γ
n
[ < A, and hence T
n
→+∞. Let v be a density
point of ¦˙ γ
n
(0)¦. Since γ
n
[
[0,t]
C
1
−→ x
v
[
[0,t]
for all t > 0, from (4.23) we
obtain that v ∈ Σ
+
(k).
By corollary 44.7, there are no weak solutions of (HJ) for k < c(L).
We complete the picture with the following:
410.2 Proposition.
If M is noncompact, then S
±
(k) = for all k ≥ c(L).
If M is compact, then S
±
(k) = for all k > c(L).
Proof: Proposition 410.1 proves the case M compact.
Now we reproduce the proof of proposition 49.7. Let w ∈ Σ
+
(k) =
, write γ(t) := x
w
(t) and
u
w
(x) := lim
t→+∞
[Φ
k
(γ(0), γ(t)) −Φ
k
(x, γ(t))]
The limit exists because by the triangle inequality for Φ
k
, the function
δ(t) = Φ
k
(γ(0), γ(t)) −Φ
k
(x, γ(t))
135
136 4. the hamiltonian viewpoint.
is nondecreasing and it is bounded above by Φ
k
(γ(0), x). The triangle
inequality implies that u ≺ L +k.
Now let x ∈ M and choose y
n
: [0, T
n
] →M a Tonelli minimizer such
that y
n
(0) = x, y
n
(T
n
) = γ(n) and
A
L+c
(y
n
[
[0,T
n
]
) ≤ Φ
k
(x, γ(n)) +
1
n
.
This implies that for all 0 < t < T
n
,
Φ
k
(x, γ(n)) ≤ A
L+k
y
n
[
[0,t]
+ Φ
k
(y
n
(t), γ(n)) ≤ Φ
k
(x, γ(n)) +
1
n
.
(4.24)
By lemma 32.3, [ ˙ y
n
[ < A. Since y
n
(T
n
) = γ(n) and by proposition4
10.1, γ(n) → ∞, then liminf
n
T
n
= +∞. Let v ∈ T
x
M be a density
point of ˙ y
n
(0).
Since y
n
[
[0,t]
C
1
−→x
v
[
[0,t]
, using (4.24) we get that
u(x
v
(t)) = lim
n
Φ
k
(γ(0), γ(n)) −Φ
k
(x
v
(t), γ(n))
= lim
n
Φ
k
(γ(0), γ(n)) −Φ
k
(x, γ(n)) +A
L+k
(y
n
[
[0,t]
)
= u(x) +A
L+k
(x
v
[
[0,t]
),
for all 0 < t < liminf
n
T
n
= +∞.
136
411. the laxoleinik semigroup. 137
411 The LaxOleinik semigroup.
The LaxOleinik semigroup was used to obtain weak KAM solutions by
Lions, Papanicolaou and Varadhan [32] on tori T
n
and later by
Fathi [21] for compact manifolds.
Nahman and Roquejoﬀre
Roquejoﬀre
Through this section we shall assume that M is compact. The
LaxOleinik semigroup 'T
−
`
t≥0
is the semigroup of operators T
−
t
:
C
0
(M, R) →C
0
(M, R) deﬁned by
T
−
t
u(x) = inf
γ
¦ u(γ(0)) +
t
0
L(γ(s), ˙ γ(s)) +c ds ¦
= min
y∈M
¦ u(y) + Φ
c
(y, x; t) ¦,
where the inﬁmum is taken on all piecewise diﬀerentiable curves γ :
[0, T] →M with γ(T) = x. Similarly, deﬁne
T
+
t
u(x) = max
z∈M
¦ u(z) −Φ
c
(x, z; t) ¦.
411.1 Proposition. If M is compact.
1. The unique c ∈ R for which the semigroup 'T
−
t
`
t≥0
has a ﬁxed
point is the critical value c = c(L).
2. u ∈ S
−
⇐⇒ u is a ﬁxed point of the semigroup 'T
−
t
`
t≥0
.
3. For all ε > 0 there exists a Lipschitz constant K = K(ε) > 0 such
that T
−
t
(C
0
(M, R)) ⊆ Lip
K
(M, R) for all t > ε.
4. T
−
t
is a weak contraction: For all t ≥ 0 and all u, v ∈ C
0
(M, R),
T
−
t
u −T
−
t
v
0
≤ u −v
0
.
5. For ε > 0 there exists K(ε) > 0 such that if u, v ∈ C
0
(M, R) and
s, t > ε, then
T
−
s
u −T
−
t
v
0
≤ u −v
0
+K(ε) [s −t[.
137
138 4. the hamiltonian viewpoint.
Proof:
1. By proposition 34.1, lim
t→+∞
Φ
k
(x, y; t) = ±∞ when k = c(L),
uniformly in x, y ∈ M. This implies that there are no ﬁxed points for
T
±
t
when c = c(L). The existence of a ﬁxed point is given by item 2.
2. Let u ∈ S
−
. For x ∈ M take v ∈ Γ
−
(u) ∩ T
x
M and let γ(t) :=
π ϕ
t
(v). For t > 0 we have that
u(x) = u(γ(−t)) +A
L+c
γ[
[−t,0]
= u(γ(−t)) + Φ
c
(γ(−t), x; t).
But since u ≺ L +c, then
u(x) ≤ u(y) + Φ
c
(y, x) ≤ u(y) + Φ
c
(y, x; t).
Hence T
−
t
u = u for all t > 0.
Now suppose that T
−
t
u = u for all t ≥ 0. Then
u(x) ≤ u(z) + Φ
c
(z, x; t) ∀t ≥ 0, ∀z ∈ M.
Taking the inﬁmum on t ≥ 0, we get that u ≺ L +c. Moreover,
u(x) ≤ min
z∈M
u(z) + Φ
c
(z, x) ≤ min
z∈M
u(z) +h
c
(z, x).
For t > 0 let z
t
be such that u(x) = u(z
t
) + Φ
c
(z
t
, x; t). Let t
n
→ +∞
be a sequence such that the limit lim
n
z
t
n
= y exists. Then
u(x) = u(y) + lim
n
Φ
c
(z
t
n
, x; t
n
) ≥ u(y) +h
c
(y, x),
where the last inequality follows from the deﬁnition of h
c
and the
Lipschitz property for Φ
c
(x, y; t) in proposition 34.1.1. Thus u(x) =
min
z∈M
u(z) +h
c
(z, x), by proposition 49.2, u ∈ S
−
.
3. Using proposition 34.1, let K = K(ε) be a Lipschitz constant
for (x, y; t) → Φ
c
(x, y; t) on M M [ε, +∞[. Given u ∈ C
0
(M, R),
138
411. the laxoleinik semigroup. 139
x, y ∈ M and t > ε, let z ∈ M be such that T
−
t
u(y) = u(z) +Φ
c
(z, y; t).
Then
T
−
t
u(x) ≤ u(z) + Φ
c
(z, x; t)
≤ u(z) + Φ
c
(z, y; t) +[Φ
c
(z, x; t) −Φ
c
(z, y; t)[
≤ T
−
t
u(y) +K(ε) d(x, y).
Changing the roles of x and y we get that T
−
t
u ∈ Lip
K
(M, R).
5. Observe that item 5 implies item 4. Let u, v ∈ C
0
(M, R), s, t > ε
and x ∈ M. Let K = K(ε) be as in item 3. Choose z ∈ M such that
T
−
s
u(x) = u(z) + Φ
c
(z, x; s). Then
T
−
t
v(x) ≤ v(z) + Φ
c
(z, x; t)
≤ T
−
s
u(x) +[v(z) −u(z)[ +[Φ
c
(z, x; t) −Φ
c
(z, x; s)[
≤ T
−
t
u(x) +u −v
0
+K(ε) [s −t[.
Changing the roles of u and v we get item 5.
411.2 Remark. Fixed Points for the LaxOleinik semigroup.
Another proof of the existence of a ﬁxed point in 411.3
We sketch here another proof of existence of a ﬁxed point for the
LaxOleinik semigroup using properties 3 and 4 of proposition 411.1.
Consider the semigroup λT
−
t
, with 0 < λ < 1 acting on the
space C
0
(M, R)/
R
of continuous functions modulo an additive con
stant. Then λT
−
t
is a contraction whose image is in the compact space
Lip
K
(M, R)/
R
, independent of λ or t. Let u
λ,t
be a ﬁxed point for λT
−
t
.
Choosing a sequence λ
n
→ 1 and a subsequence such that u
λ
n
,t
con
verges in Lip
K
(M, R)/
R
, one obtains a ﬁxed point u
t
for T
−
t
. Now let
t
n
→ 0 and choose a subsequence such that u
t
n
converges to some v in
Lip
K
(M, R)/
R
. Fix s ∈ R and let N
n
∈ Z be such that lim
n
N
n
t
n
= s.
Since by the semigroup property, u
t
n
is also a ﬁxed point for T
−
N
n
t
n
,
using proposition 411.1.5, we get that in the C
0
topology,
T
−
s
v = lim
n
T
−
N
n
t
n
u
t
n
= lim
n
u
t
n
= v.
139
140 4. the hamiltonian viewpoint.
So that v is a ﬁxed point for the whole semigroup 'T
−
t
`
t≥0
.
411.3 Theorem (Fathi [23]). If M is compact,
for all u ∈ C
0
(M, R) the uniform limit lim
t→+∞
T
−
t
u exists.
To prove this theorem we shall need some lemmas.
411.4 Lemma. If M is compact then lim
t→+∞
1
t
Φ
c
(x, y; t) = 0, uni
formly on (x, y) ∈ M M.
Proof: Write h
t
(x, y) := Φ
c
(x, y; t). Then
inf
M×M
h
t
+ 2 inf
M×M
h
1
≤ h
t+2
≤ inf
M×M
h
t
+ 2 sup
M×M
h
1
.
Then, writing C = 2 (sup h
1
−inf h
1
), we have that
∀t > 2, sup
M×M
h
t
− inf
M×M
h
t
≤ 2 C.
Now let u be a weak KAM solution. Since T
−
t
u = u, we have that for
all t > 0,
u(y) = min
x∈M
u(x) + Φ
c
(x, y; t).
Thus, for all (x, y) ∈ M M and t > 2,
[Φ
c
(x, y; t)[ ≤ sup
x∈M
u(x) − inf
x∈M
u(x) + 2 C.
This implies the lemma.
411.5 Lemma. For all ε > 0, there exists T
0
> 0 such that if T > T
0
and γ : [0, T] →M is a Tonelli minimizer, then [E(γ, ˙ γ) −c(L)[ < ε.
Proof: Let T
n
→+∞ and let γ
n
: [0, T
n
] →M be a Tonelli minimizer.
Then by lemma 32.2 there is A > 0 such that [ ˙ γ
n
[ < A for all n.
Then there exists a subsequence such that the probabilities µ
γ
n
converge
140
411. the laxoleinik semigroup. 141
weakly* to a limit µ. Since L is bounded on [v[ ≤ A, and using lemma 4
11.4,
A
L+c
(µ) = lim
n
A
L+c
(µ
γ
n
) = lim
n
1
T
n
Φ
c
(γ
n
(0), γ
n
(T
n
); T
n
) = 0.
Hence µ is a minimizing measure and E(supp µ) ≡ c(L). Thus
E(γ
n
, ˙ γ
n
) →c(L).
For f : T
∗
M →R measurable and u : M →R Lipschitz, deﬁne
ess sup
p∈T
∗
M
f = inf
A⊆T
∗
M
Leb(T
∗
M\A)=0
sup
p∈A
f(p),
and
H(f) = ess sup
x∈M
H(x, d
x
f).
411.6 Lemma. If u : M →R is Lipschitz, then u ≺ L +H(u).
Proof: Using a convolution argument as in lemma 44.5, we can approx
imate both H(u) and u in the C
0
topology by a C
∞
function. Hence we
can assume that u is C
1
. Then for all (x, v) ∈ TM,
d
x
u v ≤ L(x, v) +H(x, d
x
u) ≤ L(x, v) +H(u).
If γ ∈ ((x, y), then
u(y) −u(x) ≤
γ
du ≤
γ
L +H(u).
411.7 Lemma. If u ∈ C
0
(M, R), then lim
t→+∞
H(T
−
t
u) = c.
141
142 4. the hamiltonian viewpoint.
Proof: By proposition 411.1.3, T
−
t
u is Lipschitz. If x is a diﬀerentia
bility point for T
−
t
u, then
d
x
(T
−
t
u) = L
v
(γ(t), ˙ γ(t)),
where γ : [0, t] → M is a Tonelli minimizer satisfying γ(t) = x, and
T
−
t
u(x) = u(γ(0)) + A
L+c
(γ[
[0,t]
). Thus H(d
x
(T
−
t
u)) = E(γ(t), ˙ γ(t)).
By lemma 411.5, E(γ, ˙ γ) converges to c(L) uniformly on t.
Proof of theorem 411.3:
Let u ∈ C
0
(M, R). By proposition 411.1.2 the family 'T
−
t
u`
t≥1
is
equilipschitz. By Arzela
´
Ascoli theorem there exists a sequence t
n
→
+∞ such that the uniform limit u
∞
= lim
n
T
−
t
n
u exists. By lemma 4
11.6, T
−
t
u ≺ L + H(T
−
t
u) and by lemma 411.7, H(T
−
t
u) → c. This
implies that u
∞
≺ L +c.
Now we prove that u
∞
is a ﬁxed point of the semigroup T
−
t
. Since
u
∞
≺ L + c, then u
∞
≤ T
−
t
u
∞
for all t ≥ 0. Since T
−
t
preserves the
order, we get that u
∞
≤ T
−
s
u
∞
≤ T
−
t
u
∞
for all s ≤ t. It is enough to
show that there is a sequence s
n
→+∞ such that T
s
n
u
∞
→u
∞
. Write
s
n
= t
n+1
−t
n
, we can assume that s
n
→+∞. Then, using 411.1.4,
T
−
s
n
u
∞
−u
∞
0
≤
T
−
s
n
u
∞
−T
−
s
n
T
−
s
n
u
0
+T
−
t
n+1
u −u
∞

0
≤ u
∞
−T
t
n
u
0
+T
−
t
n+1
u −u
∞

0
→0.
It remains to show that the ﬁxed point u
∞
of T
−
t
is the limit of T
−
t
u.
But for all s ≥ 0,
T
−
t
n
+s
u −u
∞
0
=
T
−
s
T
−
t
n
u −T
−
s
u
∞
0
≤
T
−
t
n
u −u
∞
0
n
−→0.
The following corollaries are also due to A. Fathi.
411.8 Corollary. If M is compact, then lim
t→∞
Φ
c
(x, y; t) = h
c
(x, y),
uniformly in x ∈ M and also in y ∈ M.
142
411. the laxoleinik semigroup. 143
Proof: The uniform limit on y ∈ M follows from the equality
Φ
c
(x, y; t) = T
−
t−1
u(y), where u(z) = Φ
c
(x, z; 1). And the uniform
limit on x ∈ M follows from −Φ
c
(x, y; t) = T
+
t−1
v, where v(z) =
−Φ
c
(z, y; 1).
Recall that a ray is a curve γ(t) such that it is a Tonelli minimizer
for all t ≥ 0 (resp. t ≤ 0).
411.9 Corollary. If M is compact then the rays are semistatic.
Proof: For an a.c. curve γ, and s ≤ t in its domain, deﬁne
δ(s, t) := A
L+c
(γ
[s,t]
) −Φ
c
(γ(s), γ(t)).
Observe that δ(s, t) ≥ 0. We show that
[s
1
, t
1
] ⊆ [s
2
, t
2
] =⇒ δ(s
1
, t
1
) ≤ δ(s
2
, t
2
).
Indeed, by the triangle inequality,
−Φ
c
(γ(s
2
), γ(s
1
)) −Φ
c
(γ(s
1
), γ(t
1
)) −Φ
c
(γ(t
1
), γ(t
2
)) ≤ −Φ
c
(γ(s
2
, t
2
).
Adding the equality
A
L+c
(γ[
[s
2
,s
1
]
) +A
L+c
(γ[
[s
1
,t
1
]
) +A
L+c
(γ[
[t
1
,t
2
]
) = A
L+c
(γ[
[s
2
,t
2
]
)
and using that δ(s, t) ≥ 0, we get that
δ(s
1
, t
1
) ≤ δ(s
2
, s
1
) +δ(s
1
, t
1
) +δ(t
1
, t
2
) ≤ δ(s
2
, t
2
).
Now suppose that γ : [0, +∞[→M is a ray. Observe that
δ(s, t) = Φ
c
(γ(s), γ(t); t −s) −Φ
c
(γ(s), γ(t)). (4.25)
By theorem 411.3, the map t →δ(0, t) is bounded. Deﬁne
δ(s, +∞) := lim
t→+∞
δ(s, t) = sup
t≥s
δ(s, t) ≤ δ(0, t).
143
144 4. the hamiltonian viewpoint.
Since the map s →δ(s, +∞) is increasing, we have that
lim
s→+∞
δ(s, +∞) = 0.
Now we prove that the ωlimit vectors of a ray are static. Write
γ(t) = πϕ
t
(v), t ≥ 0. Suppose that t
n
→ +∞ and ˙ γ(t
n
) → w ∈ TM.
Let η(t) = π ϕ
t
(w). Since γ and η are solutions the EulerLagrange
equation, then γ[
[t
n
−s,t
n
+s]
−→
C
1
η[
[−s,s]
. We have that
A
L+c
η[
[−s,s]
+ Φ
c
(η(s), η(−s)) =
= lim
n
¸
A
L+c
(γ[
[t
n
−s,t
n
+s]
) + lim
m
Φ
c
(γ(t
n
+s), γ(t
m
−s))
¸
= lim
n
¸
A
L+c
(γ[
[t
n
−s,t
n
+s]
) + lim
m
A
L+c
(γ[
[t
n
+s,t
m
−s]
) −δ(t
n
+s, t
m
−s)
¸
= lim
n
lim
m
A
L+c
(γ[
[t
n
−s,t
m
−s]
) −δ(t
n
+s, t
m
−s)
= lim
n
lim
m
Φ
c
(γ(t
n
−s), γ(t
m
−s)) +δ(t
n
−s, t
m
+s) −δ(t
n
+s, t
m
−s)
= Φ
c
(η(−s), η(−s)) + lim
n
δ(t
n
−s, +∞) −δ(t
n
+s, +∞)
= 0.
Thus w ∈
´
Σ(L).
Finally, we prove that γ is semistatic. Let t
n
→ +∞ be such that
the limit p = lim
n
γ(t
n
) exists. Then p ∈ { = π(
´
Σ). Using (4.25), we
have that corollary 411.8, implies that
δ(s, +∞) = lim
n
δ(s, t
n
) = h
c
(γ(s), p) −Φ
c
(γ(s), p) = 0
for all s ≥ 0. Thus δ(s, t) ≡ 0 for all 0 ≤ s ≤ t, i.e. γ is semistatic.
144
412. the extended static classes. 145
412 The extended static classes.
The method in proposition 49.7 resembles the construction of Buse
mann functions in complete manifolds of nonpositive curvature. In that
case, Ballmann, Gromov and Schroeder [5] proved that the manifold can
be compactiﬁed adjoining the sphere at inﬁnity that can be deﬁned in
terms of Busemann functions.
Here we emulate that construction to obtain a compactiﬁcation of
the manifold that identiﬁes the points in the Peierls set which are in the
same static class and adjoins what we call the extended Peierls set P
∓
.
By deﬁnition of Busemann function, the extended static classes in B
∓
correspond to the αlimits (resp. ωlimits) of semistatic orbits in the
compactiﬁcation. But as we shall see in example 412.4 the classes in
P
∓
` B
∓
do not correspond to α or ω limits of orbits in TM.
On C
0
(M, R) we use the topology of uniform convergence on compact
subsets. Consider the equivalence relation on C
0
(M, R) deﬁned by f ∼ g
if f −g is constant. Let T := C
0
(M, R)/
∼
with the quotient topology.
Let M
−
be the closure in T of ¦ f(x) = Φ
c
(z, x) [ z ∈ M¦/
∼
and
M
+
the closure in T of ¦ g(x) = Φ
c
(x, z) [ z ∈ M¦. Fix a point 0 ∈ M.
We can identify
T ≈ ¦ f ∈ C
0
(M, R) [ f(0) = 0 ¦.
412.1 Lemma. M
−
and M
+
are compact.
Proof: Observe that the functions in M
−
and M
+
are dominated. By
lemma 42.1.1 the families M
−
and M
+
are equicontinuous. Since M
is separable by Arzel´aAscoli theorem M
−
and M
+
are compact in the
topology of uniform convergence on compact subsets.
Then M
−
is the closure of the classes of the functions
f
z
(x) := Φ
c
(z, x) −Φ
c
(0, x), ∀ z ∈ M
145
146 4. the hamiltonian viewpoint.
and M
+
is the closure of the classes of
g
z
(x) := Φ
c
(x, z) −Φ
c
(0, z), ∀ z ∈ M.
412.2 Lemma.
1. If f
w
(x) = f
z
(x) for all x ∈ M, then d
c
(w, z) = 0.
2. If g
w
(x) = g
z
(x) for all x ∈ M, then d
c
(w, z) = 0.
Proof: We only prove item 1. Suppose that f
z
= f
w
. From f
z
(z) =
f
w
(z) we get that
Φ
c
(w, z) = Φ
c
(w, 0) −Φ
c
(z, 0),
and from f
z
(w) = f
z
(w) we get
Φ
c
(z, w) = −Φ
c
(w, 0) + Φ
c
(z, 0).
Adding these equations we get that d
c
(z, w) = 0.
Conversely, if d
c
(z, w) = 0 and x ∈ M, then
Φ
c
(w, x) ≤ d
c
(w, z) + Φ
c
(z, x) = Φ
c
(z, x) −Φ
c
(z, w) ≤ Φ
c
(w, x).
Thus Φ
c
(w, x) = Φ
c
(w, z) + Φ
c
(z, x) for all x ∈ M. This implies that
f
z
= f
w
.
Then we have embeddings M/
d
c
→ M
−
, by z → [f
z
] ∈ T and
M/
d
c
→M
+
by z →[g
z
] ∈ T, where M/
d
c
is the quotient space under
the equivalence relation x ≡ y if d
c
(x, y) = 0. Let B
−
be the functions
deﬁned in proposition 49.7.1 and B
+
those of 49.7.2. Let B
+
= B
+
/
∼
and B
−
= B
−
/
∼
.
412.3 Remark. By proposition 49.7, if p ∈ { = then u
−
(x) :=
Φ
c
(p, x) ∈ B
−
and u
+
(x) := −Φ
c
(x, p) ∈ B
+
(modulo an additive con
stant).
146
412. the extended static classes. 147
Observe that d
c
(z, w) = 0 if and only if z = w or z, w ∈ { and
they are in the same static class. Under the identiﬁcations M → M
∓
we have that B
∓
∪ (M ` {) ⊆ M
∓
respectively. But this inclusion may
be strict as the following example shows:
412.4 Example. B
−
∪ (M ` {) = M
−
.
Let M = R and L(x, v) :=
1
2
v
2
− cos(2πx), corresponding to the
universal cover of the simple pendulum lagrangian. Then c(L) = 1,
and the static orbits are the ﬁxed points (2k + 1, 0) ∈ TR, k ∈ Z.
Moreover, H(x, p) =
1
2
p
2
+ cos(2πx) and the HamiltonJacobi equation
H(x, d
x
u) = c(L) gives d
x
u = ±2
1 −cos(2πx). The function
u(x) =
x
0
2
1 −cos(2πs) ds,
with d
x
u ≡ +2
1 −cos(2πx), is in S
−
, is the limit of u
n
(x) :=
Φ
c
(−n, x) − Φ
c
(−n, 0) but it is not a Busemann function associated
to a semistatic orbit γ because if γ(−∞) = 2k + 1 ∈ Z is the αlimit of
γ, then the Busemann function b
γ
associated to γ satisﬁes
d
x
b
γ
=
+2
1 −cos(2πx) if x ≥ γ(−∞),
−2
1 −cos(2πx) if x ≤ γ(−∞).
(4.26)
Similarly a function v : R → R with d
x
v ≡ −2
1 −cos(1πx) is in S
+
but it is not a Busemann function.
Observe that in the Busemann function in (4.26), at the point y =
γ(−∞)+3 the semistatic orbit η(t) with ˙ η(0) = Γ
−
(u)∩T
y
M has αlimit
η(−∞) = γ(−∞) + 2 = γ(−∞). Moreover, the Busemann function b
η
associated to η satisﬁes
d
x
b
η
=
+2
1 −cos(2πx) if x ≥ γ(−∞) + 2,
−2
1 −cos(2πx) if x ≤ γ(−∞) + 2;
so that b
η
= b
γ
. In fact, there is no semistatic orbit passing through y
with αlimit γ(−∞). This implies that the Busemann functions can not
147
148 4. the hamiltonian viewpoint.
be parametrized just by a (semistatic) vector based on a unique point
0 ∈ M as in the riemannian case. In particular, it may not be possible
to choose a single point q
α
≡ 0 ∈ M, ∀α ∈ B
−
in the construction for
theorem 412.7.
The functions in B
−
and B
+
are special among the weak KAM
solutions. They are “directed” towards a single static class and they are
the most regular in the following sense:
412.5 Lemma.
1. If w ∈ Σ
−
and u
w
∈ B
−
is as in proposition 49.7.1, then
u
w
(x) = max¦ u(x) [ u ∈ S
−
, u(π(w)) = 0, w ∈ Γ
−
(u) ¦.
2. If w ∈ Σ
+
and u
w
∈ B
+
is as in proposition 49.7.2, then
u
w
(y) = min¦ u(y) [ u ∈ S
+
, u(π(w)) = 0, w ∈ Γ
+
(u) ¦.
By the remark 412.3, this also holds for the functions
u
−
(x) = Φ
c
(p, x) and u
+
(x) = −Φ
c
(x, p) (modulo an additive constant),
for any p ∈ {.
Proof: We prove item 1. Let x := π(w) and v ∈ S
−
with v(x) =
u
w
(x) = 0 and w ∈ Γ
−
(v). Let x
w
(t) = π(Φ
t
(w)). Since v ≺ L + c and
w ∈ Γ
−
(v), then for t < 0, we have that
v(y) ≤ v(x
w
(t)) + Φ
c
(x
w
(t), y)
= v(x) −Φ
c
(x
w
(t), x) + Φ
c
(x
w
(t), y).
Since v(x) = u
w
(x) = 0, letting t ↓ −∞, we get that v(y) ≤ u
w
(y) for
all y ∈ M. On the other hand, u
w
is in the set of such u’s, so that the
maximum is realized by u
w
.
148
412. the extended static classes. 149
Deﬁne
P
−
:= M
−
` (M −{) , P
+
:= M
+
` (M −{).
412.6 Proposition.
The functions in P
−
and P
+
are weak KAM solutions.
Proof: Let u ∈ M
−
` (M ` {). Since u is dominated, we only
have to prove the condition 48.1.2. Adding a constant, we can as
sume that u(0) = 0. Then there is a sequence z
n
∈ M such that
u(x) = lim
n
Φ
c
(z
n
, x) − Φ
c
(z
n
, 0). Let x ∈ M and let γ
n
∈ (
T
n
(z
n
, x)
be a Tonelli minimizer such that T
n
< 0, γ
n
(0) = x, γ
n
(T
n
) = z
n
and
A
L+c
(γ
n
) ≤ Φ
c
(z
n
, x) +
1
n
. In particular
A
L+c
(γ
n
[
[t,0]
) ≤ Φ
c
(γ
n
(t), x) +
1
n
, ∀ T
n
≤ t ≤ 0.
Since u ∈ P
−
, then we can assume that either d
M
(z
n
, x) → ∞ or z
n
→
p ∈ {. Since by lemma 32.3 [ ˙ γ
n
[ < A and h
c
(p, p) = 0 for p ∈ {, in
either case we can assume that T
n
→−∞.
We can assume that ˙ γ
n
(0) →v ∈ T
x
M. Then for t ≤ 0
u(x
v
(x)) −u(x
v
(t)) = lim
n
Φ
c
(z
n
, x) −Φ
c
(z
n
x
v
(t))
= lim
n
Φ
c
(z
n
, x) −Φ
c
(z
n
, γ
n
(t)) +K d
c
(γ
n
(t), x
v
(t))
≤ lim
n
A
L+c
γ
n
[
[t,0]
+
1
n
≤ A
L+c
x
v
[
[t,x]
,
where K is a Lipschitz constant for Φ
c
.
For p ∈ B and z ∈ M let x → b
p,z
(x) be the function in the class
p ∈ B such that b
p,z
(z) = 0, i.e.
b
p,z
(x) = lim
y→p
Φ
c
(y, x) −Φ
c
(y, z).
149
150 4. the hamiltonian viewpoint.
We now give a characterization of weak KAM solutions similar to that
of corollary 49.4. For each α ∈ B
−
choose q
α
∈ M such that there is
a unique semistatic vector v ∈ Σ
−
such that π(v) = q and the αlimit
of v is in the static class α. This can be done by the graph property 4
8.3.5. Moreover, choose them such that the map B
−
÷ α → q
α
∈ M is
injective. Let P := ¦ q
α
[ α ∈ B
−
¦. We say that a function f : P →M is
strictly dominated if
f(q
a
) < f(q
β
) +b
β,q
β
(q
α
)
for all a = β in B
−
. And we say that f is dominated if f(q
a
) ≤ f(q
β
) +
b
β,q
β
(q
α
) for all a = β in B
−
.
412.7 Theorem.
The map ¦ f : P → R[ f strictly dominated ¦ → ¦u ∈
S
−
[ u[
P
strictly dominated ¦, f →u
f
, given by
u
f
(x) := inf
α∈B
−
f(q
α
) +b
α,q
α
(x),
is a bijection.
Proof: We ﬁrst prove that u
f
is bounded below. The domination con
dition implies that u
f
(q
α
) = f(q
α
) for all α ∈ B
−
. Then the same
argument as in formula (4.13), shows that u
f
≺ L+c. Fix α ∈ B
−
, then
for all x ∈ M,
u
f
(x) ≥ u
f
(q
α
) −Φ
c
(q
α
, x) = f(q
α
) −Φ
c
(q
α
, x) > −∞. (4.27)
Since u
f
> −∞ and it is an inﬁmum of weak KAM solutions, from
lemma 49.1 we get that u
f
∈ S
−
. Since u
f
(q
α
) = f(q
α
) for all α ∈ B
−
,
the map f →u
f
is injective.
We now prove the surjectivity. Suppose that u ∈ S
−
and u[
P
is
strictly dominated. Let
v(x) := min
α∈B
−
u(q
α
) +b
α,q
α
(x).
150
412. the extended static classes. 151
Observe that the domination condition implies that
v(q
α
) = u(q
a
) for all α ∈ B
−
. (4.28)
Given x ∈ M, let θ ∈ Γ
−
(u) ∩ T
x
M and let α ∈ B
−
be the αlimit of θ.
Then,
u(x) = u(x
θ
(s)) + Φ
c
(x
θ
(s), x) for all s < 0.
Since u is dominated, u(q
α
) ≤ u(x
θ
(s)) + Φ
c
(x
θ
(s), q
α
). Hence
u(x) ≥ u(q
a
) −Φ
c
(x
θ
(s), q
a
) + Φ
c
(x
θ
(s), x) for all s < 0.
Taking the limit when s →−∞, we get that
u(x) ≥ v(x) for all x ∈ M. (4.29)
Now we prove that u = v on the projection of the backward orbits of
vectors in Γ
−
(u) ending at the points q
α
, α ∈ B
−
. Let ξ ∈ Γ
−
(u)∩T
q
α
M
and let β ∈ B
−
be the αlimit of ξ. From the deﬁnition of v(x) for all
ε > 0 and s < 0 there exists γ = γ(s, ε) ∈ B
−
such that
v(x
ξ
(s)) ≥ v(q
γ
) +b
γ,q
γ
(x
ξ
(s)) −ε.
Since ξ ∈ Γ
−
(u) ∩ T
q
α
M, then for s < 0,
u(q
α
) = u(x
ξ
(s)) + Φ
c
(x
ξ
(s), q
α
) (4.30)
≥ v(x
ξ
(s)) + Φ
c
(x
ξ
(s), q
α
) by (4.29) (4.31)
≥ v(q
γ
) +b
γ,q
γ
(x
ξ
(s)) −ε + Φ
c
(x
ξ
(s), q
α
)
= lim
t→−∞
v(q
γ
) + Φ
c
(x
ξ
(t), x
ξ
(s)) + Φ
c
(x
ξ
(s), q
α
) −Φ
c
(x
ξ
(t), q
γ
) −ε
≥ v(q
γ
) + lim
t→−∞
Φ
c
(x
ξ
(t), q
α
) −Φ
c
(x
ξ
(t), q
γ
) −ε
≥ v(q
γ
) +b
γ,q
γ
(q
α
) −ε
≥ v(q
α
) −ε (4.32)
= u(q
α
) −ε. by (4.28).
151
152 4. the hamiltonian viewpoint.
Letting ε ↓ 0, from the equality between (4.31) and (4.32) we get that
v(q
α
) = v(x
ξ
(t)) + Φ
c
(x
ξ
(t), q
α
) for all t < 0. (4.33)
But then
v(q
β
) ≤ v(x
ξ
(t)) + Φ
c
(x
ξ
(t), q
β
) = v(q
α
) −Φ
c
(x
ξ
(t), q
α
) + Φ
c
(x
ξ
(t), q
β
).
Equivalently
v(q
α
) ≥ v(q
β
) + Φ
c
(x
ξ
(t), q
α
) −Φ
c
(x
ξ
(t), q
β
).
Taking the limit when t → −∞, we get that v(q
α
) ≥ v(q
β
) + b
β,q
β
(q
α
).
This contradicts the strict domination, hence β = α. Then, from the
equality between (4.30) and (4.31), we have that
u(x
ξ
(t)) = v(x
ξ
(t)) for all t < 0, and ξ ∈ Σ
−
∩ T
q
α
M, αlim(ξ) = α.
(4.34)
Now let x ∈ M and α ∈ B
−
. Let ξ ∈ Σ
−
∩T
q
a
M with αlim(ξ) = α.
Then for t < 0,
u(x) ≤ u(x
ξ
(t)) + Φ
c
(x
ξ
(t), x)
= v(x
ξ
(t)) + Φ
c
(x
ξ
(t), x) by (4.34)
= v(q
α
) −Φ
c
(x
ξ
(t), q
α
) + Φ
c
(x
ξ
(t), x), by (4.33).
Letting t →−∞, we have that
u(x) ≤ v(q
α
) +b
α,q
α
(x) = u(q
α
) +b
α,q
α
(x).
Since α ∈ B
−
is arbitrary, from the deﬁnition of v we get that u ≤ v.
412.8 Theorem.
Given u ∈ S
−
, for all α ∈ B
−
(u) choose q
α
∈ π[Λ
−
0
(α) ∩ Γ
−
0
(u)],
and let P(u) := ¦q
α
[ α ∈ B
−
(u) ¦. Then
u(x) = inf
q
α
∈P(u)
u(q
α
) +b
α,q
α
(x) for all x ∈ M.
152
412. the extended static classes. 153
Proof: Let u ∈ S
−
. For all α ∈ B
−
(u), choose q
α
∈ π(Λ
−
0
(α) ∩Γ
−
0
(u)).
Let P(u) := ¦ q
α
[α ∈ B
−
(u) ¦. We show that u[
P(u)
is dominated. Let
α, β ∈ B
−
(u) and let θ ∈ T
q
β
M ∩ Λ
−
(β) ∩ Γ
−
(u). Then for t < 0,
u(q
α
) ≤ u(x
θ
(t)) + Φ
c
(x
θ
(t), q
α
)
= u(q
β
) −Φ
c
(x
θ
(t), q
β
) + Φ
c
(x
θ
(t), q
α
).
Letting t → −∞, we get that u(q
α
) ≤ u(q
β
) + b
β,q
β
(q
α
), for all α, β ∈
B
−
(u).
Let
v(x) := inf
q
α
∈P(u)
u(q
α
) +b
α,q
α
(x). (4.35)
The same arguments as in equation (4.27) show that v > −∞ and by
lemma (49.1) v ∈ S
−
.
Given x ∈ M, let θ ∈ Γ
−
(u) ∩T
x
M and let α ∈ B
−
(u) be the αlimit
of θ. Then,
u(x) = u(x
θ
(s)) + Φ
c
(x
θ
(s), x) for all s < 0,
u(x) ≥ u(q
a
) −Φ
c
(x
θ
(s), q
a
) + Φ
c
(x
θ
(s), x) because u is dominated.
Since α ∈ B
=
(u), taking the limit when s →−∞, we get that
u(x) ≥ v(x) for all x ∈ M. (4.36)
Now let x ∈ M and q
α
∈ P(u). Let ξ ∈ Λ(α) ∩ Γ
−
(u) ∩ T
q
α
M. Then
for s < 0,
u(x) ≤ u(x
ξ
(s)) + Φ
c
(x
ξ
(s), x)
= u(q
α
) −Φ
c
(x
ξ
(s), q
α
) + Φ
c
(x
ξ
(s), x), because ξ ∈ Γ
−
(u) ∩ T
q
α
M.
Since ξ ∈ Λ
−
(α), letting s →−∞, we have that
u(x) ≤ u(q
α
) +b
α,q
α
(x).
Since q
α
∈ P(u) is arbitrary, we get that u ≤ v.
153
154 4. the hamiltonian viewpoint.
154
Chapter 5
Examples
51 Riemannian Lagrangians.
For a riemannian lagrangian L(x, v) =
1
2
v
2
x
, we have that
e
0
= c
u
= c
0
= c(L) = 0.
The static orbits are the ﬁxed points (x, 0), x ∈ M. The only weak KAM
solutions are the constant functions and so are the only subsolutions
of (HJ) for k = c(L).
Nevertheless, since the α function is superlinear, from (??) when we
add some closed forms to L we obtain c(L − ω) > c(L) = 0. Moreover,
the semistatic static orbits are not ﬁxed points because by ?? they have
energy c(L−ω) > 0. By the graph properties, the static set of L−ω, the
support of it minimizing measures and the basins of weak KAM solutions
are geodesic laminations. Since semistatic geodesics are minimizing,
they don’t have conjugate points.
52 Mechanic Lagrangians.
For a mechanic lagrangian L(x, v) =
1
2
v
x

2
−U(x), we have that
c(L) = c
0
= c
u
= e
0
= min
x∈M
E(x, 0) = max
x∈M
U(x),
155
156 5. examples
the maximum of the potential energy. The static orbits are the ﬁxed
points (x, 0), where U(x) is maximal. If U is nondegenerate at a
minimum x (i.e. d
2
x
U is nonsingular), then x is a hyperbolic sad
dle point. The constant functions are always subsolutions of (HJ) for
k = c(L) = e
0
.
53 Symmetric Lagrangians.
For a symmetric lagrangian L(x, v) = L(x, −v), and then H(x, p) =
H(x, −p). Since v →L(x, v) is convex, it attains its minimum at v = 0.
Thus, if M is compact,
L(x, v) +e
0
= L(x, v) −max
x∈M
L(x, 0) ≥ 0,
and it is 0 exactly at the ﬁxed points (x, 0) which maximize L(x, 0). This
implies that c(L) = e
0
. In fact, the constant functions are subsolutions
of the HamiltonJacobi equation (HJ) for k = e
0
. The static orbits are
the ﬁxed points (x, 0) where x maximizes L(x, 0).
If M is noncompact, the constant functions are still subsolutions
of (HJ) for k = e
0
, and thus
c(L) = c
0
= c
u
= e
0
,
but there may be no maximizers of L(x, 0) and then the static set
´
Σ(L) =
(see example 57).
54 Simple Pendulum.
The simple pendulum is the mechanic lagrangian L : TS
1
→R given by
L(θ,
˙
θ) =
1
2
[
˙
θ[
2
−cosθ,
identifying the circle S
1
= [0, 2π] mod 2π. Hence c(L) = e
0
= 1. The
static set is
´
Σ(L) = ¦(π, 0)¦. Since there is a unique static class, there
156
55. the flat torus ≈
n
. 157
are only one weak KAM solutions on S
+
and S
−
modulo an additive
constant. But S
−
= S
+
because the solutions are not diﬀerentiable.
Their cut locus is x = 0 ∈ S
1
. For u ∈ S
+
its is the whole Σ
(
L) and it
is shown in ﬁgure 54.
fig. 1: simple pendulum.
If ω = 0 is a closed 1form, c(L−ω) > e
0
= 1 and the static set
´
Σ(L−
ω) is a whole component of the energy level E = c(L), oriented by ω > 0.
Again, there is only one static class and #S
±
= 1 modulo constants.
Since π(
´
Σ(L −ω)) = S
1
, the functions on S
±
are diﬀerentiable.
55 The ﬂat Torus T
n
.
Also called the harmonic oscillator, it is the riemannian lagrangian for
T
n
= R
n
/Z
n
with the riemannian metric induced by the euclidean metric
on R
n
. The solutions of (EL) are the projection of straight lines in R
n
parametrized with constant speed. If ω(x, v) = 'A, v` with A ∈ R
n
ﬁxed, then L − ω is globally minimized at all the vectors (x, A). Since
the orbits of this vector ﬁeld are recurrent, they must be static. These
vectors are sent to 0 by the Legendre transform. By the smoothness
property 48.3.6, du = 0 and thus any weak KAM solution must be
constant. By corollary 49.5, there is only one static class.
157
158 5. examples
56 Flat domain for the βfunction.
This example is a special case of those studied by Carneiro and Lopes [8].
Let L : TT
2
→R be the lagrangian
L
(x, y), (u, v)
=
1
2
(u
2
+v
2
) +v sin(x),
where T
2
= [0, 2π]
mod 2π
. Then the lagrangian is minimized at the
vectors x = −
π
2
, v = 1 and x =
π
2
, v = −1. These vectors are tangent
to the closed curves γ
+
: x ≡ −
π
2
, ˙ y ≡ 1 and γ
−
: x ≡
π
2
, ˙ y ≡
−1. Since the curves γ
+
and γ
−
are closed, they must be static curves.
Since their tangents are exactly all the vectors which minimize L, the
Peierls set is just { =
¦−
π
2
¦ [0, 2π]
∪
¦
π
2
¦ [0, 2π]
. The invariant
measures µ
+
:= µ
γ
+
, µ
−
:= µ
γ
−
supported on the closed orbits ˙ γ
+
, ˙ γ
−
are minimizing measures with homology ρ
+
= (0,
1
2π
), ρ
−
= (0, −
1
2π
).
Since Mather’s βfunction is convex and attains its minimum at ρ
+
and ρ
−
, then the line
[ρ
−
, ρ+] =
¸
(0, t) ∈ H
1
(T
,
R) ≈ R
2
[ −
1
2π
≤ t ≤
1
2π
¸
must be a ﬂat domain for the βfunction, in fact β[
[ρ
−
,ρ
+
]
≡ c(L). Since
the static set
´
Σ(L) contains only the support of the measures µ
+
, µ
−
,
then for any homology h ∈ [ρ
−
, ρ
+
, h = tρ
−
+ (1 − t)ρ
+
, the unique
minimizing measure in homology h is ´(h) = ¦tµ
−
+ (1 −t)µ
+
¦.
57 A Lagrangian with Peierls barrier h = +∞.
Let L : TR
2
→R be L(x, v) =
1
2
[v[
2
+ ψ(x), where [ [ is the euclidean
metric on R
2
and ψ(x) is a smooth function with ψ(x) =
1
x
for [x[ ≥ 2,
ψ ≥ 0 and ψ(x) = 2 for 0 ≤ [x[ ≤ 1.
Then
c(L) = −inf ψ = 0,
158
57. a lagrangian with peierls barrier h = +∞. 159
because if γ
n
is a smooth closed curve with length (γ
n
) = 1, [γ
n
(t)[ ≥ n
and energy E(γ
n
) =
1
2
˙ γ
2
n
−ψ(γ
n
) ≡ 0, then
c(L) ≥ − inf
n>0
A
L
(γ
n
) = −
T
n
0
1
2
˙ γ
2
n
+ψ(γ
n
)
= −
1
 ˙ γ
n

0
[ ˙ γ
n
[
2
= −[ ˙ γ
n
[ ≤ −
2
n
−→0.
On the other hand,
c(L) = −inf ¦A
L
(γ) [ γ closed ¦ ≤ 0,
because L ≥ 0.
Observe that since L > 0 and on compact subsets of R
2
, L > a > 0,
then we have that
d
c
(x, y) = Φ
c
(x, y) > 0 for all x, y ∈ R
2
.
Hence
ˆ
Σ(L) = .
Suppose that h(0, 0) < +∞. Then u(x) := h(x, 0) is in S
+
. Let
ξ ∈ Γ
+
(u)∩T
0
R
2
and write x
ξ
(t) = (r(t), θ(t)) in polar coordinates about
the origin 0 ∈ R
2
. Then liminf
t→+∞
r(t) = +∞ because otherwise the
orbit of ξ would lie on a compact subset E ≡ 0 and then = ωlim(ξ) ⊆
Σ(L) = , then
[ ˙ x
ξ
(t)[ =
2
r(t)
and
L(ϕ
t
ξ) = [ ˙ x
ξ
(t)[
2
=
2
r(t)
[ ˙ x
ξ
(t)[.
Let T
n
→+∞ be such that r(T
n
) →+∞. Since L +c = L ≥ 0, then
h(0, 0) ≥
+∞
0
L(ϕ
t
(ξ)) +c(L) =
+∞
0
2
r(t)
[ ˙ x[ +r [
˙
θ[
dt
limsup
T
n
T
n
0
2
r
˙ x dt = limsup
n
r(T
n
)
0
2
r
dr = +∞.
159
160 5. examples
58 Horocycle ﬂow.
Peierls barrier 0 < h < +∞,
´
Σ = and diﬀerentiable Busemann
functions u with B
−
(u) = B
+
(u) = ¦α¦.
Let H := R]0, +∞[ with the Poincar´e metric ds
2
=
1
y
2
(dx
2
+dy
2
).
Let L : TH →R be a Lagrangian of the form
L(x, v) =
1
2
v
2
x
+η
x
(v),
where η
x
is a 1form on H such that dη(v) is the area form an [[
x
is the
Poincar´e metric. The EulerLagrange equation is
Dt
dt
˙ x = Y
x
( ˙ x) = ˙ x
⊥
, (5.1)
where Y
x
: TH →TH is a bundle map such that
dη
x
(u, v) = 'Y
x
(u), v`.
The energy function is E(x, v) =
1
2
v
2
x
. On the energy levels E <
1
2
the solutions of (5.1) are closed curves, and on E =
1
2
the solutions are
the horospheres parametrized by arc length.
Choose the form η(x, y) =
dx
y
, where (x, y) ∈ H = R]0, +∞[. Then
L
(x, y), ( ˙ x, ˙ y)
=
1
2y
2
˙ x
2
+ ˙ y
2
+
˙ x
y
160
58. horocycle flow. 161
Observe that the form η is bounded in the Poincar´e metric, so that the
Lagrangian is superlinear and satisﬁes the boundedness condition.
It can be seen directly from the EulerLagrange equation that the
curves ˙ x = −y, ˙ y = 0 are solutions with
L( ˙ x = −y, ˙ y = 0) +
1
2
≡ 0. (5.2)
The images of these curves are the stable horospheres associated to the
geodesic x = 0, ˙ y = y, parametrized by arc length.
We show that c(L) =
1
2
and h
c
< +∞. Observe that if v = ( ˙ x, ˙ y),
˙ x < 0, then
L =
1
2
v
2
−( ˙ x, 0) ≥
1
2
v
2
−v ≥ −
1
2
. (5.3)
Hence L +
1
2
≥ 0 and then c(L) ≤
1
2
.
Now ﬁx x ∈ H. For r > 0 let D
r
be a geodesic disc of radius r
such that x ∈ ∂D
r
. Let γ
r
be the curve whose image is the boundary
of D
r
oriented clockwise and with hyperbolic speed ˙ γ ≡ a. Since
E(γ) =
1
2
a
2
, then
γ
r
L +
1
2
a
2
=
γ
r
v L
v
=
γ
r
v
2
+
D
r
dA
= a length(γ
r
) −area(D
r
),
= a 2π sinh(r) −2π cosh(r),
= 2π
1
2
(a −1) e
r
−e
−r
+ 2π. (5.4)
If a < 1, for r > 0 large, formula (5.4) is negative. Hence c(L) ≥
1
2
and c(L) =
1
2
. Moreover,
h(x, x) ≤ liminf
r→+∞
A
L+
1
2
(γ
r
) = 2π < +∞.
We prove that
´
Σ = . This implies that h > 0. First observe that
if T is an isometry of H, then d(T
∗
η) is also the area form, so that T
∗
η
161
162 5. examples
is cohomologous to η. This implies that given any two points x, y ∈ D,
there is a constant b = b(x, y) ∈ R such that for all γ ∈ ((x, y),
A
L
(γ) = A
L
(T ◦ γ) +b(x, y).
In particular, the map dT leaves σ(L) and
´
Σ(L) invariant. Since a
horocycle h
1
can be sent by an isometry to another horocycle h
2
with
h
1
∩ h
2
= , then the horocycles can not be static because it would
contradict the graph property.
The constant function u : H → ¦0¦ satisﬁes u ≺ L +
1
2
because
L +
1
2
≥ 0 and by (5.2) the vectors v = (−y, 0) ∈ Γ
+
(u) = Γ
−
(u) ∈ Σ
−
are semistatic. Its derivative du = 0 is sent by the inverse of the Legendre
transform v →L
v
= 'v, `
x
+
dx
y
to
1
y
2
'v, `
eucl
= −
dx
y
,
that is v = (−y, 0). Also
H(du) =
1
2
du −
dx
y
2
=
1
2
dx
y
2
=
1
2
.
Let T : H ← be an isometry of the hyperbolic metric. Write η =
dx
y
.
Then dη = A is the hyperbolic area 2form. Since T is an isometry, then
d(T
∗
η) = T
∗
(dη) = dη.
Hence the form T
∗
η − η is exact on H and there is a smooth function
v : H →R such that
T
∗
η −η = −dv.
We show that v is a weak KAM solution. Observe that
L ◦ dT(x, v) =
1
2
v
2
x
+T
∗
η(x, v)
=
1
2
v
2
x
+η(x, v) −dv(x, v)
= L(x, v) −dv(x, v). (5.5)
162
58. horocycle flow. 163
Since by (5.3) L ◦ dT +
1
2
≥ 0, then
dv ≤ L +
1
2
. (5.6)
Hence v ≺ L +
1
2
. Moreover, the equality in (5.6) holds exactly when
L ◦ dT(x, v) +
1
2
= 0, i.e. when dT(v) = (−y, 0) ∈ T
(x,y)
H.
Since the isometries send horospheres to horospheres, they are self
conjugacies of the hamiltonian ﬂow and hence the curves γ(t) = T
−1
(x−
ty, y) realize v, i.e.
v(γ(t)) −v(γ(s)) =
γ
dv =
γ
L +
1
2
.
Here v is the Busemann weak KAM solution associated to the class
T(∞) ∈ ∂H, on the sphere at inﬁnity of H.
We now show a picture of a nonBusemann weak KAM solution. We
use the isometry T : H ←, T(z) = −
1
z
, z = x + iy ∈ C. The isometry
T = T
−1
sends the line t → −ty + iy to a horosphere with endpoint
0 ∈ C, oriented clockwise. Choose v : H → R such that dv = η − T
∗
η.
Since T leaves the line Re z = 0 invariant and η = 0 on vertical vectors,
hence v is constant on Re z = 0.
Now we describe the weak KAM solution
w(z) := min¦ u(z), v(z) ¦ ∈ S
−
.
Let γ(t) = −ty +iy. Then, using (5.5),
v(T
−1
γ(t)) = v(T
−1
γ(0)) +
T
−1
◦γ
dv
= 0 +
t
0
L ◦ dT ◦ ˙ γ +
1
2
−
t
0
L ◦ ˙ γ +
1
2
= 0 +
t
0
L ◦ dT ◦ ˙ γ +
1
2
−0.
163
164 5. examples
Since by (5.6) L(x, v) +
1
2
> 0 when v = −y + i0, then v(z) > 0 on
Re z > 0 and v(z) < 0 on Re z < 0. Thus
w(z) =
0 = u(z) if Re z > 0.
v(z) if Re z < 0.
(5.7)
The cut locus of w is Re z = 0 and the basin of w is Γ
−
(w) = A∪dT(A)
where A is the set of vectors (y, 0) ∈ T
x+iy
H.
164
Chapter 6
Generic Lagrangians.
In [38], Ma˜ n´e introduced the concept of generic property of a lagrangian
L. A property P is said to be generic for the lagrangian L if there exists
a residual set O on C
∞
(M, R) such that if ψ is in O then L+ψ has the
property P.
A set is called residual if it contains a countable intersection of open
and dense subsets. We recall which topology is used in C
∞
(M, R). Given
u ∈ C
∞
(M, R), denote by [u[
k
its C
k
norm. Deﬁne
u
∞
:=
¸
k∈N
arctan(u
k
)
2
k
.
Note that  
∞
is not a norm. Endow C
∞
(M, R) with the translation
invariant metric u −v
∞
. This metric is complete, hence the Baire
property holds: any residual subsets of C
∞
(M, R) is dense.
One of Ma˜ n´e’s objectives was to show that Mather’s the
ory of minimizing measures becomes much more accurate and
stronger if one restricts it to generic lagrangians. In this chap
ter we shall prove
60.1 Theorem.
For every lagrangian L there exist a residual set O ⊂ C
∞
(M, R) such
that if ψ ∈ O
165
166 6. generic lagrangians.
1. The lagrangian L + ψ has a unique minimizing measure µ, and
this measure is uniquely ergodic.
2. supp(µ) =
ˆ
Σ(L +ψ) = Σ(L +ψ).
3. When µ is supported on a periodic orbit or a ﬁxed point, this orbit
(point) Γ is hyperbolic and if its stable and unstable manifolds
intersect then they do it transversally.
Ma˜ n´e conjectures in [35] that there exists a generic set O such that
this unique minimizing measure is supported on a periodic orbit or an
equilibrium point.
Item 1 of theorem 60.1 was proved by Ma˜ n´e in [34]. Item 2 is proved
in [16] and item 3 in [13]. Th proof of item 1 presented here is extracted
from a more general result in [6] about families of lagrangian systems.
We will prove only the ﬁrst part and give only the ideas of the proofs
of the last part.
61 Generic Families of Lagrangians.
In this chapter the manifold M is compact. Denote by M(L) the set of
minimizing measures of the lagrangian L:
M(L) := ¦ µ ∈ ´(L) : A
L
(µ) +c(L) = 0 ¦.
Observe that M(L) is a simplex whose extremal points are the ergodic
minimizing measures (see exercises 61.2). In general M(L) may be
inﬁnite dimensional.
In this section we shall prove the following
166
61. generic families of lagrangians. 167
61.1 Theorem. Let A be a ﬁnite dimensional convex family of la
grangians. Then there exists a residual subset O of C
∞
(M, R) such that
u ∈ O, L ∈ A =⇒ dimM(L −u) ≤ dimA.
In other words, there exist at most 1 + dimA ergodic minimizing mea
sures for L −u.
61.2 Exercises:
1. Let V be a real vector space. A set Y ⊂ V is aﬃnely independent if for
any ﬁnite subset F = ¦v
0
, v
1
, , v
n
¦ ⊂ Y , every point in the convex hull
of F is uniquely expressible as a convex combination λ
0
v
0
+ +λ
n
v
n
,
0 ≤ λ
i
≤ 1 of the elements of F. Show that a ﬁnite set F is aﬃnely
independent if and only if ¦v
1
−v
0
, , v
n
−v
0
¦ is linearly independent,
and in this case the dimension of the convex hull dim(conv(F)) = n.
2. Let K ⊂ V be a closed convex subset. An extreme or extremal point of
K is a point in K which is not in the (relative) interior of any segment
contained in K. Prove that K is the convex closure of its extremal points.
A simplex is a subset Σ ⊂ V such that the set of its extremal points is
aﬃnely independent.
3. Show that if µ and ν are ergodic invariant Borel probabilities of a dynam
ical system, then they are either equal or mutually singular (i.e. there is
a Borel set A such that µ(A) = 1 and ν(A) = 0).
4. Prove any set of ergodic invariant Borel probabilities is aﬃnely indepen
dent.
5. Prove that the set of minimizing measures M(L) of a lagrangian L is
a simplex in the vector space of signed Borel measures on TM whose
extremal points are the ergodic minimizing measures.
Fix B > 0 and let H
B
be the set of holonomic measures supported
on [[v[ ≤ B]:
H
B
:= ¦ µ ∈ ((M) : µ([ [v[ > B]) = 0 ¦.
167
168 6. generic lagrangians.
Consider a lagrangian L as a functional L : H
B
→ R given by L(µ) =
A
L
(µ). Let
M
H
B
(L) := arg min
H
B
L
be the set of measures µ ∈ H
B
which minimize L[
H
B
. Let { = {(M) be
the set of probability measures on M endowed with the weak* topology.
A compatible metric (cf. on {(M) is deﬁned as follows: ﬁx a countable
dense set ¦ f
n
[ n ∈ N¦ in C
∞
(M, R) and let
d(µ, ν) =
¸
n
1
2
n
1
c
n
f
n
dµ −
f
n
dν
.
Since M is compact, under this topology the set {(M) is compact Let
π
∗
: H
B
→{ be the pushforward induced by the projection π : TM →
M. Let
M
P
(L, B) := π
∗
M
H
B
(L)
.
61.3 Proposition. Let A be a ﬁnite dimensional convex family of la
grangians on M and let B > 0. There is a residual subset O(A, B) ⊂
C
∞
(M, R) such that
L ∈ A, u ∈ O(A, B) =⇒ dimM
P
(L, B) ≤ dimA. (6.1)
Proof of theorem 61.1:
The set O(A) := ∩
B∈N
O(A, B) is residual in C
∞
(M, R). By corol
lary 36.3, E(M(L)) = ¦c(L)¦. Then inequality 1.6 implies that there is
B
0
(L) > 0 such that M(L) ⊂ H
B
for all B > B
0
(L). From theorem 2
4.1.2
M(L) = arg min
H
B
L = M
H
B
(L) for all B > B
0
(L).
By the graph property 38.1 and 36.1
dimM(L) = dimπ
∗
M(L)
= dimM
P
(L, B) for all B > B
0
(L).
These remarks, together with proposition 61.3 prove theorem 61.1.
168
61. generic families of lagrangians. 169
Proof of proposition 61.3:
Deﬁne the εneighbourhood W
ε
of a subset W ⊂ {(M) as the union
of all the open balls in {(M) which have radius ε and are centered in W.
Given D ⊂ A, k ∈ N, ε > 0, denote by O(D, ε, k) the set of potentials
u ∈ C
∞
(M, R) such that for all L ∈ D, the convex set M
P
(L, B) is
contained in the εneighbourhood of some kdimensional convex subset
of {(M).
We shall prove that the proposition holds with
O(A, B) =
¸
ε>0
O(A, ε, dimA).
Indeed, if u ∈ O(A, B) then the inequality in (6.1) holds. Because
otherwise for some L ∈ A the convex set M
P
(L, B) would contain a
ball of dimension dimA+1, and, if ε is small enough, such a ball is not
contained in the εneighbourhood of any convex set of dimension dimA.
So we have to prove that O(A, B) is residual. It is enough to prove
that O(A, ε, dimA) is open and dense for any compact subset D ⊂ A
and any ε > 0. In 61.a we prove that it is open and in 61.b that it is
dense.
61.a Open.
We prove that for all k ∈ N, ε > 0 and any compact D ⊂ A, the set
O(D, ε, k) is open. We need a lemma:
61.4 Lemma.
The set valued map (L, u) →M
H
B
(L −u) is upper semicontinuous
on A C
∞
(M, R). This means that for any open subset U of H
B
, the
set
¦(L, u) ∈ AC
∞
(M, R) : M
H
B
(L −u) ⊂ U ¦ ⊂ AC
∞
(M, R)
169
170 6. generic lagrangians.
is open in A C
∞
(M, R). Consequently, the setvalued map (L, u) →
M
P
(L −u, B) is also upper semicontinuous.
Proof: The lemma is a consequence of the continuity of the map
AC
∞
(M, R) H
B
÷ (L, u, µ) −→(L −u)(µ) = L(µ) −
u dµ.
Now let u
0
∈ O(D, ε, k). For each L ∈ D there is a kdimensional
convex subset V ⊂ {(M) such that M
P
(L − u
0
, B) ⊂ V
ε
. Then the
open sets in D C
∞
(M, R) of the form
¦ (L, u) ∈ D C
∞
(M, R) [ M
P
(L −u, B) ⊂ V
ε
¦,
where V is a kdimensional convex subset of {(M), cover the compact
subset D¦u
0
¦. This implies that there is a ﬁnite subcover of D¦u
0
¦
by open sets of the form Ω
i
U
i
, where Ω
i
is an open set in A and
U
i
⊂ O(Ω
i
, ε, k) is an open set in C
∞
(M, R) containing u
0
. We conclude
that the open set ∩U
i
is contained in O(D, ε, k) and contains u
0
.
61.b Dense.
We prove the density of O(D, ε, dimA) in C
∞
(M, R) for any ε > 0.
Let w ∈ C
∞
(M, R). We want to prove that w is in the closure of
O(D, ε, dimA). We consider a function w ∈ C
∞
(M, R) as a linear
functional w : {(M) → R as w(ν) =
wdν or w : H
B
→ R as
w(µ) =
(w ◦ π) dµ.
61.5 Lemma. There is an integer m and a continuous map
T
m
= (w
1
, . . . , w
m
) : {(M) →R
m
,
with w
i
∈ C
∞
(M, R) such that
∀x ∈ R
m
diamT
−1
m
¦x¦ < ε. (6.2)
170
61. generic families of lagrangians. 171
Proof: To each function w ∈ C
∞
(M, R) we associate the open set
U
w
=
¸
(ν, µ) ∈ {(M) {(M)
wdν =
wdµ
¸
.
The open sets U
w
cover the complement of the diagonal in {(M){(M).
One can extract a countable subcover U
w
k
, k ∈ N. This amounts to say
that the sequence w
k
separates {(M). Deﬁninig T
m
= (w
1
, . . . , w
m
), we
have to prove that (6.2) holds for m large enough. Otherwise, we would
have some ε > 0 and two sequences ν
m
and µ
m
in {(M) such that
T
m
(ν
m
) = T
m
(µ
m
) and d(ν
m
, µ
m
) ≥ ε.
By extracting a subsequence, we can assume that the sequences ν
m
and
µ
m
have diﬀerent limits ν and µ, which satisfy d(ν, µ) ≥ ε. Take m large
enough so that T
m
(ν) = T(µ). Such m exists because the sequence w
k
separates {(M). We have that
T
m
(ν
k
) = T
m
(µ
k
) for k ≥ m.
Hence at the limit T
m
(ν) = T
m
(µ). This is a contradiction.
Deﬁne the convex function F
m
: AR
m
→R ∪ ¦+∞¦ as
F
m
(L, x) = min
µ∈H
B
T
m
◦π(µ)=x
(L −w)(µ),
when x ∈ T
m
(H
B
) and F
m
(L, x) = +∞ if x / ∈ T
m
(H
B
). For y =
(y
1
, . . . , y
m
) ∈ R
m
let
M
m
(L, y) := arg min
x∈R
m
[F(L, x) −y x] ⊂ R
m
be the set of points wich minimize the function x →F(L, x) −y x. We
have that
M
P
(L −w −y
1
w
1
− −y
m
w
m
, B) ⊂ T
−1
m
(M
m
(L, y)).
171
172 6. generic lagrangians.
Let
O
m
(w) := ¦ y ∈ R
m
[ ∀L ∈ A : dimM
m
(L, y) ≤ dimA¦.
From lemma 61.5 it follows that
y ∈ O
m
(w) =⇒ w +y
1
w
1
+ +y
m
w
m
∈ O(A, ε, dimA).
Therefore, in order to prove that w is in the closure of O(A, ε, dimA) it
is enough to prove that 0 is in the closure of O
m
(w), which follows from
the next proposition.
61.6 Proposition. The set O
m
(w) is dense in R
m
.
Proof: Consider the Legendre transform of F
m
with respect to the sec
ond variable
G
m
(L, y) = max
x∈R
m
y x −F
m
(L, x) (6.3)
= max
µ∈H
B
w +y
1
w
1
+ +y
m
w
m
−L
dµ (6.4)
It follows from the second expession that the function G
m
is convex and
ﬁnitevalued, hence continuous on AR
m
.
Let ∂G
m
be the subdiﬀerential of G
m
and let
¯
Σ = ¦ (L, y) ∈ AR
m
[ dim∂G
m
(L, y) ≥ dimA+ 1 ¦ .
By proposition E.1 the Hausdorﬀ dimension
HD(
¯
Σ) ≤ (m+ dimA) −(dimA+ 1) = m−1.
Consequently, the projection Σ of the set
˜
Σ on the second factor R
m
also
has Hausdorﬀ dimension at most m−1. Therefore, the complement of
Σ is dense in R
m
. So it is enough to prove that
y / ∈ Σ =⇒ ∀L ∈ A : dimM
m
(L, y) ≤ dimA.
172
61. generic families of lagrangians. 173
Since we know by deﬁnition of Σ that if y / ∈ Σ, dim∂G
m
(L, y) ≤ dimA,
it is enough to observe that
dimM
m
(L, y) ≤ dim∂G
m
(L, y).
The last inequality follows from the fact that the set M
m
(L, y) is con
tained in the subdiﬀerential of the convex function
R
m
÷ z →G
m
(L, z)
at the point y.
173
174 6. generic lagrangians.
174
Chapter 7
Generic Lagrangians.
In [34], Ma˜ n´e introduced the concept of generic property of a lagrangian
L. A property P is said to be generic for the lagrangian L if there exists
a generic set O (in the Baire sense) on the set C
∞
(M, R) such that if ψ
is in O then L+ψ has the property P. One of Ma˜ n´e’s objectives was to
show that Mather’s theory of minimizing measures becomes much more
accurate and stronger if we restrict ourselves to generic lagrangians. The
main purpose of this chapter is to proof the following
70.1 Theorem. For every lagrangian L there exists a generic set O ⊆
C
∞
(M, R) such that
(A) If ψ is in O then L+ψ has a unique minimizing measure, µ and
this measure is uniquely ergodic.
(B)Moreover supp(µ) =
ˆ
Σ(L +ψ) = Σ(L +ψ).
(C) When µ is supported on a periodic orbit or a ﬁxed point, this orbit
(point) Γ is hyperbolic and its stable and unstable manifolds if intersect
they do it transversally.
On Ma˜ n´e [35], it is conjectured that there exists a generic set O such
that this unique minimizing measure is supported on a periodic orbit or
an equilibrium point.
175
176 7. generic lagrangians.
The ﬁrst statement of this was proved by Ma˜ n´e in [34]. The second
statement is proved in [16] and the third one in [13]. We will prove only
the ﬁrst part and give only the ideas of the proofs of the last part.
71 Generic Lagrangians.
Proof of (A)
Given a potential ψ on C
∞
(M, R) deﬁne
m(ψ) = min
ν∈C
L +ψdν,
M(ψ) = ¦ν ∈ C :
L +ψdν = m(ψ)¦
Where C is the set of holonomic measures. For > 0 let
O
= ¦ψ : diam M(ψ) < ¦
This set is open, in fact if ν
n
is in M(ψ
n
) then
L +ψdν
n
≤ m(ψ) + 2[[ψ −ψ
n
[[
C
0 (7.1)
So by theorem 24.1 in chapter 2 if ψ
n
→ψ then the sequence ν
n
is
precompact and the limit is in M(ψ). From this follows that O
is open.
It remains to prove that it is also dense.
Given a compact convex set K
0
on C and potential ψ on C
∞
(M, R)
deﬁne
m
0
(ψ) = min
ν∈K
0
ψdν,
M
0
(ψ) = ¦ν ∈ K
0
:
ψdν = m
0
(ψ)¦
176
71. generic lagrangians. 177
71.1 Lemma. Let K
0
as before then if µ is an extremal point of K
0
,
for all > 0there exists ψ on C
∞
(M, R) such that
diam M
0
(ψ) <
d(µ, M
0
(ψ)) <
Proof
Denote by D the diagonal of K
0
K
0
for each pair (µ, ν) in K
0
K
0
−
D take a potential ψ
(µ,ν)
such that
ψ
(µ,ν)
dµ =
ψ
(µ,ν)
dν, then there
is a neigbourhood U(µ, ν) contained on K
0
K
0
such that
ψ
(µ,ν)
dµ
=
ψ
(µ,ν)
dν
for every (µ
, ν
) in U(µ, ν).
Take a covering ¦U(µ
n
, ν
n
)¦ of K
0
K
0
− D and set ψ
n
= ψ
(µ
n
,ν
n
)
then if (µ, ν) in K
0
K
0
−D there exist n such that
ψ
n
dµ =
ψ
n
dν (7.2)
Deﬁne T
n
: C →R
n
as
T
n
(µ) = (
ψ
1
dµ, ...,
ψ
n
dµ)
Using (7.2) and the compactness of K
0
K
0
it is easy to see that
given there exist δ > 0 and n > 0 such that
S ⊂ R
n
, diam S < δ ⇒diam T
−1
S < (7.3)
Let B = T
n
(k
0
) then B is a compact convex set, let f : R
n
→ R a
linear function such that its minimum restricted to B is attained in only
one point p.
Deﬁne ψ =
¸
i
λ
i
ψ
i
where f =
¸
i
λ
i
p
i
, then
f ◦ T
n
=
¸
i
λ
i
ψ
i
177
178 7. generic lagrangians.
and so
M
0
(ψ) = T
−1
n
(p)
Then by (7.3) we get
diam M
0
(ψ) <
The following lemma proves the density of O
and hence the ﬁrst
part of (A).
71.2 Lemma. If ψ is on C
∞
(M, R) and µ is an extremal point of
M(ψ) then for every neighbourhood U of ψ and every > 0 there
exists ψ
1
on U such that
diam M(ψ
1
) =
Proof
For K
0
= M(ψ) applying the previous lemma, we can ﬁnd given a
ψ
1
such that
ψ
1
dν attains its minimum, say m
1
for all measures ν on
K
0
= M(ψ) on a set S = M
0
(ψ
1
) such that d(µ, S) < . Set
m
0
= m(ψ)
f
0
(ν) =
L +ψ −m
0
dν
f
1
(ν) =
ψ
1
−m
1
dν
Then
f
1
(ν) = f
0
(ν) if ν ∈ S (7.4)
f
1
(ν) ≥ 0 if ν ∈ M(ψ) (7.5)
ν ∈ M(ψ), f
1
(ν) = 0 ⇒ν ∈ S (7.6)
178
71. generic lagrangians. 179
ν ∈ C, f
0
(ν) = 0 ⇒ν ∈ M(ψ) (7.7)
For λ > 0 deﬁne
f
λ
= f
0
+λf
1
and set
m(λ) = min
ν∈C
f
λ
(ν),
M(λ) = ¦ν ∈ C : f
λ
(ν) = m(λ)¦.
We claim that
lim
λ→0
diam (M(λ), ¦µ¦) ≤ (7.8)
This proves the lemma since
M(λ) = M(ψ +λψ
1
)
Proof of the claim
Suppose otherwise that there exist λ
n
→0 and µ
λ
n
, ν
λ
n
on M(λ
n
) =
M(ψ+λ
n
ψ
1
) such that d(µ
λ
n
, ν
λ
n
) > . Then by (7.1) ¦µ
λ
n
¦ and ¦ν
λ
n
¦
are precompact and as in the proof of the open property we may assume
that µ
λ
n
→µ ∈ M(ψ) and ν
λ
n
→ν ∈ M(ψ). Naturally d(µ, ν) ≥ .
Now Because of (7.4) we have that m(λ) ≤ 0
0 ≥ m(λ
n
) =
= f
0
(µ
λ
n
) +λf
1
(µ
λ
n
)
≥ λ
n
f
1
(µ
λ
n
)
So f
1
(µ
λ
n
) ≤ 0 and hence f
1
(µ) ≤ 0, since µ ∈ M(ψ) by (7.5)
f
1
(µ) = 0 and then by (7.6) µ is in S. Similarly ν is in S. This is a
contradiction with the fact that the diameter of S is less than .
The fact that µ is uniquely ergodic follows from the fact that ergodic
components of a minimizing measure are also minimizing. And the proof
of (A) is complete.
179
180 7. generic lagrangians.
It is worth to remark that the proof presented here is a particular
case of Ma˜ n´e’s original [38] more general setting:
Let E, F be real convex spaces, K contained on F a metrizable
convex subset and φ : E →F
, L : F →R linear maps satisfying
• (a) The map EF →R deﬁned by (w, x) →φ(w)(x) is continuous.
• (b) For any x = y in K there exists w in E such that φ(w)(x) =
φ(w)(y).
• (c) For all w in E and c in R the set
¦x ∈ K : L(x) +φ(w)(x) ≤ c¦
is compact.
Denote by
m(w) = min
x∈K
L(x) +φ(w)(x)
which exists by (c). And
M(w) = ¦x ∈ K : L(x) +φ(w)(x) = m(w)
71.3 Proposition. If E is a Frechet space then there exists a residual
set O contained on E such that if w is on E then M(w) has only one
element.
The reader can verify that with the following choices, we get the
desired result.
• (1) Let E be the Banach space C
∞
(M, R)
• (2) As in section 23 let C
0
be the set of continuous functions
f on TM such that sup
f(x,v)
1+v
≤ ∞, and C the set of holonomic
probabilities. Let K be C and F be the subspace of (C
0
)
∗
spanned
by C.
180
71. generic lagrangians. 181
• (3) Finally let L : F → R is the linear map such that if µ is in C
then L(µ) =
Ldµ; and for ψ in C
∞
(M, R) φ(ψ) is the restriction
to F of the linear map on (C
0
)
∗
such that w →< w, ψ >.
This general setting has some other applications see theorems A, C
D in [35] and also [11]
Proof of (C)
Let O be the residual given by (A). Let / be the subset of O of
potentials ψ for which the measure on ´(L + ψ) is supported on a
periodic orbit. Let B := O ` / and let /
1
be the subset of / on
which the minimizing periodic orbit is hyperbolic. We prove that /
1
is
relatively open on /. For, let ψ ∈ /
1
and
´(L +ψ) = ¦µ
γ
¦
where µ
γ
is the invariant probability measure supported on the hyper
bolic periodic orbit γ for the ﬂow of L + ψ. We claim that if φ
k
∈ /,
φ
k
→ ψ and ´(L + φ
k
) = ¦µ
η
k
¦, then η
k
→ γ. Indeed, since L is
superlinear, the velocities in the support of the minimizing measures
µ
k
:= µ
η
k
are bounded (cf. corollary 36.3 and inequality 1.6), and
hence there exists a subsequence µ
k
→ν converging weakly* to a some
invariant measure ν for L +ψ. Then if ν = µ
γ
,
lim
k
S
L+φ
k
(µ
k
) = S
L+ψ
(ν) > S
L+ψ
(µ
γ
) . (7.9)
Thus if δ
k
is the analytic continuation of the hyperbolic periodic orbit
γ to the ﬂow of L + φ
k
in the original energy level c(L + ψ), since
lim
k
S
L+ψ
k
(µ
δ
k
) = S
L+ψ
(µ
γ
), for k large we have that,
S
L+φ
k
(µ
δ
k
) < S
L+φ
k
(µ
η
k
) ,
which contradicts the choice of η
k
. Therefore ν = µ
γ
. For energy levels
h near to c(L + ψ) and potentials φ near to ψ, there exist hyperbolic
periodic orbits γ
φ,h
which are the continuation of γ. Now, on a small
181
182 7. generic lagrangians.
neighbourhood of a hyperbolic orbit there exists a unique invariant mea
sure supported on it, and it is in fact supported in the periodic orbit.
Thus, since η
k
→ γ, then η
k
is hyperbolic. Hence φ
k
∈ /
1
and /
1
contains a neighbourhood of φ in /.
Let  be an open subset of C
∞
(M, R) such that /
1
=  ∩ /. We
shall prove below that /
1
is dense in /. This implies that /
1
∪ B is
generic. For, let ψ := int (C
∞
(M, R) ` ), then  ∪ ψ is open and
dense in C
∞
(M, R). Moreover, ψ ∩ / = because / ⊆ /
1
⊆  and
ψ ∩ / ⊆ /`  = . Since O = /∪ B is generic and
( ∪ ψ) ∩ (/∪ B) = ( ∩ /) ∪
( ∪ ψ) ∩ B
⊆ /
1
∪ B,
then /
1
∪ B is generic.
The perturbation to achieve hyperbolicity in a ﬁxed point is easy. Is
very much as the mechanic case: L =
1
2
< v, v >
x
−U(x). The reader
can verify that if x
0
= max U then the Dirac measure supported on the
point (x
0
, 0) is minimizing. And it is well known that this critical point
is hyperbolic if and only if the maximum has
non degenerate quadratic form.
In fact, from the EulerLagrange equation (EL) we get that
L
x
(x
0
, 0) = 0. Diﬀerentiating the energy function (1.3) we see that
(x
0
, 0) is a singularity of the energy level c(L). Moreover, the min
imizing property of µ implies that x
0
is a minimum of the function
x →L
xx
(x, 0). In particular, L
xx
(x
0
, 0) is positive semideﬁnite in linear
coordinates in T
x
0
M. And it is hyperbolic if and only if it is positive
deﬁnite. So to achieve hyperbolicity we must just add a small quadratic
form.
The perturbation needed in the case of a periodic orbit the same
spirit; Because of the graph property the projection of the orbit Γ, π(Γ)
is a simple closed curve. We add a C
∞
small non negative potential ψ,
which is zero if and only if x is on π(Γ) that is nondegenerate in the
transversal direction. It follows that Γ is also a minimizing solution of
the perturbed lagrangian L +ψ.
182
71. generic lagrangians. 183
To prove that it actually is hyperbolic is much more diﬃcult. The
reason is that the linearization of the ﬂow (the Jacobi equation) is along
the periodic orbit and hence non autonomous as in the case of a singu
larity.
To explain the idea of the proof we need some deﬁnitions. Let H
be the associated hamiltonian by the Legendre transformation on T
∗
M
and ψ its ﬂow. Denote by π : T
∗
M → M be the canonical projection
and deﬁne the vertical subspace on θ ∈ T
∗
M by ψ(θ) = ker(dπ). Two
points θ
1
, θ
2
∈ T
∗
M are said to be conjugate if θ
2
= ψ
τ
(θ
1
) for some
τ = 0 and dψ
τ
(ψ(θ
1
)) ∩ ψ(θ
2
) = ¦0¦.
A basic property of orbits without conjugate points is given by the
following
71.4 Proposition. Suppose that the orbit of θ ∈ T
∗
M does not contain
conjugate points and H(θ) = e is a regular value of H. Then there exist
two ϕinvariant lagrangian subbundles E, F ⊂ T(T
∗
M) along the orbit
of θ given by
E(θ) = lim
t→+∞
dψ
−t
ψ(ψ
t
(θ))
,
F(θ) = lim
t→+∞
dψ
t
ψ(ψ
−t
(θ))
.
Moreover, E(θ)∪F(θ) ⊂ T
θ
Σ, E(θ)∩ψ(θ) = F(θ)∩ψ(θ) = ¦0¦, 'X(θ)` ⊆
E(θ)∩F(θ) and dimE(θ) = dimF(θ) = dimM, where X(θ) = (H
p
, −H
q
)
is the hamiltonian vector ﬁeld and Σ = H
−1
¦e¦.
These bundles where constructed for disconjugate geodesics of rie
mannian metrics by Green [27] and of Finsler metrics by Foulon [24]. In
the general case where constructed in [13]
We will only sketch the proof;
Fix a riemannian metric on M and the corresponding induced metric
on T
∗
M. Then T
θ
T
∗
M splits as a direct sum of two lagrangian sub
spaces: the vertical subspace ψ(θ) = ker (dπ(θ)) and the horizontal sub
183
184 7. generic lagrangians.
space H(θ) given by the kernel of the connection map. Using the isomor
phism K : T
θ
T
∗
M →T
π(θ)
M T
∗
π(θ)
M, ξ →(dπ(θ) ξ , ∇
θ
(π ξ)), we can
identify H(θ) ≈ T
π(θ)
M¦0¦ and ψ(θ) ≈ ¦0¦T
∗
π(θ)
M ≈ T
π(θ)
M. If we
choose local coordinates along t →π ψ
t
(θ) such that t →
∂
∂q
i
(π ψ
t
(θ)) are
parallel vector ﬁelds, then this identiﬁcation becomes ξ ↔(dq(ξ), dp(ξ)).
Let E ⊂ T
θ
T
∗
M be an ndimensional subspace such that E∩ψ(θ) = ¦0¦.
Then E is a graph of some linear map S : H(θ) → ψ(θ). It can be
checked that E is lagrangian if and only if in symplectic coordinates S
is symmetric.
Take θ ∈ T
∗
M and ξ = (h, v) ∈ T
θ
T
∗
M = H(θ) ⊕ψ(θ) ≈ T
π(θ)
M ⊕
T
π(θ)
M. Consider a variation
α
s
(t) =
q
s
(t), p
s
(t)
such that for each s ∈] −ε, ε[, α
s
is a solution of the hamiltonian H such
that α
0
(0) = θ and
d
ds
α
s
(0)[
s=0
= ξ.
Writing dψ
t
(ξ) =
h(t), v(t)
, we obtain the hamiltonian Jacobi
equations
˙
h= H
pq
h +H
pp
v ,
˙ v = −H
qq
h −H
qp
v , (7.10)
where the covariant derivatives are evaluated along π(α
o
(t)), and H
qq
,
H
qp
, H
pp
and H
qq
are linear operators on T
π(θ)
M, that in local co
ordinates coincide with the matrices of partial derivatives
∂
2
H
∂q
i
∂q
j
,
∂
2
H
∂q
i
∂p
j
,
∂
2
H
∂p
i
∂p
j
and
∂
2
H
∂p
i
∂q
j
. Moreover, since the hamiltonian H is
convex, then H
pp
is positive deﬁnite.
We derive now the Riccati equation. Let E be a lagrangian subspace
of T
θ
T
∗
M. Suppose that for t in some interval ] − ε, ε[ we have that
dψ
t
(E) ∩ ψ(ψ
t
(θ)) = ¦0¦. Then we can write dψ
t
(E) = graph S(t),
where S(t) : H(ψ
t
θ) → ψ(ψ
t
θ) is a symmetric map. That is, if ξ ∈ E
then
dψ
t
(ξ) =
h(t), S(t) h(t)
.
184
71. generic lagrangians. 185
Using equation (7.10) we have that
˙
Sh +S(H
pq
h +H
pp
Sh) = −H
qq
h −H
qp
Sh.
Since this holds for all h ∈ H(ψ
t
(θ)) we obtain the Riccati equation:
˙
S +SH
pp
S +SH
pq
+H
qp
S +H
qq
= 0 . (7.11)
Let K
c
(θ) : H(θ) → ψ(θ) be the symmetric linear map such that
graph(K
c
(θ)) = dψ
−c
(ψ(ψ
c
(θ))). Deﬁne a partial order on the sym
metric isomorphisms of T
π(θ)
M by writing A ~ B if A − B is positive
deﬁnite.
The following proposition based essentially on the convexity of H
proves 71.4.
71.5 Proposition. For all ε > 0,
(a) If d > c > 0 then K
−ε
~ K
d
~ K
c
.
(b) If d < c < 0 then K
ε
≺ K
d
≺ K
c
.
(c) lim
d→+∞
K
d
= S, lim
d→−∞
K
d
= U.
(d) S U.
(e) The graph of S is the stable green bundle E and the graph of U is
the unstable green bundle F
An example of the relationship between the transversality of the
Green subspaces and hyperbolicity appears in the following
71.6 Proposition. Let Γ be a periodic orbit of ψ
t
without conju
gate points. Then Γ is hyperbolic (on its energy level) if and only if
E(θ) ∩F(θ) = 'X(θ)` for some θ ∈ Γ, where 'X(θ)` is the 1dimensional
subspace generated by the hamiltonian vector ﬁeld X(θ). In this case E
and F are its stable and unstable subspaces.
185
186 7. generic lagrangians.
This proposition follows ideas of Eberlein [18] and Freire, [25].
It is known that minimizing orbits do not have conjugate points. So
by proposition 71.4 and 71.6 to prove the density of hyperbolicity it is
enough to perturb to make the Green bundles transverse. This is done
using two formulas for the index. One in the lagrangian setting and
another one in the hamiltonian setting.
Let Ω
T
be the set of continuous piecewise C
2
vector ﬁelds ξ along
γ
[0,T]
. Deﬁne the index form on Ω
T
by
I(ξ, η) =
T
0
˙
ξ L
vv
˙ η +
˙
ξ L
vx
η +ξ L
xv
˙ η +ξ L
xx
η
dt , (7.12)
which is the second variation of the action functional for variations f(s, t)
with
∂f
∂s
∈ Ω
T
. For general results on this form see Duistermaat [17].
From this formula it is easy to compare the index of the original and
the perturbed lagrangian along the same solution Γ.
Finally we use the following transformation of the index form. It is
taken from Hartman [29] and originally due to Clebsch [10] see also [13].
Let θ ∈ T
∗
M and suppose that the orbit of θ, ψ
t
(θ), 0 ≤ t ≤ T does not
have conjugate points. Let E ⊂ T
θ
T
∗
M be a lagrangian subspace such
that dψ
t
(E) ∩ ψ(ψ
t
(θ)) = ¦0¦ for all 0 ≤ t ≤ T.
Let E(t) := dψ
t
(E) and let H(t), ψ(t) be a matrix solution of the
hamiltonian Jacobi equation (7.10) such that det H(t) = 0 and E(t) =
Image (H(t), ψ(t)) ⊂ T
ψ
t
(θ)
(T
∗
M) is a lagrangian subspace. For ξ =
Hζ ∈ Ω
T
, η = Hρ ∈ Ω
T
, we obtain (see [13])
I(ξ, η) =
T
0
(Hζ
)
∗
(H
pp
)
−1
(Hρ
) dt + (Hζ)
∗
(V ρ)[
T
0
. (7.13)
Deﬁne N(θ) := ¦ w ∈ T
π(θ)
M[ 'w, ˙ γ` = 0 ¦. Then N(θ) is the sub
space of T
π(θ)
M generated by the vectors
∂
∂x
2
,. . .,
∂
∂x
n
. Let v
o
∈ N(θ),
[v
o
[ = 1 and let ξ
T
(t) :=
¯
Z
T
(t) v
o
. Denote by
˜
I
T
and I
T
be the index
forms on [0, T] for
¯
L and L respectively. Using the solution (
¯
Z
T
,
¯
V
T
) on
186
71. generic lagrangians. 187
formula (7.13), we obtain that
˜
I
T
(ξ
T
, ξ
T
) = −(
¯
Z
T
(0) v
o
)
∗
(
¯
V
T
(0) v
o
) = −v
o
∗
˜
K
T
(0) v
o
. (7.14)
Moreover, in the coordinates (x
1
, . . . , x
n
;
∂
∂x
1
, . . . ,
∂
∂x
n
) on TU we have
that
˜
I
T
(ξ
T
, ξ
T
) =
T
0
˙
ξ
T
¯
L
vv
˙
ξ
T
+ 2
˙
ξ
T
¯
L
xv
ξ
T
+ξ
T
¯
L
xx
ξ
T
dt
=
T
0
˙
ξ
T
L
vv
˙
ξ
T
+ 2
˙
ξ
T
L
xv
ξ
T
+ξ
T
L
xx
ξ
T
dt
+
T
0
ε
n
¸
i=2
ξ
T
i
2
dt .
(7.15)
We have that
¯
Z
T
(0) = I and for all t > 0, lim
t→∞
¯
Z
T
(t) =
¯
h(t)
with
¯
h(t) the solution of the Jacobi equation for
˜
H corresponding to the
stable Green bundle. Writing π
N
(ξ) = (ξ
2
, ξ
3
, . . . ξ
n
) then
π
N
¯
h(0) v
0
=
[v
0
[ = 1 because v
0
∈ N(θ). Hence there exists λ > 0 and T
0
> 0
such that
π
N
ξ
T
(t)
=
π
N
¯
Z
T
(t) v
0
>
1
2
for all 0 ≤ t ≤ λ and T > T
0
.
Therefore
˜
I
T
(ξ
T
, ξ
T
) ≥ I
T
(ξ
T
, ξ
T
) +
ελ
4
. (7.16)
Let (h(t), v(t)) = dψ
t
◦
dπ[
E(θ)
−1
be the solution of the Jacobi
equation for H corresponding to the stable Green subspace E and let
S(ψ
t
(θ)) = v(t) h(t)
−1
be the corresponding solution of the Riccati equa
tion, with graph[S(ψ
t
(θ)] = E(ψ
t
(θ)). Using formula (7.13), and writing
ξ
T
(t) = h(t) ζ(t), we have that
I
T
(ξ
T
, ξ
T
) =
T
0
(h
˙
ζ)
∗
H
−1
pp
(h
˙
ζ) dt + 0 −
h(0) ζ(0)
∗
v(0) ζ(0)
I
T
(ξ
T
, ξ
T
) ≥ −v
∗
o
S(θ) v
o
. (7.17)
From (7.14), (7.16) and (7.17), we get that
v
∗
o
S(θ) v
o
≥ v
∗
o
˜
K
T
v
o
+
ελ
4
.
187
188 7. generic lagrangians.
From proposition 71.5, we have that lim
T→+∞
˜
K
T
(0) =
¯
S(θ), where
graph(
¯
S(θ)) =
¯
E(θ), the stable Green bundle for
˜
H. Therefore
v
∗
o
S(θ) v
o
≥ v
∗
0
¯
S(θ) v
o
+
ελ
4
. (7.18)
Similarly, for the unstable Green bundles we obtain that
v
∗
o
U(θ) v
o
+λ
2
≤ v
∗
o
¯
U(θ) v
o
for v
o
∈ N(θ), [v
o
[ = 1. (7.19)
for some λ
2
> 0 independent of v
o
.
From proposition 71.5 we have that U(θ) _ S(θ). From (7.18) and
(7.19) we get that
¯
U[
N
~ U[
N
_ S[
N
~
¯
S[
N
. Since
¯
E(θ) = graph(
¯
S(θ))
and
¯
F(θ) = graph(
¯
U(θ)), we get that
¯
E(θ) ∩
¯
F(θ) ⊆ '
¯
X(θ)`. Then
proposition B shows that Γ is a hyperbolic periodic orbit for L +φ.
This proves that /
1
is dense in /.
Let /
2
be the subset of /
1
of potentials ψ for which the minimizing
hyperbolic periodic orbit Γ has transversal intersections. The proof that
/
2
is open and dense in /
1
is similar to the previous proof, see [13].
72 Homoclinic Orbits.
Assume in this section that
´
Σ contains only one static class. By theo
rem 70.1.(A), this is true for generic lagrangians. By proposition 311.4,
the static classes are always connected, thus if we assume that there is
only one static class,
´
Σ must be connected.
Given ε > 0, let U
ε
be the εneighbourhood of π(
´
Σ). Since
´
Σ is
connected, the open set U
ε
is connected for ε suﬃciently small. Let
H
1
(M, U
ε
, R) denote the ﬁrst relative singular homology group of the
pair (M, U
ε
) with real coeﬃcients.
We shall say that an orbit of L is homoclinic to a closed invariant
set K ⊂ TM if its α and ωlimit sets are contained in K.
Observe that to each homoclinic orbit x : R →M to the set of static
orbits
´
Σ we can associate a homology class in H
1
(M, U
ε
, R). Indeed,
188
72. homoclinic orbits. 189
since there exists t
0
> 0 such that for all t with [t[ ≥ t
0
, x(t) ∈ U
ε
, the
class of x[
[−t
0
,t
0
]
deﬁnes an element in H
1
(M, U
ε
, R). Let us denote by
H the subset of H
1
(M, U
ε
, R) given by all the classes corresponding to
homoclinic orbits to
´
Σ.
72.1 Theorem. Suppose that
´
Σ contains only one static class. Then
for any ε suﬃciently small the set H generates over R the relative homol
ogy H
1
(M, U
ε
, R). In particular, there exist at least dimH
1
(M, U
ε
, R)
homoclinic orbits to the set of static orbits
´
Σ.
Let U
ε
be an εneighbourhood of supp(µ). From theorems 72.1
and 70.1 we obtain:
72.2 Corollary. Given a lagrangian L there exists a generic set
O ⊂ C
∞
(M, R) such that if ψ ∈ O the lagrangian L + ψ has a unique
minimizing measure µ in ´
0
(L + ψ) and this measure is uniquely er
godic. For any ε suﬃciently small the set H of homoclinic orbits to
supp(µ) generates over R the relative homology H
1
(M, U
ε
, R). In partic
ular, there exist at least dimH
1
(M, U
ε
, R) homoclinic orbits to supp(µ).
To prove theorem 72.1 we consider ﬁnite coverings M
0
of M whose
group of deck transformations is given by the quotient of the torsion free
part of H
1
(M, U
ε
, Z) by a ﬁnite index subgroup. Using that the lifted
lagrangian L
0
has the same critical value as L, we conclude that the
number of static classes of L
0
must be ﬁnite. Hence we can apply theo
rem 311.1 to L
0
to deduce that the group generated by the homoclinic
orbits to the set of static orbits of L coincides with H
1
(M, U
ε
, R).
We note that the homoclinic orbits that we obtain in theorem 72.1
and corollary 72.2 have energy c but they are not semistatic orbits of
L. However, they are semistatic for lifts of L to suitable ﬁnite covers.
Combining corollary 72.2, theorem 70.1 and lemma 72.4, we obtain
72.3 Corollary. Let M be a closed manifold with ﬁrst Betti number
≥ 2. Given a lagrangian L there exists a generic set O ⊂ C
∞
(M, R)
such that if ψ ∈ O the lagrangian L+ψ has a unique minimizing measure
189
190 7. generic lagrangians.
in ´
0
(L+ψ) and this measure is uniquely ergodic. When this measure
is supported on a periodic orbit, this orbit is hyperbolic and the stable
and unstable manifolds have transverse homoclinic intersections.
72.4 Lemma. Let M be a closed manifold with ﬁrst Betti number
b
1
(M, R) ≥ 2. Then if A ⊂ M is a closed submanifold diﬀeomorphic to
S
1
and U
ε
denotes the ε neighborhood of A, we have that H
1
(M, U
ε
, R)
is non zero for all ε suﬃciently small.
Proof: Since A is diﬀeomorphic to a circle, the singular homology of
the pair (M, U
ε
) coincides with the singular homology of the pair (M, A)
and therefore the vector space H
1
(M, U
ε
, R) must have dimension ≥
b
1
(M, R) −1 ≥ 1.
For the proof of theorem 72.1 we shall need the following lemma:
72.5 Lemma. Let p : M
1
→M
2
be a covering such that c(L
1
) = c(L
2
).
Then any lift of a semistatic curve of L
2
is a semistatic curve of L
1
.
Also the projection of a static curve of L
1
is a static curve of L
2
. If in
addition, p is a ﬁnite covering, then any lift of a static curve of L
2
is a
static curve of L
1
.
Proof: Observe ﬁrst that for any k ∈ R we have that
Φ
1
k
(x, y) ≥ Φ
2
k
(px, py),
for all x and y in M
1
. Hence if we write c = c(L
1
) = c(L
2
) we have
Φ
1
c
(x, y) ≥ Φ
2
c
(px, py), (7.20)
for all x and y in M
1
.
Suppose now that x
2
: R → M
2
is a semistatic curve of L
2
and let
x
1
: R →M
1
be any lift of x
2
to M
1
. Using (7.20) and the fact that x
2
is semistatic we have for s ≤ t,
190
72. homoclinic orbits. 191
Φ
1
c
(x
1
(s), x
1
(t)) ≤ A
L
1
+c
(x
1
[
[s,t]
) = A
L
2
+c
(x
2
[
[s,t]
)
= Φ
2
c
(x
2
(s), x
2
(t)) ≤ Φ
1
c
(x
1
(s), x
1
(t)).
Hence x
1
is semistatic for L
1
.
Suppose now that x
1
: R → M
1
is a static curve of L
1
and let
x
2
: R → M
2
be p ◦ x
1
. Using (7.20) and the fact that x
1
is static we
have for s ≤ t,
−Φ
1
c
(x
1
(t), x
1
(s)) = Φ
1
c
(x
1
(s), x
1
(t)) = A
L
1
+c
(x
1
[
[s,t]
) = A
L
2
+c
(x
2
[
[s,t]
)
≥ Φ
2
c
(x
2
(s), x
2
(t)) ≥ −Φ
2
c
(x
2
(t), x
2
(s)) ≥ −Φ
1
c
(x
1
(t), x
1
(s)).
Hence x
2
is static for L
2
.
Suppose now that p is a ﬁnite covering and let x
2
: R → M
2
be a
static curve of L
2
. Let x
1
: R →M be any lift of x
2
to M
1
. Since x
2
is
static, given s ≤ t and ε > 0, there exists a curve α : [0, T] → M
2
with
α(0) = x
2
(t), α(T) = x
2
(s) such that
A
L
2
+c
(x
2
[
[s,t]
) +A
L
2
+c
(α) ≤ ε.
Since p is a ﬁnite covering, there exists a positive integer n, bounded
from above by the number of sheets of the covering, such that the nth
iterate of x
2
[
[s,t]
∗ α lifts to M
1
as a closed curve. Hence, there exists a
curve β joining x
1
(t) to x
1
(s) such that
A
L
1
+c
(x
1
[
[s,t]
) +A
L
1
+c
(β) ≤ nε,
and thus x
1
is static for L
1
.
Proof of theorem 72.1:
Let U
def
= U
ε
denote the εneighborhood of π(
´
Σ(L)), where
´
Σ(L) is
the set of static vectors of L. Since we are assuming that
´
Σ(L) contains
only one static class, the set U is also connected for small ε. Let i : U →
M be the inclusion map. The vector space H
1
(M, U, R) is isomorphic
to the quotient of H
1
(M, R) by i
∗
(H
1
(U, R)).
191
192 7. generic lagrangians.
Let H denote the torsion free part of H
1
(M, Z) and let K denote
the torsion free part of i
∗
(H
1
(U, Z)). Let us write G
def
= H/K = Z ⊕
k
. . . ⊕ Z where k = dimH
1
(M, U, R). Let J be a ﬁnite index subgroup
of G. There is a surjective homomorphism j : G → G/J given by the
projection.
If we take the Hurewicz map
π
1
(M) →H
1
(M, Z),
and we compose it with the projections H
1
(M, Z) → H, H → G and
j : G →G/J, we obtain a surjective homomorphism
π
1
(M) →G/J,
whose kernel will be the fundamental group of a ﬁnite covering M
0
of M with covering projection map p : M
0
→ M and group of deck
transformations given by the ﬁnite abelian group G/J.
Since J is a subgroup of G = H/K, G/J acts transitively and freely
on the set of connected components of p
−1
(U) which coincides with the
set of connected components of p
−1
(π(
´
Σ(L))). Therefore we have
72.6 Lemma. There is a one to one correspondence between elements
in G/J and connected components of p
−1
(π(
´
Σ(L))).
Observe that to each homoclinic orbit x : R → M to
´
Σ(L) we can
associate a homology class in H/K. Indeed, since there exists t
0
> 0
such that for all t with [t[ ≥ t
0
, x(t) ∈ U, the class of x[
[−t
0
,t
0
]
deﬁnes an
element in H
1
(M, U, Z). Let us denote by H the subset of H/K given
by all the classes corresponding to homoclinic orbits to
´
Σ(L).
72.7 Lemma. For any J as above, the image of 'H` under j is precisely
G/J.
Proof: Let L
0
denote the lift of the lagrangian L to M
0
. Observe ﬁrst
that by proposition 27.2, c(L) = c(L
0
) and therefore by lemma 72.5
we have
π
0
(
´
Σ(L
0
)) = p
−1
(π(
´
Σ(L))), (7.21)
192
72. homoclinic orbits. 193
fig. 1: Creating homoclinic connections with ﬁnite coverings.
where π
0
: TM
0
→M
0
is the canonical projection of the tangent bundle
TM
0
to M
0
.
Let us prove now that L
0
satisﬁes the hypothesis of theorem 311.1,
that is, the number of static classes of L
0
is ﬁnite. In fact, we shall
show that the projection to M
0
of a static class of L
0
coincides with
a connected component of p
−1
(π(
´
Σ(L))). Using (7.21) and proposition
311.4 we see that the projection of a static class of L
0
to M
0
must
be contained in a single connected component of p
−1
(π(
´
Σ(L))). Hence,
it suﬃces to show that if x and y belong to a connected component
of p
−1
(π(
´
Σ(L))) then d
0
c
(x, y) = 0. Since we are assuming that
´
Σ(L)
contains only one static class we have that d
c
(px, py) = 0. Since p :
M
0
→ M is a ﬁnite covering there are lifts x
1
of px and y
1
of py such
that d
0
c
(x
1
, y
1
) = 0. Since static classes are connected x
1
and y
1
must
belong to the same connected component of p
−1
(π(
´
Σ(L))) and thus there
is a covering transformation taking x
1
into x and y
1
into y which implies
that d
0
c
(x, y) = 0 as desired.
193
194 7. generic lagrangians.
Now theorem 311.1 and (7.21) imply that every covering transfor
mation in G/J can be written as the composition of covering transfor
mations that arise from elements in H, that is, j('H`) = G/J.
We shall need the following algebraic lemma.
72.8 Lemma. Let G = Z ⊕
k
. . . ⊕ Z. Given a ﬁnite index subgroup
J ⊂ G let us denote by j : G →G/J the projection homomorphism.
Let A be a subgroup of G. If A has the property that for all J as
above j(A) = G/J, then A = G.
Proof: The hypothesis readily implies that
A/A∩ J is isomorphic to G/J (7.22)
• If the rank of A is strictly less than the rank of G, one can easily
construct a subgroup J ⊂ G with ﬁnite index such that A ⊆ J and
G/J = ¦0¦. But this contradicts (7.22) because A/A∩ J = ¦0¦.
• If the rank of A equals the rank of G, then A has ﬁnite index in
G and by (7.22) G/A = ¦0¦ and thus G = A.
Observe now that any set H of a free abelian group G of rank k
such that the group generated by H is G must have at least k elements.
Therefore if we combine lemma 72.7 and lemma 72.8 with 'H` = A
we deduce that the set H of classes corresponding to homoclinic orbits
generates G and must have at least k elements thus concluding the proof
of theorem 72.1.
194
Appendix.
A Absolutely continuous functions.
A.1 Deﬁnition. A function f : [a, b] → R is absolutely continuous, if
∀ ε > 0 ∃ δ > 0 such that
N
¸
i=1
[t
i
−s
i
[ < δ =⇒
N
¸
i=1
[f(t
i
) −f(s
i
)[ < ε,
whenever ]s
1
, t
1
[, . . . , ]s
N
, t
N
[ are disjoint intervals in [a, b].
A.2 Proposition.
The function f : [a, b] →R is absolutely continuous if and only if
(i) The derivative f
(t) exists for a.e. t ∈ [a, b].
(ii) f
∈ L
1
([a, b]).
(iii) f(t) = f(a) +
t
a
f
(s) ds.
Proof: Deﬁne
µ([s, t]) := f(t) −f(s).
We claim that µ deﬁnes a ﬁnite signed Borel measure on [a, b]. Indeed,
let / be the algebra of ﬁnite unions of intervals. The function µ can be
extended to a σadditive function on /. Moreover, if B is a Borel set
195
196 appendix.
and ¦A
n
¦
n∈N
⊂ / is a family with A
n
↓ B, then µ(B) := lim
n
µ(A
n
)
exists because µ(B
n
` B
m
) →0 when n, m →+∞.
Observe that the properties of external measures and the absolute
continuity of f imply that µ < m, where m is the Lebesgue measure.
Let g =
dµ
dm
be the RadonNikodym derivative. Then g ∈ L
1
and
f(t) −f(a) = µ([a, t]) =
t
a
g(s) ds.
By the Lebesgue diﬀerentiation theorem
lim
h→0
f(t +h) −f(t)
h
= lim
h→0
1
h
t+h
t
g = g(t) for a.e. t ∈ [a, b].
Conversely, suppose that (i)(iii) hold. Using (ii), let µ(A) =
A
f
dm. Then by (iii),
µ([s, t]) = f(t) −f(s) for s, t ∈ [a, b].
Then µ <m implies
1
that f is absolutely continuous.
The Lebesgue diﬀerentiation theorem gives the following characteri
zation.
A.3 Corollary.
The function f : [a, b] →R is absolutely continuous if and only if
there exists g ∈ L
1
([a, b]) such that f(t) = f(a) +
t
a
g(s) ds.
1
If µ is ﬁnite, then µ m is equivalent, using the BorelCantelli lemma, to
∀ e > 0 ∃ δ > 0, m(A) < δ =⇒ µ(A) < ε.
196
B. measure theory 197
B Measure Theory
B.1 Riesz Theorem. [28]
Let X be a locally compact Hausdorﬀ topological space and let C
b
(X)
the vector space of continuous functions f : X → R with compact sup
port. Then any positive linear functional I : P(X) →R deﬁnes a unique
Borel measure µ on X such that I(f) =
f dµ for all f ∈ C
b
(X).
197
198 appendix.
C Convex functions.
A function f : R
n
→R is said convex if
f
λx + (1 −λ) y
≤ λf(x) + (1 −λ) f(y)
for all 0 ≤ λ ≤ 1 and x, y ∈ R
n
. Equivalently, if the set ¦ (x, r) ∈
R
n
R[ r ≥ f(x) ¦ is convex.
For x
0
∈ R
n
the subdiﬀerential of f at x
0
is the set
∂f(x
0
) := ¦ p ∈ R
n
∗
[ f(x) ≥ p (x −x
0
) +f(x
0
) ¦.
Its elementst are called subderivatives or subgradient of f at x
0
, and the
planes
¦ (x, r) ∈ R
n
R[ r = p (x −x
0
) +f(x
0
) ¦
are called supporting hyperplanes for f at x
0
. The functional p ∈ R
n
∗
is called the slope of the hyperplane.
For the proof of the following proposition see Rockafellar [62].
C.1 Proposition.
(a) ∂f(x) = for every x ∈ Dom(f).
(b) A ﬁnite convex function is continuous and Lebesgue almost every
where diﬀerentiable.
(c) If ∂f(x) = ¦p¦ then f is diﬀerentiable at x and f
(x) = p.
198
D. the fenchel and legendre transforms. 199
D The Fenchel and Legendre Transforms.
Given a convex function f : R
n
→ R the Fenchel Transform (or the
convex dual of f is the function f : (R
n
)
∗
→R ∪ ¦+∞¦ deﬁned by
f
∗
(p) = max
x∈R
n
[ p x −f(x) ] (D.1)
The function f admits a supporting hyperplane with slope p ∈ R
n
∗
if
and only if f
∗
(p) = +∞. If f is superlinear, then f
∗
is ﬁnite on all R
n
.
D.1 Proposition.
1. If f is convex then f
∗
is convex.
2. If f and f
∗
are superlinear then f
∗∗
= f.
3. f is superlinear if and only if f
∗
is bounded on balls, more explic
itly,
f(x) ≥ A [x[−B(A) , ∀x ∈ R
n
⇐⇒ f
∗
(p) ≤ B([p[) , ∀p ∈ R
n
∗
4. If f is superlinear, the maximum D.1 is attained at some point
x ∈ R
n
.
Proof:
1. Given 0 ≤ λ ≤ 1 and p
1
, p
2
∈ R
n
∗
we have that
f
∗
λp
1
+ (1 −λ) p
2
= max
x∈R
n
λp
1
+ (1 −λ) p
2
x −f(x)
≤ λ max
x∈R
n
[ p
1
x −f(x) ] + (1 −λ) max
x∈R
n
[ p
2
x −f(x) ]
= λf
∗
(p
1
) + (1 −λ) f
∗
(p
2
).
2. From (D.1) we get that
f(x) ≥ p x −f
∗
(p) for all x ∈ R
n
, p ∈ (R
n
)
∗
199
200 appendix.
Hence,
f(x) ≥ sup
p∈R
n
∗
[ p x −f
∗
(p) ] = f
∗∗
(x).
Let p
x
∈ ∂f(x) = . Then
f(y) ≥ f(x) +p
x
(y −x), ∀ y ∈ R
n
.
Hence
f
∗
(p
x
) = max
y∈R
n
p
x
y −f(y)
= p
x
x −f(x).
And
f(x) = p
x
x −f
∗
(p
x
) ≤ max
p∈R
n
∗
p x −f(x)
= f
∗∗
(x).
3. We have that
f
∗
x
(p) = max
v∈R
n
[ p v −f
x
(v) ]
≤ max
v∈R
n
[ p v −[p[ v ] +B([p[)
= B([p[).
Conversely, suppose that f
∗
x
(p) ≤ B([p[). Given A ∈ R and x ∈ R
n
there exists p
x
∈ R
n
∗
such that [p
x
[ = A and p
x
v = [p
x
[ [x[ = A[x[.
Then
f(x) = max
p∈R
n
∗
[ p x −f
∗
(p) ]
≥ p
x
v −B([p
x
[) = A[x[ −B(A).
4. By item 3, f
∗
is ﬁnite. Let p ∈ R
n
∗
. If b > 0 is such that f(x) >
([p[ + 1) [x[ −b, then
p x −f(x) < b −[x[ < f
∗
(p) −1 for [x[ > b + 1 −f
∗
(p).
Hence
f
∗
(p) = max
x≤b+1−f
∗
(p)
[ p x −f(x) ],
200
D. the fenchel and legendre transforms. 201
and the maximum is attained at some interior point x
p
in the
closed ball [x[ ≤ b + 1 −f
∗
(p).
D.2 Corollary.
If f : R
n
→R is convex and superlinear then so is f
∗
: R
n
∗
→R. In
this case f
∗∗
= f.
Observe that in this case we have
f
∗
(0) = − min
x∈R
n
f(x) and f(0) = − min
p∈R
n
∗
f
∗
(p).
If f : R
n
→ R is convex and superlinear we deﬁne the Legendre
Transform L : R
n
→2
R
n
∗
of f, by
L(x) = ¦ p ∈ R
n
∗
[ p x = f(x) +f
∗
(p) ¦, (D.2)
D.3 Proposition. If f : R
n
→R is C
2
and there is a > 0, such that
y f
(x) y ≥ a [y[
2
for all x, y ∈ R
n
,
then the Legendre transform L : R
n
→R
n
∗
is a C
1
diﬀeomorphism given
by L(x) = d
x
f.
Proof: The function f is convex and it is superlinear because
f(x) = f(0) +
1
0
f
(sx) ds
= f(0) +
1
0
1
0
sxf
(ts x) x dt ds
≥ f(0) +
1
2
a [x[
2
.
201
202 appendix.
From (D.1) we get that
p x ≤ f(x) +f
∗
(p) for all x ∈ R
n
, p ∈ R
n
∗
. (D.3)
By proposition D.2, f
∗
is superlinear. Then item 4 in proposition D.2
implies that
L(x) = arg max
p∈R
n
∗
¦ p x −f
∗
(p) ¦ = .
Moreover, from (D.3), if p ∈ L(x) then x = arg max
x∈R
n¦ p x − f(x) ¦.
Thus p = d
x
f = L(x). This proves that L diﬀerentiable and singled
valued. Moreover since d
x
L = f
(x) is nonsingular, then L is a local
C
1
diﬀeomorphism.
Since
(y −x) [ d
y
f −d
x
f ] =
1
0
f
sx + (1 −s)y
ds > 0,
then x → d
x
f = L(x) is injective. We now prove that L is surjective.
By item 4 in proposition D.1 the maximum
f
∗
(p) = max
x∈R
n
[ p x −f(x) ]
is attained at some x
p
∈ R
n
. Then p ∈ L(x
p
).
202
E. singular sets of convex funcions. 203
E Singular sets of convex funcions.
Let f : R
n
→ R be a convex function. Recall that its subdiﬀerential at
x ∈ R
n
is the set
∂f(x) := ¦ : R
n
→R linear [ f(y) ≥ f(x) + (y −x) , ∀y ∈ R
n
¦.
Then the sets ∂f(x) ⊂ R
n
are convex. If k ∈ N, let
Σ
k
(f) := ¦ x ∈ R
n
[ dim∂f(x) ≥ k ¦.
E.1 Proposition. If f : R
n
→ R is a convex function then for all
0 ≤ k ≤ n the Hausdorﬀ dimension HD(Σ
k
(f)) ≤ n −k.
We recall here an elegant proof due to Ambrosio and Alberti, see [3].
More can be said on the structure of Σ
k
, see [2, 66] for example.
By adding [x[
2
if necessary (which does not change Σ
k
) we can as
sume that f is superlinear and that
f(y) ≥ f(x) +(y −x) +
1
2
[y −x[
2
∀x, y ∈ R
n
, ∀ ∈ ∂f(x). (E.4)
E.2 Lemma. ∈ ∂f(x),
∈ ∂f(x
) =⇒ [x −x
[ ≤  −
.
Proof: From inequality (E.4) we have that
f(x
) ≥ f(x) +(x
−x) +
1
2
x
−x
2
,
f(x) ≥ f(x
) +
(x −x
) +
1
2
x −x
2
.
Then
0 ≥ (
−)(x −x
) +
x −x
2
(E.5)
−
x −x
≥ ( −
)(x −x
) ≥
x −x
2
. (E.6)
Therefore  −
 ≥ [x −x
[.
Since f is superlinear, the subdiﬀerential ∂f is surjective and we
have:
203
204 appendix.
E.3 Corollary.
There exists a Lipschitz function F : R
n
→R
n
such that
∈ ∂f(x) =⇒ x = F().
Proof of Proposition E.1:
Let A
k
be a set with HD(A
k
) = n − k such that A
k
intersects any
convex subset of dimension k. For example
A
k
= ¦x ∈ R
n
[ x has at least k rational coordinates ¦.
Observe that
x ∈ Σ
k
=⇒ ∂f(x) intersects A
k
=⇒ x ∈ F(A
k
).
Therefore Σ
k
⊂ F(A
k
). Since F is Lipschitz, we have that HD(Σ
k
) ≤
HD(A
k
) = n −k.
204
F. symplectic linear algebra. 205
F Symplectic Linear Algebra.
205
206 appendix.
206
Bibliography
[1] Ralph Abraham and Jerrold E. Marsden, Foundations of Mechanics, Benjamin
/ Cummings Publishing Co., Inc., Advanced Book Program, Reading, Mass.,
1978.
[2] Giovanni Alberti, On the structure of singular sets of convex functions, Calc.
Var. & Partial Diﬀerential Equations 2 (1994), no. 1, 17–27.
[3] Giovanni Alberti and Luigi Ambrosio, A geometrical approach to monotone func
tions in
n
, Math. Z. 230 (1999), no. 2, 259–316.
[4] J. Ball and V. Mizel, One dimensional variational problems whose minimizers
do not satisfy the eulerlagrange equation, Arch. Ration. Mech. Anal. 90 (1985),
325–388.
[5] Werner Ballmann, Mikhael Gromov, and Viktor Schroeder, Manifolds of non
positive curvature, Progress in Mathematics, vol. 61, Birkhauser, Boston, 1985.
[6] Patrick Bernard and Gonzalo Contreras, A generic property of families of la
grangian systems, to appear in Ann. Math., 2007.
[7] C´esar Camacho and Alcides Lins Neto, Geometric theory of foliations,
Birkhauser, Boston, 1985.
[8] M. J. Dias Carneiro and Arthur Lopes, On the minimal action function of au
tonomous Lagrangians associated to magnetic ﬁelds, Ann. Inst. H. Poincar´e Anal.
Non Lin´eaire 16 (1999), no. 6, 667–690.
[9] Mario Jorge Dias Carneiro, On minimizing measures of the action of autonomous
Lagrangians, Nonlinearity 8 (1995), no. 6, 1077–1085.
[10] Alfred Clebsch, ¨ uber die reduktion der zweiten variation auf ihre einfachste form,
J. Reine Angew. Math. 55 (1958), 254–276.
[11] G. Contreras, A. O. Lopes, and Ph. Thieullen, Lyapunov minimizing measures
for expanding maps of the circle, Ergodic Theory Dynam. Systems 21 (2001),
no. 5, 1379–1409.
207
208 BIBLIOGRAPHY
[12] Gonzalo Contreras, Jorge Delgado, and Renato Iturriaga, Lagrangian ﬂows:
the dynamics of globally minimizing orbits. II, Bol. Soc. Brasil. Mat. (N.S.) 28
(1997), no. 2, 155–196, special issue in honor of Ricardo Ma˜ n´e.
[13] Gonzalo Contreras and Renato Iturriaga, Convex Hamiltonians without conju
gate points, Ergodic Theory Dynam. Systems 19 (1999), no. 4, 901–952.
[14] Gonzalo Contreras, Renato Iturriaga, Gabriel P. Paternain, and Miguel Pater
nain, Lagrangian graphs, minimizing measures and Ma˜ n´e’s critical values, Geom.
Funct. Anal. 8 (1998), no. 5, 788–809.
[15] Gonzalo Contreras, Renato Iturriaga, and H´ector S´anchezMorgado, On the cre
ation of conjugate points for Hamiltonian systems, Nonlinearity 11 (1998), no. 2,
355–361.
[16] Gonzalo Contreras and Gabriel P. Paternain, Connecting orbits between static
classes for generic Lagrangian systems, Topology 41 (2002), no. 4, 645–666.
[17] J. J. Duistermaat, On the Morse index in variational calculus, Advances in Math.
21 (1976), no. 2, 173–195.
[18] Patrick Eberlein, When is a geodesic ﬂow of anosov type? i., J. Diﬀerential
Geometry 8 (1973), 437–463.
[19] Laurence. C. Evans and Ronald F. Gariepy, Measure theory and ﬁne properties
of functions, CRC Press, New York, 1992.
[20] Albert Fathi, Solutions KAM faibles conjugu´ees et barri`eres de Peierls, C. R.
Acad. Sci. Paris S´er. I Math. 325 (1997), no. 6, 649–652.
[21] , Th´eor`eme KAM faible et th´eorie de Mather sur les syst`emes lagrangiens,
C. R. Acad. Sci. Paris S´er. I Math. 324 (1997), no. 9, 1043–1046.
[22] , Orbites h´et´eroclines et ensemble de Peierls, C. R. Acad. Sci. Paris S´er.
I Math. 326 (1998), no. 10, 1213–1216.
[23] , Sur la convergence du semigroupe de LaxOleinik, C. R. Acad. Sci.
Paris S´er. I Math. 327 (1998), no. 3, 267–270.
[24] Patrick Foulon, Estimation de l’entropie des syst`emes lagrangiens sans points
conjugu´es, Ann. Inst. Henri Poincar´e 57 (1992), 117–146.
[25] Alexandre Freire, Fluxos geod´esicos em variedades riemannianas compactas de
curvatura negativa, Master’s thesis, IMPA, Rio de Janeiro, 1981.
[26] Alexandre Freire and Ricardo Ma˜ n´e, On the entropy of the geodesic ﬂow in
manifolds without conjugate points, Invent. Math. 69 (1982), no. 3, 375–392.
[27] L. W. Green, A theorem of E. Hopf, Michigan Math. J. 5 (1958), 31–34.
[28] Paul R. Halmos, Measure theory, SpringerVerlag, 1950.
[29] Philip Hartman, Ordinary Diﬀerential Equations, John Wiley & Sons, Inc., New
YorkLondonSydney, 1964.
208
BIBLIOGRAPHY 209
[30] Renato Iturriaga and H´ector SanchezMorgado, Finsler metrics and action po
tential, Proc. Amer. Math. Soc 128 (2000), 3311–3316.
[31] Wilhelm Klingenberg, Lectures on closed geodesics, Grundlehren der Mathema
tischen Wissenschaften, Vol. 230, SpringerVerlag, BerlinNew York, 1978.
[32] P.L. Lions, G. Papanicolau, and Varadhan S.R.S, Homogenization of hamilton
jacobi equations, preprint, unpublished, 1987.
[33] Ricardo Ma˜ n´e, Introdu¸c˜ ao ` a teoria erg´ odica, Instituto de Matem´ atica Pura e
Aplicada, Rio de Janeiro, 1983.
[34] , Ergodic theory and diﬀerentiable dynamics, SpringerVerlag, BerlinNew
York, 1987, Translated from the Portuguese by Silvio Levy.
[35] , On the minimizing measures of Lagrangian dynamical systems, Nonlin
earity 5 (1992), no. 3, 623–638.
[36] , Global Variational Methods in Conservative Dynamics, 19
o
Coloquio
Bras. Mat., IMPA, Rio de Janeiro, 1993.
[37] , Ergodic variational methods: new techniques and new problems, Pro
ceedings of the International Congress of Mathematicians, Vol. 1, 2 (Z¨ urich,
1994), Birkh¨ auser, Basel, 1995, pp. 1216–1220.
[38] , Generic properties and problems of minimizing measures of Lagrangian
systems, Nonlinearity 9 (1996), no. 2, 273–310.
[39] , Lagrangian ﬂows: the dynamics of globally minimizing orbits, Interna
tional Conference on Dynamical Systems (Montevideo, 1995), Longman, Har
low, 1996, Reprinted in Bol. Soc. Brasil. Mat. (N.S.) 28 (1997), no. 2, 141–153.,
pp. 120–131.
[40] Daniel Massart, Normes stables des surfaces, C. R. Acad. Sci. Paris S´er. I Math.
324 (1997), no. 2, 221–224.
[41] , Stable norms of surfaces: local structure of the unit ball of rational
directions, Geom. Funct. Anal. 7 (1997), no. 6, 996–1010.
[42] John N. Mather, Modulus of continuity for Peierls’s barrier, Periodic solutions
of Hamiltonian systems and related topics (Il Ciocco, 1986), Reidel, Dordrecht
Boston, MA, 1987, pp. 177–202.
[43] , Minimal action measures for positivedeﬁnite Lagrangian systems, IXth
International Congress on Mathematical Physics (Swansea, 1988), Hilger, Bris
tol, 1989, pp. 466–468.
[44] , Minimal measures, Comment. Math. Helv. 64 (1989), no. 3, 375–394.
[45] , Diﬀerentiability of the minimal average action as a function of the ro
tation number, Bol. Soc. Brasil. Mat. (N.S.) 21 (1990), no. 1, 59–70.
[46] , Action minimizing invariant measures for positive deﬁnite Lagrangian
systems, Math. Z. 207 (1991), no. 2, 169–207.
209
210 BIBLIOGRAPHY
[47] , Variational construction of orbits of twist diﬀeomorphisms, J. Amer.
Math. Soc. 4 (1991), no. 2, 207–263.
[48] , Variational construction of connecting orbits, Ann. Inst. Fourier (Greno
ble) 43 (1993), no. 5, 1349–1386.
[49] John N. Mather and Giovanni Forni, Action minimizing orbits in Hamiltonian
systems, Transition to chaos in classical and quantum mechanics (Montecatini
Terme, 1991), Springer, Berlin, 1994, Lectures by Bellissard, J. and Degli Es
posti, M. and Forni, G. and Graﬃ, S. and Isola, S. and Mather, J. N. given at
the Third C.I.M.E. Session held in Montecatini Terme, July 6–13, 1991, Edited
by S. Graﬃ, pp. 92–186.
[50] Marston Morse, A fundamental class of geodesics on any closed surface of genus
greater than one, Trans. Amer. Math. Soc. 26 (1924), 25–60.
[51] , Calculus of variations in the large, SpringerVerlag, Berlin, 1994, Lec
tures given at the Third C.I.M.E. Session held in Montecatini Terme, July 6–13,
1991, Edited by S. Graﬃ.
[52] G. Namah and J.M. Roquejoﬀre, Comportement asymptotique des solutions
d’une classe d’equations paraboliques et de hamiltonjacobi, C. R. Acad. Sci.,
Serie I, 324 (1997), 1367–1370.
[53] Richard S. Palais, Morse theory on Hilbert manifolds, Topology 2 (1963), 299–
340.
[54] Gabriel P. Paternain, Entropy and completely integrable Hamiltonian systems,
Proc. Amer. Math. Soc. 113 (1991), no. 3, 871–873.
[55] , On the topology of manifolds with completely integrable geodesic ﬂows,
Ergodic Theory Dynamical Systems 12 (1992), no. 1, 109–121.
[56] , Hyperbolic dynamics of EulerLagrange ﬂows on prescribed energy levels,
S´eminaire de Th´eorie Spectrale et G´eom´etrie, No. 15, Ann´ee 1996–1997, Univ.
Grenoble I, SaintMartind’H`eres, 1997, pp. 127–151.
[57] Gabriel P. Paternain and Miguel Paternain, On Anosov energy levels of convex
Hamiltonian systems, Math. Z. 217 (1994), no. 3, 367–376.
[58] , On the topological entropy of autonomous lagrangian systems, preprint,
1994.
[59] , Anosov geodesic ﬂows and twisted symplectic structures, International
Conference on Dynamical Systems (Montevideo, 1995), Longman, Harlow, 1996,
pp. 132–145.
[60] , Critical values of autonomous Lagrangian systems, Comment. Math.
Helv. 72 (1997), no. 3, 481–499.
[61] , First derivative of topological entropy for Anosov geodesic ﬂows in the
presence of magnetic ﬁelds, Nonlinearity 10 (1997), no. 1, 121–131.
210
BIBLIOGRAPHY 211
[62] R. Tyrrell Rockafellar, Convex analysis, Princeton Mathematical series, 21,
Princeton University Press, New Jersey, 1970.
[63] J.M. Roquejoﬀre, Comportement asymptotique des solutions d’equations de
hamiltonjacobi monodimensionnelles, C. R. Acad. Sci. Serie I,326 (1998), 185–
189.
[64] Rafael Oswaldo Ruggiero, On the creation of conjugate points, Math. Z. 208
(1991), 41–55.
[65] Sol Schwartzman, Asymptotic cycles, Ann. of Math. (2) 66 (1957), 270–284.
[66] Ludˇek Zajˇcek, On the points of multiplicity of monotone operators, Comment.
Math. Univ. Carolinae 19 (1978), no. 1, 179–189.
211
212 BIBLIOGRAPHY
212
Index
A
L
(γ), 9
A
L
(µ), 31, 44
C(x, y), 23
C
0
, 28
C
k
(q
1
, q
2
; T), 9
C
T
(x, y), 10, 55
S(x, y; T), 38
S
+
(γ), 38
[h]
Lip
, 40
((M), 30, 44
Γ
+
, 84
Γ
+
0
, 84
Γ
±
(u, k), 134
Lip
K
(M, R), 137
Φ
k
(x, y), 23
Φ
k
(x, y; T), 62
Σ
+
(L), 71
Σ
−
(L), 71
Σ
±
(k), 134
α(v), 72
α(ω) = β
∗
(ω), 50
H(u), 141
β(h), 50
/, 71
H, 39
´, 51
´(L), 31, 44
´(h), 50
´
, 28
¯
^ = Σ(L), 71
{ = π(
´
Σ(L)), 71, 79
S
±
(k), 134
δ(s, t), 143
ess sup , 141
´
Σ(L), 71
 
0
, 129

, 28

∞
, 165
µ
γ
, 30, 44
ω(v), 72
∂f(x), 198
≺, 99, 127, 141
, 89
ρ(µ), 47
c(L), 23, 44, 50
c
0
(L), 51, 52
c
a
(L), 52
c
u
(L), 45, 53
d
1
(γ
1
, γ
2
), 38, 56
d
k
(x, y), 24
e
0
, 12, 88
f
(x), 198
213
214 INDEX
h(x, y), 79
h
T
(x, y) = Φ
c
(x, y; T), 79
x
v
, 8
A
abelian cover, 52
real –, 53
absolutely continuous
functions, 195–196
metric, 38, 56
absolutely equicontinuous, 57
action
of a curve, 9
of a measure, 31, 44
potential, 23
of ﬁnite time, 62
action potential
Lipschitz property, 24
triangle inequality, 24
aﬃnely independent, 167
alfa limit, 188
of a semistatic, 73, 90
alpha function, 50
alpha limit, 72
Anosov energy level, 118
asymptotic cycle, 47
Aubry set, 71
AubryMather theory, 47–51
B
barrier
Peierls’, 79–81
basin
of a static class, 84
of a weak KAM, 121
beta function, 50
bound
a priori – for speed, 62, 63
for L
v
, 16
for the energy, 13
for the Legendre transform, 16
for the Peierls barrier, 79
lower – for a lagrangian, 23, 44
boundedness, 8, 15
Busemann weak KAM, 130
C
chain
recurrent, 90
transitive, 90
class
static, 84
closed form, 8, 47, 98
coboundary
Lipschitz, 86
properties, 86–87
condition
boundedness, 8, 15
convexity, 7
superlinearity, 7
conﬁguration space, 48
continuity
of the action potential, 24
of the critical value, 27
of the homology function ρ, 47
contraction
by a vector ﬁeld, 14
weak, 137
214
INDEX 215
convex
dual, 50, 199
function, 50, 171, 198–204
set, 204
convexity, 7
of the α function, 50
of the β function, 50
convolution, 106, 141
covering, 45, 52–53, 105
abelian, 52
ﬁnite, 52, 189, 191, 192
properties, 88, 122, 134
real abelian, 53
universal, 53, 118
critical value, 23, 24, 44, 50
e
0
≤ c
u
≤ c
a
≤ c
0
≤ c(L), 88
of a covering, 52
of the abelian cover (c
a
), 52
of the universal cover, 45, 53
strict, 88
strict (c
0
), 51, 52
curve
semistatic, 70
energy of, 71
static, 70
cut locus
of a point, 100
of a static class, 84
of a weak KAM, 121
D
deck transformation, 189
deformation retract, 48
distance
absolutely continuous, 38, 56
dominated, 127
dominated (≺), 99
dominated function, 99–102
domination
properties, 122, 134
dual
convex –, 50, 199
E
energy
Finsler, 115
function, 12–13
kinetic, 17
level
Anosov, 118
compactness, 13
regular, 15, 118
of a minimizing measure, 76
of a semistatic curve, 71
of a timefree minimizer, 65
potential –, 17
equicontinuous
absolutely, 57
ergodic
uniquely, 166
ess sup, 141
EulerLagrange
equation, 8
magnetic lagrangian, 19
mechanic lagrangian, 17
riemannian lagrangian, 17
ﬂow, 8
reversible, 18
215
216 INDEX
example
c(L) > c
0
(L) > e
0
, 53
c
0
(L) > c
u
(L), 53
extremal point, 167
extreme point, 167
F
Fenchel transform, 15, 199
ﬁnite
– time potential, 62
covering, 189, 191
ﬁnite covering, 192
Finsler
energy, 115
lagrangian, 17
metric, 17, 113–117
ﬁxed point
LaxOleinik semigroup, 137,
139
ﬂow
embedded, 21
EulerLagrange, 8
Finsler geodesic –, 17, 114
reversible, 18
riemannian geodesic, 17
twisted geodesic, 19
form
canonical symplectic –, 14
closed, 8, 47, 98
Liouville’s 1form, 14
function
convex, 50, 171, 199, 204
(, 198
G
generic
point, 48
geodesic ﬂow
Finsler, 114
of a Finsler metric, 17
riemannian, 17
twisted –, 19
gradient, 18
graph
lagrangian, 98
properties, 82–85, 122, 134
submanifold, 98
growth
linear, 28
superlinear, 7, 23, 44, 50, 199
H
HamiltonJacobi
equation
weak solution, 103–104
theorem, 97
hamiltonian, 14
ﬂow, 14
riemannian, 17
vector ﬁeld, 14
holonomic measure, 30
homoclinic orbit
homology of –, 188
to a static class, 188
homology
of a homoclinic orbit, 188
of a measure, 47
Horocycle ﬂow, 160
216
INDEX 217
Hurewicz homomorphism, 52, 53,
192
hyperplane
slope of –, 198
supporting, 198, 199
hypothesis on L, 7, 8, 15
I
independent
aﬃnely, 167
inequality
c(L) ≥ e
0
, 88
invariant measures ´(L), 31, 44
isotropic subspace, 97
K
KAM
solution, 121
kinetic energy, 17
L
lagrangian
Finsler, 17
ﬂow, 8
reversible, 18
graph, 98
cohomology class of, 98
exact, 98
lower bound, 23, 44
magnetic, 18
mechanic, 17
natural, 17
riemannian, 17
submanifold, 97
subspace, 97
symmetric, 18, 53
LaxOleinik semigroup, 137
convergence, 140
ﬁxed point, 137, 139
Lipschitz images, 137
weak contraction prop., 137
Legendre transform, 15, 100, 201
Lie derivative, 14
Liouville’s 1form, 14
Lipschitz
coboundary, 86
graph, 82
images of LaxOleinik, 137
inverse, 84
properties
weak KAM solution’s, 122
property
action potential’s, 24
ﬁnitetime potential’s, 67
Peierls barrier’s, 79
weak KAM solution’s, 134
Rademacher’s theorem, 106
smallest – constant, 40
subsolutions of HJ, 109
local
static curve, 72
lower bound
for a lagrangian, 23, 44
M
Ma˜ n´e
set, 71
magnetic lagrangian, 18
217
218 INDEX
EL equation, 19
Mather
beta function, 50
crossing lemma, 82
minimizing measures, 51
set, 51, 71
theory, 47–51
Mather set, 71
measure
holonomic, 47
Mather minimizing, 51
minimizing, 44, 75
invariance of –, 31
mechanic lagrangian, 17
EL equation, 17
metric
absolutely continuous, 38, 56
Finsler, 17
minimizing measure, 44, 75
energy of –, 76
invariance of –, 31
support of, 75
molliﬁcation, 106, 141
N
natural lagrangian, 17
O
omega limit, 72, 188
of a semistatic, 73, 90
operator
LaxOleinik, 137
convergence, 140
ﬁxed point, 139
Lipschitz images, 137
weak contraction, 137
P
partial order, 89
Peierls barrier, 79–81
characterization
with action potential, 81
with KAM, 129
inﬁnite, 130, 158
is weak KAM, 127
Lipschitz property, 79
Peierls set {, 71, 79, 82
chain recurrency, 90
empty, 158, 160
nonemptyness, 76
phase space, 48
plane
supporting, 198, 199
point
extremal, 167
extreme, 167
generic, 48
potential
action –, 23
properties, 24
energy, 17
a priori bound for speed, 62, 63
probability
holonomic, 47
invariant, 31, 44
properties
coboundary –, 86–87
covering –, 88, 122, 134
218
INDEX 219
domination –, 122, 134
graph –, 82–85, 122, 134
Lipschitz –, 122, 134
recurrence –, 89–95
smoothness –, 122, 134
weak contraction –, 137
R
Rademacher’s theorem, 106
ray, 73, 143
real abelian cover, 53
recurrence
properties, 89–95
regular
energy level, 15
residual, 165
reversible ﬂow, 18
riemannian
hamiltonian, 17
lagrangian, 17
riemannian lagrangian
EL equation, 17
Riesz theorem, 32, 197
rotation of a measure, 47
S
Schwartzman, 47
semicontinuous, 50
semigroup
LaxOleinik, 137
convergence, 140
ﬁxed point, 137, 139
Lipschitz images, 137
weak contraction, 137
semistatic curve, 70
α and ωlimits, 73, 90
energy of, 71
lift of, 190
set
Aubry, 71
convex, 204
Ma˜ n´e, 71
Mather, 51, 71
Peierls, 71, 79, 82
chain recurrency, 90
simplex, 167
slope of a hyperplane, 198, 199
smoothness
properties, 122, 134
solution
weak – of HJ equation, 103–
104
static
class, 73, 84, 90
curve
local, 72
orbit, 70
energy of, 71
existence, 76
lift of, 190
strict critical value (c
0
), 51, 52
subderivative, 198
subdiﬀerential, 172, 198, 203–204
subgradient, 198
submanifold
graph, 98
lagrangian, 97
219
220 INDEX
subsolution of the HamiltonJacobi
equation, 105
subspace
isotropic, 97
lagrangian, 97
superlinear, 199
superlinearity, 7, 23, 44
of convex duals, 199
of the α function, 50
of the β function, 50
supporting hyperplane, 198, 199
surjectivity
of the homology function ρ, 49
of the projection of an energy
level, 12
symmetric lagrangian, 18, 53
symplectic form, 14
canonical, 97
T
theorem
HamiltonJacobi, 97
Tonelli’s, 55
timefree minimizer
energy of, 65
Tonelli
minimizer, 38, 55
theorem, 55
torsion, 53
transform
Fenchel, 15, 199
Legendre, 15, 100, 201
triangle inequality
for the action potential, 24
for the Peierls barrier, 79
twisted geodesic ﬂow, 19, 160
uniquely ergodic, 166
vector ﬁeld
hamiltonian, 14
W
weak
contraction, 137
KAM solution, 121
Busemann –, 130
solution of HJ, 103–104
subsolution of HJ, 109
220
ii
ii
Contents
1 Introduction. 11 Lagrangian Dynamics. . . . . 12 The EulerLagrange equation. 13 The Energy function. . . . . . 14 Hamiltonian Systems. . . . . 15 Examples. . . . . . . . . . . . 7 7 9 12 14 17 23 23 27 28 31 44 47 47 47 50 52
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
2 Ma˜´’s critical value. ne 21 The action potential and the critical value. . 22 Continuity of the critical value. . . . . . . . . 23 Holonomic measures. . . . . . . . . . . . . . . 24 Invariance of minimizing measures. . . . . . . 25 Ergodic characterization of the critical value. 26 The AubryMather Theory. . . . . . . . . . . 26.a Homology of measures. . . . . . . . . . 26.b The asymptotic cycle. . . . . . . . . . 26.c The alpha and beta functions. . . . . 27 Coverings. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
3 Globally minimizing orbits. 55 31 Tonelli’s theorem. . . . . . . . . . . . . . . . . . . . . . . . 55 32 A priori compactness. . . . . . . . . . . . . . . . . . . . . 62 iii
iv
CONTENTS
33 34 35 36 37 38 39 310 311 4 The 41 42 43 44 45 46 47 48 49
Energy of timefree minimizers. . . . . . . The ﬁnitetime potential. . . . . . . . . . Global Minimizers. . . . . . . . . . . . . . Characterization of minimizing measures. The Peierls barrier. . . . . . . . . . . . . . Graph Properties. . . . . . . . . . . . . . Coboundary Property. . . . . . . . . . . . Covering Properties. . . . . . . . . . . . . Recurrence Properties. . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
65 67 70 75 79 82 86 88 89 97 97 99 103 105 111 113 118 121 125 126 127 130 134 137 145
Hamiltonian viewpoint. The HamiltonJacobi equation. . . . . . . . . . . Dominated functions. . . . . . . . . . . . . . . . . Weak solutions of the HamiltonJacobi equation. Lagrangian graphs. . . . . . . . . . . . . . . . . . Contact ﬂows. . . . . . . . . . . . . . . . . . . Finsler metrics. . . . . . . . . . . . . . . . . . . . Anosov energy levels. . . . . . . . . . . . . . . . . The weak KAM Theory. . . . . . . . . . . . . . . Construction of weak KAM solutions . . . . . . . 49.a Finite Peierls barrier. . . . . . . . . . . . 49.b The compact case. . . . . . . . . . . . . . 49.c Busemann weak KAM solutions. . . . . . 410 Higher energy levels. . . . . . . . . . . . . . . . . 411 The LaxOleinik semigroup. . . . . . . . . . . . . 412 The extended static classes. . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
5 Examples 155 51 Riemannian Lagrangians. . . . . . . . . . . . . . . . . . 155 52 Mechanic Lagrangians. . . . . . . . . . . . . . . . . . . . 155 53 Symmetric Lagrangians. . . . . . . . . . . . . . . . . . . 156 iv
. 175 71 Generic Lagrangians. . . .
. . . . . . . . 195 . . . . . . .
. . . .
. . . . . . F Symplectic Linear Algebra. . . . . . . . . . . . . . . . . .
. . . . E Singular sets of convex funcions. . 203 . . . . . . The ﬂat Torus Tn . . . . . . . . . .
. . . . . . . . . A Lagrangian with Peierls barrier h = +∞. . Horocycle ﬂow. . . . . . . . . . . . . . . . .
v
. . . . .a Open. . . Flat domain for the βfunction. . 61 Generic Families of Lagrangians. 169 . . . .
. . . . . . . .
. . . . . . . . .
. 197 . 170
7 Generic Lagrangians. . . . .
. . . . . . . . . . . . . . . .
. . . 61. . . .
. . . 205 207 213
. . . . . . . . B Measure Theory . . . . . .b Dense. . . . . . . . . . . . . . . 176 72 Homoclinic Orbits. . . 166 . . .
. . . . . . . . . . . . . . . . . . . . . . . . . 199 . . . . . . . . . Index. . . . . . . . . . C Convex functions. . . . A Absolutely continuous functions. . . D The Fenchel and Legendre Transforms. . .
165 . . . . . . . . . . . . . . . .CONTENTS
v
54 55 56 57 58
Simple Pendulum. . . . . . . . .
.
156 157 158 158 160
6 Generic Lagrangians. . . . 195 . . 188 Appendix. . . . . . . . Bibliography. . 198 . . . . . . . . . . . . . . 61.
. . .
. . . . . .
vi
CONTENTS
vi
.
v) = +∞. there is A > 0 such that The Hessian w · Lvv (x. v) ∈ T M. for all A ∈ R there is B ∈ R such that L(x.
Let M be a boundaryless ndimensional complete riemannian manifold. i.Chapter 1
Introduction. An (autonomous) Lagrangian on M is a smooth function L : T M → R satisfying the following conditions: (a) Convexity: ∂2L (x. v) · w ≥ A w2 (b) Superlinearity: L(x.
equivalently. v) ≥ A v − B 7 for all (x.
11 Lagrangian Dynamics. v v→+∞ lim uniformly on x ∈ M.e. for all (x. v) ∈ T M . v). calculated in linear co∂vi ∂vj ordinates on the ﬁber Tx M . is uniformly positive deﬁnite for all (x. v) ∈ T M and w ∈ Tx M .
.
x).2.v)≤r
The EulerLagrange equation associated to a lagrangian L is (in local coordinates) d ∂L ∂L x. v) < +∞.v)∈T M.
The Boundedness condition (c) is equivalent to the condition that the associated hamiltonian H is convex and superlinear.
w=1
(x. x) = ˙ (x. the new lagrangian L + ω also satisﬁes the hypothesis (a)(c) and has the same EulerLagrange equation as L. Observe that when we add a closed 1form ω to the lagrangian L.1) (1. This condition is immediate when the manifold M is compact.
L(x. ˙ We shall be interested on coverings p : N → M of a compact manifold M and the lifted Lagrangian L = L ◦ dp : T N → R of a convex superlinear lagrangian L on M . This can also be seen using the variational interpretation of the EulerLagrange equation (see 12. xv (t) . ˙ (EL) dt ∂v ∂x The condition (c) implies that the EulerLagrange equation (EL) deﬁnes a complete ﬂow ϕt on T M (proposition 13. (r) = g(r) = sup
(x. called the EulerLagrange ﬂow. where xv : R → M is the ˙ solution of (EL) with xv (0) = x0 and xv (0) = v0 . introduction.
(1.3). by setting ϕt (x0 .2).8
1. v) · w < +∞. see remark 14.
(c) Boundedness1 : For all r ≥ 0. The lagrangian L then satisﬁes (a)–(c) and its ﬂow ψt is the lift of ϕt . v0 ) = xv (t).
1
8
.2)
v≤r
sup w · Lvv (x.
If a curve x(t) in the space C k (q1 . . γ(t)) dt ˙
One of the main problems of the calculus of variations is to ﬁnd and to study the curves that minimize the action. . then x satisﬁes the EulerLagrange equation d Lv (x(t). q2 . suﬃciently small the curve yε = x + εh is on C k (q1 . the eulerlagrange equation. T ] → M such that γ(0) = q1 and γ(T ) = q2 . xn ) about x(t). x(t)) = Lx (x(t). q2 . . q2 . T ) and contained in the coordinate system. Consequently. 12. this equation does not depend on the coordinate system. T ) the set of C k diﬀerentiable curves γ : [0.1 Proposition. Then for every ε.
Proof: Choose a coordinate system (x1 .12. T ] → M is deﬁned by
T
AL (γ) =
0
L(γ(t). T ). T ) is a critical point of the action functional on C k (q1 . x(t)) ˙ ˙ dt
(EL)
in local coordinates. .
The action of a diﬀerential curve γ : [0. Deﬁne
g(ε) = AL (yε ) 9
. Let h(t) a diﬀerentiable curve such that h(0) = h(T ) = 0. q2 .
9
12
The EulerLagrange equation. Denote by C k (q1 .
v) = (v.
Then g has a minimum in zero and g(ε) − g(0) = lim ε→0 ε→0 ε lim
T
=
0 T
˙ L(x + εh. ·). A priori minimizers do not have to be diﬀerentiable and there are examples where they are not. T ] → M such that γ(0) = p and γ(T ) = q.2 Remark. x(t)) h dt. 12. The Euler Lagrange equation is a second order diﬀerential equation on M . This implies that x(t) satisﬁes the Euler Lagrange equation (EL).10
1. However 10
. see Ball & Mizel [4]. q). ˙ dt
for any function h ∈ C k (0. the set of absolutely continuous curves γ : [0. T ). Observe that X is of the form X(x. x + εh) − L(x. 0. introduction.
T 0
=
0
(Lx −
Hence 0=
0
T
Lx (x(t). ˙ v = (Lvv )−1 (Lx − Lvx v). It is possible to do the same thing in the space CT (p. but the convexity hypothesis (Lvv invertible) implies that this equation can also be seen as a ﬁrst order diﬀerential equation on T M : x = v. x) ˙ ˙ dt ε 0 ˙ εLx h + εLv h + o(ε) lim dt ε→0 ε
T
=
0 T
˙ Lx h + Lv h dt (Lx −
d dt Lv ) h d dt Lv ) h
=
0 T
dt + Lv h dt. x(t)) − ˙
d Lv (x(t). ˙ The associated vector ﬁeld X on T M is called the lagrangian vector ﬁeld and its ﬂow ϕt the lagrangian ﬂow.
every absolutely continuous minimizers is C 2 and satisﬁes the EulerLagrange equation. the eulerlagrange equation. x2 . This implies that the EulerLagrange equations for L and L + ω are the same.
γ
because the curve η is homologous to γ.2). the critical points for AL+ω and for AL are the same. the lagrangian L + ω also satisﬁes the hypothesis (a)(c).12. But since the values of AL+ω and AL are diﬀerent.
11
. 12. See Mather [46].
11
when the lagrangian ﬂow is complete (cf. Moreover. minimizers of these two actions may be diﬀerent. Therefore. proposition 13. T ) satisﬁes AL+ω (η) = AL (η) + ω. the action functional AL+ω on a neighbourbood of a curve γ ∈ C k (x1 . If we add a closed 1form ω to the lagrangian L.3 Remark.
v) · v > 0. then d d ˙ ˙ dt E(x. invariant function2 ) for the lagrangian ﬂow ϕt and its level sets. ∂v (1.e. v). x) = dt Lv − Lx · x = 0. sv)s=1 = v · Lvv (x.1) and (1. By the uniform convexity.4)
by the superlinearity e0 > −∞. 0) = −L(x. timeindependent) lagrangians.5)
w · Lvv (x. 0) = − min L(x. then e0 = min { k ∈ R  π : E −1 {k} → M is surjective }. 0). Hence E : T M → R is an integral (i.
The energy function of the lagrangian L is E : T M → R. Moreover.6)
The energy is invariant only for autonomous (i.
12
. v) = E(x. 0) > −∞.e.12
1. the convexity implies that
d ds
E(x. called energy levels are invariant under ϕt .3)
Observe that if x(t) is a solution of the EulerLagrange equation (EL).
(1. v) · w > 0.
w=1
and then using (1. 0) +
0
d ds
v E x. v) = ∂L (x.
Write e0 := max E(x.
v
E(x. introduction.
Thus
v∈Tx M
min E(x.
x∈M x∈M
(1. and the boundedness condition. A := inf
(x. v) · v − L(x. deﬁned by E(x. v) = E(x.v)∈T M
(1.
13
The Energy function.2). s v ds
≥ − (0) +
2
1 2
A v2 .
γ(t)).1 Remark.6). we can extend the interval of deﬁnition ]α. β[ is the maximal interval of deﬁnition of t → ϕt (v). Let k = E(v). x) ≤ a β − α + 1 . there is a > 0 such that 0 ≤ ϕt (v) ≤ a for α ≤ t ≤ β. Since ϕt (v) is of the form (γ(t).7)
Hence 13. Then by the theory of ordinary diﬀerential equations. then E −1 {k} ∩ TK M is compact. then ϕt (v) remains in the interior of the ˙ compact set Q := (y. Since E(ϕt (v)) ≡ k. w ≤ a + 1 . v) ≤ e0 + g(v) v. E(x.2). 13.
13
Similarly.2 Proposition. If k ∈ R and K ⊆ M is compact. using (1. The EulerLagrange vector ﬁeld is uniformly Lipschitz on Q. β[ of t → ϕt (v). The EulerLagrange ﬂow is complete. and −∞ < α or β < +∞. by (1. the energy function.13. Proof: Suppose that ]α. w) ∈ T M d(y. (1.
where x = π(v).
13
.
.
14
.
LX is the Lie derivative.8)
In local charts. xn .9) p = − Hx . . (1. . . because3
∗ d dt (ψt ω)
3
= LXH ω = d iXH ω + iXH ω = d(dH) + iXH (0) = 0. p1 . introduction. . .
Let T ∗ M be the cotangent bundle of M . In these coordinates the forms Θ and ω are written Θ = p · dx =
i
pi dxi .14
1. it preserves the symplectic form ω. . .
14
Hamiltonian Systems. because
d dt H
= Hx x + Hp p = 0. The hamiltonian vector ﬁeld XH associated to H is deﬁned by ω(XH . dpi ∧ dxi . pn ) of T ∗ M writing p ∈ T ∗ M as p = Σi pi dxi . . the hamiltonian vector ﬁeld deﬁnes the diﬀerential equation x= ˙ Hp . Let ψt be the hamiltonian ﬂow. (1. where ψt is the ﬂow of X. The canonical symplectic form on T ∗ M is deﬁned as ω = dΘ.
where π : T ∗ M → M is the projection. ·) is the contraction by X. Deﬁne the Liouville’s 1form Θ on T ∗ M as Θp (ξ) = p (dπ ξ) for ξ ∈ Tp (T ∗ M ). · ) = dH. . xn ) of M induces a local chart (x. The Lie derivative satisﬁes LX η = dt ψt ηt=0 .
i
ω = dp ∧ dx =
A hamiltonian is a smooth function H : T ∗ M → R. where d ∗ iX η = η(X. deﬁned on forms η by LX η = diX η + iX dη. . ˙ ˙
Moreover. . ˙ where Hx and Hp are the partial derivatives of H with respect to x and p. Observe that it preserves H. A local chart x = (x1 . p) = (x1 .
v)) is the Legendre transform of L.
We say that an energy level H −1 (k) is regular. Using that L∗ = H and H ∗ = L. The Legendre transform L : T M → T ∗ M . p) = max p v − L(x. p). v) = (x.3) and L(x. hamiltonian systems. Proof: By corollary D. v) = (x. and the EulerLagrange equation x= ˙ p= ˙
d dt d dt
x =v =
Hp . v) − L(x.14. p) = k. Hp (x. p)). dH(x.
v∈Tx M
Observe that H = E ◦ L−1 .
Lv = Lx = −Hx . So if p = Lv (x. Moreover 14.2 Remark. if k is a regular value of H. With this notation: H(x.2. L(x. where E is the energy function (1.1 Proposition.e. from proposition D. p) = 0 whenever H(x.
is the same as the hamiltonian equations. v)) is a conjugacy between the lagrangian ﬂow and the hamiltonian ﬂow. the convexity and superlinearity hypothesis imply that L = L∗∗ = H ∗ . p) − L(x. Lv (x. i. v) = E ◦ L−1 = p · Hp (x. 15
. Thus Hx = −Lx . Lv (x. p) = v · Lv (x. v) then v = Hp (x. v).2 in the appendix we obtain that the boundedness condition is equivalent to (c) Boundedness: H = L∗ is convex and superlinear.
15
We shall be specially interested in hamiltonians obtained by the Fenchel transform of a lagrangian: H(x. 14.
16
1. (1. ∞[→ R+ such that Lv (x. v) + max{ L(x. v) ∈ T M . introduction. There is a function f : [0. if H(x. v) ≤ E(x.
Applying this inequality to −w. Then. L(x. v). w)
w∈Tx M
is attained at w = v0 with p = Lv (x. Lv (x. v)) = E(x. G : T ∗ M → R are two hamiltonians with H −1 (k) = G−1 ( ) and k. v). v) ≤ f (v) for all (x. Thus there exists λ(x. e0 + g(r)r} +  (r) =: f (r). p) = k. 14. p) = max p · w − L(x.p) H −1 (k) = T(x. v). Lv (x. p) XG when H(x. w ∈ Tx M. We shall need the following estimate on the norm of the partial derivative Lv (x.p) H = T(x. Equation (1. −w).3 Proposition. v) − L(x.p) G. w). v) · w ≤ E(x. v) + L(x. ∀v. Thus using (1.8) implies that XH = λ(x. Two hamiltonian ﬂows restricted to a same regular energy level are reparametrizations of each other.p) H −1 (k) = ker d(x. ∀x ∈ M. we have Lv (x.4 Lemma.
14. v) · w ≥ −E(x. p) d(x. ker d(x. 2
16
.p) H = λ(x. p) > 0 such that d(x.7) and (1. we get that Lv (x.1).p) G.6). for v ≤ r. p) = k. Since H(x. v0 ). Proof: Suppose that H. Proof: The convexity condition implies that the maximum in H(x. are regular values for H and G respectively. −v) } ≤ max{ (0) + 1 Ar2 .
15. v) =
1 2
v
2 x. ˙
(1.
17
15
Examples. the riemannian lagrangian on M is given by the kinetic energy L(x. given also by formula (1. v) = Its EulerLagrange equation is
D dt
4
1 2
v
2 x
− U (x). ˙
i.
We give here some basic examples of lagrangians. The EulerLagrange ﬂow of a Finsler lagrangian is called the geodesic ﬂow of the Finsler metric · x .
(1. Riemannian Lagrangians: Given a riemannian metric g = ·. but where · x is a Finsler metric. is given by the kinetic energy minus the potential energy U : M → R.10). x 2
Analogous to the riemannian lagrangian is the Finsler lagrangian.
(1. examples. p) = 1 p 2 . also called natural lagrangian. · x is a (non necessarily symmetric4 ) norm on Tx M which varies smoothly on x ∈ M . L(x.11) Its corresponding
and its EulerLagrange ﬂow is the geodesic ﬂow. Mechanic Lagrangians: The mechanic lagrangian.e.12)
x = − U (x).e. λ v
x
=λ v
x
only for λ ≥ 0
17
. hamiltonian is H(x.10)
Its EulerLagrange equation (EL) is the equation of the geodesics of g:
D dt
x ≡ 0. i. · x on T M .
This can be seen by ﬁrst observing that the solutions of the EulerLagrange equation are the critical points of the action functional on curves on C(x. L(x. We call a magnetic lagrangian a lagrangian of the form L(x. + U (x). y. the action functional of L and L on C(x. v) + ωx (v).14) 2 18
.
Symmetric Lagrangians. i. T ) (with ﬁxed time interval and ﬁxed endpoints). These are the lagrangians which satisfy L(x. Magnetic Lagrangians. If one adds a closed 1form ω to a lagrangian. But adding a nonclosed 1form to a lagrangian does change the EulerLagrange ﬂow.
Its energy function and its hamiltonian are given by the kinetic energy plus potential energy: E(x. (1. The symmetric lagrangians is a class of lagrangian systems which includes the riemannian and mechanic lagrangians. (1. Since ω is closed. −v) for all (x.
D where dt is the covariant derivative and respect to the riemannian metric g. v) = H(x.e. v
x
for all (x. v) ∈ T M.18
1. v) = L(x. p) =
1 2 1 2
v p
2 x 2 x
+ U (x). T ) diﬀer by a constant and hence they have the same critical points. the EulerLagrange ﬂow does not change.13)
Their EulerLagrange ﬂow is reversible in the sense that ϕ−t (v) = −ϕt (−v). y. introduction.
U is the gradient of U with
dx U (v) =
U (x). v) = 1 v x + ηx (v) − U (x). v) = L(x. v) ∈ T M.
v x . and U : M → R a smooth function. 19
. ζ) = dπ ξ .
(1. Twisted geodesic ﬂows. Let K : T T M → d T M be the connection map Kξ = x v.15. Fix a riemannian metric .
19
where · x is a riemannian metric. Let ˙ π : T M → M be the canonical projection. The twisted geodesic ﬂows correspond to the motion of a particle under the eﬀect of a magnetic ﬁeld with no potential energy. v then the EulerLagrange equation of (1. Let ω0 be the symplectic form in T M obtained by pulling back the canonical symplectic form via the Legendre transform associated to the riemannian metric. Kζ − dπ ζ .14) is
D dt
x = Yx (x) − ˙ ˙
U (x). v) = Y (u). where dη ≡ 0. The energy functional is the same as that of the mechanical lagrangian but its hamiltonian changes because of the change in the Legendre transform: E(x. and a 2form Ω on M . v) = H(x. i. examples. v) = 2 1 2 v x + ηx (v).e. where ξ = dt x(t).
2 x
p − A(x)
+ U (x). A generalization of these ﬂows can be made using a nonzero 2form Ω instead of dη and not requiring Ω to be exact. If Y : T M → T M is the bundle map such that dη(u. This can be modeled as the EulerLagrange ﬂow of a lagrangian of the form L(x. This is better presented in the hamiltonian setting.15)
This models the motion of a particle with unit mass and unit charge under the eﬀect of a magnetic ﬁeld with Lorentz force Y and potential energy U (x). ω0 (ξ. η is a 1form on M with dη = 0. Kξ . p) =
1 2 1 2
v
2 x
+ U (x). But the EulerLagrange equations depend only on the riemannian metric and dη.
where A : M → T M is the vector ﬁeld given by ηx (v) = A(x). v(t) .
20
1. introduction.
The coordinates Tθ T M ξ ←→ (dπ ξ, Kξ) ∈ Tπ(θ) M ⊕ Tπ (θ)M = H(θ) ⊕ V (θ) are the standard way of writing the horizontal and vertical components of a vector ξ ∈ Tθ T M for a riemannian manifold M (see Klingenberg [31]). Deﬁne a new symplectic form ωΩ on T M by ωΩ = ω0 + π ∗ Ω . This is called a twisted symplectic structure on T M . Let H : T M → R be the hamiltonian H(x, v) = 1 v 2 . x 2 Consider the hamiltonian vector ﬁeld XF corresponding to (H, ωΩ ), i.e. ωΩ XΩ (θ), · = dH . Deﬁne Y : T M → T M as the bundle map such that Ωx (u, v) = Y (u), v
x.
(1.16)
(1.17)
The hamiltonian vector ﬁeld XΩ (θ) ∈ Tθ T M is given by XΩ (θ) = (θ, Y (θ)) ∈ H(θ) ⊕ V (θ). Hence the hamiltonian equation is
D dt
x = Yx (x), ˙ ˙
recovering equation (1.15) with U ≡ 0, but where Ω doesn’t need to be exact. If H 1 (M, R) = 0, both approaches coincide, and any twisted geodesic ﬂow is the lagrangian ﬂow of a magnetic lagrangian of the form L(x, v) = 2 1 2 v x + ηx (v), with dη = Ω. For example if N is a compact manifold Ω is a 2form in N and M is the abelian cover or the universal cover of N ; if Ω is not exact, then the corresponding twisted geodesic ﬂow is a lagrangian ﬂow on M but not on N (where it is locally a lagrangian ﬂow). This lagrangian ﬂow on M is actually the lift of the twisted geodesic ﬂow on N .
20
15. examples.
21
Embedding ﬂows: There is a way to embed the ﬂow of any bounded vector ﬁeld on a lagrangian system. Given a smooth bounded vector ﬁeld F : M → T M , let L(x, v) = 1 v − F (x) 2 . (1.18) x 2 Since F (x) is bounded, then the lagrangian L is convex, superlinear and satisﬁes the boundedness condition. The lagrangian L on a ﬁber Tx M is minimized at (x, F (x)), hence the integral curves of the vector ﬁeld, x = F (x), are solutions to the EulerLagrange equation. ˙
21
22
1. introduction.
22
Chapter 2
Ma˜´’s critical value. ne
21 The action potential and the critical value.
We shall be interested on action minimizing curves with free time interval. Unless otherwise stated, all the curves will be assumed to be absolutely continuous. For x, y ∈ M , let C(x, y) = { γ : [0, T ] → M  T > 0, γ(0) = x, γ(T ) = y }. For k ∈ R deﬁne the action potential Φk : M × M → R ∪ {−∞}, by Φk (x, y) = inf
γ∈C(x,y)
AL+k (γ).
Observe that if there exists a closed curve γ on N with negative L + k action, then Φk (x, y) = −∞ for all x, y ∈ N , by going round γ many times. Deﬁne the critical level c = c(L) as c(L) = sup{ k ∈ R  ∃ closed curve γ with AL+k (γ) < 0 }. Observe that the function k → Φk (x, y) is increasing. The superlinearity implies that L is bounded below. Hence there is k ∈ R such that 23
y). If γ ∈ C(x. ∀ x ∈ M . y) + Φk (y.e. but items 2. mane’s critical value. 24
. (a) For k < c(L). 4 imply that dk (x. y) ∈ R ∪ {−∞}. y ∈ M . Φk (x. 21. x) ≥ 0 ∀ x. ∀ x.1 Proposition. y. Φk (x. y) + Φk (y. 3. We ﬁrst prove 2 for all k ∈ R. For k > c(L). Proof: 2. the inequality in item 2 makes sense for all k ∈ R. z). (b) For k ≥ c(L). The action potential Φk is not symmetric in general. y) = −∞ for all x. For k ≥ c(L) the action potential Φk is Lipschitz. Since k → AL+k (γ) is increasing for any γ.24
˜´ 2. η ∈ C(y. x) = 0. we have that c(L) = inf{ k ∈ R  AL+k (γ) ≥ 0 ∀ closed curve γ }.
L + k ≥ 0. Φk (x. y) + Φk (y. we obtain 2. y) = Φk (x.
21. perhaps dc (x. For k ≥ c(L). 2. Thus c(L) < +∞. 4. y ∈ M . y) and η ∈ C(y.2 Remark. y) + Φk (y. Φk (x. Φk (x.
5. y) = 0 for some x = y and c = c(L)]. y ∈ M . 1. Φk (x. z) and hence Φk (x. y) ∈ R for all x. Taking the inﬁma on γ ∈ C(x. z). z). then γ ∗ η ∈ C(x. z) ≤ AL+k (γ ∗ η) ≤ AL+k (γ) + AL+k (η). x) is a metric for k > c(L) and a pseudometric for k = c(L) [i. Since Φk (x. z) ≤ Φk (x. z ∈ M . 3. x) > 0 if x = y.
Since the function k → Φk (x. z) + Φk (z. 5.ε] + AL+k γ(t − ε)[0. This proves item 1(b). 3. But the deﬁnition of c(L) and the monotonicity of k → Φk (x. Thus there is γ ∈ C(x. y) is increasing. z) ≤ lim AL+k (γ ∗ · · · ∗ γ) = lim N AL+k (γ) = −∞.
N →∞ N N
For x. y ∈ M . x) ≥ 0 for all k ≥ c(L).
25
1. y) = −∞ actually implies that k < c(L). (a) If γ is a closed curve with AL+k (γ) < 0 and γ(0) = z. x2 ∈ M we have that Φk (x1 . Then Φk (x. y) + Φk (y. x) with AL+k (γ) < 0. v) + k ≤ Q for v ≤ 2. item 2 implies that Φk (x. z) + Φk (z. Let k ∈ R by the boundedness condition there exists Q > 0 be such that L(x. Letting ε → 0 we get that Φk (x. (2. the action potential and the critical value.ε] ≤ 2 Q ε. (b) Conversely. then Φk (z. Given x1 . ε] → M be a diﬀerentiable curve with γ ≡ 1 and ˙ γ(0) = x. y ∈ M . x2 ) ≤ AL+k (γ) ≤ Q dM (x1 . x) = −∞. if Φk (x. then Φk (x. y) = −∞ for some k ∈ R and x.21.1) Now let γ : [0. x) ≤ Φk (x. y) = −∞. then item 1(a) follows. x2 ). x) imply that Φk (x. γ(ε)) + Φk (γ(ε). x) ≤ 0. Then k ≤ c(L). Observe that the set { k ∈ R  AL+k (γ) < 0 for some closed curve γ } is open. y) ≤ Φk (x. 25
. x) ≤ AL+k γ[0. Hence Φk (x. x) ≤ Φk (x. Let k ≥ c(L).
y2 ∈ M . Now let ηn : [0. γn ∈ C(x. y1 ) ≤ Q [dM (x1 . v) ∈ T M . x = y and dk (x. mane’s critical value. v) ≥ A v−B. Tn ] → M . d(x1 . y) + Φk (y. Let A > 0.1). ∀ (x. y) be such that Φk (x. x) ≤ lim AL+c (γn ∗ ηn )
n n
≤ lim AL+k (γn ) + (c − k)T + AL+k (ηn ) + (c − k)S ≤ lim Φk (x. Φc (x. γn ) + k ˙ γ + (k − B) Tn ˙
≥ lim A
n
= A dM (x. y) = lim
n 0
L(γn . Now suppose that k > c(L). from the superlinearity there is B > 0 such that L(x. Then
TN
Φk (x. Indeed.26
˜´ 2.
where γ : [0. y) = limn AL+k (γn ). ηn ∈ C(y. x2 )] → N is a unit speed minimizing geodesic joining x1 to x2 and Q > 0 is from (2. x) + (c − k)(T + S)
n
≤ (c − k)(T + S) < 0. y2 )]. y) = +∞ which is false. We claim that Tn is bounded below. If y1 . 4. x2 ) + dM (y1 . Let γn : [0. y) = 0. Sn ] → M . y) Letting A → +∞ we get that Φk (x. Changing the roles of (x1 . suppose that limn Tn = 0. x). Then for c = c(L) < k. which contradicts item 3. y1 ) and (x2 . y2 ) ≤ Φk (x1 . x2 ) + Φk (y2 . x) with limn AL+k (ηn ) = Φk (y. The ﬁrst part of item 4 follows from items 2 and 3. y1 ) − Φk (x2 .
26
. y2 ) we get item 5. then the triangle inequality implies that Φk (x1 . Choose 0 < T < lim inf n Tn and 0 < S < lim inf n Sn .
which yields a contradiction to the deﬁnition of the critical value c. Take ε such that c < c + ε < lim supn cn . Fix ε > 0. R) ψ → c(L + ψ) is continuous in the topology induced by the supremum norm.2) Since ψn → ψ.2). Suppose that c < lim supn cn . Since ε was arbitrary we have that c ≤ lim inf n cn . there exists m ≥ n0 such that c < c + ε < cm .22.
22.1 Lemma.3)
. T ] → M such that AL+ψ+c−ε (γ) < 0. We show now that lim supn cn ≤ c.3) we have AL+ψ+c (γ) ≤ AL+ψm +c+ε (γ) < 0. By the deﬁnition of critical value there exists a closed curve γ : [0. We will prove that cn → c. By (2. (2. and thus c − ε ≤ lim inf n cn . continuity of the critical value. 27 (2. hence for all n suﬃciently large AL+ψn +c−ε (γ) < 0. T ] → M such that AL+ψm +c+ε (γ) < 0. This proof also shows that L → c(L) is continuous if we endow the set of lagrangians L with the topology induced by the supremum norm on compact subsets of T M . Therefore for n suﬃciently large c − ε < cn . −ε ≤ ψ − ψn ≤ ε. and hence using (2. Proof: Suppose that ψn → ψ and let cn := c(L + ψn ) and c := c(L + ψ). there exists n0 such that for all n ≥ n0 . by the deﬁnition of critical value there exists a closed curve γ : [0. The function C ∞ (M.
27
22
Continuity of the critical value. Since c − ε < c.
1 Exercises: 1.4)
for all f ∈ C 0 . f (x.
Let C 0 be the set of continuous functions f : T M → R having linear growth. (x. i. Then M is naturally embedded in (C 0 ) and its topology coincides with that induced by the weak* topology on (C 0 ) . v).
28
.
23.
TM
endowed with the topology such that limn µn = µ if and only if lim
n
f dµn =
f dµ
(2. mane’s critical value. ·) on M by d(µ1 . µ2 ) = v dµ1 − v dµ2 +
n
1 1 2n cn
fn dµ1 −
fn dµ2 (2.e. v) f := sup < +∞.
23
Holonomic measures.v)∈T M 1 + v Let M be the set of Borel probabilities µ on T M such that v dµ < +∞.28
˜´ 2. Construct µ ∈ M such that v2 dµ = +∞. Show that the ﬁrst term in (2. We shall see that this topology is metrizable. Let (C 0 ) the dual of C 0 . Deﬁne a metric d(·. Let {fn } be a sequence of functions with compact support on C 0 which is dense on C 0 in the topology of uniform convergence on compact sets of T M .
2.5) is necessary.5)
where cn = sup(x.v) fn (x.
v dµ.
∀ m. n ≥ N . Then f dµn −→
Km Km
f dµ. The metric d(·. 4 T M −Km(ε) 29
. v dµ < ε.
29
23.6)
Given ε > 0.
(2. µ) → 0.
v dµn −→ This implies that
n→∞ T M −K m
lim
v dµn =
T M −Km
v dµ. choose m(ε) > 0 such that ε (1 + v) dµ < . then for n > N we have that
M 1 2m
d(µn .
1 m≥M 2m
Given ε > 0. µ) → 0. ·) generates the weak* topology on M . Proof: We prove that d(·.
for 0 ≤ m ≤ M. Now suppose that d(µn . Let Km be compact sets such that Km ⊂ Km+1 and that T M = ∪Km .
= 1. ·) induces the weak* topology on M ⊂ (C 0 ) . Suppose that f dµn → f dµ. for n ≥ N.
∀m. holonomic measures. and choose
fm dµ < ε.2 Proposition.23.
Thus d(µn . µ) ≤ ε +
m=1
·ε+
m≥M +1
1 2m
·2·
fm cm
= 3 ε.
∀f ∈ C 0 . choose M > 0 such that N > 0 such that fm dµn − v dµn − Since
fn ∞ cn
· 2 < ε. ∀f ∈ C 0 .
30
˜´ 2.
30
.
If γ : [0. Observe that the set C(M ) is convex.
f dµ < ε.
∀ n > N.
Using a similar estimate for µ we obtain that f dµn − f dµ ≤ ε + f
ε ε ( 2 + 4 ). We call C(M ) the set of holonomic measures on M . mane’s critical value. 2 T M −Km(ε) Fix f ∈ C 0 . Choose N > 0 such that f dµn −
Km(ε) Km(ε)
∀n > N. Let C(M ) be the set of such µγ ’s and let ˙ C(M ) be its closure in M . let µγ ∈ M be deﬁned by 1 T f dµγ = f γ(t). γ(t) dt ˙ T 0 for all f ∈ C 0 . Observe that µγ ∈ M because if γ is absolutely continuous then γ(t) dt < +∞.
Then f  dµn ≤ f
T M −Km(ε)
(1 + v) dµn ≤ f
T M −Km(ε)
ε .
and N such that ε (1 + v) dµn < . 2
∀ n > N. T ] → M is a closed absolutely continuous curve.
The inclusion M(L) ⊆ C(M ) follows from Birkhoﬀ’s ergodic theorem and the fact that C(M ) is convex.
31
24
Invariance of minimizing measures.
Given a Borel probability measure µ in T M deﬁne its action by AL (µ) =
TM
L dµ.4) we see that M is closed. then µ ∈ M(L). Let k > 0 and deﬁne Lk := min{L. 24.1 Theorem (Ma˜´ [38].24. it is enough to prove that the set A(a) := µ ∈ M  AL (µ) ≤ a
is compact in M .1).
Since by the superlinearity the lagrangian L is bounded below. 3.1. so that C(M ) ⊆ M . Proof of item 24. M(L) ⊆ C(M ) ⊆ M . ne 1.3: Since C(M ) is closed. then the set { µ ∈ C(M )  AL (µ) ≤ a } is compact. If M is compact and a ∈ R. If µ ∈ C(M ) satisﬁes AL (µ) = min{ AL (ν)  ν ∈ C(M ) }. 1. 1. Taking f = v in equation (2. Observe that item 3 implies the existence of a minimizer as in item 2. k}. Let Bk := µ∈M  31 Lk dµ ≤ a . 1. prop. Let M(L) be the set of ϕt invariant probabilities on T M . this action is well deﬁned. 2. Observe that L ∈ C 0 and that AL (µ) = +∞ for / some µ ∈ M (see exercise 23.3. invariance of minimizing measures. First we prove that A(a) is closed.1.
.2).
Then v dµ ≤ AL (µ) + B ≤ a + B < +∞. by the monotone convergence theorem.8)
Let B > 0 be such that v < L(x. v) ∈ T M .
v>r
32
.
Similarly. Since A(a) = ∩k>0 Bk .1. then A(a) is closed. taking a subsequence we can assume that there exists a measure µ on the Borel σalgebra of T M such that fi dµn −→ fi dµ. ·). Approximating Lk by functions fi we have that Lk dµ = lim
n
Lk dµn ≤ lim inf
n
L dµn ≤ a. we get that AL (µ) ≤ a. Let ε > 0.7)
for every fi in the sequence used for the deﬁnition of d(·.32
˜´ 2. (2. then Bk is closed in M . Then v dµn ≤
v>r
ε a
L dµn ≤
v>r
ε a
L dµn ≤ ε. In order to prove the compactness. We now prove that limn v dµn −→ v dµ. v dµ ≤ ε. Applying the Riesz’ theorem B.9)
So that µ ∈ M .8). (2. consider a sequence {µn } ⊂ A(a). By adding a constant we may assume that L > 0. v) + B for all (x. Approximating the function 1 by the functions fi we see that µ is a probability.
Letting k ↑ +∞. mane’s critical value. (2.
Since Lk ∈ C 0 . v) > a ε−1 v for all v > r. by (2. Choose r > 0 such that L(x.
1.7) we obtain that there is N > 0 such that v dµ −
v≤r v≤r
v dµn < ε.24.1.3 Remark.
The proof of item 24.
33
From (2.2). It is easy to see that the function AL : C(M ) → R is always lower semicontinuous (see the last argument of the proof of 24. Given µ ∈ C(M ). It is possible to give a sequence µγn ∈ C(M ) such that µγn → µ in C(M ) but lim inf n AL (µγn ) > AL (µ) for a quadratic lagrangian L.
for n > N. 24.2 Proposition.2 is proved at the end of the section. there are µηn ∈ C(M ) such that µηn → µ and lim
n
L dµηn =
L dµ.
24. invariance of minimizing measures. Item 24. The statement of proposition 24. This can be made by calibrating the high speeds in γn so that [v>R] v dµγn → 0 but a := lim inf n [v>R] L dµγn > 0. Then the limit measure µ will have support on [v ≤ R] and “will not see” the remnant a of the action.
33
.2 requires some preliminary results which we present now.
Adding these inequalities we get that v dµn − v dµ ≤ 3 ε.2 is not trivial. The following proposition is needed to show that the minimum of the action in C(M ) is the same as the minimum on C(M ). but in general it is not continuous.
so that s (t) = v(t) = w(s(t)) 1
v(t) A
when v(t) ≤ A. S] → M such that ˙ η = γ when γ < A and η = γ A when γ > A. A
S(T ) =
[v(t)≤A]
dt +
[v(t)≥A]
S(T ) = µγ ([v ≤ A]) + T S(T ) −1 ≤ T ≤2
[v>A]
[v≥A]
dµγ +
[v>A] [v≥A]
v dµγ A (2. A v dµγ .10)
v dµγ . We reparametrize γ to a curve η : [0. γ(t(s)) A) ds = γ(t(s)) ˙
[γ(t)>A] ˙
˙ f γ(t). A
34
. Since v(t) > A then
ds dt
=
v(t) A
when
[γ(t(s))>A] ˙
˙ f (η(s).
Proof: Let A > 1 and let γ : [0. So that η ≤ A. w(s) = η(s) and v(t) = γ(t). mane’s critical value. T ] → M be a closed absolutely continuous curve.
Suppose that f : T M → R is µη integrable. Write ˙ ˙ ˙ ˙ ˙ ˙ γ ˙ η(s(t)) = γ(t). We want ˙ ˙
s(t) t
w(s) ds =
0 0
v(t) dt. when v(t) ≥ A.34
˜´ 2.
Then v(t) dt. γ(t) A γ(t) ˙
γ(t) ˙ dt.
(2. γ(t)) dt + ˙
[γ≤A] ˙ ˙ f η(s). (2.
35
Then f dµη = = 1 S(T ) T S(T ) 1 S(T ) f (η(s). fA dµγ .11)
[v>A]
Deﬁne fA (v) := Then f dµη =
T S(T )
f (v) v f v A
v A
if v ≤ A.
(2. if v > A. η(s)) ds ˙ f (γ(t).14)
∞
f  dµγ +
[v>A]
f − fA  dµγ
≤1+
[v>A]
v dµγ ≤ 1 + ε.11).24. A 35
. we have that
T S(T )
− 1 ≤ 4 ε. γ(t(s)) A ds γ(t(s)) ˙
[γ(t(s))>A] ˙
=
f (v) dµγ (v) +
[v≤A] [v>A]
f
v v A
v dµγ A
For A > 1 big enough. then fA  dµγ ≤
[v≤A]
fA dµγ ≤
T S(T )
−1
fA  dµγ ≤ 4 ε
fA  dµγ .10) and (2.12)
Observe that from (2.
(2.
1 v dµγ < ε < 4 .13)
Then f dµη − If f ≤ 1. invariance of minimizing measures.
f dµη −
f dµγ = ≤
T S(T ) T S(T )
fA dµγ − −1
f dµγ f − fA  dµγ
fA  dµγ +
≤ 4 ε (1 + ε) + ε ≤ 6 ε. then LR (v) ≤ L(v) for all v ∈ T M. We have f (s) = v · Lv (sv) − L(v) f (s) = v · Lvv (sv) · v > 0. using (2.
Proof: If v ≤ R then LR (v) = L(v). deﬁne LR (v) := L(v) v L v R
v R
v dµγ =
T S(T )
−1
v dµγ ≤ 4 ε
v dµγ .
Claim: If E(v) > 0 for all v ≥ R.16)
if v ≤ R. Also. Suppose that v = R.14).36
˜´ 2. 36
. v dµη − Hence dM (µη . For s ≥ 1 let f (s) := L(sv) − LR (sv) = L(sv) − sL(v). It is enough to prove that f (s) ≥ 0 for all s ≥ 1.
(2. mane’s critical value. if v > R.15)
(v + 1) dµ. µγ ) ≤ 6 ε Now let µ ∈ C(M ).
(2.
(v + 1) dµγ . Let K := 1 + For R > 0.
37
We have that f (1) = 0.
(2.
37
. Then from (2. µ) < LR(N ) dµγN ≤ and v dµγN ≥
[v≤R(N )] [v≤R(N )] 1 N. f (s) > 0 for all s ≥ 1.
(2. from (2. choose R = R(N ) > 1 such that E(v) > 0 if v ≥ R and
v>R
v dµ <
1 N. (2.21)
Construct ηN as above for γN and A = R(N ). µ) <
N 18 N
K+
1 N.18).21). Moreover. From (2.17)
Observe LR(N ) has linear growth.19)
Then (v + 1) dµγN ≤ K v dµγN <
[v>R(N )] 3 N
from (2. from (2.12). dM (µηN .16) and (2. (2. ♦ Given N > 0.20) and (2. (2.11). invariance of minimizing measures.
LR(N ) dµ +
v dµ −
1 N. L dµηN =
TN S(TN ) 12 N
LR(N ) dµγN ≤ L dµ +
1 N
TN S(TN )
LR(N ) dµ +
1 N
≤ 1+
.18) and (2.20) (2.19).18)
1 N.
Thus µηN −→ µ in C(M ).13). µγN ) < 18 K. f (1) = E(v) > 0. (2. Choose µγN ∈ C(M ) such that dM (µγN . (2.17). This implies that f (s) ≥ 0 for all s ≥ 1.18). N dM (µηN .15).24.11) and the claim.
(2.
(2.
4 Proposition. γ2 ) := sup d γ1 (t). γ2 ) is deﬁned by
T
d1 (γ1 . T ]. Observe that a Tonelli minimizer is a solution of (EL). mane’s critical value. Given a compact subset K ⊆ M and given C. M ). Therefore lim inf
N
L dµηN ≥ lim
N
LR dµηN =
LR dµ. deﬁne S + (γ) := AL (γ) − S(γ(0). γ2 (t)] dt. the absolutely continuous distance d1 (γ1 .38
˜´ 2. T ]. T ).
Given x. ε > 0 there exist δ > 0 such that if γ : [0. Then LR has linear growth and by the claim LR ≤ L. by the dominated convergence theorem we get that lim inf
N
L dµηN ≥
L dµ. T ) > −∞ because L is bounded below. deﬁne S(x. y. γ(T ). T ) := inf
γ∈CT (x. T ] → M is absolutely continuous and satisﬁes 38
.
Letting R ↑ +∞. γ2 ∈ C ac ([0.
Fix R > 0 such that E > 0 on v > R. [γ2 (t). If γ ∈ C ac ([0. y ∈ M . γ2 (t) +
t∈[0. γ1 (t)].y)
AL (γ). The absolutely continuous curves γ with S + (γ) = 0 are called Tonelli minimizers.
Hence lim sup
N
L dµηN ≤
L dµ. ˙ ˙
24.
Observe that S(x. Given γ1 . M ). y.T ] 0
dT M [γ1 (t).
S + (γ) ≤ δ. γn → γ0 in the d1 topology. Then we can assume that there is γ0 ∈ CT (x. Tn ]. We can assume that Tn → T . Since the action functional is lower semicontinuous. y) such that γn → γ0 in the C 0 topology.
Then either γ([0. γn (t)) ≤ d(γn (s0 ). Then we have that γn ([0. we can assume that Tn ≡ T . T ]) ∩ K = or there exists a Tonelli minimizer γ0 : [0.
1 ≤ T ≤ C. By proposition 31.3. M ) such that γn ([0. because S + (γn ) → 0. AL (γn ) ≤ C and d1 (γn . Moreover. ˙
Let Q := { y ∈ M  d(y.2.24. invariance of minimizing measures. γn (0) → x ∈ Q and γn (Tn ) → y ∈ Q. Tn ] such that γn (s0 ) ∈ K. γn (t)) ≤
[s0 . η) ≥ ε for any Tonelli minimizer η. Choose s0 ∈ [0. Tn ]) ∩ K = . iii. γn (0) ≡ x and γn (T ) ≡ y.
Let H := h : TM → R f
∞
≤ 1. h with compact support . then AL (γ0 ) ≤ lim inf n AL (γn ) = S(x. γn ) + B ≤ C + B C. y. ii. Tn ]) ⊆ Q. we have that AL (γn ) → AL (γ0 ). Proof: If such δ does not exists then there is a sequence γn ∈ C ac ([0.
39
i. v) > v − B for all (x. S + (γn ) → 0. 39
. T ] → M such that d1 (γ0 . the set A(b) = γ ∈ CT (x. y)  AL (γ) ≤ b is compact in the C 0 topology. K) ≤ C+B C }. Let B > 0 be such that L(x. [h]Lip ≤ 1. 1 ≤ Tn ≤ C. T ). Adding a constant we can assume that L > 0. Moreover. Then d(K. Thus γ0 is a Tonelli minimizer. v) ∈ T M .t]
γn  ˙
≤
[s0 . By theorem 31.t]
L(γn . γ) ≤ ε. AL (γ) ≤ C.
˙
40
.
where [h]Lip =
h(x.4. γ) ≡ 0 and h◦ϕ1 (γ. T ])∩K = . γ0 ) ≤ 1 · 1 · ε. w) (x.4. 24. This implies (2.4. Given C > 0 and ε > 0 let δ = δ(C.4.i. ϕ−1 (supp(h)). v).iii we have that either γ([0. (y.4. T ] → M satisﬁes conditions 24. T ] → M satisﬁes conditions 24.22).w) dT M (x.4 then if γ : [0.v)=(y. γ0 ) dt ≤ 2. T ]) ∩ K = . v) − h(y. then h(γ. We have that h−
γ γ0
h ≤ [h]Lip d1 (γ.i.ii. ε) > 0 and A > 0 be given by proposition 24. 24.ii.40
˜´ 2. γ) ≤ ε.iii then h−
γ γ
h ◦ ϕ1 ≤ 5. h) > 0 such that if γ : [0. 24. or we can take γ0 minimizing such that d1 (γ0 . γ0 ) ≤ 1 · [ϕ1 Q(h) ]Lip · ε. Let Q(h) :=
where [h]Lip is the smallest Lipschitz constant of h. 24. Suppose then that d1 (γ0 . then h ◦ ϕ1 −
γ γ0
h ◦ ϕ1 ≤ [h]Lip [ϕ1 Q(h) ]Lip d1 (γ. Given h ∈ H and C > 0 there exist δ = δ(C. γ) ≤ ε. Observe that if γ([0.5 Corollary. γ0 (t) − h γ0 (t + 1). γ) ≡ ˙ ˙ 0.22)
Proof: Let K = π supp(h) ∪ ϕ−1 (supp(h)) . we have that
T
h−
γ0 γ0
h ◦ ϕ1 =
0 1
h γ0 (t). γ0 ) dt + ˙
T
h(γ0 . γ0 (t + 1) dt ˙ ˙
T +1
≤
0
h(γ0 .4.
(2. w) sup
is the smallest Lipschitz constant for h.
Since γ0 is a solution of (EL). 24. mane’s critical value.
Proof of item 24.
AL (γn.
Thus
#Bn (C) 2k ≤ .j ) ≥ C #Bn (C). invariance of minimizing measures.j := γn [jN.
AL (γn.
41
Hence h−
γ γ
h ◦ ϕ1 ≤ ε 1 + [ϕ1 Q(h) ]Lip + 2. (2. we can assume that every Tn is a multiple of N and that limn→∞ Tn = +∞. By joining a constant curve if necessary. γn ) dt ≤ 2 k ˙
0
∀n.2. Take an integer N > 0.1.24.25)
. adding a constant we can assume that L > 0.(j+1)N ] . there exists a sequence µγn ∈ C(M ) such that µγn → µ and lim AL (µγn ) = AL (µ) = min AL (ν)  ν ∈ C(M )
n
=: k.2: Observe that to prove that µ is invariant it is enough to prove that h dµ = h d(ϕ∗ µ) 1 for all h ∈ H. Then we can assume that AL (µγn ) = Hence 2 k Tn ≥
j∈Bn (C)
j ∈ N1 ≤ j ≤
Tn N . Given C > 0 let Bn (C) := where γn.24)
Let Tn be a period of the curve γn : R → M .
(2. By the superlinearity L is bounded below.
1 Tn
Tn
L(γn . Tn C 41
(2.j ) ≥ C .23)
By proposition 24.
j γn.
. Since sup h ≤ 1. h) > 0 from corollary 24. then h dµγn − h d(ϕ∗ µγn ) ≤ 1 1 Tn h−
j ∈Bn / γn.j ) ≥ δ #Bn (δ). Then h dµγn − h d(ϕ∗ µγn ) 1 1 ≤ Tn
(Tn /N )−1
1 + 1 S(γn (0).26) and corollary 24. Tn Tn
(2. mane’s critical value. Tn δ Now ﬁx h ∈ H. Tn ) = AL (µγn ) − S (γn ).j
1 2N #Bn .26)
h−
j=0 γn.25).42
˜´ 2. Using equations (2.j
h ◦ ϕ1 .j
h ◦ ϕ1 +
γn.5 we obtain that h dµγn − h d(ϕ∗ µγn ) ≤ 1 1 Tn 5 − #Bn + 2N #Bn Tn N Tn 5 2k 1 ≤ + 2N + AL (µγn ) − k N C δ 42
. (2. let Bn (δ) := Then
(Tn /N )−1
j ∈ N1 ≤ j ≤
Tn N
− 1.
Given δ > 0. γ(Tn ).j ) > δ . k≤ Hence S + (γn ) ≤ Tn AL (µγn ) − k .
Moreover. Therefore #Bn (δ) 1 ≤ AL (µγn ) − k .
S + (γn ) ≥
j=1
S + (γn. Tn
Now choose C ≥ N 2 and δ = δ(C. S + (γn.
Denote Bn := Bn (C) ∪ Bn (δ).5.
N N
Since N is arbitrary. h ◦ ϕ1 ∈ C 0 (because they have compact support).
43
. we obtain that h dµ − h d(ϕ∗ µ) ≤ 1 5 4k + . this diﬀerence is zero and we get (2.24) and that C ≥ N 2 .
43
Now let n → ∞. Using equation (2.23).24. µγn → µ and h. invariance of minimizing measures.
3 Deﬁnition. 25. this action is well deﬁned. T ] → M is a closed absolutely continuous curve. γ(t) dt ˙
0
for all f ∈ C 0 .
Since by the superlinearity the lagrangian L is bounded below.2 Theorem. theorem 25.2 below. Recall that if γ : [0. then ne c(L) = − min{ AL (µ)  µ ∈ M(L) }. mane’s critical value. If M is compact.27)
25. the measure µγ ∈ M is deﬁned by f dµγ = 1 T
T
f γ(t).1 Theorem (Ma˜´ [35]).44
˜´ 2. We will obtain theorem 25. 25. If M is noncompact. We say that a holonomic measure µ ∈ C(M ) is (globally) minimizing if AL (µ) = −c(L). Let M(L) be the set of ϕt invariant probabilities on T M . which also applies to the noncompact case. and that C(M ) is the closure of the set of such µγ ’s in M.1 may not hold. as seen in example 57. (2.1 from theorem 25. 44
.
Given a Borel probability measure µ in T M deﬁne its action by AL (µ) =
TM
L dµ.
25
Ergodic characterization of the critical value. c(L) = − inf{ AL (µ)  µ ∈ C(M ) } = − inf{ AL (µ)  µ ∈ C(M ) }.
2 is nontrivial and follows from proposition 24.1 imply theorem 25. 1.
45
25.1(3) allows to obtain a minimum on (2. while the minimizing measures for M have action cu < c0 .25. 5.1 imply that c(L) = − min{ AL (µ)  µ ∈ M(L) ∩ dp∗ C(N ) }. then theorems 25. 4.1. Theorems 25. Here dp∗ C(N ) is the set of probabilities µγ on T M where γ is a curve on M whose lifts to N are closed.28) allows to obtain minimizing measures which don’t appear in the Mather’s theory.2.27) which may not be attained in the noncompact case. The drawback of this approach is that we obtain honest minimizing invariant measures on T M which may not lift to ﬁnite measures on the covering T N . if a minimizing measure exists.2 and 24. If p : N → M is a covering. The statement for coverings in equation (2.4 Remarks.3).
45
. 3.2 and 24. For example if cu is the critical value of the universal cover M of M and c0 is the critical value of the abelian cover M of M . Recall that by theorem 24.1. The compactness property on theorem 24. (2.31) and proposition 27.28) instead of the inﬁmum on (2. the minimizing measures on a ﬁxed homology class (corresponding to Mather’s theory) all have action A(µ) ≥ −c0 . 2. M is compact and L = L ◦ dp is the lifted lagrangian.28)
by noticing that AL (dp∗ ν) = AL (ν) for ν ∈ C(N ). The measures for M correspond to “minimizing in the zero homotopy class” while the measures for M are minimizing in the zero homology class. ergodic characterization of the critical value. then it is invariant under the lagrangian ﬂow. (see equation (2. The equality between the two inﬁma in theorem 25.
Thus −c(L) ≤ inf{ AL (µ)  µ ∈ C(M ) } = inf{ AL (µ)  µ ∈ C(M ) }. Now let k ↑ c(L).2: If µγ ∈ C(M ).2.
Proof of theorem 25. Hence AL (µγ ) ≥ −c(L). then AL+c(L) (µγ ) ≥ 0. mane’s critical value.46
˜´ 2. If k < c(L) then there is a closed absolutely continuous curve γ on M such that AL+k (γ) < 0. Thus µγ ∈ C(M ) and −k > AL (µγ ) ≥ inf{ AL (µ)  µ ∈ C(M )}.
46
. where the last equality follows from proposition 24.
29) depends only on the cohomology class of ω.a
Homology of measures. R)∗ .29) shows how the homology class ρ(µ) acts on H 1 (M. we have that 26. . H 1 (M. R) is continuous. .
Given a diﬀerentiable ﬂow ϕt on a compact manifold N and a ϕt invariant probability µ. Here we have used the identiﬁcation1 H1 (M. where [ω] ∈ H 1 (M.
Observe that since M is compact.b
The asymptotic cycle. R).
1
47
. an holonomic probability µ ∈ C(M ) satisﬁes T M v dµ < +∞ and df dµ = 0
TM
for all f ∈ C ∞ (M.26.
(2.
26. the Schwartzman’s [65] asymptotic cycle of an
` ´ In fact. Using a ﬁnite basis { [ω1 ]. R) is a ﬁnite dimensional vector space and hence it is naturally isomorphic to its double dual H 1 (M. the integral in (2. . Since µ is holonomic. [ω] =
TM
ω dµ . R) ≈ H 1 (M. R). The map ρ : C(M ) → H1 (M. R) is the cohomology class of ω. R). [ωk ] } for H 1 (M. . R) and the topology of C(M ). R) ≈ hom H1 (M. R)∗ by the universal coeﬃcient theorem. R = H1 (M. The class ρ(µ) is called the homology of µ or the rotation of µ by analogy to the twist map theory. By deﬁnition. Since M is compact. the aubrymather theory.
26. R) ≈ H 1 (M.
47
26
The AubryMather Theory. then H1 (M.1 Lemma.
Through this section we shall assume that M is compact. R)∗ and equation (2. R)∗ by ρ(µ) .
Then we can deﬁne its homology class as ρ(µ) ∈ H1 (M.29)
for any closed 1form ω on M . any 1form is in C 0 .
R)∗ such that A(µ) . This integral depends only on the cohomology class of ω because the integral of a coboundary by an invariant measure is zero: in fact. R) ≈ H 1 (N.
lim 1 T →+∞ T
RT
0
f (ϕt x) dt =
R
f dµ for all f ∈ C 0 (N. In the case of a lagrangian ﬂow. the curve δT is a unit speed geodesic from ϕT (x) to x. T ].
0
Applying this to a basis {ω1 .
for any closed 1form ω.
invariant probability µ is deﬁned to be the homology class A(µ) ∈ H1 (N. . but it has the same homotopy type as the conﬁguration space M because M is a deformation retract of T M (contracting T M along the
2
i. [ω] = lim 1 T →+∞ T
T
ω X(ϕt x) dt . ωk } for H 1 (N. then
If µ is ergodic and x ∈ N is a generic point A(µ) . where [ω] ∈ H 1 (N. . R).
where γT (t) = ϕt (x). if df is an exact 1form. R). [ω] =
N
ω(X) dµ .e.
by Birkhoﬀ’s theorem. R) is the cohomology class of ω and X is the vector ﬁeld of ϕt . . the phase space N = T M is not compact. then deﬁne F (y) := lim 1 T →+∞ T
T
df X(ϕt y) dt = lim
0
1 T →+∞ T
f (ϕt y) − f (y) = 0.48
˜´ 2. we get that A(µ) = lim
1 T →+∞ T
[ γT ∗ δT ] ∈ H1 (N.
48
. R). R). . mane’s critical value. t ∈ [0. and the limit is on the ﬁnite dimensional vector space H1 (N. df (X) dµ =
N N 2
F dµ = 0. for µ.
which is a compact submanifold of T M by remark 13. R) is the map induced by the projection π TM → M. the ergodic components of an invariant measure of a lagrangian ﬂow are contained in a unique energy level. Then π∗ A(µ) .3 Lemma. the aubrymather theory. We see that the homology of an invariant probability and its asympπ∗ totic cycle coincide under the identiﬁcation H1 (T M. R) is surjective. v) = (v. γ is a periodic orbit for the lagrangian ﬂow with period 1. Let γ be a minimizer of the action of L among the set of absolutely continuous curves [0. Proof: Let h ∈ H1 (M. 1] → M be a closed curve with homology class h.2 Proposition. R) → H1 (M.
where π∗ : H1 (T M. then (π ∗ ω)(X(x. Z).
because the lagrangian vector ﬁeld X has the form X(x. Proof: If ω is a closed 1form on M .2.1. The map ρ is aﬃne and M(L) is convex. The map ρ : M(L) → H1 (M. Moreover. 49
. R) ≈ H1 (M. π ∗ [ω] =
TM
(π ∗ ω)(X) dµ
=
TM
ω dµ = ρ(µ) .26. Then by remark 12. The invariant measure µγ satisﬁes ρ(µγ ) = h. R). 1] → M with the same homotopy class as η.
49
ﬁbers to the zero section M × 0). Z) be an integer homology class. R) ⊆ ρ M(L) . in particular. v) = ωx (v) . π∗ A(µ) = ρ(µ) for all µ ∈ M(L). Let η : [0. H1 (M.
26. Thus. hence ρ M(L) is convex and. [ω] . v)) = ω dπ X(x. it contains the convex hull of H1 (M. 26. [ω] = A(µ) . ∗).
Hence we can deﬁne the Mather’s beta function β : H1 (M.R) µ∈M(L) µ∈M(L)
[ω] . α is superlinear. h2 ∈ H1 (M.
= − min
= − min AL−ω (µ) = c(L − ω) . mane’s critical value.3. ρ is surjective. 2. µ2 ∈ M(L) be such that ρ(µi ) = hi and AL (µi ) = β(hi ) for i = 1.
The action functional AL : M(L) → R is lower semicontinuous3 and the sets M(h) := { µ ∈ M(L)  ρ(µ) = h } are closed. Hence β λ h1 +(1 − λ) h2 ≤ AL λ µ1 +(1 − λ) µ2 = λ β(h1 ) + (1 − λ) β(h2 ).30)
Observe that since L − ω is also a convex superlinear lagrangian.50
˜´ 2. ρ(µ) using 26. By proposition 26. h − β(h) AL (µ) − [ω] . The α and β functions are convex and superlinear.1. R) → R.c
The alpha and beta functions. as β(h) := min AL (µ).
µ∈M(h)
We shall prove below that the βfunction is convex.
50
. then α([ω]) is ﬁnite. (2. Let h1 . by 25. The Mather’s alpha function α = β ∗ : H 1 (M. R) and 0 ≤ λ ≤ 1. α and β are
3
AL is lower semicontinuous iﬀ lim inf n AL (νn ) ≥ AL (µ) when νn → µ.4 Theorem. By D. R) → R is the convex dual of the βfunction: α([ω]) = max
h∈H1 (M. Proof: We ﬁrst prove that β is convex.3. 26.1 on the appendix.
26.1. Let µ1 . and hence β is ﬁnite. The probability ν = λ µ1 + (1 − λ) µ2 satisﬁes ρ(ν) = λ h1 + (1 − λ) h2 . By proposition D.
the Mather set.R)
Mω (L). R) and ω ∈ H 1 (M. if ω ∈ ∂β(h). Thus Mh (L) =
h∈H1 (M. these minimal energy Mather minimizing measures have trivial homology. Formula (2. R).3 the strict critical value is the lowest energy level which supports Mather minimizing measures and since c0 (L) = −β(0).30).
51
convex. implies that α is ﬁnite and then by D. since by corollary D. For h ∈ H1 (M.2 α∗ = β.
51
. the aubrymather theory. then Mω (L) ⊆ Mh (L) if h ∈ ∂α(ω).R)
min
α(ω) (2.R)
We call these measures Mather minimizing measures and the set M := M0 (L) = {µ ∈ M(L)  AL (µ) = −c(L) }.R)
min
c(L − ω) =
ω∈H 1 (M. By corollary 36.1. write Mh (L) := { µ ∈ M(L)  ρ(µ) = h. Mω (L) := { µ ∈ M(L)  AL−ω (µ) = −c(L − ω) }. Conversely. then Mh (L) ⊆ Mω (L). Since the βfunction has a supporting hyperplane at each homology class h.26.31)
= −β(0).
ω∈H 1 (M. Deﬁne the strict critical value as c0 (L) : =
ω∈H 1 (M. β is superlinear. AL (µ) = β(h) } .
Particularly interesting are the abelian cover M . If M1 −→ M is a covering.2 Proposition.
p
Proof: We know that c(L1 ) ≤ c(L2 ). Then we have cu (L) ≤ ca (L) = c0 (L) ≤ c(L − ω) 52 ∀[ω] ∈ H 1 (M. Its deck transformation group is H1 (M.1 Lemma. A closed curve in M projects to a closed curve in M with trivial homology. Hence there exists a closed curve γ in M2 with negative (L2 + k)action.3 Proposition. Suppose that the strict inequality holds and let k be such that c(L1 ) < k < c(L2 ).52
˜´ 2. When π1 (M ) is abelian. If M1 is a ﬁnite covering of M2 then c(L1 ) = c(L2 ). p If M1 −→ M is a covering. 27.
We shall deal mainly with compact manifolds M . but there are some important noncompact cases. M = M . Since M1 is a ﬁnite covering of M2 some iterate of γ lifts to a closed curve in M1 with negative (L1 + k)action which contradicts c(L1 ) < k. Z) and H 1 (M . mane’s critical value. Proof: The lemma follows form the fact that closed curves on N project to closed curves on M .
27
Coverings. denote by L1 := L ◦ dp : T M1 → R the lifted lagrangian to T M1 . Z). Z) = {0}.
. [60] c0 (L) = ca (L) = critical value of the abelian cover. then c(L1 ) ≤ c(L). The abelian cover M of M is the covering whose fundamental group is the kernel of the Hurewicz homomorphism π1 (M ) → H1 (M. the universal cover M and the ﬁnite coverings. R). for example the coverings of M . 27. 27.
ρ(µ) = 0 } because M(L) ⊂ C(M )
ˇ The real abelian cover is the covering M of M with h : π(M ) → H1 (M. Paternain [60] give an example in which cu (L) < ca (L). −ca (L) = inf AL (µ)  µ ∈ C(M ) b by proposition 24. Hence M → M is a ﬁnite cover so that they ˇ the torsion 1 have the same critical value ca (L) = c0 (L). For symmetric lagrangians c(L) = e0 = c0 (L) = cu (L). R) is the Hurewicz homomorphism. Z).
= inf{ AL (µγ )  µγ ∈ C(M ) } b = inf{ AL (µγ )  µγ ∈ C(M ).
53
where cu is the critical value of the lift of the lagrangian to the universal cover. by 24.e. Z). Hence ca (L) ≤ Moreover.2.
because a closed curve γ on M has homology [γ] = 0 if and only if it has a closed lift to M . the method in equation (2. the elements of ﬁnite order Zn1 ⊕ · · · ⊕ Znk ⊆ H1 (M.
53
.
4
i. Paternain and M. = −β(0) = c0 (L). G. Then −ca (L) = inf{ AL (µ)  µ ∈ C(M ). the lift ω of ω to M is exact. R) = {0}. coverings. When cu (L) < c0 (L).2 and 26.R)
min
c(L − ω) = c0 (L). Proof: Let ω be a closed form in M . ρ(µ) = 0 }.28) gives some minimizing measures which are not Mather minimizing. then ca (L) := c(L) = c(L − ω) ≤ c(L − ω). Ma˜´ [39] gives an example in ne which e0 < ca (L) = c0 (L) < c(L).1
≤ min{ AL (µ)  µ ∈ M(L). ρ(µ) = 0 }. It is an intermediate covering ˇ ˇ M → M → M and the deck transformations of M → M are given by 4 of H (M.
ω∈H 1 (M. Since H1 (M.27.
mane’s critical value.
54
.54
˜´ 2.
y). y ∈ M and T > 0 there exists a Tonelli minimizer on CT (x. An independent proof of this corollary is given in remark 31.
Given x. M )  γ(0) = x. y)  AL (γ) ≤ c } 55 (3. γ(T ) = y . The idea of Tonelli’s theorem is to prove that the sets A(c) := γ ∈ CT (x. y) is a Tonelli minimizer if AL (γ) = min
η∈CT (x.1 Tonelli’s Theorem. The only diﬀerence in the proof of this theorem when M is noncompact is corollary 31.9. T ]. For all x.1)
.
31 Tonelli’s theorem.
In this section we shall prove 31. y) := γ ∈ C ac ([0.y)
AL (η) .8.
We say that γ ∈ CT (x. let CT (x. y ∈ M and T > 0.Chapter 3
Globally minimizing orbits.
γ1 (t)]. An addendum to Tonelli’s theorem due to Mather [46] states that these sets are compact in the topology of absolutely continuous curves. y)  AL (γ) ≤ b is compact in the C 0 topology. T ].56
3. converges in the d1 topology to γ. proposition 24. [γ2 (t). In fact. y) a priori it is not possible to ensure the convergence of the action in a C 0 convergent sequence obtained from theorem 31. . the set A(b) in proposition 31. γ2 . is a sequence in ac ([a. for example. γ2 . b ∈ R. if the action is converging to the minimal action (e. .12. (Mather [46]) If N ⊆ M is a compact subset and γ1 . globally minimizing orbits. .
t∈[0. . γ2 (t)] dt. γ2 (t) + dT M [γ1 (t). 31. γ2 ) := sup dM γ1 (t). We quote here the following proposition (addendum on page 175 of Mather [46]). y ∈ M . M ) deﬁne their absolutely continuous distance by
T
˙ ˙ d1 (γ1 . . the set A(b) := γ ∈ CT (x.
c≥α
where α = inf η∈CT (x. Then a Tonelli minimizer will be a curve in A(c) = .g.4). C L i L then γ1 .y) AL (η) ≥ inf L > −∞.T ] 0
31. Given γ1 . Unless. b].
56
. For any x.2 Theorem (Mather [46]). . T > 0. N ) which converges C 0 to γ and A (γ ) converges to A (γ). γ2 ∈ C ac ([0. Since the action functional is only lower semicontinuous on CT (x.2 is not compact in the d1 topology as the following example shows.3 Proposition.2. This theorem is proved in 31.
are compact in the C 0 topology.
57
. xti ) < ε. A(b) is not compact in the d1 topology. 31. t ∈ [0. the family F(c) := γ ∈ C ac ([0. ]sN . γn → η in the C 0 topology. 0) and γn (t) = (t.
whenever ]s1 . n sin(2πnt)). since a reparametrization preserves length. . 31. T ]. . b].
57
31. Moreover. Hence (γn ) ≥ 2n 4/n2 + 1/(4n2 ) > 4. tonelli’s theorem. (i) An absolutely equicontinuous family is equicontinuous. b]. . Let η(t) = (t.31. The length of γn is bounded below by a polygonal curve joining the maxima and minimums of its second component. t1 [. 1]. M )  AL (γ) ≤ c
is absolutely equicontinuous. and thus γn → η in the d1 topology.6 Remark. there is no reparametrization of the γn ’s which converges in the d1 topology to η. A family F ⊆ C 0 ([a.
We shall split the proof of Tonelli’s theorem in several parts: 31.7 Lemma. .4 Example. tN [ are disjoint intervals in [a. Let L = 1 v2 be the riemannian lagrangian on R2 with the ﬂat 2 1 metric. M ) is absolutely equicontinuous if ∀ ε > 0 ∃ δ > 0 such that
N N
ti − si  < δ
i=1
=⇒
i=1
d(xsi . Therefore (γn ) → (η). (ii) A uniform limit of absolutely equicontinuous functions is absolutely continuous.5 Deﬁnition. The action AL (γn ) = 1 + 2π 2 is bounded. For all c ∈ R and T > 0.
58
. B(x. v (3. A(c) ⊆ C ac [0. R) .3)
This implies the absolute equicontinuity of F(c). R) := { z ∈ M  dM (x. For a > 0 let K(a) := inf L(x. y ∈ M . x) + a · K(a) m(J) ˙ (because L ≥ 0).
≤ c + a · K(a) m(J).58
3.2)
The superlinearity implies that lima→+∞ K(a) = +∞. K(a) 2 Let 0 ≤ s1 < t1 ≤ · · · ≤ sN < tN ≤ T . ti ] and E := i=1 J ∩ x > a . then L(xs . J := ∪N [si .8 Corollary. xt ) ≤ K(a)
E
x + K(a) ˙
J\E
x ˙
≤
E
L(x. where m is the Lebesgue measure on [0. by adding a constant we may assume that L ≥ 0. T ]. the lagrangian L is bounded below. globally minimizing orbits. v) ∈ T M. xti ) ≤
i=1
J
xs  ≤ ˙
c K(a)
+ a m(J) ≤
ε 2
+ a m(J). v ≥ a .
In order to apply the ArzelaAscoli theorem we need a compact range. z) ≤ R }. where B(x.
Proof: Since by the superlinearity. Given ε > 0 let a > 0 be such that c ε < . We have that ˙ ˙ ˙
N
K(a)
i=1
d(xs . xs ) ≥ K(a) xs  for s ∈ E. v) (x.
(3. T ]
N
d(xsi . for this we have: 31. For all c ∈ R and T > 0 there is R > 0 such that for all x.
w) ≥ L(x. 31. T ]) is contained in a compact neighbourhood K of γ([0. x) ≤ BT + c. There is B > 0 such that L(x. T ]). v) ∈ T M . Another proof for corollary 31.5 that a uniform limit of absolutely equicontinuous curves is absolutely continuous. ε It is enough to take R = 2 + a T . xt ) ≤
s
x ≤ BT + ˙
s
L(x.4)
Proof of theorem 31. We shall need the following lemma.31. v) (w − v) − ε. v) ≥ v − B for all (x.3) for N = 1 and J = [s.11. M )  AL (γ) ≤ c
31. then by lemma 31.5)
. t] is d(xs .
ε Proof: Inequality (3. Let ε > 0 and E = γ ≤ a . x − y ≤ δ.10 Theorem.9 Remark. then L(y.8 is the only diﬀerence for the proof of Tonelli’s theorem when M is noncompact. ˙ L(γ) + Lv (γ)(γn − γ) − ε ≤ ˙ ˙ ˙ ˙
E E
L(γn ) ≤ c ˙ 59
(since L ≥ 0). a > 0 and ε > 0. By the superlinearity we may assume that L ≥ 0. T ]. If γn ∈ F(c) and γn → γ in the uniform topology.
59
31. xt ) ≤ 2 +a t−s.
Corollary 31. then γ ∈ F(c).10: It is immediate from the deﬁnition 31.8 is the following: Adding a constant we may assume that L ≥ 0. We may assume that γn ([0. tonelli’s theorem. (3.11 Lemma. we have that
t t
d(xs . v ≤ a and w ∈ Rn . Then for 0 ≤ s ≤ t ≤ T . (3. v) + Lv (x. ˙
Recall that F(c) := γ ∈ C ac ([0. We may assume that M = Rn . for n large. Given K compact. there exists δ > 0 such that if x ∈ K.
zn = γn −γ and 1E is the characteristic ˙ function of E. We now prove the claim. Let B = [ψ − f  > ε]. given ε > 0 there is δ > 0 such that m(D) < δ =⇒
D
zn  < ε.6)
where ψ = Lv (γ)·1E is bounded. ˙ ˙ ˙
Letting n → +∞ on (3.7)
Since γ is absolutely continuous. ˙
(3. we have that L(γ) − ε T ≤ c. From γ ∈ L1 and inequality ˙ ˙ (3. We have to prove that lim
n
ψ zn = 0. then m(B) < ε. (ψ − f ) zn ≤ ε ˙
Bc
 zn  + 3 ψ ˙ 60
∞
zn  ≤ h(ε).8) imply that (3. γ ∈ L1 . ˙
Now let ε → 0.7) holds for any Borel set A.3). Let f be a simple function such that f ∞ ≤ 2 ψ ∞ and ψ − f  < ε2 . if A = [a.7) and (3.5).60
3. then
T
L(γ) = lim ˙
0
a→+∞ E
L(γ) ≤ c + ε T. Hence (3. ˙
∀n ∈ N. globally minimizing orbits.
(3.8)
Equations (3. approximating A by a ﬁnite union of intervals. ˙
E
Since E ↑ [0. T ] when a → +∞ and L ≥ 0.
(3. b] is an interval then
b
lim
n A
zn = lim zn ˙
n
a
= 0.
Claim: lim
n E
Lv (γ) (γn − γ) = 0. Since zn ∞ → 0. ˙
B
.6) holds if ψ is a simple function k ak 1Ak .
Then if y ∈ M and w > b.11: Let C1 := sup Lv (x. A(c) is compact.31. v) · (w − v) This gives (3. Then a γ∈
c≥inf CT (x. Proof of Tonelli’s theorem: By lemma 31.7. v ≤ a . Then lim
n
ψ zn ≤ lim ˙
n
f zn + lim sup ˙
n
(ψ − f ) zn ≤ 0 + h(ε) −→ 0. v ≤ a C2 := sup L(x.1) is equicontinuous. Proof of lemma 31. v) + Lv (x.8. L(x. v) · (w − v) ∀ w ∈ Rn .4) holds. ˙
ε
31. the curves in A(c) have a uniform compact range.
Then there is δ > 0 such that for x − y ≤ δ. v) · w ≥ L(x. tonelli’s theorem. w) ≥ K(b) w ≥ C2 + C1 w ≥ C2 + Lv (x.
for w ≥ b. w) ≥ L(x.
61
By (3. v) · v  x ∈ K. v) − Lv (x.
where K(b) is from (3. v) + Lv (x.10. Let b > 0 be such that K(b) · r ≥ C2 + C1 r for all r ≥ b.8) one can take h(ε) with limε h(ε) = 0. By Arzel´Ascoli’s theorem and theorem 31.12. y). Since L is convex.y) AL
A(c)
is a Tonelli’s minimizer on CT (x. the family A(c) in (3. v ≤ a and w ≤ b inequality (3.
61
. v)  x ∈ K.4) when w ≥ b. L(y. and by corollary 31.2).
R) + k] T + 1.6) imply that there is A = A(C) > 0 such that γ ≤ A.
(ii) If γ ∈ CT (x.y)
AL+k (γ). then γ ≤ A(ε. R) + k] T .1 Lemma. v) ∈ T M .
The following lemma. ˙
Proof: By the superlinearity there is D > 0 such that L(x. v) ≥ v − D for all (x. T ) ≤ [C(ε.2 Corollary. due to Mather [46] for Tonelli minimizers in the nonautonomous case. y) < R and k ∈ R. 32. B(ε). there are constants A(ε). C(ε) > 0 such that if T ≥ ε. ˙
For k ≥ c(L) and x. T ]. γ) ≤ B(ε. R).62
3. ˙ The conservation of the energy and the uniform bounds (1. R) and ˙ E(γ. then (i) Φk (x.
32
A priori compactness. ˙ 62
. y ∈ M .
32. y. then γ(t) < A for all t ∈ [0. deﬁne Φk (x. y ∈ M and γ ∈ CT (x. T [ such that γ(t0 ) ≤ D + C. For C > 0 there exists A = A(C) > 0 such that if x. Since AL (γ) ≤ C T .7) and (1. y) is a solution of the EulerLagrange equation such that AL+k (γ) ≤ [C(ε. will be very useful. In the autonomous case its proof is very simple. globally minimizing orbits. T ) := inf
γ∈CT (x. y) is a solution of the EulerLagrange equation with AL (γ) ≤ C T . d(x. Given ε > 0. y. the mean value theorem implies that there is t0 ∈]0.
7) we obtain B(ε.32.we obtain A(ε. then (a) T >
1 A
dM (x. y) ˙
≥ D d(x. R). y) − B T − d(x. y)
T
∀(x. y) is a solution of the EulerLagrange equation with AL+c (γ) ≤ Φc (x. we get (i) with C(ε. y). R). a priori compactness.
63
Proof: Comparing with the action of a geodesic on CT (x. R) and inequality (1.
(b) γ(t) < A for all t ∈ [0. ˙
Proof: Let η : [0. y ∈ M and γ ∈ CT (x.y) ≤ R . Using A(ε. y) ≥ 63
1 B
d(x. v) ∈ T M. y)] → M be a minimal geodesic with η ≡ 1. d(x.
(3. y) + dM (x. Then [ (1) + c] d(x. ε ε Using (i) and lemma 32. y) ≥ AL+c (η) ≥ Φc (x.3 Lemma.
32.1) and D = (1) + c + 2. Hence T ≥
D− −c−1 B
d(x. v) + c > D v − B. ˙ Let (r) be from (1. R) = sup L(v) v ≤ d(x. y).
. y). y) ≥ AL+c (γ) − d(x. y).1. y).10)
≥
0
D γ − B dt − d(x. From the superlinearity condition there is B > 0 such that L(x. T ].9) (3. There exists A > 0 such that if x.
y) − c T. globally minimizing orbits.
64
.
From (3. B [ (1) + c + 1 ] − c T.10). we get that AL (γ) ≤ ≤ (1) + c + 1 d(x.1 completes the proof.
Then lemma 32.64
3.9) and (3.
˙
Proof: Since xλ (t) = λ x(λt).2 must be zero. x) − k dt. x) − k dt.1. ˙
Diﬀerentiating A(λ) and evaluating at λ = 1. γ) ≡ k.2 Lemma. y) is a global minimizer or time free minimizer for L + k if k ≥ c(L) and AL+k (γ) = Φk (x. we have that Lv x − L dt ˙
0 T T
= −T k +
0
E(x. 33. A timefree minimizer for L + k has energy E ≡ k. We need the following 33. For λ > 0. Since it minimizes with free time.
Integrating by parts the term (Lx x + Lv x) t = ˙ ¨ A (1) = −T L(x(T ).33. Since γ is a solution of the EulerLagrange equation its energy E(γ. So that E(γ. y).1 Proposition. T ] → M be an absolutely continuous curve and k ∈ R. γ) is constant. we have that
T
A (1) = −T L(x(T ). ˙
Proof of proposition 33. then ˙ ˙ A(λ) =
0
T λ
L(x(λt). let xλ (t) := x(λt) and A(λ) := AL+k (xλ ). λ x(λt)) + k dt.
A curve γ ∈ C(x. x(T )) + k + ˙
0
Lx t x + Lv (x + t x) ˙ ˙ ¨
d dt L T
dt. Then
T
A (1) =
0
E(x. the derivative ˙ in lemma 33. x(T )) + k + L t T + ˙ 0
t. ˙ 65
. x) dt = ˙
0
E(x.
65
33
Energy of timefree minimizers. energy of timefree minimizers. Let : x : [0.
2. ˙
0 d dS T = − S2 d dλ . t Let x ∈ C ac ([0. 1].
on lemma 33. we have that 1 T
d dλ λ=1 A
Thus
=−
=−
1 T
T
E(yT . yT ) − k dt.3 Corollary. yT ) − k dt. Then B (T ) = − Proof: Using λ =
d dS S=T B T S
1 T
T
E(yT . write yT (t) = x( T ) : [0. T ] → M and B(T ) = AL+k (yT ). globally minimizing orbits. M ) and k > 0. ˙
0
66
.
33. For T > 0.66
3.
for k ≥ c(L). T ) = +∞. T ) = Φc (x. v) > A v − B. 1.34. 2. x. for k < c(L).
67
34
The ﬁnitetime potential. 67
. x = y. Then
ε
Φk (x. x.1 Proposition. +∞[ for any ε > 0. For k ∈ R. y. lim inf ε→0+ Φk (x. Given A > 0. y).
Here we shall prove the following proposition. t) → Φk (x. ε) = +∞. x. y. Φk (x. y. 4. the finitetime potential. y. y. let B > 0 be such that L(x. y ∈ M. y) + (k − B) ε. y. x.
6.y)
AL+k (γ).y)
AL+k (γ) ≥ inf
γ 0
A γ − B + k ˙
≥ A d(x. When M is compact. T ) = −∞. for k > c(L). T ) := inf
γ∈CT (x. limT →+∞ Φk (x. y). y ∈ M . 5. ε) =
inf
γ∈Cε (x. ε) ≥ A d(x. 34. the function (x. limε→0+ Φk (x. a compact subset R ⊆ M and ε > 0. t) is Lipschitz on R × R × [ε. limT →+∞ Φk (x. Now let A → +∞. T ) + (k − c) T .8. Φk (x. Thus. the limits in items 4 and 5 are uniform in (x. 3. 3. y ∈ M. We now compute the limits.
Proof: Item 2 follows from the fact that both action potentials satify an equivalent variational principle. y. y. y. See also corollary 411. y ∈ M and T > 0.
Recall that if k ≥ c(L).
for k ≥ c(L). y.
4. ˙ T If γs (t) := γ( s t). R) + k. T ) ≥ lim
T →+∞
Φc (x.
x. y.2 and inequality (1. then h(s) ≤ AL+k (γs ) =: B(s). y) on compact subsets follows from the Lipschitz condition of item 1 that we now prove. Denote h(s) := Φk (x. S) ≤ Φk (x. y. τ ) + n AL+c (γ) + Φk (z.
T →+∞
lim Φk (x. y. s). y) is a Tonelli minimizer. 2 2 This implies that limT →+∞ Φk (x. +∞[ t → Φk (x. 68
. s]. t) is uniformly Lipschitz. Observe that if k > c(L).6) there exists D = D(ε. Fix ε > 0. y.3 we have that f (T ) : = lim sup
δ→0
h(T + δ) − h(T ) δ 1 T
T
≤ B (T ) =
k − E(γ. we have that Φk (x. T ) ≥ inf Φc (x.68
3. Using corollary 33. y. T ) ≤ Φk (x.y∈M
5. with S ≤ τ < 2S. R) + k T − S. γ) dt ˙
0
≤ D(ε. If T > ε and γ ∈ CT (x. T ) + D(ε. τ ). If S. this limit is uniform because Φk (x. S > 0. T ) = −∞. y. For k < c(L) there exists a closed curve γ ∈ CS (z. The uniformity of this limit for (x. z). from corollary 32. T > ε we have that
S
Φk (x. R) > 0 such that E(γ. T ) +
T
f (t) dt
≤ Φk (x. z. globally minimizing orbits. γ) − k ≤ D(ε. z ∈ M such that AL+c (γ) < 0. R) + k . Then for T = nS + τ . y. y. We prove now that the function ]ε. t ∈ [0. y. y.
When M is compact. y) + (k − c) T = +∞. y) + (k − c) T.
˙
= Lv γ(s). T ). Now we prove the Lipschitz condition on a neighbourhood of y0 for y → Φk (x0 .
69
Since we can reverse the roles of S and T . Lv (x. T )
By Weierstrass theorem ??. Let λ ∈ CT (x0 . t). t) are Tonelli minimizers — and thus they realize Φk (y0 . y0 ) be a Tonelli minimizer.T ] . η(s. R)) d(y. Let γ : [0. T )) ≤ f (A(δ. 1]. Since R is compact. Then by corollary 32.4. s ∈ [0. 1] → M be a length minimizing geodesic joining y0 to y. T ). with γ ≡ 1. y0 . y ∈ R. T ) +
0 1
E (s) ds f (A(δ. y. y. R)) is a local Lipschitz constant for y → Φk (x.34. ˙ t ∈ [T − δ. T ) + f (A(δ. ε Fix 0 < δ < 2 small. y0 . y. T ) ≤ Φk (x0 . this is enough to show that the function is Lipschitz on R × R×]ε. t) = λ(t). if δ and d(y. T ) < ∂t A(δ. R). By lemma 14. ∂η (s. Thus ∂t
1
Φk (x0 .T −δ] ∗ η(s. T0 ). Changing the roles of y and y0 we obtain that f (A(δ. T ) +
= Φk (x0 . T ) = γ(s) and η(0. x. Observe that λ realizes Φk (x0 . +∞[. T ). ·)[T −δ. y.
T −δ ∂η ∂t (s. R) > 0 such that  ∂η (s. y. δ). t) T
dt =
T −δ
Lx
d dt Lv ∂η ∂s
∂η ∂s
+ Lv
∂2η ∂s ∂t
= Lv
+
Lx − · γ(s). y0 . The value od δ can be taken locally constant on a neighbourhood of y0 . Let E(s) := AL+k λ[0. there exists A(δ. R)) γ(s) ds ˙
0
≤ Φk (x0 . y0 ) are small enough. 69
. T − ε) = λ(T − δ). the curves t → η(s. the finitetime potential. y0 ).2. The proof for a neighbourhood of x0 is similar. Then E (s) = d ds
T T
L
T −δ ∂η ∂s T −δ T
∂η ∂t (s. T ] be a variation by solutions of (EL) such that η(s. y0 . R)). this implies the uniform Lipschitz condition for T → Φk (x. Let η(s.
y ∈ M . then there is γ ∈ C(x. T ). Now take a Tonelli minimizer γ on CT (x.1. We now study minimizers at k = c(L). y) = inf t>0 Φk (x. y). t) = Φk (x. and semistatic orbits lie in the stable or unstable manifolds of those ﬁxed points. From lemma 33. An absolutely continuous curve γ ∈ C(x. By proposition 34. y.
Here we construct curves that realize the action potential.
35
Global Minimizers. y. If k > c(L) and x. 35. static orbits are the ﬁxed points (x0 . y). the energy of γ is E(γ. (3. Hence it attains its minimum at some T > 0.70
3. y) is said static if AL+c (γ) = −Φc (y.1 Proposition. so there are no minimizers. Moreover. we have that AL+c (γ) ≥ Φc (x. y). globally minimizing orbits. Set c = c(L). For k < c(L). Observe that for c = c(L) and any absolutely continuous curve γ ∈ C(x. x = y. γ) ≡ k. y) ≥ −Φc (y. Moreover. Φk ≡ −∞. y) such that AL+k (γ) = Φk (x. 70
. the energy of γ is k.2 Deﬁnition. x). x). These names are justiﬁed by the following remark: For mechanic 1 lagrangians L = 2 v2 − U (x). y). f (t) is continuous and f (t) → +∞ when t → 0+ or t → +∞. t). An absolutely continuous curve γ ∈ C(x. ˙ Proof: Let f (t) := Φk (x. y) is said semistatic if AL+c (γ) = Φc (x.11)
35. Φk (x. 0) x of the lagrangian ﬂow where U (x) is maximal.2. y.
Semistatic curves have energy E ≡ c(L). N the Ma˜´ set. b] → M is equivalent to AL+c x[s. (3. set Using the characterization of minimizing measures 36.1 and corollary 35. (3. +∞[→ M is semistatic
We call M the Mather set.3 Corollary.11) implies that static curves are semistatic. M = ∪ supp(µ) µ ∈ M(L). global minimizers.t] = Φc x(s). A is the Aubry set. y) = Φc (x. N is the Ma˜´ set and ne ´ is the energy level E = [E ≡ c(L)]. x) = 0.5. AL (µ) = −c(L) N = Σ(L) := A = Σ(L) := Σ (L) := Σ+ (L) :=
−
w ∈ T M xw : R → M is semistatic w ∈ T M xw : R → M is static w ∈ T M xw :] − ∞.12)
Inequality (3.1 we get 35. The typographical relationship was observed by Albert Fathi.
71
By the triangle inequality for Φc the deﬁnition of semistatic curve x : [a.
71
. ∀ a ≤ s ≤ t ≤ b. 0] → M is semistatic w ∈ T M xw : [0. a curve γ ∈ C(x. and dc (x.3 we have that2 ´ M⊆A⊆N ⊆E . P = π(Σ(L)) the Peierls ne 1 and A = Σ(L) the Aubry set. Moreover. 35. All these inclusions can be made ´ E
1 2
The name is justiﬁed by proposition 37.
From proposition 33.35.13)
where M is the Mather set. x(t) .1. y) is static if (a) (b) γ is semistatic. y) + Φc (y.4 Deﬁnition.
ε] ). η(a)) = 0. 72
.b] is static then v ∈ Σ(L) (i. ˙ ˙ Let ξ(s) = π ϕs (w). η(a)) be solutions of (EL) with
1 AL+c (γn ) < Φc (η(b).
By the a priori bounds 32. We can assume that γn (0) → w. ξ(ε)) < AL+c (η[b−ε.ε] ) + lim inf AL+c (γn [ε. globally minimizing orbits. the whole orbit is static). which contradicts proposition 21.
Φc (η(a).e. Thus w = η(b) and similarly ˙ limn γn (Tn ) = η(a).3 γn  < A.72
3.e. A local static is a global static. ξ(ε)) + Φc (ξ(ε). Thus Φc (η(b − ε). η(a)) ≤ Φc (η(a).b] ) + AL+c (ξ[0. ˙ and hence by remark 12. Denote by α(v) and ω(v) the α and ωlimits of v under the EulerLagrange ﬂow. η(a)) + Φc (η(b).ε] is not C 1 . ˙ ˙ If lim sup Tn < +∞.Tn ]
n
≤ −Φc (η(b).18) and adding a properly chosen potential φ(x).b−ε] ) + AL+c (η[b−ε. η(a)) + n . i.b] ∗ ξ[0.
proper constructing examples of embedded ﬂows as in equation (1. η(a)) < AL+c (η[a.ε] ∗ γn [ε.1(3). In this case η is a (semistatic) periodic orbit of period τ + b − a and then it is static. we can assume that τ = limn Tn > 0 exists. 35.b] ) + AL+c (ξ[0.5 Proposition.2. it can not be a Tonelli minimizer.Tn ] )
n
≤ AL+c (η[a. If w = η(b) then the curve η[b−ε. η(b − ε)) + Φc (η(b − ε). Proof: Let η(t) = π ϕt (v) and let γn ∈ CTn (η(b).b] ) + lim γn [0. if xv [a.
35.
35.tn +s] −→ η[−s.6 Deﬁnition. We shall see in corollary 411.b+s] ) + Φc (η(b + s). we have that AL+c (η[a−s. η(a − s)) ≤ ≤ lim AL+c (γn [Tn −s. Proof: We prove only that ω(v) ⊂ Σ. η(a − s)) ≤ Φc (η(a). γ(t). Let γ(t) = π ϕt (v).
Clearly a semiinﬁnite semistatic curve is a ray.Tn −s] ) + AL+c (γn [Tn −s. Thus η[a−s. +∞[→ M (resp.0] ) is a Tonelli minimizer for all t > 0. then γ[tn −s.35. η(b)) + lim AL+c (γn [0. 0] → M ) is a ray if γ[0. Suppose that tn → +∞ and γ(tn ) → w ∈ T M .Tn ] )
n
≤ Φc (η(a).9 that rays are semistatic.
73
Now suppose that limn Tn = +∞. If v ∈ Σ is semistatic. An absolutely continuous curve γ : [0. global minimizers.
C1
73
. η(a)) = 0. AL+c (γ[0.7 Proposition. i.e. then α(v) ⊂ Σ(L) and ω(v) ⊂ Σ(L).s] )
n
+ Φc (η(b + s).b+s] is static. η :]−∞.s] . Since γ and η ˙ are solutions the EulerLagrange equation.Tn ] ) + AL+c (η) + AL+c (γn [0.s] ) + AL+c (γn [s. Moreover α(v) and ω(v) are each contained in a static class.t] (resp. If s > 0.t] ) = Φc (γ(0). Set c = c(L). η(b)) + Φc (η(b). t) for all t > 0. η[−t. Let η(t) = π ϕt (w).
πu) = Φc (πw.sn ] ) + AL+c (γ[sn . πw) = 0.74
3. πu) + Φc (πu.tn+1 ] )
n n
= lim AL+c (γn [tn .
Then AL+c η[−s. Thus w ∈ Σ(L). We may assume that γ(sn ) → u with ˙ tn < sn < tn+1 .
74
. globally minimizing orbits. πw) = lim AL+c (γ[tn .s] + Φc (η(s). Let u ∈ ω(v).tn +s] ) + lim AL+c (γ[tn +s. γ(tm − s))
n m
= Φc (η(−s). Thus w and u are in the same static class.tm −s] )
n m
= lim lim AL+c (γ[tn −s. Then dc (πw.tn+1 ] ) = Φc (πw. η(−s)) = = lim AL+c (γ[tn −s.tm −s] )
n m
= lim lim Φc (γ(tn − s). η(−s)) = 0.
ϕT θ) = 0 and L + c dµ = lim
M
1 T →+∞ T
1 T →+∞ T
T
L(ϕt θ) + c dt
0
≤ lim inf
Φc π(ϕT θ). it is enough to prove the theorem for ergodic measures.1 they are invariant. π ϕTn θ)
n Tn
≤ lim
n 0
L(ϕt θ) + c dt = 0.36.
75
. ne µ ∈ M(L) is a minimizing measure if and only if supp(µ) ⊆ Σ(L). ϕTn θ) → 0 and
Tn
lim
n 0
L(ϕt θ) + c dt = 0. Applying corollary 36. Suppose that µ ∈ M(L) is ergodic and supp(µ) ⊆ Σ(L). πϕ1 θ) = lim Φc (π θ. Since Σ(L) is closed and A is dense in supp(µ) then supp(µ) ⊆ Σ(L). by Birkhoﬀ’s theorem there is a set of total µmeasure A such that if θ ∈ A then lim inf T →+∞ dT M (θ. Proof: Since Σ is closed. π ϕ1 θ) + Φc (π ϕ1 θ. we get that there is a set A ⊂ T M of total µmeasure such that if θ ∈ A then there is a sequence Tn → +∞ such that d(θ. Since µ is ﬁnite.
Recall that minimizing measures µ satisfy AL (µ) = −c(L) and that by theorem 24.
75
36
Characterization of minimizing measures.
This implies that θ is static.5 to F = L + c and X = T M . Then L + c dµ = 0.
Then 0 ≤ dc (π θ. π(θ) = 0. characterization of minimizing measures.
Now suppose that µ ∈ M(L) is minimizing. 36.1 Theorem (Ma˜´ [39]).
For
SN F (p) :=
n=0
F f n (p) . f an ergodic measure preserving map and F : X → R an integrable function.1 that 36.4 Lemma. 36.1 with corollary 35. we get 36. 76
. Let (X.3 Corollary (Dias Carneiro [9]).
It follows from theorem 25.
Combining theorem 36.2 Corollary. ν) be a probability space.76
3. Deﬁne δ(k) by N1 < N2 < · · · be the integers for which f
Nk δ(k) = SNk F (x) . Given ˆ A ∈ B with ν(A) > 0 denote by A the set of point p ∈ A such that for all ε > 0 there exists an integer N > 0 such that f N (p) ∈ A and
N −1 j=0
F f j (p) − N
F dν < ε . If µ is a minimizing measure then it is supported in the energy level E(supp(µ)) = c(L).
ˆ Then ν(A) = ν(A).
Let A(ε) := p ∈ A  ∃N > 0 such that f N (p) ∈ A and SN F (p) < ε . Take x ∈ A ∩ X and let Ni (x) ∈ A. because A = n A( n ). globally minimizing orbits.
1 ˆ It is enough to prove that ν(A(ε)) = ν(A). If M is compact then Σ(L) =
. Let ˆ X be the set of points for which the Birkhoﬀ’s theorem holds for F and ˆ the characteristic functions of A and of A(ε).3. B. Proof: Without loss of generality we can assume that p ∈ X denote
N −1
F dν = 0.
/ This implies that ν A − A(ε) = lim 1 # 0 ≤ j < Nk  f j (x) ∈ A − A(ε) k→+∞ Nk 1 ≤ #S(k) . Hence j ∈ S(k) =⇒ f Nj (x) ∈ A(ε) . for k ∈ K.
From (3. Assume that S is inﬁnite.
1 2
(2ε) #S(k) ≤ ck  = δ(k) Nk .14) Nk
If S is ﬁnite.36. / so that SN −Nj F f Nj (x) = c − cj  ≤ ε . Choose an inﬁnite sequence K in S such that for all k ∈ K we have. then there is an > j such that c − cj  ≤ ε. Set ck := SNk F (x) . This implies that the set { ck  k ∈ S } is unbounded. c − ck  ≥ ε . ck  = max cj .
j∈S(k)
Then.
Observe that if j ∈ S(k). characterization of minimizing measures. 1≤j ≤k−1 j ∈S . S := S(k) := k ∈ N ∀ > k.14) we get that ν A − A(ε) = 0. we get that ν A − A(ε) ≤ lim δ(k) = 0.14).
77
ˆ Since x ∈ X we have that limk→+∞ δ(k) = 0. (3. ε
k∈K
77
. concluding the proof of the lemma. from inequality (3.
X is a complete separable metric space.
36. Then ˆ 1 the full measure subset ∩ ∪ Vn ( m ) satisﬁes the required property.4. x) < ε and
N −1 j=0
F f j (x) − N
F dν < ε
Proof: Given ε > 0 let {Vn (ε)} be a countable basis of neighbourhoods ˆ with diameter < ε and let Vn be associated to Vn as in lemma 36. globally minimizing orbits. If besides the hypothesis of lemma 36.e.5 Corollary.
m n
78
. and B is its Borel σalgebra.4.78
3. x ∈ X the following property holds: for all ε > 0 there exists N > 0 such that d(f N (x). then for a.
37. y ∈ M deﬁne hT (x. 5. y. ∀ x ∈ M . z) ≤ h(x. h is Lipschitz. ∀ x. p) + Φc (p. y) + hS (y. ∀ x. 3. Deﬁne the Peierls barrier as h(x.
4. x) ≥ Φc (x.1 Proposition. T ) := inf
γ∈CT (x. If Σ =
. q) + Φc (q. y).
. 6. y). in particular h(x. hT +S (x.
T →+∞
The diﬀerence between the action potential and the Peierls barrier is that in the Peierls barrier the curves must be deﬁned on large time intervals. y) ≤ inf p∈π(Σ) Φc (x.
For T > 0 and x. y ∈ M . y). Clearly h(x. 2. 37. 79 for all S > 0. y. h(x.y)
AL+c (γ). z) ≤ hT (x. h(x. y) := lim inf hT (x. If h : M × M → R is ﬁnite. h(x. y) + hS (y. Observe that for all S. y) are the Tonelli minimizers on CT (x. p) + h(p. h(x. then 1. z). p. y). the peierls barrier. Taking lim inf T →+∞ we get that h(x. b
Proof: Item 2 is trivial. y. z). z).
79
37
The Peierls barrier. x) = 0 ⇐⇒ x ∈ π(Σ) = P. z ∈ M . y). ∀ x. y) ≤ Φc (x. y) = Φc (x. x) ≥ 0.
So that the curves which realize hT (x. h(x. y). y) + h(y. y) ≥ Φc (x. q ∈ M . T > 0 and y ∈ M . z) ≤ h(x.
if h(x. x) ≤ lim AL+c γnk [0.
Taking lim inf T →+∞ we get item 4. y) + Φc (y.3. y) ≤ Φc (x. y) + A dM (y. z) ≤ h(x. Then 0 ≤ h(p.s] + Φc (π ϕs v. Then if γnk (0) → v. we obtain that h is Lipschitz. z). x. z) ≤ h(x. z. y) + Φc (y. q) + Φc (q. z ∈ M.s] + AL+c γnk [s. Let v be an accumulation point ˙ of γn (0) and η(t) := π ϕt (v). then h(p. p) = 0. 80
. 1. p) ≤ h(p. Take v ∈ Σ such that π(v) = p and y ∈ π(ωlimit(v)). p) + Φc (y. Taking the inﬁmum on S > 0. where A is a Lipschitz constant for Φc . p)
n n
≤ lim −Φc (γ(tn ). for any s > 0 we have ˙ ˙ that 0 ≤ Φc (x. 5. x) = 0. y. p) = 0. globally minimizing orbits. By lemma 32.
Taking lim inf S→+∞ . Conversely. 4. γ is uniformly bounded. Let γ(t) := π ϕt (v) and choose tn ↑ +∞ such that γ(tn ) → y. x). π ϕs v) + Φc (π ϕs v. p) ≤ lim AL+c (γ[0. We ﬁrst prove that if p ∈ P = π(Σ). we obtain item 3. Thus v ∈ Σ. x) ≤ AL+c η[0.Tn ]
k k
= 0. Observe that
S>T
∀ x. Tn ) with Tn → +∞ and AL+c (γn ) → 0. then there exists a sequence of Tonelli n minimizers γn ∈ C(x. we get that h(x. y. p) + hT (p.
inf hS (x. Changing the roles of x.80
3.tn ] ) + Φc (y.
y) ≤ inf
b p∈π(Σ)
Φc (x.
81
6. y) < +∞ for all x. y) < +∞. y) + 2A d(qn .
81
.
Proof: From proposition 37. Then. Using items 4 and 5.
In particular h(x.6 we have that h(x.
37. y) ≤ inf
b p∈π(Σ)
Φc (x. Tn ]) be such that limn qn = q. y). the peierls barrier.37. q) + Φc (q.1. y) . Then T AL+c (γn ) → 0.2 Proposition. y) . we get that h(x. Let q ∈ π supp(µ) and qn ∈ γn ([0. y) = inf
b p∈π(Σ)
Φc (x. y) . p) + 0 + Φc (p. p) + Φc (p. Now let γn ∈ CTn (x. Φc (x. p) + Φc (p. q). we get that Φc (x. q) + Φc (q. Then µ is minimizing. y) ≤ Φc (x. Let µ be a weak limit of a subsequence of the measures µγn . y) ≤ h(x. qn ) + Φc (qn . y) 1 with Tn → +∞ and AL+c (γn ) → h(x. y ∈ M . then h(x. q) ≤ AL+c (γn ) + 2A d(qn . Letting n → ∞. If M is compact.
xi (a) = vi . .a+ε] ) > AL (γ1 ) + AL (γ2 )
3 4
This name is justiﬁed by proposition 37. (Ma˜´) [39] ne For all p ∈ π(Σ) there exists a unique ξ(p) ∈ Tp M such that (p.
38
Graph Properties.
AL (x1 [a−ε. ξ(p)) ∈ Σε . 2 satisfy d(p1 . The reader may provide those proofs as exercises. the map ξ : π(Σ) → Σ is Lipschitz. γ1 (a + ε) = x2 (a + ε) . (pi . write xv (t) = π ϕt (v).
In this section we shall see that the projection π : Σ → M is injective. (p2 . The proof of the Lipschitz condition need the following lemma. i = 1. Given ε > 0. ε) → M or xw : (−ε. For v ∈ T M . 2. a + ε] → M of L with 0 < ε < ε1 . The proofs of the injectivity of π in this book only need that the solutions of the EulerLagrange equation are diﬀerentiable4 . p2 ) then. p2 ) < δ and d((p1 . 0] → M is semistatic }. due to Mather. let Σε := {w ∈ T M  xw : [0. 38. there exist ˙ solutions γi : [a − ε. if a ∈ R and xi : R → M .a+ε] ) + AL (x2 [a−ε. v1 ).
82
. i = 1. are the solutions of L with xi (a) = pi . v2 )) ≥ K −1 d(p1 . and hence a nondiﬀerentiable curve can not be a Tonelli minimizer. ξ(p)) ∈ Σ and Σ = graph(ξ). For the proof see [46] or Ma˜´ [36]. Moreover. ne 38. [46] Given A > 0 there exists K > 0 ε1 > 0 and δ > 0 with the following property: if vi  < A. γ2 (a + ε) = x1 (a + ε) .1 Theorem.2 Mather’s Crossing lemma. globally minimizing orbits. vi ) ∈ T M . satisfying γ1 (a − ε) =x1 (a − ε) γ2 (a − ε) =x2 (a − ε) . 3 Thus the projection πΣ gives b an identiﬁcation P ≈ Σ. We shall call P := π(Σ) the Peierls set.1(5).82
3. in particular (p.
p) = Φc (q−ε . q) < δ. p) + Φc (p. ε] → M such that α(−ε) = xw (−ε) =: q−ε .0] ) + AL (xv [−ε. graph properties. v). β(−ε) = xv (−ε) =: pε . β : [−ε. β(0) = q . v) ∈ Σ. q) . q) ≤ Φc (q−ε . and d(p. p−ε ) Thus α(0) = p . (q.2 there exist α. (q. p) + Φc (p−ε . w) < K dM (p. Φc (q−ε . Suppose it is false.1: We prove that if (p. q) + Φc (p−ε . and AL (α) + AL (β) < AL (xw [−ε.0] ).
83
Proof of theorem 38. we only prove the case in which xv [−ε. So Φc (q−ε .38.0] is semistatic. then dT M (p. Then by lemma 38.
fig. 1: graph property. Observe that this implies the theorem. q) < Φc (q−ε . q) − Φc (p. For simplicity. q) which is a contradiction. p−ε ) + Φc (p−ε . 83
. q) < Φc (q−ε . w) ∈ Σε .
w) ≤ K dM (π(v). Then there are v. theorem 70.84
3.5. w ∈ Γ+ then 0 dT M (v. The continuity of the pseudometric dc implies that a static class is closed.1 we can deﬁne an equivalence relation on Σ by u.
The set π(Γ+ \ Γ+ ) is called the cut locus of Γ+ .+∞[ is semistatic and ω(v) ⊆ Γ . if v . there exist α ∈ C2ε (xv (−ε). But π may not be injective on Γ+ \ Γ+ even for generic lagrangians. Let Γ be a static class. The equivalence classes are called static classes. Let ε > 0 be such that xv [−ε.+∞[ and xw [−ε. π(v)) = 0. 0 38. the projection π : Γ+ → M is injective with 0 Lipschitz inverse. v ∈ Σ.3 Theorem. π(w)). By lemma 38. the set Γ+ = v ∈ Σ+ (L)  ω(v) ⊆ Γ is called the (forward) basin of Γ. (3. But when M is compact for generic lagrangians π(Γ+ ) = π(Γ− ) = M because there is only one static class (cf.15)
Suppose it is false. (Ma˜´)[39] ne For every static class Γ. Let Γ+ = 0 =
t>0 ε>0
ϕt (Γ+ ) v ∈ T M  xv ]−ε. Clearly Γ+ is forward invariant. For v ∈ T M denote by ω(v) its ωlimit. globally minimizing orbits. u ≡ v ⇐⇒ dc (π(u).1.ε] ) + AL+c (xw [−ε. 0 Proof: We prove that for K as in lemma 38.ε] ). xv (ε)) such that AL+c (α) + AL+c (β) + δ < AL+c (xv [−ε. The projection π : Γ+ → M may not be surjective.15) 0 does not hold. 84
. w ∈ Γ+ such that inequality (3.+∞[ are semistatic.2.2. xw (ε)) and β ∈ C2ε (xw (−ε).
Using the graph property 38. and it is invariant by proposition 35.(B)).
Letting n → ∞ and adding dc (p.sn ] ) + AL+c (xw [−ε. xw (tn ) → q. q) + Φc (xw (−ε). q) = 0. xv (sn )) + Φc (xw (−ε). q) < Φc (xv (−ε). p) + Φc (p.tn ] ) = Φc (xv (−ε). q) ≤ Φc (xv (−ε).tn ] ) + AL+c (β ∗ xv [ε. q).
85
.38. p) + Φc (q.
85
for some δ > 0. Then Φc (xv (−ε). xw (tn )) + Φc (xw (−ε). Let p. q ∈ π(Γ) and sn tn → +∞ be such that xv (sn ) → +∞ and xv (sn ) → p. xv (sn )) + δ ≤ AL+c (α ∗ xw [ε. p) + Φc (xw (−ε). we have that Φc (xv (−ε). p) + Φc (xw (−ε).sn ] ) + δ < AL+c (xv [−ε. graph properties. xw (tn )).
(Ma˜´) [39] ne If c = c(L). π(v)). globally minimizing orbits.16)
. π(ϕh w) G(ϕh w) = Φc p. More precisely.h]
h
= lim =
1 h→0 h lim 1 h→0 h
= lim
h→0
1 h
L xw (s). b taking any p ∈ M and deﬁning G : Σ → R by G(w) = Φc p. 39. 86 (3.
39
Coboundary Property. Ma˜´ in [38] and ne further developed in [39] and by A. Fathi. then (L + c) where 1 dG (w) := lim G(ϕh (w) − G(w) . π(ϕh w) = Φc p. π(w) + Fw ϕh (w) . π(w) . We claim that for any p ∈ M and any w ∈ Σ.86
3. G(ϕh w) = Φc p. πw) SL+c xw [0. We have that dFw dϕ = lim
w 1 h→0 h b Σ
=
dG . then (L+c) Σ is a Lipschitz coboundary.
The coboundary property was ﬁrst presented by R. h→0 h dϕ Proof: Let w ∈ Σ and deﬁne Fw (v) := Φc (π(w). π(w) + Φc π(w). πϕh w) − Φc (πw. df
Fw (ϕh w) − Fw (w) Φc (πw.1 Theorem. xw (s) + c ds ˙
0
= L(w) + c. h ∈ R.
x). then Φc (x. x). x) − Φc (q. (3.
w
and G is Lipschitz by proposition 21. We now prove (3.
87
This is enough to prove the theorem because then dG dϕ =
w
d Fh (ϕh w) dh
=
h=0
Fw dϕ
= L(w) + c. q).1.17).39. This implies (3. coboundary property. q) = −Φc (q.
87
.17) Since the points q and x can be joined by the static curve xw [0. q) + Φc (q. x := π(ϕh w). x) + Φc (x. q) ≤ Φc (p. x) = Φc (p. Let q := π(w).h] . We have to prove that Φc (p. Using twice the triangle inequality for Φc we get that Φc (p.16). q) = Φc (p. x) ≤ Φc (p.
310. But in general π : Σ+ (L) → M is not injective. γnk (s) + = Φc η(0).+∞[ is semistatic and v ∈ Σ+ .1 Theorem. The rest of the ˙ proof is similar to the case above. γn  < A.3.s] = lim AL+c γnk [0. take a Tonelli minimizer γn ∈ CTn (x. and hence Tn → +∞. hence S = +∞. Given x ∈ M \ π(Σ). Thus η[0.88
3.3. p) such that
1 AL+c (γn ) < Φc (x. If S < +∞ then η(S) = limk γnk (Tk ) = p. Take p ∈ π(Σ).
Then by lemma 32. Since x ∈ π(Σ). this contradicts the graph / property 38. yn ) be a Tonelli minimizer such that
1 AL+c (γn ) < Φc (x. S]. p). M is noncompact. Let γn ∈ CTn (x.3 E(Σ+ ) = c(L). π(Σ+ (L)) = M . yn ) → +∞.1. Proof: First suppose that Σ(L) = .18)
88
. p) + n . η(s) .5) we get that c(L) ≥ e0 and then e0 ≤ cu ≤ ca ≤ c0 ≤ c(L). yn ) + n . using (1.
1 nk
Then η is semistatic on [0.2. If Σ = . where S = lim inf k Tnk .s]
k k
≤ lim Φc γnk (0). (3. By corollary 35.
310
Covering Properties. γn  < A and Tn > A d(x. we have that
AL+c η[0. then by corollary 36. ˙ Then. globally minimizing orbits. Let η(t) := π ϕt (v). Let x ∈ M and yn ⊆ M such that dM (x. 1 By the a priori bounds 32. Let v = ˙ ˙ limk γnk (0) be an accumulation point of γn (0) . if 0 < s < lim inf k Tnk .
311. .311. . In other words. 89
. recurrence properties. between two static classes there exists a chain of static classes connected by heteroclinic semistatic orbits (cf. . 2: connecting orbits between static classes. is transitive. .1 could be restated by saying that if the cardinality of Λ is ﬁnite. vn−1 ∈ Σ such that for all 1 ≤ k ≤ n − 1 we have that α(vk ) ⊆ Λk and ω(vk ) ⊆ Λk+1 . we have that Λi Λj . . . . then given two static classes Λi and Λj there exist classes Λi = Λ1 . Λn = Λj and semistatic vectors v1 .
89
311
Recurrence Properties. Then given Λi and Λj in Λ.
(c) If there is v ∈ Σ with the αlimit set α(v) ⊆ Λi and ωlimit set ω(v) ⊆ Λj .1 Theorem. . ﬁgure 2). The three closed curves represent the static classes and the other curves represent semistatic orbits connecting them. Deﬁne a reﬂexive partial order Λ by (a) (b) is reﬂexive.
in
Let Λ be the set of static classes. Theorem 311.
fig. then Λi Λj . Suppose that M is compact and the number of static classes is ﬁnite.
Σ(L) is chain recurrent. Now we proceed to prove theorem 311.s] + Φc (η(s). η(−s)) = = lim AL+c (γ[tn −s.
A proof of the following theorem can be found in [12] 311. Σ(L) is chain transitive. η(−s)) = 0.s] . then 1.tm −s] )
n m
= lim lim AL+c (γ[tn −s. then γ[tn −s. 311.tn +s] −→ η[−s. πw) = 0.
C1
Then AL+c η[−s. We may assume that γ(sn ) → u with ˙ tn < sn < tn+1 .1.2 Theorem. Then dc (πw.tn +s] ) + lim AL+c (γ[tn +s. πw) = lim AL+c (γ[tn .tm −s] )
n m
= lim lim Φc (γ(tn − s).90
3. then α(v) ⊂ Σ(L) and ω(v) ⊂ Σ(L). 2. Proof: We prove only that ω(v) ⊂ Σ. πu) = Φc (πw. Assume for the rest of this section that M is compact. πu) + Φc (πu. Let γ(t) = π ϕt (v).3 Proposition. Suppose that tn → +∞ and γ(tn ) → w ∈ T M . If M is compact. Let u ∈ ω(v).sn ] ) + AL+c (γ[sn . Thus w ∈ Σ(L).tn+1 ] )
n n
= lim AL+c (γn [tn . Let η(t) = π ϕt (w). γ(tm − s))
n m
= Φc (η(−s). globally minimizing orbits.tn+1 ] ) = Φc (πw. Moreover α(v) and ω(v) are each contained in a static class. 90
. Thus w and u are in the same static class. If v ∈ Σ is semistatic. Since γ and η ˙ are solutions the EulerLagrange equation.
91
.19)
Let u be a limit point of vn . xu (s)) + Φc (xu (s). xu (1)) + Φc (xu (1). p1 ) = 0. xvn (t))+Φc (xvn (t). p2 ) + n . xu (t)) ≤ ≤ Φc (xu (s). bn ] → M . xvn (an ) = p1 . Since U1 and U2 are disjoint sets we can take a solution xvn : [an . So that u ∈ Σ. p2 )+Φc (p2 . where in the last inequality we used (3. xu (s))] = 0. p2 ) = 0. p2 ) + Φc (p2 . therefore dc (p1 . an < 0 < bn of (EL) such / that xvn (0) ∈ π(U1 ∪ U2 ).4 Proposition. xu (t)) + [Φc (xu (t). xvn (s))+Φc (xvn (s). xvn (s)) + Φc (xvn (s). t = 1: dc (xu (0). 2. Hence xu (0) ∈ π(Λ). for s = 0. 2. Combining the last equation with the triangle inequality we obtain dc (xu (s). Let U1 . Hence Φc (p1 . recurrence properties. Let pi ∈ π(Ui ∩ Λ). for an ≤ s ≤ t ≤ bn .
91
311. i = 1.311. xvn (t)) + Φc (xvn (t). i = 1. Every static class is connected. p2 ) ≤ Φc (p2 . xu (t)) + Φc (xu (t). p2 ) + Φc (p2 . p2 )]
n
≤ Φc (p2 . p2 ) + Φc (p2 . then xu : R → M is semistatic (see the proof of claim 2 item (a)). dc (p1 . Proof: Let Λ be a static class and suppose that it is not connected. This / contradicts the fact that Λ ⊆ U1 ∪ U2 . p2 ) ≤ Φc (p1 . p1 ) + lim inf AL+c (xvn )
n
≤ dc (p1 . p1 ).19). p1 ) + Φc (p1 . Moreover.
(3. On the other hand xu (0) ∈ π(U1 ∪ U2 ). p1 ) ≤ ≤ Φc (p1 . p1 ) + lim inf [Φc (p1 . U2 be disjoint open sets such that Λ ⊆ U1 ∪ U2 and Λ ∩ Ui = . p1 )] = 0. xu (0)) + [Φc (xu (0). xvn (bn ) = p2 and
1 AL+c (xvn ) ≤ Φc (p1 . Then.
Σ) < ε }. Hence if we assume that there are only ﬁnitely many of them. Given v ∈ T M denote by α(v) and ω(v) its α and ωlimits respectively.
Given v ∈ Σ with α(v) ⊆ A and 0 < δ < ε. Let s1 (v) := inf{ s ∈ R  fs (v) ∈ A(ε) } ∈ R ∪ {+∞}. i. / If sk (v) < +∞. globally minimizing orbits.
B :=
{ Λj ∈Λ  Λi Λj }
Λj .4 the static classes are connected. where Λi (δ) are disjoint open sets containing exactly one static class and the number of components N (ε) is ﬁxed for all 0 < δ < ε. Observe that sk (v) < +∞ implies that Tk (v) < +∞ because by the deﬁnition of B and the transitivity of we have that ω(v) ⊆ A. tk (v). k ≥ 1. sk (v) < +∞ }. 92
. Λk ∈ Λ be such that Λi Λk . deﬁne tk (v) := sup{ t < sk (v)  ft (v) ∈ A(δ) }.e. For ε > 0.1. Tk (v) as follows. Fix ε > 0 small enough such that the connected components of Σ(ε) are the εneighborhoods of the static classes. Now suppose that the theorem is false. Tk (v) := inf{ t > sk (v)  ft (v) ∈ A(δ) }.92
3. Let Λi . So that for 0 < δ < ε. Let A :=
{ Λj ∈Λ  Λi Λj }
Λj
.
Proof of theorem 311. let Σ(ε) be the εneighborhood of Σ. deﬁne inductively sk (v). Σ(ε) := { v ∈ T M  dT M (v. α(v) ⊆ A. Deﬁne Ak = Ak (δ) := sup{  Tk (v) − tk (v)  : v ∈ Σ. By proposition 311. the connected components of Σ are ﬁnite and must coincide with the static classes. Σ(δ) = N (ε) i=1 Λi (δ).
If u1 ∈ Λj ⊆ A. πu2 ) = Φc (πu1 . πu2 ) + Φc (πu2 .1 into the following claims: Claim 1. . πu1 )
n
= dc (πu1 .
93
.
(b) lim supk Ak (δ) = supk Ak (δ) = +∞.1. (a) M ∩ B = . obtaining a contradiction and thus proving theorem 311. πu2 ) = 0. / Observe that sk (v). πu2 ) + Φc (πu2 . write A (δ) ≡ 0 for all ≥ k. 0 < sn < tn such that vn → u1 ∈ A. . fsn (vn ) → u2 ∈ A(ε). 2. Claim 3. / We now use claim 3 to complete the proof of theorem 311. and all 0 < δ < ε. Ak (δ) < +∞ for all k = 1. Now set: sk+1 (v) := inf{ s > Tk (v)  ft (v) ∈ A(ε) }. πu1 ) ≤ lim Φc (πvn . . πu3 ) + Φc (πu3 . It is enough to show that dc (πu1 . where the fourth equation holds because vn is a semistatic vector. α(v) ⊆ A }. recurrence properties. πfsn (vn )) +Φc (πfsn (vn ). πu1 ) = lim Φc (πvn . Indeed dc (πu1 . πftn (vn ))
n
+ Φc (πu3 . πu1 ) ≤ Φc (πu1 . πftn (vn )) + Φc (πu3 . Deﬁne M := { v  v ∈ Σ. we shall prove that u2 ∈ Λj \A(ε). πu3 ) = 0. Claim 2. We split the rest of the proof of theorem 311. There exist sequences vn ∈ Σ.
93
if sk (v) = +∞ for all v ∈ Σ with α(v) ⊆ A.311. πu3 ) = 0. tk (v) and Tk (v) are invariant under ft .1. ftn (vn ) → u3 ∈ A and dc (πu1 .
Proof of claim 2: (a) Let p ∈ πA. For n > 0. Let u = limn vn ∈ π −1 (∂ π B(δ)). .
Proof of claim 1: Suppose that Ai < +∞ for i = 1. . Then for all n. α(u) ⊂ A. Since ft (vn ) ∈ A(ε) for 0 < t < Tk (vn ) and Tk (vn ) → +∞. Taking limits in (3. k − 1 and Ak = +∞. and that the sequence vn converges (cf. .t] is semistatic for all lim inf n an ≤ s ≤ t ≤ lim supn bn . The case k = 1 is similar. q ∈ πB. 94
1 n
(3.
where m is the Lebesgue measure on R.t] ≤ Φc xvn (s). xvn (t) + for all an ≤ s ≤ t ≤ bn . we have that
k−1
m{ t < 0  ft (vn ) ∈ A(ε) } ≤ /
i=1
Ai . lemma 32. Let u = limn vn ∈ ∂A(δ). q) + n . xvn (bn ) = q and
1 AL+c (xvn ) ≤ Φc (p. let xvn : [an . But then the orbit of u contradicts the deﬁnition of B.20) we obtain that xu [s.3.20)
.3). with α(vn ) ⊂ A and Tk (vn ) − tk (vn ) → +∞. globally minimizing orbits. This implies that
k−1
m{ t < 0  ft (u) ∈ A(ε) } ≤ /
i=1
Ai
and hence α(u) ∩ A(ε) = . .94
3. By proposition 311.
This implies that AL+c xvn [s. bn ] → M be a solution of (EL) such that xvn (an ) = p. then ft (u) ∈ A(ε) / / for all t > 0 and hence ω(u) ⊆ B. We can assume that inf{ s > an  xvn (s) ∈ B(δ) } = 0. Then there exists vn ∈ Σ. We can assume that tk (vn ) = 0 and that vn converges (Σ is compact).
πu3 ) = 0. then ft (u) ∈ B(δ) for all t < 0. tn := τ2 (δn ) and choose a subsequence such that vn . u). Since u ∈ π −1 (∂ π B(δ)) there exists z ∈ B such that dM (π(u). Hence u ∈ Σ. T ). }.
(b) By claim 1 it is enough to show that supk Ak (δ) = +∞. ˙ and by the graph property (theorem 38. there exist τ1 (δ) < s(δ) < τ2 (δ) with fτ1 (v) ∈ Λj (δ). z) ≤ K δ. recurrence properties.
. u3 ∈ Λj .
95
Any limit point w of xvn (an ) = fan (vn ) satisﬁes π(w) = p ∈ πA. Hence / α(u) ⊆ A and thus u ∈ M. Hence there is v = vδ ∈ Σ with α(v) ⊂ A. then M ⊆ M(δ.e. Since k > N (ε) there is one component Λj (δ) ⊆ A(δ) with two of these returns. Since ft (vn ) ∈ B(δ) for / all an ≤ t < 0 and an → −∞. T )∩B = This contradicts item (a). Thus u ∈ M ∩ B(Kδ). If supk Ak (δ) < T . then / limn an = −∞. there exists a sequence such that the repeated component Λj ⊂ Λj (δn ) is always the same.311. fs (v) ∈ A(ε) and / fτ2 (v) ∈ Λj (δ). . u3 = limn ftn (vn ) ∈ Λj and u2 = limn fsn (vn ) ∈ A(ε). we obtain that M ∩ B = . Now. any limit point of fbn (vn ) is in B. Letting δ → 0.1). Similarly. because B is invariant and B∩A(δ) =
Proof of claim 3: Given 0 < δ < ε. Let sn := s(δn ). limn bn = +∞. by claim 2(b) there exists k > N (ε) such that Ak (δ) > 0. such that the orbit of v leaves A(ε) and returns to A(δ) at least k times.T ] (v) ∩ A(δ) = Then M∩B ⊆ M(δ. w ∈ A. / then dc (πu1 . Since u1 . i. Let u1 = limn vn ∈ ∩n Λj (δn ) = Λj . fsn (vn ) and ftn (vn ) converge. (π(z). where M(δ. π(z)) ≤ δ. We can choose vδ so that τ1 (δ) = 0. T ) = { v ∈ Σ  f[−T. by theorem 38.
95
.1 we have that dT M (π(u). Since A ∪ B is invariant and u ∈ A ∪ B. Since z ∈ Σ and u ∈ Σ.
96
3. globally minimizing orbits.
96
.
half of the dimension of T ∗ M . Since the tangent spaces to N are lagrangian. then N is invariant under the hamiltonian ﬂow. Isotropic spaces of dimension n are called lagrangian subspaces We say that a submanifold W ⊂ T ∗ M is lagrangian if at each point θ ∈ W . Since ω(X. ξ) = 0 for all ξ ∈ T N . ·) = dH. its tangent space Tθ W is a lagrangian subspace of Tθ T ∗ M .
41 The HamiltonJacobi equation. Proof: We only have to show that the hamiltonian vector ﬁeld X is tangent to N . In particular. dim W = dim M = n. Since H is constant on N . the isotropic subspaces have dimension ≤ n. 97
.Chapter 4
The Hamiltonian viewpoint. therefore X ∈ TN.1 Theorem (HamiltonJacobi). then ω(X. A subspace λ of Tp T ∗ M is called isotropic if ω(X. then dHT N ≡ 0. Since ω is nondegenerate. 41.
Let ω be the canonical symplectic form on T ∗ M . If the hamiltonian H is constant on a lagrangian submanifold N . Y ) = 0 for all X. Y on λ. they are maximal isotropic subspaces.
. Gη is a lagrangian graph if and only if the form η is closed: Gη is lagrangian ⇐⇒ dη ≡ 0 Proof: Choose local coordinates q1 .98
4. Lagrangian graphs with zero cohomology class are the graphs of the exact 1forms: Gdf . (4.
∂pi ∂qj
ω(Ei . In fact. Applying ω = dp ∧ dq. A basis of the tangent space to the graph Gη is given by Ei =
∂ ∂qi . the hamiltonian viewpoint. with η = df and f : M → R a smooth function.1)
where ηx is a 1form on M . qn of M . . dx u) = k. 41.
−
∂pj ∂qi
dqj ∧ dqi . Then η(q) = k pk (q) dqk .
then ωT Gη ≡ 0 ⇐⇒ dη ≡ 0. R) to each lagrangian graph Gη . A lagrangian graph is a lagrangian graph submanifold.
Thus. 98
. (HJ)
Thus a smooth solution of the HamiltonJacobi equation corresponds to an exact invariant lagrangian graph. ηx )  x ∈ M } ⊂ T ∗ M. we can associate a cohomology class [η] ∈ H 1 (M.2 Lemma. which are of the form Gη = { (x. These are called exact lagrangian graphs. ∂pk ∂ k ∂qi ∂pk
. Some distinguished ndimensional manifolds on T ∗ M are the graph submanifolds. . The HamiltonJacobi equation for autonomous hamiltonians is H(x. . Ej ) = Since dη =
i<j ∂pi ∂qj
−
∂pj ∂qi . u : M → R.
and write u u(y) − u(x) ≤ Φk (x. y) are dominated. we get that u is Lipschitz. dx u) ≤ k at any diﬀerentiability point x of
Proof: 1. 1. We have that u(y) − u(x) ≤ Φc (x. y) for all x. x) and v(x) = −Φk (x.
L + k if
We say that a function u is dominated by L + k. Changing the roles of x and y. a family of dominated functions is equicontinuous. We have that u(y) − u(x) ≤
γ
L(γ. then u is Lipschitz with the same Lipschitz constant as Φc . In particular. dx u) ≤ k. If u u. dominated functions. 42. L + k then H(x.1 Lemma. v)  v ∈ Tx M }. 2. Since H(x. where A is a Lipschitz constant for Φc . dx u) = sup{ dx u · v − L(x.
The triangle inequality implies that the functions u(x) = Φk (y. y).42. y ∈ M. 99
. y). This implies that dx u · v ≤ L(x.
99
42
Dominated functions. for any y ∈ M . y) ≤ A dM (x. v) + k for all v ∈ Tx M when u is diﬀerentiable at x ∈ M . 2. If u L + k. then H(x. γ) + k ˙
for all curves γ ∈ C(x.
(4. Equation dx u = Lv (x.5 Remarks. then they are diﬀerentiable at any point which is not at the (backward) (resp. 2. we say that an absolutely continuous curve γ : [a. then u is diﬀerentiable at γ(0).t] ). 0] → M or γ : [0. Fix x0 ∈ M . R) such that f (x0 ) = 0 and H(dfk ) < k (cf. then H(x.e. v(x) = −Φc (x. Then u < +∞. b] → M realizes u. In particular. ε[→ M realizes u. then dx u = Lv (x. x) (resp.2)
Observe that such realizing curves must be global minimizers. For k > c(L) choose fk ∈ C ∞ (M. the hamiltonian viewpoint.4 Proposition. If γ :]−ε. if u(γ(t)) − u(γ(s)) = AL+k (γ[s. dx u) = k. 100
. γ(0)) of any a. since the functions u(x) = Φk (p. 44. If u : M → R is diﬀerentiable. dx u) ≤ c(L) for a.3 Deﬁnition. 1. p)) are dominated. they are semistatic. 42. Let u(x) := lim supk→c(L) fk (x). Given a dominated function u L + k. ˙ 42.4). x ∈ M . Suppose that u L + k. u is Lipschitz and H(x. forward) (L + k)cut locus of p. γ(0)) means that the tangent vector ˙ (x.
42.c. for all a ≤ s ≤ t ≤ b. In particular for k = c(L). curve γ realizing u is sent by the Legendre ˙ transform to dx u.2 Exercises: 1. γ(0)) and H(x. dx u) ≤ k ⇐⇒ u L + k. If γ :] − ε.
2.
42. ε[→ M realizes u and u is diﬀerentiable at x = γ(0).
1. The following proposition shows that we actually get a solution of (HJ) if there are (semistatic) curves which realize a dominated function u. 2.100
4.
Then. 0)) − u(x) ≤ A (0). γ(0)) · w . Hence (4. ˙ lim sup 1 u(x) − u(η(s. Observe that the energy E(x.
Proof: 1. t) be a variation of γ ﬁxing the ∂ endpoints γ(−ε). i. s
Similarly.ε] . 0)) − u(x) = ≤
u(η(s.4)
From (4. dx u). 0)) − u(x) ≥ Lv (x. s
s→0
Hence lim inf
s→0 1 s
u(η(s. ˙ 101
.
0
A (0) = Lv ξ0 + −ε where ξ(t) :=
1 s ∂ ∂s η(0. 0)) − u(γ(−ε)) + u(γ(−ε)) − u(γ(0)) [ A(s) − A(0) ] . t) = γ(t) and ∂s η(0.
101
3. γ(0)). Deﬁne
0
A(s) :=
−ε
∂ L( ∂t η(s. if B(s) := AL+k η(s. γ(0)) = H(x.7. t)) + k dt. we show that any semistatic orbit realizes some dominated function. γ(0)) · w.4) we get that u is diﬀerentiable at x = γ(0). ·)[0. ˙
(4.
Lx −
−ε
d dt Lv
ξ dt = Lv (x. then u(γ(ε)) − u(η(s. Let w ∈ Tx M and let η(s. γ(0)) · w.2). t). ˙
Also
1 s 1 s
u(η(s.e. that the semistatic orbits have energy c(L). 0)) − u(γ(ε)) + u(x) ≤ B(s) − B(0). Thus we obtain another proof for Σ ⊂ E −1 {c}. γ(ε) such that η(0. 0) = w. integrating by parts and using the EulerLagrange equation (EL). dominated functions. and dx u = Lv (x.3)
lim sup 1 u(η(s. In proposition 4˙ 9. 0)) ≤ B (0) = −Lv (x.
where we used that u
s→0
L + k and (4.42.3) and (4.
γ(0)) · w ˙ for all w ∈ Tx M. the hamiltonian viewpoint. ˙ ˙ Hence H(x. γ(s)) + k ds.1. v) } ≥ k.
v∈Tx M
102
. ˙
then dx u · γ(0) = L(γ(0). since u L + k.3 shows that dx u · w ≤ Lv (x. γ(0)) + k. Now assume that γ :] − ε. 0] → M realizes u and that u is diﬀerentiable at x = γ(0). Since
0
u(γ(0)) − u(γ(t)) = AL+k (γ[t. The same argument as in 4.
Applying this inequality to −w and combining both inequalities we get that dx u = Lv (x. dx u) = sup { dx u · v − L(x. γ(0)).0] ) =
t
L(γ(s).
2.102
4. H(x. ˙ Now. dx u) ≤ k. by lemma 42.
4. as example ?? shows. weak solutions of the hamiltonjacobi equation.43. x). By the triangle inequality. Thus
1 u(x) − u(γn (s)) − AL+c (γn [s. Let γn ∈ C(y. x) satisﬁes H(x. On the other hand. If k = c(L) then u may be realized by an inﬁnite semistatic orbit as follows. then for any y ∈ M . So that δ(t) ≡ 0 and hence γ backwardrealizes u at x.t] ) − Φk (y.7 that there are no weakly diﬀerentiable subsolutions of (HJ) for k < c(L).s] ) ≤ Φc (y. This implies that
1 Φc (y.0] ) ≤ n .
103
43
Weak solutions of the HamiltonJacobi equation. Also δ(t) ≥ 0 and δ(T ) = 0.
We shall see in corollary 44. the function δ(t) = AL+k (γ[0. x ∈ M . x) ≤ AL+c (γn ) ≤ Φc (y.
for all −Tn ≤ s ≤ 0. When M is noncompact there may be diﬀerentiable solutions on k > c(L). γ(s)) + n . dx u) = k for a.4 we show that when M is compact the only energy level that supports a differentiable solution is k = c(L). 43. x = y. Since u is dominated by proposition 42. x) with AL+k (γ) = Φk (y.2. γn : [−Tn . γ(t)) is increasing. In the next proposition we show that when k ≥ c(L) there are always Lipschitz solutions of (HJ).5)
103
. x) + n . for all x ∈ M . γ(s)) ≤ AL+c (γn [−Tn .e. the function u(x) = Φk (y. by by Rademacher’s theorem [19].1.1 Proposition. x). in theorem 48. by proposition 35. If k ≥ c(L).
(4. Proof: Since u is Lipschitz. it is diﬀerentiable at Lebesguealmost every point. 0] → M be a Tonelli minimizer such that 1 Φc (y. it is enough to see that u is onesided realized at every point. there exists a ﬁnitetime global minimizer γ ∈ CT (y. If k > c(L).
0] → λ[−ε.3. γn  < A for all n.
104
. lim inf n Tn > 0.104
4. Let λ(t) = ˙ π ϕt (v). we get that λ realizes u. Then γn [−ε.0] in the C 1 topology. if x = y.5). We can assume that γn (0) → v ∈ Tx M and Tn > ε > 0. the hamiltonian viewpoint. Hence.
By proposition 32. Letting n → ∞ on (4.
k) contains an exact lagrangian graph. In exercise 42. dx f (v) − L(x.
We say that a function u : M → R is a subsolution of the HamiltonJacobi equation if H(x. dx f )
= inf{k ∈ R : there exists f ∈ C ∞ (M.
105
44
Lagrangian graphs. then c(L) =
f ∈C ∞ (M. then k ≥ c(L). This is a very geometric way of describing the critical value.
v∈Tx M
Since H(df ) ≤ k it follows that for all (x. lagrangian graphs. R) such that H(df ) < k}. If there exists a C 1 function f : M → R such that H(df ) ≤ k. Hence 44.R) x∈M
inf
sup H(x.1 Theorem. Proof: Recall that H(x. where H is the hamiltonian associated with L. v) ≤ k. Theorem 44. there is an elementary construction of a weak subsolution for k = c(L).2 and proposition 44. 105
.44.1 could be restated by saying that c(L) is the inﬁmum of the values of k ∈ R for which H −1 (−∞.1 is an immediate consequence of lemma 44. 44. v)}.4 below. We shall prove that for k > c(L) there is always a C ∞ subsolution of the HamiltonJacobi equation and for k < c(L) there are no (weakly) diﬀerentiable subsolutions. p) = max {p(v) − L(x. Theorem 44.2. dx u) ≤ k.2 Lemma. If M is any covering of a closed manifold. v) ∈ T M .
[19. γ) + k) dt = ˙
0 0
L(γ.106
4.
44. its α and β functions are translates of those for L. ˙ ˙
and thus k ≥ c(L).1. We proceed to regularize u. The static set Σ(L) is contained in the level set L = 0. for any C ∞ function ϕ : U → R with compact support. x). R) such that H(df ) < k.2. Let M be a riemannian covering of a compact manifold and suppose that supv≤k ∂L (x. x ∈ M.3]). Section 4. if γ : [0. By the triangle inequality.
Therefore. dx u) ≤ k for a. If H(du) ≤ c(L).e. H(dx u) ≤ c at any point x ∈ M where u(x) is diﬀerentiable.4 Proposition. v) − dx u(v) + c(L) ≥ 0 The new lagrangian L is positive. has the same minimizing measures as L. For any k > c(L) there exists f ∈ C ∞ (M. Fix q ∈ M and let u(x) := Φc (q. that is. v) := L(x. The next lemma completes the proof. The utility of a diﬀerentiable subsolution of the HamiltonJacobi equation can be seen in lemma 44.3 Remark. the hamiltonian viewpoint.2. By lemma 42. If u : M → R is weakly ∂x diﬀerentiable and H(x. [19]) it is diﬀerentiable at Lebesgue almost every point. γ) + k − dγ f (γ) dt ≥ 0. T ] → M is any absolutely continuous closed curve with T > 0. 44. then the lagrangian L can be replaced by the lagrangian L(x. v) < +∞. we have
T T
(L(γ. Since u is Lipschitz.2.5 Lemma. u L + c. by Rademacher’s theorem (cf. equation (4. 106
. Moreover it is weakly diﬀerentiable (cf.6) holds. Proof: Set c = c(L). 44.
At any point of diﬀerentiability of u. and H(dz u) = H(dρ(z) u) ≤ k. (b) If x ∈ M . H(y.44. Extend u(x) to U by u(z) = u ρ(z) . Then u(z) is also weakly diﬀerentiable. Let ψ : R → R be a C ∞ function such that ψ(x) ≥ 0. then H(x. p) < k + δ. 44. v) = 1 v2 + sin(x2 ) on R shows. Let Nε be the εneighbourhood of M in RN . Let ε > 0 be such that (a) The 3εneighbourhood of M in RN is contained in U . Extend the lagrangian to U by L(z. In general it may not be true as the ∂x lagrangian L(x. and ρ : U → M a C ∞ projection along the normal bundle. We can assume that M ⊂ RN . p) ≤ k and dRN (x. then the condition on ∂L follows. R) such that H(x. Let K : RN × RN → R be K(x. ε) and RN ψ(x) dx = 1.
Then the corresponding hamiltonian satisﬁes H(z. v) = L ρ(z). support(ψ) ⊂ (−ε. lagrangian graphs. (y. 2 The condition on M can be replaced by some bounds in the riemannian metric of M . then for all δ > 0 there exists f ∈ C ∞ (M. y) < ε.6 Remark. Deﬁne f : 107
. dx f ) < k+δ for all x ∈ M . Let U be a tubular neighbourhood of M in RN . see appendix ??. If M is a covering M → N of a compact manifold N and the lagrangian L on M is lifted from a lagrangian on N : L = ◦dρ. p ◦ dz ρ) = H ρ(z). Proof: We shall explain ﬁrst how to prove the proposition in the case in which M is compact and then we will lift the construction to an arbitrary covering M . we have that dz u = dρ(z) u ◦ dz ρ. y) = ψ(x − y). p ∗ for p ∈ Tρ(z) M . p) ∈ T ∗ RN = R2N . dz ρ(v)) +
1 2 ρ
107
v − dz ρ(v)2 .
dy u) K(x.108
4. y) dy. Since u(y) is weakly diﬀerentiable.
Now. y) dy. suppose that M is a covering of a compact manifold N with covering projection p.
for all x ∈ M . since dx f =
RN
u(y) ∂x K(x.
Then f is C ∞ on Nε . 108
. y) dy < k + δ . y) dy u dy. y) dy =
RN
K(x. y) dy u dy. Assume that N ⊆ RN . y) = −∂y K(x. Now. the hamiltonian viewpoint. We can regularize our function u similarly as we shall now explain.
we obtain dx f =
RN
K(x. dy u) < k + δ for almost every y ∈ suppK(x. Since K(x.
From the choice of ε > 0 we have that H(x. y). by Jensen’s inequality H(dx f ) ≤ H(dx f ) ≤
RN
H(x.
RN
(4. ·) and x ∈ M .6)
Hence −
RN
u(y) ∂y K(x. y) dy . Observe that ∂x K(x. for any C ∞ function ϕ : U → R with compact support (ϕ du + u dϕ) dx = 0 . y) dy is a probability measure.
Nε → R by f (x) =
RN
u(y) K(x. For x ∈ M let x be the projection of x to N and let µx be the Borel probability measure on N deﬁned by ϕ dµx =
N RN
(ϕ ◦ ρ)(y) K(x.
we obtain Mather’s alpha function: 44.30) we have that α(κ) = c(L − ω0 ). Let µx be the Borel probability measure on M uniquely deﬁned by the e conditions: supp(µx ) ⊂ {y ∈ M : dM (x.5 with lemma 44. y) < ε} and p∗ µx = µx . Then e e we have d dy ϕ dµx (y). lagrangian graphs. There are no weakly diﬀerentiable subsolutions of (HJ) for k < c(L). Then the support of µx satisﬁes supp(µx ) ⊂ {y ∈ N : dN (x.44. Hence. ω(x)).8 Corollary. By equality (2. If we consider lagrangian graphs with other cohomology classes.7 Corollary. e Combining lemma 44. it suﬃces to show that c(L − ω0 ) = inf sup H(x.7)
109
. y) < ε}. ω(x)). e
M
satisﬁes H(dx f ) < k. Then the same arguments ∂x as above show that f (x) = u(y) dµx (y). α(κ) = inf sup H(x.
[ω]=κ x∈M
In particular the critical value of the abelian cover c0 = minκ α(κ) is the inﬁmum of the energy levels which contain a lagrangian graph of any cohomology class in is interior. The condition (b) above is now granted by the bound on ∂L .
[ω]=κ x∈M
(4. we obtain: 44. ϕ dµx = b e e dx M M for any weakly diﬀerentiable function ϕ : M → R. Proof: Let us ﬁx a closed one form ω0 such that [ω0 ] = κ.2.
109
for any continuous function ϕ : N → R. If M is compact.
110
4.
110
. p + ω0 (x)). equality (4.7) is now an immediate consequence of theorem 44. Since all the closed one forms in the class κ are given by ω0 + df where f ranges among all smooth functions.
It is straightforward to check that the hamiltonian associated with L−ω0 is H(x. the hamiltonian viewpoint.1.
0) lies on the exterior of Σx and (x.45. p) is Hp (x. i. p)) = 0.
Let N be a 2n+1 smooth manifold and let α be a nondegenerate 1form.
Thus a reparametrization of ψt preserves Θ if and only if its vector ﬁeld 1 is a constant multiple of Y = Θ(X) X. ∀w ∈ Tx N }. p) is the tangency point of a tangent line to Σx passing through (x. Observe that
d ∗ dt ψt Θ
= LX ΘΣ = iX dΘ + d iX Θ = dHΣ + dΘ(X) = dΘ(X)Σ . Observe that Θ(X) = p · Hp .
45.
111
45
Contact ﬂows. w) = 0. dim ker dα ≡ 1. where ker dx α := { v ∈ Tx N  dx α(v.e. p). 0). the sets Σx := Tx M ∩ Σ have compact convex interiors.
¿Tambi´n se tiene desigualdad estricta cuando M no es come pacta?
Proof: Let Σ = [H ≡ k]. The hamiltonian ﬂow on the energy level [H ≡ k] is a reparametrization of a Θpreserving ﬂow if and only if k > c(L). etc x ∈ N. Such reparametrization exists if and only if Θ(X) = 0 on Σ. we have that Θ(X(x. α ∧ (dα)n is a volume form for N . Then if the point (x. contact ﬂows. Thus Θ(X) = 0 on Σ implies that 111
. ∗ Since H is convex. Deﬁne a vector ﬁeld Y on N by Y (x) ∈ ker dx α.1 Proposition. The outward normal vector to Σx at (x.
By theorem 44. by theorem 44. Thus. there exists f ∈ C ∞ (M. Moreover. then Θ(X) > 0. Then
112
. p) := H(x.1. Then the energy level [H ≡ k] contains the zero section. then Θ(X) > 0. k ≥ c(L). R) with H(df ) < k. the zero section of T ∗ M lies in H < k. Now suppose that k > c(L). Let H be the hamiltonian of L and X its hamiltonian ﬂow. Let L := L − df . Since the zero section is the derivative of a constant function. Deﬁne the new convex hamiltonian H(x. if the zero section lies inside H < k.112
4.1. p + dx f ). the hamiltonian viewpoint. if X is its hamiltonian vector ﬁeld.
Their action potentials are related by Pk (x. and the same action functional on closed curves and invariant measures.1 Theorem. y) = inf
γ∈C(x. the subsolution H(df ) < k obtained in proposition 44.y)
lF (γ). if f ∈ C ∞ (M. This allows to borrow theorems from Finsler geometry. 30] If k > c(L) then the lagrangian ﬂow on the energy level E ≡ k is a reparametrization of the geodesic ﬂow of a Finsler metric on its unit tangent bundle. deﬁne its Finsler length as lF (γ) = F (γ). R) is such that H(df ) < k. the Finsler lagrangian F can be taken to be F (x.
In this section we prove that if k > c(L) then the EulerLagrange ﬂow on the energy level E ≡ k is a reparametrization of the geodesic ﬂow on the unit tangent bundle of a Finsler metric. We give ﬁrst a lagrangian proof and afterwards a hamiltonian proof. finsler metrics. y) + f (y) − f (x). Moreover. [14. v) + k − dx f (v) 113
. Deﬁne the Finsler distance as DF (x. y) = Φk (x.46. then L + k > 0. √ Given a Finsler metric F and an absolutely continuous curve γ. These lagrangians have the same energy function and equivalent variational principles.
113
46
Finsler metrics. minimizing orbits. Hence they have the same lagrangian ﬂow. the deﬁnition does not depend on the parametrization of the curve. When k > c(L). v) = L(x. with a ﬂavor of symplectic geometry. ˙
Observe that since the Finsler metric is homogeneous of degree one.
46.4 can be used to replace the lagrangian L by the lagrangian L = L − df .
y) + f (y) − f (x). By the homogeneity of F . R) such that H(df ) < k. / M(x) = { y ∈ M  Pk (x. Let L = L − df . we only need to prove that suﬃciently small EulerLagrange solutions with energy k are √ geodesics of F . Since k > c(L) = c(L − df ). So it is enough to prove that suﬃciently small orbits with energy k are global minimizers. and then Φk (x. Thus Φk − ∆f = DF .114
4. y ∈ M.
Then M(x) is a neighborhood of N (x) with M(x) ⊂ N (x).8) implies that any √ (L + k)global minimizer is a geodesic for F . v) = − inf L(x. with energy k and joining x to y. for any x = y there exists a global minimizer on C(x.1. Let Pk be the action potential for L and let ε = inf{ Pk (x. by proposition 35.
on E(x.8)
0 · v − L(x. Observe that H(x. The equality (4. we can restrict the curves in the deﬁnition of DF to those with energy k. Proof: By theorem 44.1. By the triangle inequality. If k > − inf L then f can chosen f ≡ 0. To show that the lagrangian ﬂow on E ≡ k is√ reparametrization a of the geodesic ﬂow on the unit tangent bundle of F .
v∈Tx M
So that if k > − inf L. z)  z ∈ N (x) } > 0. y) = DF (x. deﬁne a Finsler metric on T M by F = L+k−df on E −1 {k} and extend it by homogeneity. 0) = max
v∈Tx M
for all x. L + k − df > 0. y) for L + k − df which √ has energy k. So that the zero section M × 0 ⊂ E −1 {k} ⊂ T M √ Also. y) <
ε 2
}. v) = k. Fix x ∈ M and a small neighbourhood N (x) of x such that for all y ∈ N (x) there exists a unique EulerLagrange solution contained in N (x). if k > c(L) then there exists f ∈ C ∞ (M. we can choose f ≡ 0. Then we can . v). any (L + k)global minimizer joining x to a point y ∈ 114
. the hamiltonian viewpoint.
(4. Observe that if H(df ) < k then k > c(L) > e0 .
Observe that f ∗ ◦ Lf (p) = f (p) · p − f (p) =
d dt f (tp) t=1 − f (p)
= 2 f (p) − f (p) = f (p). Let f ∈ C 2 (M. 115
. We say that a function f : Rn → R is a Finsler energy if f (x) > 0 when x = 0 and f is positively homogeneous of order 2. then. Proof: Suppose that H −1 {k} = G−1 { } = Σ. these derivatives are nonzero. If p ∈ Σ. finsler metrics. ·) So that XH = λ(p) XG . Thus dp H = λ(p) dp G for some λ(p) > 0. Moreover. By proposition 35. λ > 0. Hence all the small solutions contained in N (x) joining x to points y ∈ M(x) are global minimizers.
115
M(x) must be contained in N (x).
So that f ∗ > 0. Since Σ is a regular energy level.46. writing v = λ w. ω(XH .1 such minimizer exists.
v w
Let Lf be the Legendre transform of f . then dp H = ker dp G = Tp Σ. First we need a hamiltonian characterization of Finsler lagrangians. R) be strictly convex and superlinear. Observe that L is a Finsler lagrangian if and only if it is a Finsler energy on each tangent space Tx M 46. Now we shall give a hamiltonian proof of theorem 46.3 Lemma. since H and G are convex. then their hamiltonian vector ﬁelds are parallel on Σ.2 Lemma. they are positive on vectors pointing outwards Σ. ·) = dp H = λ(p) dp G = λ(p) ω(XG . 46. Also. we have f ∗ (λ p) = max λp v − f (v) = max λ2 p w − λ2 f (w) = λ2 f ∗ (p). Then f is a Finsler energy if and only if its convex dual f ∗ is a Finsler energy. Proof: If f is homogeneous. Thus the argument above shows that if f ∗ is homogeneous then so is f . then f ∗∗ = f . Since f is strictly convex and superlinear. If two convex hamiltonians have a common regular level set Σ.1.
3 that the hamiltonian ﬂows of G and H−1 {k} coincide up to reparametrization on the energy level G−1 { 1 } = H−1 {k}.2.4 the orbits of the EulerLagrange ﬂow of L on E −1 {k} are reparametrization of unit speed geodesic of G∗ .
Hp · p = Θ(X) > 0. By lemma 46. p) > 0.
If X is the hamiltonian ﬂow of G. Let H(x. 4 The Legendre transforms LG (x. The result follows because the EulerLagrange ﬂow of F is the hamiltonian ﬂow of its energy function on T M with respect to the symplectic form L∗ (ω). p) for some λ(x. Since G(x. by theorem 44. λp
+ sw)
s=1
=
d 2 ds λ G(x. Hp ) on H−1 {k} satisfy Gp · w = 0 = Hp · p. then dλp G(w) =
d ds G(x. k[) contains the zero section of T ∗ M .1: If k > c(L). Gp ) and LH (x. p + dx f ). F where ω is the canonical symplectic form on T ∗ M . p) = λ LH (x. then the orbits of its lagrangian ﬂow are reparametrizations of the unit speed geodesics of F .1. R) with H(df ) < k. p). Gp · p = 2 G(p) =
1 2 ∗ for all w ∈ Tp (TX M ∩ H−1 {k}). it follows from lemma 46. the convex dual G∗ of G is a Finsler metric on T M . def
> 0.1). λp) = λG(x. ·) = dp G. Let G be a Finsler hamiltonian. p) = H(x. Then H−1 (] − ∞. p
+ sw) λ
s=1
= λ dp G(w).
(4. Also LH conjugates the hamiltonian ﬂow on H−1 {k} to the lagrangian ﬂow of L on E −1 {k}. p) for all λ ≥ 0. Hamiltonian proof of theorem 46. the hamiltonian viewpoint.4 Lemma. λp) = λ2 G(x. 116
. there exists f ∈ C ∞ (M. Deﬁne a new hamiltonian G : T ∗ M −M ×0 → R by G ≡ 1 on H−1 {k} and G(x. p) = (x. Proof: We prove it for a Finsler hamiltonian G. then ω(X(p). If F is a Finsler lagrangian. 1 Since by deﬁnition G−1 { 4 } = H−1 {k}.116
4.9)
So that LG (x.
46. Therefore X(λ p) = λX(p). By 4 lemma 46. p) = (x.
(by 45.
Hp = 2 Θ(X) Gp . p+dx f ) are related by h ◦ φt = ψt ◦ h. finsler metrics. p) = H(x. We now compute G∗ .
117
.46. From (4. p + dx f ). p) = (x.9). On the other hand. Hp ) = G∗ (x. Gp ) = 4 Θ(X)2 G(p) = Θ(X)2 = (p · Hp )2 = (Lv · v)2 = (L + k)2 . thus G∗ (x.
117
Let h : T ∗ M → T ∗ M be the map h(x. a lagrangian energy level [E = k] with k > e0 always contains the zero section. Thus. the hamiltonian ﬂow of H is conjugate to a Finsler hamiltonian ﬂow. Then the hamiltonian ﬂows φt of H and ψt of Hdf (x. 2 Θ(X) Gp ) = 4 Θ(X)2 G∗ (x. A hamiltonian level set can be made a Finsler level set if and only if it contains the zero section.
Paternain and M. each leaf W s (x. Let N be the universal covering of N . p) intersects each ﬁber of the ﬁbration πΣ : Σ → N at just one point. If the energy level E−1 (k) is Anosov.p) is in fact a diff feomorphism and W s (x. The foliation W s is also transverse to the ﬁbration πΣ : Σ → N .
47
Anosov energy levels. πW s (x. On the other hand it is well known that the weak stable leaves of an Anosov 118
def
. p) ∈ Σ the map πW s (x. f is a covering map. W s (x. p) is the graph of a one form. Consequently. We also have a ﬁbration by (n − 1)spheres πΣ : Σ → N . e In other words. [7]) implies that for every (x. 47.118
4. Paternain proved in [57] that Σ must project onto the whole manifold N and that the weak stable foliation W s of φ∗ is transverse to the ﬁbers t of the ﬁbration by (n − 1)spheres given by πΣ : Σ → N. Since the ﬁbers are compact a e result of Ehresmann (cf. e Let W s be the lifted foliation which is in turn a weak stable foliation for the hamiltonian ﬂow of H restricted to Σ. Since N is simply connected.p) : W s (x.1 Theorem. Proof: Suppose that the energy level k is Anosov and set Σ = H−1 (k). the hamiltonian viewpoint. Let π : T ∗ N → N denote the canonical projection. p) is simply connected. p) → N .P. G.
An Anosov energy level is a regular energy level on which the ﬂow φt is an Anosov ﬂow. then k > cu (λ). Let Σ denote the energy level k of the lifted hamiltonian H.
It is proved in [15] the following 47.47. Given a convex superlinear lagrangian L. O.2 Theorem. The theorem now follows from lemma 44. it follows that each leaf W s (x. This theorem is an extension to the Hamiltonian setting of a result of R. let Ae be the set of φ ∈ C ∞ (M ) such that the ﬂow of H + φ is Anosov in (H + φ)−1 (e) and let Be be the set of φ ∈ C ∞ (M ) such that (H + φ)−1 (e) contains no conjugate points. Proof: If an orbit does not have conjugate points then there exist along it two subbundles called the Green subbundles. k < cu (L) and ε > 0 there exists a smooth function ψ : N → R with ψC 2 < ε and such that the energy level k of L + ψ possesses conjugate points. they are contained in the same energy level as the orbit and they do not intersect the vertical subbundle (cf. the energy level k of λ + ψ has no conjugate points.1. 47. lagrangian and they have dimension n = dim N . Paternain [54] have shown that Ae is contained in Be . k + ε) the energy level k is Anosov. If k is a regular value of the energy such that k < e. and M. For e ∈ R.3 Corollary. They have the following properties: they are invariant. anosov energy levels. Since any closed one form in the universal covering must be exact. p) is an exact lagrangian graph.
119
energy level are lagrangian submanifolds. Theorems 47. Moreover. On the other hand G. The main result in [15] says that in this case the energy level k of λ must be Anosov thus contradicting theorem 47. Ruggiero for the geodesic ﬂow [64].1 have as corollary: 47. Proof: Suppose now that there exists > 0 such that for every ψ with ψC 2 < . [13]).2 and the fact that by structural stability there exists ε > 0 such that for all k ∈ (k − ε. If k is 119
. As is well known Ae is open in C k topology and Be is closed. then the energy level k has conjugate points.2 and 47. The interior of Be in the C 2 topology is Ae .4 Proposition.
120
4. the hamiltonian viewpoint.
a regular value of the energy with k < e, then π(E−1 (k)) is a manifold with boundary and at the boundary the vertical subspace is completely contained in the energy level. Therefore the orbits that begin at the boundary must have conjugate points, because at the boundary two ndimensional subspaces contained in the energy level (which is (2n − 1)dimensional) must intersect.
120
48. the weak kam theory.
121
48
The weak KAM Theory.
In the rest of this chapter we develop the theory of global weak KAM solutions, discovered by Albert Fathi. Recall that for an autonomous hamiltonian H : T ∗ M → R, the HamiltonJacobi equation is H(x, dx u) = k, (HJ)
where u : U ⊆ M → R. Here we are interested on global solutions of (HJ), i.e. u : M → R satisfying (HJ). It may not be possible to obtain a smooth global solution of (HJ). Instead, for certain values of k, we shall ﬁnd weak solutions of (HJ), which are Lipschitz. By Rademacher’s theorem [19], a Lipschitz function is Lebesgue almost everywhere diﬀerentiable so that (HJ) makes sense in a.e. point. In fact, we have seen in corollary 44.7 that there are no weakly differentiable1 global solutions for k ≤ c(L). In theorem 48.4 we shall see that when M is compact, there are no C 1+Lip solutions unless k = c(L). In proposition 49.7 we show that there are always Lipschitz solutions for k = c(L), and in proposition 410.2 we show that when M is noncompact there are Lipschitz solutions for k > c(L). Given a dominated function u L + c deﬁne the sets Γ+ (u) := { v ∈ Σ+  u(xv (t)) − u(xv (0)) = Φc (xv (0), xv (t)), ∀t > 0 }, 0 Γ− (u) := { v ∈ Σ−  u(xv (0)) − u(xv (t)) = Φc (xv (t), xv (0)), ∀t < 0 }, 0 Γ+ (u) :=
t>0
ϕt Γ+ (u)
,
Γ− (u) :=
ϕt Γ+ (u) ,
t<0
where xv (t) = π ϕt (v). We call Γ+ (u) (resp. Γ+ (u)) the basin of u and π Γ+ (u) \ Γ+ (u) (resp. π Γ+ (u) \ Γ+ (u) ) the cut locus of u. 0 0 48.1 Deﬁnition. A function u− : M → R is a backward weak KAM solution of (HJ) if
1
In particular, by Rademacher’s theorem, there are no Lipschitz global solutions.
121
122
4. the hamiltonian viewpoint.
1. u−
L + c.
2. π Γ− (u− ) = M . 0 A function u+ : M → R is a forward weak KAM solution of (HJ) if 1. u+ L + c.
2. π Γ+ (u+ ) = M . 0 48.2 Remark. From the domination condition it follows that u is Lipschitz and that the curve γ is semistatic. From proposition 42.4, at an interior point x of such a curve γ, the function u is diﬀerentiable and H(x, dx u) = c. Moreover, item 2 in proposition 42.4 shows that if u is diﬀerentiable at an endpoint of a curve γ, then H(x, dx u) = c. By Rademacher’s theorem [19], u is diﬀerentiable at (Lebesgue) almost every point in M . So that u is indeed a weak solution of the HamiltonJacobi equation for k = c(L). 48.3 Theorem. If u ∈ S+ (resp. u ∈ S− ) is a weak KAM solution, then 1. u is Lipschitz and hence diﬀerentiable (Lebesgue)almost everywhere. 2. u L + c.
3. H(x, dx u) = c(L) at any diﬀerentiability point x of u. 4. Covering Property: π(Γ+ (u)) = M . 0 5. Graph Property: π : Γ+ (u) → M is injective and its inverse is Lipschitz, with Lipschitz constant depending only on L. 6. Smoothness Property: u is diﬀerentiable on Γ+ (u) and its derivative dx u is the image of (πΓ+ (u) )−1 (x) under the Legendre transform L of L. In particular, the energy of Γ+ (u) is c(L). 0 122
48. the weak kam theory.
123
Proof: Items 2 and 4 are the deﬁnition of u ∈ S+ . Item 1 follows from proposition 42.1.1 and Rademacher’s theorem [19]. Item 3 follows from proposition 42.4.2 and the fact that by item 4, u is onesided realized at every point. Item 6 follows from proposition 42.4 and remark 42.5. We prove item 5. Let (z1 , v1 ), (z2 , v2 ) ∈ Γ+ (u) and suppose that dT M (v1 , v2 ) > K dM (z1 , z2 ), where K is from lemma 38.2 and the A that we input on lemma 38.2 is from lemma 32.3. Let 0 < ε < ε1 , (with ε1 from lemma 38.2) be such that ϕ−ε (zi , vi ) ∈ Γ+ (u). Let yi = xvi (ε), i = 1, 2, then u(yi ) = u(xi ) + Φc (xi , yi ), i = 1, 2. Then lemma 38.2 implies that Φc (x1 , y2 ) + Φc (x2 , y1 ) < Φc (x1 , y1 ) + Φc (x2 , y2 ). Adding u(y1 ) + u(y2 ) and using that u L + c, we get that
u(x1 ) + u(x2 ) ≤ Φc (x1 , y2 ) + u(y2 ) + Φc (x2 , y1 ) + u(y1 ) < Φc (x1 , y1 ) + u(y1 ) + Φc (x2 , y2 ) + u(y2 ) = u(x1 ) + u(x2 ), which is a contradiction. This proves item 5.
fig. 1: graph property.
123
Observe that the measure ν := ξ∗ (µ) is holonomic and L + k = du on supp(ν). If u ∈ C 1+Lip (M. Γ (u) = M and by 48.10)
Then k = c(L) and u is a weak KAM solution u ∈ S− ∩ S+ . dx u) = k.
t
u ψt (x) = u ψs (x)) +
s
d L(ψτ (x). Proof: From (4.11).6. Let µ be and invariant measure for the ﬂow ψt .4 Theorem (Fathi [23]). L + k − du ≥ 0. dτ ψτ (x)) + k dτ. v) = k. if u ∈ S+ ∩ S− . ξ is Lipschitz. Conversely.e. Then (L + k) dν = (du) dν = 0.
(4. u ∈ S− ∩ S+ .12) From (4.5.3. From (4. Thus it can be integrated to obtain a ﬂow ψt on M . This implies that c(L − du) = c(L) ≥ k.
48. R). then u ∈ C 1+Lip . the hamiltonian viewpoint.12). Conversely.11).11)
The strict convexity of L on Tx M implies that the maximum is attained at a unique vector ξ(x) ∈ Tx M . u is diﬀerentiable and −1 (d u) = ξ(x). Moreover.
Hence k ≤ c(L) and thus k = c(L). by 48. i. Then by (4. The implicit function theorem implies that the vector ﬁeld x → ξ(x) is Lipschitz.3. ∀x ∈ M. M is compact and H(x.10) we get that
v∈Tx M
max dx u · v − L(x. ∀x ∈ M.124
4. L x
124
. if u ∈ S− ∩ S+ . ∀s < t.
(4. u L + k and the ﬂow lines of ψt realize L + k. (4.
49. construction of weak kam solutions
125
49
Construction of weak KAM solutions
In this section we present three ways to construct weak KAM solutions: when the Peierls set is nonempty (in remark 49.3.4), when the Peierls barrier is ﬁnite (in proposition 49.2), and the general case (in proposition 49.7). In the horocycle ﬂow (example 58), the Peierls barrier is ﬁnite, but the Peierls set is empty. In example 57, hc = +∞ and also P= . When M is compact we characterize all weak KAM solutions in terms of their values on each static class. We begin by observing that 49.1 Lemma. 1. If U ⊆ S− is such that v(x) := inf u∈U u(x) > −∞, for all x ∈ M ; then v ∈ S− . 2. If U ⊆ S+ is such that v(x) := supu∈U u(x) < +∞, for all x ∈ M ; then v ∈ S+ . Proof: We only prove item 1. Since u
u∈U u∈U
L + c for all u ∈ U, then (4.13)
v(y) = inf u(y) ≤ inf u(x) + Φc (x, y) = v(x) + Φc (x, y).
Thus v L + c. Let x ∈ M and choose un ∈ U such that un (x) → v(x). Choose wn ∈ Γ− (un ) ∩ Tx M . Since by lemma 32.3 wn  < A, we can assume that wn → w ∈ Tx M . By lemma 42.1.1, all the functions u ∈ U have the same Lipschitz constant K as Φc . For t < 0, we have that v(xw (t)) ≤ lim inf un (xwn (t)) + K dM (xw (t), xwn (t))
n n
= lim inf un (x) − Φc (xw (t), x) + K dM (xw (t), xwn (t)) = v(x) − Φc (xw (t), x) ≤ v(xw (t)), Hence w ∈ Γ− (v). 125 because v L + c.
126
4. the hamiltonian viewpoint.
49.a
Finite Peierls barrier.
49.2 Proposition. If hc < +∞ and f : M → R is a continuous function. Suppose that v− (x) := inf f (z) + hc (z, x) > −∞,
z∈M
v+ (x) := sup f (z) − hc (x, z) > −∞.
z∈M
Then v− ∈ S− and v+ ∈ S+ . Proof: We only prove that v− ∈ S− . By lemma 49.1 it is enough to prove that the functions u(x) → hc (z, x) are in S− for all z ∈ M . By proposition 37.1.4, u L + c. Now ﬁx x ∈ M . Choose Tonelli minimizers γn : [Tn , 0] → M such that γn ∈ C(z, x), Tn < −n and
1 AL+c γn [Tn ,0] ≤ hc (z, x) + n .
By lemma 32.3, γn (0) < A for all n. We can assume that γn (0) → ˙ ˙ w ∈ Tx M . If −n ≤ s ≤ 0, then s > Tn and AL+c γn [Tn ,s] + Φc (γn (s), x) ≤ ≤ AL+c γn [Tn ,s] + AL+c γn [s,0] ≤ hc (z, x) +
1 n 1 ≤ hc (z, γn (s)) + Φc (γn (s), x) + n ,
n
for − n ≤ s < 0.
Taking lim inf n→∞ , we get that hc (z, xw (s)) + AL+c xw [s,0] = hc (z, x). Hence w ∈ Γ− (u). 49.3 Remarks.
126
49. construction of weak kam solutions
127
1. Observe that, since Φc (x, x) = 0, u L + c ⇐⇒ u(x) = inf u(z) + Φc (z, x).
z∈M
2. Item 49.3.1 implies that the function hc in proposition 49.2 can not be replaced by Φc . In fact, the function uz (x) = Φc (z, x) satisﬁes uz L + c, but in general u ∈ S− , if z is not properly / chosen. 3. For any z ∈ M the function uz (x) = hc (z, x) ∈ S− and vz (x) := −hc (x, z) ∈ S+ . 4. If p ∈ P then up (x) := Φc (p, x) ∈ S− , because Φc (p, x) ≤ hc (p, x) ≤ hc (p, p) + Φc (p, x) ≤ Φc (p, x). Similarly, vp (x) := −Φc (x, p) ∈ S+ .
49.b
The compact case.
In the next theorem we characterize all weak KAM solutions when M is compact. They are determined by their values on one point of each static class. Let Γ = P/dc be the set of static classes of L. For each γ ∈ Γ choose pΓ ∈ Γ and let P = { pΓ  Γ ∈ Γ }. We say that a function f : P → R is dominated (f L + c) if f (p) ≤ f (q) + Φc (q, p), for all p, q ∈ P. 49.4 Theorem. If M is compact, the maps { f : P → R  f
p∈P
L + c } → S− ,
f −→ uf (x) := inf f (p) + Φc (p, x), and { f : P → R  f L + c } → S+ ,
p∈P
f −→ vf (x) := sup f (p) − Φc (x, p), are bijective isometries in the sup norm. 127
128
4. the hamiltonian viewpoint.
Proof: We only prove it for f → uf . The domination condition f L + c implies that uf > −∞. Then remark 49.3.4 and lemma 49.1.1 imply that uf ∈ S− . The injectivity follows from uf (p) = min f (q) + Φc (q, p) = f (p)
q∈P
∀p ∈ P,
because f is dominated. To prove the surjectivity, let u ∈ S− and let f = uP . Given x ∈ M choose w ∈ Γ− (u) ∩ Tx M and let γ(t) = π ϕt (w). So that 0 u(x) − u(γ(t)) = AL+c (γ[t,0] ) = Φc (γ(t), x) for t < 0. (4.14)
Choose q ∈ π[αlim(v)] ⊂ P (by proposition 311.3), and tn → −∞ such n that γ(tn ) → q. Using t = tn on equation (4.14), in the limit we have that u(x) = u(q) + Φc (q, x). (4.15) Now take p ∈ P such that dc (p, q) = 0. Since u L + c, then
u(q) ≤ u(p) + Φc (p, q) ≤ u(q) + Φc (q, p) + Φc (p, q) = u(q). So that u(q) = u(p) + Φc (p, q). By the triangle inequality Φc (p, x) ≤ Φc (p, q) + Φc (q, x) ≤ Φc (p, q) + Φc (q, p) + Φc (p, x) = Φc (p, x). So that Φc (p, x) = Φc (p, q) + Φc (q, x). Combining equalities (4.15), (4.17) and (4.16), we have that u(x) = u(p) + Φc (p, x), 128 (4.17) (4.16)
then h(x. There is only one static class if and only if S− (resp. p) + Φc (p. x)
n
≤ f −g
0. Since uf P = f and ug P = g. u− (y) ≤ u− (p) + Φc (p.
Changing the roles of f and g we get that uf − ug 0 ≤ f − g 0 . y) = sup { u− (y) − u+ (x) }. from the domination we get u+ (p) ≤ u= (x) + Φc (x.5 Corollary. x). p). Now we see that f → uf is an isometry in the supremum norm 0. y).49.1. 129
. Adding these equations and using that u= (p) = u− (p). (Fathi [20]) If M is compact. construction of weak kam solutions
129
with p ∈ P. So that uf ≤ u.6 Corollary. x) − g(pn ) − Φc (pn . S+ ) is unitary modulo an additive constant. 49. Given x ∈ M . Since uf L + c.
u− ∼u+
u ∈S
Proof: If u+ ∼ u− and p ∈ P. using remark 49. then uf − ug 0 ≥ f − g 0 . This characterization of weak KAM solutions allows us to recover the following theorem: We say that two weak KAM solutions u− ∈ S− and u+ ∈ S+ are conjugate if u− = u+ on P and denote it by u− ∼ u+ . y). we have that u ≤ uf . But since u L + c and f = uP . 49.3. choose pn ∈ P such that ug (x) = limn f (pn ) + Φc (pn . we have that uf (x) − ug (x) ≤ lim inf f (pn ) + Φc (pn . we get u− (y) − u+ (x) ≤ Φc (x.
We call the functions of proposition 49. z) }
q∈P q∈P
(4. let u= (z) := −h(z. 1.
Taking inf p∈P and then supu= ∼u− we obtain sup
u+ ∼u−
u− (y) − u+ (x)
≤ h(x. γ(0))]
t<0
= lim [Φc (γ(t).
On the other hand. y) and u− (z) : = min u= (q) + Φc (q.19). y) + Φc (q.7 weak KAM Busemann functions.3. y) + Φc (q. resembling the constructions of Busemann functions in riemannian geometry.4. the hamiltonian viewpoint.130
4. Since u+ is dominated.1. x) − Φc (γ(t).18)
= min = min
q∈P
− h(q.
49. 130
.18) we get that u+ ∼ u− . z) − φc (q. If w ∈ Σ− (L) and γ(t) = xw (t). Σ+ = and Σ− = even when M is noncompact. u− (y) = 0 and hence u− (y) − u+ (x) = h(x. By proposition 310. we use another method to obtain weak KAM solutions. 49.
When hc = +∞. γ(0))]
t→−∞
is in S− .c
Busemann weak KAM solutions.19)
From remark 49. u± ∈ S± .3 and corollary 49. x) − Φc (γ(t). then uw (x) = inf [Φc (γ(t). y). z) } (4.7 Proposition. y). from (4. From (4.
x) + Φc (x. γ(t))]
t>0
= lim [Φc (γ(0). y) − Φc (γ(t). x). γ(t)) − Φc (x. where the last inequality follows from the triangle inequality applied to the triple (γ(s). x) ≥ Φc (y. x). Since u(y) = inf Φc (γ(t). Proof: We only prove item 1. By the triangle inequality. y). δ(t) ≤ Φc (γ(0). x) − Φc (γ(s). We start by showing that the function δ(t) = Φc (γ(t). γ(0)) − Φc (γ(t). p) = 0. x) − Φc (γ(t). x) + Φc (γ(s). γ(0))
t<0
= u(x) + Φc (x. y) − Φc (γ(t). If s < t. If w ∈ Σ+ (L) and γ(t) = xw (t). Let p = xv (t) and y ∈ M . γ(t)) ≥ 0. x) + Φc (x. 131
. γ(0))
t<0
≤ inf Φc (γ(t). then δ(t) − δ(s) = Φc (γ(t). p) + Φc (p. p) + Φc (p. γ(t). Since dc (x. x) = Φ(y. then Φc (y. γ(t)) − Φc (x. Let (x. then u L + c. x) − Φc (γ(s). Suppose that x ∈ P = . v) ∈ Σ and t < 0. x) + Φc (γ(s). x) ≥ Φc (y. 0) is increasing. x). construction of weak kam solutions
131
2. γ(t))]
t→+∞
is in S+ . γ(0)) = Φc (γ(t). then uw (x) = sup [Φc (γ(0).49. hence limt↓−∞ δ(t) = inf t<0 δ(t) and this limit is ﬁnite.
x) ≤ Φc (γ(n).
Since y → Φc (z. p) + Φc (p. x) − Φc (γ(n). xv (t))] = Φc (xv (t). x) = u(xv (s)) + AL+c (xv [s. x) = Φc (y. x) + n . for simplicity. xv (s)) + Φc (xv (s). We can assume that yn (0) → v ∈ Tx M .0] .
Now we prove that limn Tn = −∞.0] ) ≤ Φc (γ(n).0] ) ≤ Φc (γ(n).
(4. γ(0))
n
= lim Φc (γ(n). Suppose.0] )
for all s < 0. γ(0))
n
= u(xv (s)) + Φc (xv (s).
This implies that
1 AL+c (yn [s.
n
(4. 0] → M a Tonelli minimizer such that yn (Tn ) = γ(n). Then v ∈ Σ− (L). x) − Φc (γ(n). Now let x ∈ M \ P and choose yn : [Tn .
Hence Φc (y. yn  < A.0] ) = Φc (γ(t). x) + n . Observe that for Tn ≤ s ≤ 0 we have that Φc (γ(n). the hamiltonian viewpoint. ˙ ˙ Then AL+c (xv [t. yn (0) = x and
1 AL+c (yn [Tn .132
4. For y = γ(s) (and p = xv (t)).21)
We prove below that limn Tn = −∞.0] and ˙ 132
C1
. yn (t)) + n . Since yn (0) → v. y) is uniformly Lipschitz.3. x) − Φc (γ(s). x)
1 ≤ AL+c (yn [Tn . then yn [Tn . x) for lim inf Tn ≤ t ≤ 0. we have that u(x) − u(xv (t)) = lim [Φc (γ(s). x)
s→+∞
= AL+c xv [t.0] −→ xv [T0 . that limn Tn = T0 > −∞.t] ) ≤ Φc (yn (s).
for Tn ≤ s < t ≤ 0. yn (s)) + Φc (yn (s). we obtain that u(x) = lim Φc (γ(n).20)
By lemma 32. x). because v ∈ Σ− .
49.3 γ is ˙ bounded.21).8 Corollary. Since π(w1 ) = xv (T0 ) = ˙ ˙ p.3. ˙ w1 ∈ αlim(γ) ⊆ Σ. construction of weak kam solutions
133
hence γ(n) = yn (Tn ) → xv (T0 ) =: p.
133
. xv (T0 ) ∈ Σε . we can assume that limn γ(n) = (p.1 implies that xv (T0 ) ∈ Σ. then ˙ v ∈ Σ and hence x = π(v) ∈ π(Σ) = P.49. By lemma 311. Since by lemma 32. then lemma 38. Since Σ is invariant. From (4. S− = and S+ = . This contradicts the hypothesis x ∈ M \ P. w1 ).
22)
Γ− (u.22) are deﬁned in the whole ray ] − ∞.
410
Higher energy levels. αlim(v) = ).0] = Φk xv (t). xv (0) .e. then ωlim(v) = (resp. covering.
Deﬁne similarly Σ+ (k).
The method in proposition 49. Proof: Suppose that v ∈ Σ+ (k) = and that w ∈ ωlim(v) = . Then the functions u ∈ S± satisfy Lipschitz.18). x). 410. k > c(L) ≥ e0 . the hamiltonian viewpoint.
L + k deﬁne v ∈ T M  AL+k xv [t. Γ+ (u. Let Σ− (k) := For u v ∈ T M  AL+k xv [t. Let p = π(w) and choose s > 0 such that 134
. k) and S+ (k).3. 0]. If M is noncompact then Σ± (k) = for all k > c(L). if v ∈ Σ± (k). If we look for weak solutions of (HJ) with realizing curves deﬁned only on ﬁnite intervals then the action potential u(x) = Φk (p. v(x) = −Φk (p. (4. x) give such examples. Observe that the orbit of w can not be a ﬁxed point (i.0] = u xv (0) − u xv (t) . ∀t ≤ 0 .134
4.1 Proposition. Moreover. graph and smoothness properties analogous to theorem 48. w = 0) because by (3. Observe that we are requiring that the global minimizers in (4. If M is compact. k) = Let
S− (k) =
u : M → Ru
L + k and π Γ− (u. k) = M .7 allows us to obtain analogous weak KAM solutions on energy levels k > c(L) when M is noncompact. In the following lemma we show that there are not such weak KAM solutions when M is compact and k > c(L). ∀t ≤ 0 . then Σ± (k) = and hence S± (k) = for all k > c(L).
We complete the picture with the following: 410. Let γn ∈ CTn (x. If M is compact.1 proves the case M compact. Hence S± (k) = when M is compact. Let x ∈ M and yn ⊆ M such that dM (x.3 shows that dk (p. Since γn [0. Now we reproduce the proof of proposition 49. Let w ∈ Σ+ (k) = . p) = 0.1.23)
Then by lemma 32. γ(t)) − Φk (x.
C1
By corollary 44.
(4. γ(t))]
t→+∞
The limit exists because by the triangle inequality for Φk . and hence Tn → +∞.4. yn ) → +∞. Let v be a density ˙ point of {γn (0)}.2 Proposition.7. γ(t)) 135
. higher energy levels. Proof: Proposition 410. then S± (k) = for all k > c(L).410. k) ⊆ Σ+ (k).
135
q = π(ϕs (w)) = p. Then the same argument as in proposition 311. then S± (k) = for all k ≥ c(L). γn  < A. the function δ(t) = Φk (γ(0).3. Observe that if u+ ∈ S+ (k) then = Γ+ (u. We show that Σ+ (k) = 0.t] −→ xv [0.t] for all t > 0. from (4.23) we ˙ obtain that v ∈ Σ+ (k).7. If M is noncompact. γ(t)) − Φk (x. write γ(t) := xw (t) and uw (x) := lim [Φk (γ(0). This proves that Σ± (k) = when M is compact. yn ) + n . there are no weak solutions of (HJ) for k < c(L). yn ) be a Tonelli minimizer such that
1 AL+c (γn ) < Φk (x. q) + Φl (q. which contradicts proposition 21. q) = Φk (p. Assume now that M is noncompact.
Let v ∈ Tx M be a density point of yn (0). The triangle inequality implies that u L + k. (4.t] ) = u(x) + AL+k (xv [0. for all 0 < t < lim inf n Tn = +∞. γ(n)) + n .t] + Φk (yn (t). γ(n)) + AL+k (yn [0.Tn ] ) ≤ Φk (x. using (4. then lim inf n Tn = +∞.t] ). ˙ Since yn [0.t] −→ xv [0. yn (Tn ) = γ(n) and
1 AL+c (yn [0. γ(n)) ≤ AL+k yn [0. Now let x ∈ M and choose yn : [0. γ(n)) − Φk (xv (t). yn  < A. γ(n)) − Φk (x.3.
This implies that for all 0 < t < Tn .
is nondecreasing and it is bounded above by Φk (γ(0). γ(n)) + n . γ(n))
n n C1
= lim Φk (γ(0).t] .
136
. Since yn (Tn ) = γ(n) and by proposition4˙ 10. γ(n) → ∞. x).
1 Φk (x. the hamiltonian viewpoint.24)
By lemma 32.1. Tn ] → M a Tonelli minimizer such that yn (0) = x.136
4. γ(n)) ≤ Φk (x.24) we get that u(xv (t)) = lim Φk (γ(0).
For ε > 0 there exists K(ε) > 0 such that if u. x. where the inﬁmum is taken on all piecewise diﬀerentiable curves γ : [0. T ] → M with γ(T ) = x. R)) ⊆ LipK (M. then Ts u − Tt− v 0 ≤ u − v 0 + K(ε) s − t. t) }. Tt− u − Tt− v 0 ≤ u − v 0 . R) → C 0 (M. 5. v ∈ C 0 (M. Similarly.
t≥0
1. t > ε. Nahman and Roquejoﬀre Roquejoﬀre Through this section we shall assume that M is compact.
137
411
The LaxOleinik semigroup. γ(s)) + c ds } ˙
0
= min { u(y) + Φc (y. Papanicolaou and Varadhan [32] on tori Tn and later by Fathi [21] for compact manifolds.
3. v ∈ C 0 (M. R). 137
. u ∈ S− ⇐⇒ u is a ﬁxed point of the semigroup Tt−
t≥0 .
has a ﬁxed
2. z. For all ε > 0 there exists a Lipschitz constant K = K(ε) > 0 such that Tt− (C 0 (M. the laxoleinik semigroup.
z∈M
411. 4. R) deﬁned by Tt− u(x) = inf { u(γ(0)) +
γ y∈M t
L(γ(s).1 Proposition. R) for all t > ε.411. The LaxOleinik semigroup T − t≥0 is the semigroup of operators Tt− : C 0 (M. The unique c ∈ R for which the semigroup Tt− point is the critical value c = c(L). Tt− is a weak contraction: For all t ≥ 0 and all u. R) and − s. deﬁne Tt+ u(x) = max { u(z) − Φc (x.
The LaxOleinik semigroup was used to obtain weak KAM solutions by Lions.
If M is compact. t) }.
R).
n
where the last inequality follows from the deﬁnition of hc and the Lipschitz property for Φc (x.138
4. y ∈ M . Then u(x) ≤ u(z) + Φc (z.1. limt→+∞ Φk (x. t). x) ≤ u(y) + Φc (y. u ∈ S− . uniformly in x. 3. Let u ∈ S− . we get that u
z∈M z∈M
u(x) ≤ min u(z) + Φc (z. Using proposition 34. t) in proposition 34. But since u L + c.
Proof: 1.1. For t > 0 we have that u(x) = u(γ(−t)) + AL+c γ[−t. y. x).1. Given u ∈ C 0 (M. the hamiltonian viewpoint. let K = K(ε) be a Lipschitz constant for (x. L + c. Moreover. x). then u(x) ≤ u(y) + Φc (y. Let tn → +∞ be a sequence such that the limit limn ztn = y exists. x. Hence Tt− u = u for all t > 0. Now suppose that Tt− u = u for all t ≥ 0. x. x). y. t). t) = ±∞ when k = c(L).0] = u(γ(−t)) + Φc (γ(−t). t) → Φc (x.1.
Taking the inﬁmum on t ≥ 0. by proposition 49. x. For t > 0 let zt be such that u(x) = u(zt ) + Φc (zt . tn ) ≥ u(y) + hc (y. 138
. y. By proposition 34. t) on M × M × [ε. The existence of a ﬁxed point is given by item 2. 2.2. Then u(x) = u(y) + lim Φc (ztn . ∀z ∈ M. x. For x ∈ M take v ∈ Γ− (u) ∩ Tx M and let γ(t) := π ϕt (v). t). This implies that there are no ﬁxed points for Tt± when c = c(L). t) ∀t ≥ 0. x. Thus u(x) = minz∈M u(z) + hc (z. y. +∞[. x) ≤ min u(z) + hc (z.
1. Now let tn → 0 and choose a subsequence such that utn converges to some v in LipK (M. x. Changing the roles of x and y we get that Tt− u ∈ LipK (M.1. Fix s ∈ R and let Nn ∈ Z be such that limn Nn tn = s. t) − Φc (z. Then λTt− is a contraction whose image is in the compact space LipK (M. Let uλ. t) − Φc (z. x. x.
− − Ts v = lim TNn tn utn = lim utn = v. R)/R of continuous functions modulo an additive constant. utn is also a ﬁxed point for TNn tn .411. y). independent of λ or t. t > ε and x ∈ M . n n
139
. let z ∈ M be such that Tt− u(y) = u(z) + Φc (z. s)
≤ Tt− u(x) + u − v
0
+ K(ε) s − t. Choose z ∈ M such that − Ts u(x) = u(z) + Φc (z. y. y ∈ M and t > ε. Fixed Points for the LaxOleinik semigroup.3 We sketch here another proof of existence of a ﬁxed point for the LaxOleinik semigroup using properties 3 and 4 of proposition 411. with 0 < λ < 1 acting on the space C 0 (M. we get that in the C 0 topology. t)
− ≤ Ts u(x) + v(z) − u(z) + Φc (z. 411. R)/R . using proposition 411.2 Remark.
Changing the roles of u and v we get item 5. x. Observe that item 5 implies item 4.
139
x. x. Consider the semigroup λTt− . 5. Let u.t converges in LipK (M. Then Tt− u(x) ≤ u(z) + Φc (z. R). s. x. s). Let K = K(ε) be as in item 3. v ∈ C 0 (M. Another proof of the existence of a ﬁxed point in 411.t be a ﬁxed point for λTt− . Choosing a sequence λn → 1 and a subsequence such that uλn . Then Tt− v(x) ≤ v(z) + Φc (z. t) + Φc (z. R)/R . R)/R . one obtains a ﬁxed point ut for Tt− . y. t) ≤ Tt− u(y) + K(ε) d(x. t). t) ≤ u(z) + Φc (z. the laxoleinik semigroup.5. − Since by the semigroup property. y. R).
If M is compact. Proof: Write ht (x.
x∈M x∈M
This implies the lemma.
411.4 Lemma. t). y. T ] → M is a Tonelli minimizer. for all (x. for all u ∈ C 0 (M. Since Tt− u = u. t). Then
M ×M
inf ht + 2 inf h1 ≤ ht+2 ≤ inf ht + 2 sup h1 . R) the uniform limit lim Tt− u exists. we have that for all t > 0. we have that ∀t > 2. then E(γ. If M is compact then limt→+∞ 1 Φc (x. ˙ Then there exists a subsequence such that the probabilities µγn converge 140
.
t→+∞
To prove this theorem we shall need some lemmas. u(y) = min u(x) + Φc (x.
411. the hamiltonian viewpoint. unit formly on (x. ˙ Proof: Let Tn → +∞ and let γn : [0.2 there is A > 0 such that γn  < A for all n. y. there exists T0 > 0 such that if T > T0 and γ : [0. y) := Φc (x. t) ≤ sup u(x) − inf u(x) + 2 C. y) ∈ M × M .
So that v is a ﬁxed point for the whole semigroup Tt−
t≥0 . For all ε > 0.3 Theorem (Fathi [23]). Then by lemma 32. 411. y. Φc (x.
M ×M M ×M M ×M
Then. writing C = 2 (sup h1 − inf h1 ).
M ×M M ×M
Now let u be a weak KAM solution. γ) − c(L) < ε.5 Lemma. y) ∈ M × M and t > 2. y.140
4.
x∈M
Thus. Tn ] → M be a Tonelli minimizer. sup ht − inf ht ≤ 2 C. t) = 0.
γn ) → c(L). If γ ∈ C(x.
141
weakly* to a limit µ. γn (Tn ). ˙
Thus
For f : T ∗ M → R measurable and u : M → R Lipschitz. R).4. the laxoleinik semigroup. v) + H(u). dx u) ≤ L(x. n
n n
Hence µ is a minimizing measure and E(supp µ) ≡ c(L). then lim H(Tt− u) = c. E(γn . and using lemma 411. deﬁne ess sup f =
p∈T ∗ M
A⊆T ∗ M Leb(T ∗ M \A)=0
inf
sup f (p).
Proof: Using a convolution argument as in lemma 44. v) ∈ T M . Tn ) = 0. we can approximate both H(u) and u in the C 0 topology by a C ∞ function. v) + H(x.
411.411.
p∈A
and H(f ) = ess sup H(x. then u(y) − u(x) ≤
γ
du ≤
γ
L + H(u). AL+c (µ) = lim AL+c (µγn ) = lim T1 Φc (γn (0). y).7 Lemma.
t→+∞
141
.6 Lemma. If u ∈ C 0 (M. Hence we can assume that u is C 1 . then u
L + H(u). Since L is bounded on v ≤ A. dx f ).
x∈M
411.5. dx u · v ≤ L(x. Then for all (x. If u : M → R is Lipschitz.
Since u∞ L + c. using 411. 411. t] → M is a Tonelli minimizer satisfying γ(t) = x. we get that u∞ ≤ Ts u∞ ≤ Tt− u∞ for all s ≤ t.6.7. By ArzelaAscoli theorem there exists a sequence tn → +∞ such that the uniform limit u∞ = limn Tt− u exists. Since Tt− preserves the − order. γ) converges to c(L) uniformly on t. It is enough to show that there is a sequence sn → +∞ such that Tsn u∞ → u∞ . γ(t)).
It remains to show that the ﬁxed point u∞ of Tt− is the limit of Tt− u.1. But for all s ≥ 0.1.
Proof: By proposition 411.t] ). then dx (Tt− u) = Lv (γ(t). Tt− u is Lipschitz. y). the hamiltonian viewpoint. ˙ where γ : [0. ˙ Proof of theorem 411.3: Let u ∈ C 0 (M. By lemma 4n 11. then lim Φc (x. γ(t)). Thus H(dx (Tt− u)) = E(γ(t). 142
. H(Tt− u) → c. E(γ.3.2 the family Tt− u t≥1 is ´ equilipschitz. By proposition 411. Write sn = tn+1 − tn . Now we prove that u∞ is a ﬁxed point of the semigroup Tt− . R).8 Corollary. If M is compact. y. t) = hc (x.142
4. Tt− u L + H(Tt− u) and by lemma 411. we can assume that sn → +∞.
t→∞
uniformly in x ∈ M and also in y ∈ M .
− Tsn u∞ − u∞ 0 − − − ≤ Tsn u∞ − Tsn Tsn u 0 Tt− u n+1
+ Tt− u − u∞ n+1 − u∞
0
0
≤ u∞ − Ttn u
0
+
→ 0. then u∞ ≤ Tt− u∞ for all t ≥ 0. ˙ By lemma 411. Tt−+s u − u∞ n
0 − − = Ts Tt− u − Ts u∞ n 0
≤ Tt− u − u∞ n
0
−→ 0.4. If x is a diﬀerentiability point for Tt− u. Fathi. This implies that u∞ L + c.1.
n
The following corollaries are also due to A. and Tt− u(x) = u(γ(0)) + AL+c (γ[0.5. Then.
t) ≥ 0. t1 ] ⊆ [s2 .9 Corollary. t) = Tt−1 v. we get that δ(s1 .
t→+∞ t≥s
(4. the map t → δ(0. t) = Φc (γ(s). t) is bounded. where v(z) = −Φc (z. −Φc (γ(s2 ). And the uniform + limit on x ∈ M follows from −Φc (x. t) = sup δ(s.
143
Proof: The uniform limit on y ∈ M follows from the equality − Φc (x. deﬁne δ(s. By theorem 411. t1 ) + δ(t1 . Deﬁne δ(s. z. t) := AL+c (γ[s. 1). curve γ. Observe that δ(s. γ(t2 )) ≤ −Φc (γ(s2 .t1 ] ) + AL+c (γ[t1 . t). t2 ).411. 1). y. +∞[→ M is a ray. and s ≤ t in its domain. +∞) := lim δ(s. where u(z) = Φc (x. γ(t)). y. If M is compact then the rays are semistatic. 411.s1 ] ) + AL+c (γ[s1 .3. γ(t). t) ≤ δ(0. t) ≥ 0. γ(t1 )) − Φc (γ(t1 ). by the triangle inequality. γ(t)). t − s) − Φc (γ(s). t1 ) ≤ δ(s2 . t2 ] =⇒ δ(s1 .t] ) − Φc (γ(s). Observe that δ(s. y. t ≤ 0). t2 ) ≤ δ(s2 . Recall that a ray is a curve γ(t) such that it is a Tonelli minimizer for all t ≥ 0 (resp. Adding the equality AL+c (γ[s2 . t2 ). Now suppose that γ : [0. Proof: For an a.c. t1 ) ≤ δ(s2 .25)
143
. We show that [s1 . s1 ) + δ(s1 . γ(s1 )) − Φc (γ(s1 ). Indeed.t2 ] ) and using that δ(s. the laxoleinik semigroup. t2 ).t2 ] ) = AL+c (γ[s2 . t) = Tt−1 u(y).
144
4. Since γ and η are solutions the EulerLagrange equation. η(−s)) = = lim AL+c (γ[tn −s. tm − s)
n m
= Φc (η(−s). t) ≡ 0 for all 0 ≤ s ≤ t. we have that corollary 411.25). the hamiltonian viewpoint. we have that
s→+∞
lim δ(s. we prove that γ is semistatic. Let tn → +∞ be such that the limit p = limn γ(tn ) exists. Write γ(t) = πϕt (v).8.tm −s] ) − δ(tn + s. +∞) = lim δ(s.e.tn +s] ) + lim AL+c (γ[tn +s. p) = 0
n
for all s ≥ 0. Then p ∈ P = π(Σ). Using (4. Thus δ(s. +∞)
n
= 0. tm + s) − δ(tn + s.s] . γ(tm − s)) + δ(tn − s. i. +∞) = 0. We have that
C1
AL+c η[−s. η(−s)) + lim δ(tn − s. tm − s)
n n
= lim lim AL+c (γ[tn −s. t ≥ 0. Suppose that tn → +∞ and γ(tn ) → w ∈ T M . γ is semistatic.tn +s] −→ η[−s.tn +s] ) + lim Φc (γ(tn + s). Finally. then γ[tn −s. implies that δ(s. +∞) − δ(tn + s. p) − Φc (γ(s). ˙ Let η(t) = π ϕt (w). Thus w ∈ Σ(L).
Now we prove that the ωlimit vectors of a ray are static. γ(tm − s))
n m m
= lim AL+c (γ[tn −s. tm − s)
m
= lim lim Φc (γ(tn − s).tm −s] ) − δ(tn + s. +∞) is increasing.
Since the map s → δ(s.s] + Φc (η(s).
144
. tn ) = hc (γ(s).
But as we shall see in example 412. M− and M+ are compact.
145
412
The extended static classes. the extended static classes. Proof: Observe that the functions in M− and M+ are dominated. Gromov and Schroeder [5] proved that the manifold can be compactiﬁed adjoining the sphere at inﬁnity that can be deﬁned in terms of Busemann functions. By deﬁnition of Busemann function. Then M− is the closure of the classes of the functions fz (x) := Φc (z. R) we use the topology of uniform convergence on compact subsets.4 the classes in P \ B do not correspond to α or ω limits of orbits in T M . z)  z ∈ M }.7 resembles the construction of Busemann functions in complete manifolds of nonpositive curvature. We can identify F ≈ { f ∈ C 0 (M. On C 0 (M.1.
The method in proposition 49. Let F := C 0 (M. Let M− be the closure in F of { f (x) = Φc (z. R) deﬁned by f ∼ g if f − g is constant. In that case. By lemma 42. Ballmann. Consider the equivalence relation on C 0 (M. R)/∼ with the quotient topology.1 Lemma. R)  f (0) = 0 }. Fix a point 0 ∈ M . Here we emulate that construction to obtain a compactiﬁcation of the manifold that identiﬁes the points in the Peierls set which are in the same static class and adjoins what we call the extended Peierls set P .1 the families M− and M+ are equicontinuous. Since M is separable by Arzel´Ascoli theorem M− and M+ are compact in the a topology of uniform convergence on compact subsets. 145 ∀z∈M
. x) − Φc (0. the extended static classes in B correspond to the αlimits (resp. ωlimits) of semistatic orbits in the compactiﬁcation. 412. x)  z ∈ M }/∼ and M+ the closure in F of { g(x) = Φc (x. x).412.
.7. Conversely. the hamiltonian viewpoint.2.7. Suppose that fz = fw . x) = Φc (z. z). x) ∈ B and u+ (x) := −Φc (x. x) for all x ∈ M .
and M+ is the closure of the classes of gz (x) := Φc (x. Then we have embeddings M/dc → M− . 0). Adding these equations we get that dc (z. where M/dc is the quotient space under the equivalence relation x ≡ y if dc (x. then dc (w. x) ≤ dc (w. 0) + Φc (z. 2. if dc (z. 0) − Φc (z. by z → [fz ] ∈ F and M/dc → M+ by z → [gz ] ∈ F. z) + Φc (z. z) = 0. 146 ∀ z ∈ M.1 and B+ those of 49. y) = 0. w) = 0 and x ∈ M . This implies that fz = fw . If fw (x) = fz (x) for all x ∈ M .146
4. p) ∈ B (modulo an additive constant). w) = 0. x) = Φc (w. then dc (w. 412. z) = 0. w) = −Φc (w. By proposition 49. if p ∈ P = then u− (x) := − + Φc (p. z) − Φc (0. Thus Φc (w. If gw (x) = gz (x) for all x ∈ M . z) + Φc (z. and from fz (w) = fz (w) we get Φc (z. w) ≤ Φc (w. then Φc (w. z) = Φc (w. From fz (z) = fw (z) we get that Φc (w. x) − Φc (z.7. x).3 Remark. 1. 0).2 Lemma. Proof: We only prove item 1. Let B+ = B+ /∼ and B− = B− /∼ . 412. Let B− be the functions deﬁned in proposition 49.
p) = 2 p2 + cos(2πx) and the HamiltonJacobi equation H(x. B− ∪ (M \ P) = M− . the extended static classes. if x ≤ γ(−∞) + 2. But this inclusion may be strict as the following example shows: 412. In fact. is in S− . dx u) = c(L) gives dx u = ±2 1 − cos(2πx). (4. Moreover. if x ≤ γ(−∞). The function x
u(x) =
0
2
1 − cos(2πs) ds.
so that bη = bγ . 0) but it is not a Busemann function associated to a semistatic orbit γ because if γ(−∞) = 2k + 1 ∈ Z is the αlimit of γ. 1 Moreover. then the Busemann function bγ associated to γ satisﬁes dx bγ = +2 −2 1 − cos(2πx) 1 − cos(2πx) if x ≥ γ(−∞).26).
1 Let M = R and L(x. at the point y = γ(−∞)+3 the semistatic orbit η(t) with η(0) = Γ− (u)∩Ty M has αlimit ˙ η(−∞) = γ(−∞) + 2 = γ(−∞). Observe that in the Busemann function in (4.26)
Similarly a function v : R → R with dx v ≡ −2 1 − cos(1πx) is in S+ but it is not a Busemann function.
with dx u ≡ +2 1 − cos(2πx). v) := 2 v 2 − cos(2πx). H(x. 0) ∈ T R.412. there is no semistatic orbit passing through y with αlimit γ(−∞). w ∈ P and they are in the same static class. corresponding to the universal cover of the simple pendulum lagrangian. k ∈ Z. w) = 0 if and only if z = w or z. This implies that the Busemann functions can not 147
. x) − Φc (−n. is the limit of un (x) := Φc (−n.4 Example.
147
Observe that dc (z. and the static orbits are the ﬁxed points (2k + 1. Then c(L) = 1. Under the identiﬁcations M → M we have that B ∪ (M \ P) ⊆ M respectively. the Busemann function bη associated to η satisﬁes dx bη = +2 −2 1 − cos(2πx) 1 − cos(2πx) if x ≥ γ(−∞) + 2.
then uw (y) = min{ u(y)  u ∈ S+ . 1. then for t < 0.7. u(π(w)) = 0.
be parametrized just by a (semistatic) vector based on a unique point 0 ∈ M as in the riemannian case. x) and u+ (x) = −Φc (x. we have that v(y) ≤ v(xw (t)) + Φc (xw (t). By the remark 412. letting t ↓ −∞. p) (modulo an additive constant).2. Proof: We prove item 1.3. 148
. y). it may not be possible to choose a single point qα ≡ 0 ∈ M. x) + Φc (xw (t). In particular.1.148
4. uw is in the set of such u’s. u(π(w)) = 0. w ∈ Γ+ (u) }. the hamiltonian viewpoint. w ∈ Γ− (u) }. Since v(x) = uw (x) = 0.7. we get that v(y) ≤ uw (y) for all y ∈ M . Let x := π(w) and v ∈ S− with v(x) = uw (x) = 0 and w ∈ Γ− (v).5 Lemma. The functions in B− and B+ are special among the weak KAM solutions. On the other hand. ∀α ∈ B− in the construction for theorem 412. If w ∈ Σ− and uw ∈ B− is as in proposition 49. y) = v(x) − Φc (xw (t). this also holds for the functions u− (x) = Φc (p. for any p ∈ P. 2. Let xw (t) = π(Φt (w)). so that the maximum is realized by uw . then uw (x) = max{ u(x)  u ∈ S− . They are “directed” towards a single static class and they are the most regular in the following sense: 412. Since v L + c and w ∈ Γ− (v). If w ∈ Σ+ and uw ∈ B+ is as in proposition 49.7.
Then there is a sequence zn ∈ M such that u(x) = limn Φc (zn . γn (t)) + K dc (γn (t). bp.
y→p
149
.1. γn (0) = x. We can assume that γn (0) → v ∈ Tx M . i. in ˙ either case we can assume that Tn → −∞. Since u is dominated. Adding a constant. x) − Φc (y. x) − Φc (zn . x) + n .0] ) ≤ Φc (γn (t).2.
For p ∈ B and z ∈ M let x → bp. then we can assume that either dM (zn .e. z).z (x) be the function in the class p ∈ B such that bp. the extended static classes. x) − Φc (zn .
149
Deﬁne P− := M− \ (M − P) . x) be a Tonelli minimizer such that Tn < 0.
412. x) − Φc (zn xv (t))
n n
= lim Φc (zn .6 Proposition. 0). Let x ∈ M and let γn ∈ CTn (zn .412. x) + n . Then for t ≤ 0 ˙ u(xv (x)) − u(xv (t)) = lim Φc (zn . we only have to prove the condition 48. The functions in P− and P+ are weak KAM solutions. x) → ∞ or zn → p ∈ P.z (z) = 0. p) = 0 for p ∈ P.x] . Proof: Let u ∈ M− \ (M \ P).z (x) = lim Φc (y. In particular
1 AL+c (γn [t. P+ := M+ \ (M − P).3 γn  < A and hc (p. Since by lemma 32. we can assume that u(0) = 0. γn (Tn ) = zn and 1 AL+c (γn ) ≤ Φc (zn .
∀ Tn ≤ t ≤ 0.0] +
n 1 n
≤ AL+c xv [t. where K is a Lipschitz constant for Φc .
Since u ∈ P− . xv (t)) ≤ lim AL+c γn [t.
the map f → uf is injective.7 Theorem.150
4.
α∈B
−
→
{u
∈
is a bijection. from lemma 49. Let P := { qα  α ∈ B− }. Then the same argument as in formula (4. The map { f : P → R  f strictly dominated } S−  uP strictly dominated }. Fix α ∈ B− .
We now give a characterization of weak KAM solutions similar to that of corollary 49.qα (x). 412.1 we get that uf ∈ S− .
α∈B
−
150
. x) > −∞. choose them such that the map B− α → qα ∈ M is injective. Moreover. The domination condition implies that uf (qα ) = f (qα ) for all α ∈ B− .13).3. For each α ∈ B− choose qα ∈ M such that there is a unique semistatic vector v ∈ Σ− such that π(v) = q and the αlimit of v is in the static class α. We say that a function f : P → M is strictly dominated if f (qa ) < f (qβ ) + bβ.qβ (qα ) for all a = β in B− .27)
Since uf > −∞ and it is an inﬁmum of weak KAM solutions. shows that uf L + c. Since uf (qα ) = f (qα ) for all α ∈ B− . Proof: We ﬁrst prove that uf is bounded below. We now prove the surjectivity. Suppose that u ∈ S− and uP is strictly dominated. Let v(x) := min u(qα ) + bα. (4. f → uf . This can be done by the graph property 48. uf (x) ≥ uf (qα ) − Φc (qα . given by uf (x) := inf f (qα ) + bα.5.4. x) = f (qα ) − Φc (qα . And we say that f is dominated if f (qa ) ≤ f (qβ ) + bβ.qα (x). the hamiltonian viewpoint. then for all x ∈ M .qβ (qα ) for all a = β in B− .
30) (4.28).
151
Observe that the domination condition implies that v(qα ) = u(qa ) for all α ∈ B− . (4.28)
Given x ∈ M . qα ) = lim v(qγ ) + Φc (xξ (t). (4. qα ). qγ ) − ε
t→−∞
(4.
Now we prove that u = v on the projection of the backward orbits of vectors in Γ− (u) ending at the points qα . u(qα ) = u(xξ (s)) + Φc (xξ (s). qα ) − Φc (xξ (t). x) Taking the limit when s → −∞. Then. Let ξ ∈ Γ− (u)∩Tqα M and let β ∈ B− be the αlimit of ξ. xξ (s)) + Φc (xξ (s). by (4. then for s < 0. we get that u(x) ≥ v(x) for all x ∈ M. qα ) ≥ v(xξ (s)) + Φc (xξ (s). qα ) by (4. α ∈ B− . x) for all s < 0. Since ξ ∈ Γ− (u) ∩ Tqα M . From the deﬁnition of v(x) for all ε > 0 and s < 0 there exists γ = γ(s.29) ≥ v(qγ ) + bγ. 151 (4. u(qα ) ≤ u(xθ (s)) + Φc (xθ (s).412. the extended static classes. qa ) + Φc (xθ (s).qγ (qα ) − ε ≥ v(qα ) − ε = u(qα ) − ε. ε) ∈ B− such that v(xξ (s)) ≥ v(qγ ) + bγ.qγ (xξ (s)) − ε. Since u is dominated. qγ ) − ε
t→−∞
≥ v(qγ ) + bγ. let θ ∈ Γ− (u) ∩ Tx M and let α ∈ B− be the αlimit of θ. u(x) = u(xθ (s)) + Φc (xθ (s). Hence u(x) ≥ u(qa ) − Φc (xθ (s).qγ (xξ (s)) − ε + Φc (xξ (s).32)
.29) for all s < 0. qα ) − Φc (xξ (t).31)
≥ v(qγ ) + lim Φc (xξ (t).
qα ) But then v(qβ ) ≤ v(xξ (t)) + Φc (xξ (t). the hamiltonian viewpoint. This contradicts the strict domination. qα ) − Φc (xξ (t).
inf
u(qα ) + bα. Since α ∈ B− is arbitrary.32) we get that v(qα ) = v(xξ (t)) + Φc (xξ (t). qβ ).31). Equivalently v(qα ) ≥ v(qβ ) + Φc (xξ (t). Then.33)
by (4.
Letting ε ↓ 0. Taking the limit when t → −∞.
.8 Theorem. from the deﬁnition of v we get that u ≤ v. Now let x ∈ M and α ∈ B . we have that u(x) ≤ v(qα ) + bα.qα (x) 152
for all x ∈ M. 412.31) and (4. qα ) + Φc (xξ (t).qβ (qα ). x). αlim(ξ) = α. qβ ) = v(qα ) − Φc (xξ (t). Given u ∈ S− . we have that u(xξ (t)) = v(xξ (t)) for all t < 0. from the equality between (4. u(x) ≤ u(xξ (t)) + Φc (xξ (t).qα (x). Then u(x) =
qα ∈P(u)
for all t < 0. qα ) + Φc (xξ (t).
(4.34) by (4.34) − − ∩ T M with αlim(ξ) = α. x) = v(xξ (t)) + Φc (xξ (t). 0 0 and let P(u) := {qα  α ∈ B− (u) }. for all α ∈ B− (u) choose qα ∈ π[Λ− (α) ∩ Γ− (u)]. Letting t → −∞. hence β = α.152
4.qα (x) = u(qα ) + bα.30) and (4. we get that v(qα ) ≥ v(qβ ) + bβ. from the equality between (4. Let ξ ∈ Σ qa Then for t < 0. (4.33). qβ ). and ξ ∈ Σ− ∩ Tqα M. x) = v(qα ) − Φc (xξ (t).
27) show that v > −∞ and by lemma (49. For all α ∈ B− (u). (4. Then. 0 0 Let P(u) := { qα α ∈ B− (u) }. qa ) + Φc (xθ (s).1) v ∈ S− . β ∈ B− (u) and let θ ∈ Tqβ M ∩ Λ− (β) ∩ Γ− (u). we have that u(x) ≤ u(qα ) + bα. Letting t → −∞. Since ξ ∈ Λ− (α).qα (x).
u(x) ≥ u(qa ) − Φc (xθ (s). we get that u(x) ≥ v(x) for all x ∈ M.qβ (qα ). Given x ∈ M .qα (x). x) because u is dominated. Since α ∈ B= (u). x) = u(qα ) − Φc (xξ (s). taking the limit when s → −∞. choose qα ∈ π(Λ− (α) ∩ Γ− (u)). Since qα ∈ P(u) is arbitrary. Then for s < 0. qα ). x). Let (4. x) for all s < 0. We show that uP(u) is dominated. the extended static classes. Then for t < 0. qβ ) + Φc (xθ (t).412. for all α. Let α.35) v(x) := inf u(qα ) + bα.
153
.36)
Now let x ∈ M and qα ∈ P(u). because ξ ∈ Γ− (u) ∩ Tqα M .
qα ∈P(u)
The same arguments as in equation (4. Let ξ ∈ Λ(α) ∩ Γ− (u) ∩ Tqα M . u(qα ) ≤ u(xθ (t)) + Φc (xθ (t). let θ ∈ Γ− (u) ∩ Tx M and let α ∈ B− (u) be the αlimit of θ. we get that u ≤ v. u(x) = u(xθ (s)) + Φc (xθ (s). u(x) ≤ u(xξ (s)) + Φc (xξ (s). qα ) = u(qβ ) − Φc (xθ (t). qα ) + Φc (xξ (s). we get that u(qα ) ≤ u(qβ ) + bβ.
153
Proof: Let u ∈ S− . letting s → −∞. β ∈ B− (u).
the hamiltonian viewpoint.
154
.154
4.
x∈M
155
. 0) = max U (x). the semistatic static orbits are not ﬁxed points because by ?? they have energy c(L−ω) > 0. the static set of L−ω.
we have that
e0 = cu = c0 = c(L) = 0. The static orbits are the ﬁxed points (x.
52
Mechanic Lagrangians. The only weak KAM solutions are the constant functions and so are the only subsolutions of (HJ) for k = c(L). Since semistatic geodesics are minimizing. from (??) when we add some closed forms to L we obtain c(L − ω) > c(L) = 0. the support of it minimizing measures and the basins of weak KAM solutions are geodesic laminations. Moreover. since the α function is superlinear. By the graph properties.
1 2
For a mechanic lagrangian L(x.Chapter 5
Examples
51
Riemannian Lagrangians. we have that
x∈M
c(L) = c0 = cu = e0 = min E(x. x ∈ M . v) =
v
2 x.
1 2
For a riemannian lagrangian L(x. Nevertheless. v) =
vx
2
− U (x). they don’t have conjugate points. 0).
The static set is Σ(L) = {(π.e. Since v → L(x. 0) ≥ 0. 0) which maximize L(x. v) = L(x.
53
Symmetric Lagrangians. if M is compact.
54
Simple Pendulum. v) is convex. and thus c(L) = c0 = cu = e0 . In fact. −v).
The simple pendulum is the mechanic lagrangian L : T S 1 → R given by ˙ L(θ. there 156
. 0). The static orbits are the ﬁxed points (x.
identifying the circle S 1 = [0. If U is nondegenerate at a minimum x (i. the constant functions are subsolutions of the HamiltonJacobi equation (HJ) for k = e0 . 0). 0) where x maximizes L(x. 2π] mod 2π. The static orbits are the ﬁxed points (x. v) − max L(x. If M is noncompact. Since there is a unique static class. L(x. but there may be no maximizers of L(x. d2 U is nonsingular).
For a symmetric lagrangian L(x. The constant functions are always subsolutions of (HJ) for k = c(L) = e0 . the constant functions are still subsolutions of (HJ) for k = e0 .156
5.
x∈M
and it is 0 exactly at the ﬁxed points (x. v) + e0 = L(x. 0) and then the static set Σ(L) = (see example 57). then x is a hyperbolic sadx dle point. Thus. examples
the maximum of the potential energy. 0)}. where U (x) is maximal. θ) =
1 2
˙ θ2 − cosθ. 0). and then H(x. −p). This implies that c(L) = e0 . p) = H(x. Hence c(L) = e0 = 1. it attains its minimum at v = 0.
The solutions of (EL) are the projection of straight lines in Rn parametrized with constant speed. the functions on S± are diﬀerentiable. By corollary 49. If ω = 0 is a closed 1form. the flat torus ≈n .
55
The ﬂat Torus Tn . These vectors are sent to 0 by the Legendre transform.5. there is only one static class. For u ∈ S+ its is the whole Σ( L) and it is shown in ﬁgure 54. 1: simple pendulum. then L − ω is globally minimized at all the vectors (x. du = 0 and thus any weak KAM solution must be constant. v) = A. Their cut locus is x = 0 ∈ S 1 .
Also called the harmonic oscillator.3.6. there is only one static class and #S± = 1 modulo constants. v with A ∈ Rn ﬁxed. they must be static. A). By the smoothness property 48. If ω(x.
157
. Again. it is the riemannian lagrangian for Tn = Rn /Zn with the riemannian metric induced by the euclidean metric on Rn . But S− = S+ because the solutions are not diﬀerentiable. c(L−ω) > e0 = 1 and the static set Σ(L− ω) is a whole component of the energy level E = c(L).55. Since π(Σ(L − ω)) = S 1 .
157
fig. Since the orbits of this vector ﬁeld are recurrent. are only one weak KAM solutions on S+ and S− modulo an additive constant. oriented by ω > 0.
µ− := µγ− supported on the closed orbits γ+ . y). R) ≈ R2  −
1 2π
≤t≤
1 2π
must be a ﬂat domain for the βfunction. Since the curves γ+ and γ− are closed. 2π ). Then c(L) = − inf ψ = 0. v) = 1 (u2 + v 2 ) + v sin(x). (u. ρ+] = (0. in fact β[ρ− . ρ− = (0. 158
. then the line [ρ− . Since the static set Σ(L) contains only the support of the measures µ+ .
57
A Lagrangian with Peierls barrier h = +∞. where  ·  is the euclidean 2 1 metric on R2 and ψ(x) is a smooth function with ψ(x) = x for x ≥ 2. they must be static curves. the Peierls set is just P = {− π } × [0. v = 1 and x = π .
This example is a special case of those studied by Carneiro and Lopes [8]. ρ+ . examples
56
Flat domain for the βfunction. These vectors are tangent 2 2 to the closed curves γ+ : x ≡ − π . h = tρ− + (1 − t)ρ+ . Since Mather’s βfunction is convex and attains its minimum at ρ+ and ρ− .158
5. − 2π ). 2π] . Then the lagrangian is minimized at the vectors x = − π .
Let L : T R2 → R be L(x. 2 where T2 = [0. y ≡ 1 and γ− : x ≡ π . Let L : T T2 → R be the lagrangian L (x. µ− . t) ∈ H1 (T. 2π] mod 2π . The invariant 2 2 measures µ+ := µγ+ . the unique minimizing measure in homology h is M(h) = {tµ− + (1 − t)µ+ }. v = −1. γ− ˙ ˙ 1 1 are minimizing measures with homology ρ+ = (0. ψ ≥ 0 and ψ(x) = 2 for 0 ≤ x ≤ 1. 2π] ∪ { π } × [0. y ≡ 2 ˙ 2 ˙ −1.ρ+ ] ≡ c(L). Since their tangents are exactly all the vectors which minimize L. then for any homology h ∈ [ρ− . v) = 1 v2 + ψ(x).
0) is in S+ . θ(t)) in polar coordinates about the origin 0 ∈ R2 .
159
.57. Since L + c = L ≥ 0. then xξ (t) = ˙ and L(ϕt ξ) = xξ (t)2 = ˙
2 r(t) 2 r(t)
xξ (t). ˆ Hence Σ(L) = . Suppose that h(0.
159
because if γn is a smooth closed curve with length (γn ) = 1. γn (t) ≥ n 1 2 and energy E(γn ) = 2 γn − ψ(γn ) ≡ 0. ˙
Let Tn → +∞ be such that r(Tn ) → +∞. a lagrangian with peierls barrier h = +∞. then ˙
Tn
c(L) ≥ − inf AL (γn ) = −
n>0 0
1 2
γn + ψ(γn ) ˙2
2 n
=−
0
1 γn  ˙
γn 2 = −γn  ≤ − ˙ ˙
−→ 0. L > a > 0. y) = Φc (x. 0) ≥
0
L(ϕt (ξ)) + c(L) =
0 Tn 2 r
2 r(t) r(Tn )
˙ x + r θ dt ˙
2 r
lim sup
Tn 0
x dt = lim sup ˙
n 0
dr = +∞. Let ξ ∈ Γ+ (u)∩T0 R2 and write xξ (t) = (r(t). c(L) = − inf {AL (γ)  γ closed } ≤ 0. Then u(x) := h(x. because L ≥ 0. y ∈ R2 . then
+∞ +∞
h(0.
On the other hand. 0) < +∞. then we have that dc (x. Observe that since L > 0 and on compact subsets of R2 . y) > 0 for all x. Then lim inf t→+∞ r(t) = +∞ because otherwise the orbit of ξ would lie on a compact subset E ≡ 0 and then = ωlim(ξ) ⊆ Σ(L) = .
where (x. Σ = and diﬀerentiable Busemann functions u with B− (u) = B+ (u) = {α}. examples
58
Horocycle ﬂow. +∞[. v . Then y
L (x. y) = ˙ ˙ 160
1 x ˙ x2 + y 2 + · ˙ ˙ 2y 2 y
.
where ηx is a 1form on H such that dη(v) is the area form an ·x is the Poincar´ metric. v) = Yx (u). y). Let H := R×]0. y) = dx .
v
2 x
+ ηx (v). +∞[ with the Poincar´ metric ds2 = e Let L : T H → R be a Lagrangian of the form L(x. Choose the form η(x.1)
where Yx : T H → T H is a bundle map such that dηx (u.
Peierls barrier 0 < h < +∞.
1 The energy function is E(x. (x.160
5.1) are closed curves. and on E = 2 the solutions are the horospheres parametrized by arc length. y) ∈ H = R×]0. On the energy levels E < 1 x 2 1 the solutions of (5. The EulerLagrange equation is e
Dt dt
x = Yx (x) = x⊥ . v) =
1 2 1 y2
(dx2 + dy 2 ). ˙ ˙ ˙
(5. v) = 2 v 2 .
Observe that if v = (x. 2
(5. h(x. 0) ≥ ˙
1 2
v
2
− v ≥ −1.
161
Observe that the form η is bounded in the Poincar´ metric. so that T∗ η 161
. horocycle flow.
(5. ˙ 1 We show that c(L) = 2 and hc < +∞. formula (5. then d(T∗ η) is also the area form. Let γr be the curve whose image is the boundary of Dr oriented clockwise and with hyperbolic speed γ ≡ a. First observe that if T is an isometry of H.3)
1 Hence L + 2 ≥ 0 and then c(L) ≤ 1 . parametrized by arc length. ˙ ˙ x < 0. for r > 0 large. y). x) ≤ lim inf AL+ 1 (γr ) = 2π < +∞. This implies that h > 0. Since ˙ 1 E(γ) = 2 a2 . Moreover. y = y. so that the e Lagrangian is superlinear and satisﬁes the boundedness condition. = a · 2π sinh(r) − 2π cosh(r). y = 0) + ˙ ˙
1 2
≡ 0.
r→+∞
2
We prove that Σ = .
(5.2)
The images of these curves are the stable horospheres associated to the geodesic x = 0. For r > 0 let Dr be a geodesic disc of radius r such that x ∈ ∂Dr .4)
1 2
If a < 1.4) is negative. 2 Now ﬁx x ∈ H.58. y = 0 are solutions with ˙ ˙ L(x = −y. then ˙ L=
1 2
v
2
− (x. = 2π
1 2
(a − 1) er − e−r + 2π. It can be seen directly from the EulerLagrange equation that the curves x = −y. Hence c(L) ≥ 1 and c(L) = 2 . then
γr
L + 1 a2 = 2
v · Lv =
γr γr
v
2
+
Dr
dA
= a · length(γr ) − area(Dr ).
y ∈ D. In particular. Since a horocycle h1 can be sent by an isometry to another horocycle h2 with h1 ∩ h2 = . there is a constant b = b(x. AL (γ) = AL (T ◦ γ) + b(x.162
5. then the horocycles can not be static because it would contradict the graph property. This implies that given any two points x. y). v). 0) ∈ Γ+ (u) = Γ− (u) ∈ Σ− are semistatic.5)
= L(x. 2
Let T : H ← be an isometry of the hyperbolic metric. y
that is v = (−y. the map dT leaves σ(L) and Σ(L) invariant. v) + η(x. v) = =
1 2 1 2
v v
2 x 2 x
+ T ∗ η(x. y Then dη = A is the hyperbolic area 2form. Its derivative du = 0 is sent by the inverse of the Legendre transform v → Lv = v. examples
is cohomologous to η. v) − dv(x. y) ∈ R such that for all γ ∈ C(x. 0). The constant function u : H → {0} satisﬁes u L + 1 because 2 1 L + 2 ≥ 0 and by (5. We show that v is a weak KAM solution. Also H(du) =
1 2
du −
dx 2 y
=
1 2
dx 2 y
= 1. Since T is an isometry. v) (5. · x + dx to y
1 y2
v. Observe that L ◦ dT (x. ·
eucl
= − dx .2) the vectors v = (−y. then d(T ∗ η) = T ∗ (dη) = dη. y). Hence the form T ∗ η − η is exact on H and there is a smooth function v : H → R such that T ∗ η − η = −dv. v) − dv(x. Write η = dx . 162
.
then
1 dv ≤ L + 2 .
Here v is the Busemann weak KAM solution associated to the class T (∞) ∈ ∂H. hence v is constant on Re z = 0.
163
Since by (5. We 1 use the isometry T : H ← . when dT (v) = (−y.6) holds exactly when 1 L ◦ dT (x. Since the isometries send horospheres to horospheres. on the sphere at inﬁnity of H. Since T leaves the line Re z = 0 invariant and η = 0 on vertical vectors.
163
. they are selfconjugacies of the hamiltonian ﬂow and hence the curves γ(t) = T −1 (x− ty.3) L ◦ dT +
1 2
≥ 0. horocycle flow.58.
v(γ(t)) − v(γ(s)) =
γ
dv =
γ
1 L + 2. The isometry T = T −1 sends the line t → −ty + iy to a horosphere with endpoint 0 ∈ C. oriented clockwise.
(5.y) H. using (5. T (z) = − z . v) + 2 = 0.5). i. Choose v : H → R such that dv = η − T ∗ η. the equality in (5. Then. We now show a picture of a nonBusemann weak KAM solution. Moreover.6)
1 Hence v L + 2 . y) realize v. 0) ∈ T(x. Let γ(t) = −ty + iy. Now we describe the weak KAM solution w(z) := min{ u(z). z = x + iy ∈ C. v(T −1 γ(t)) = v(T −1 γ(0)) +
T −1 ◦γ t t
dv
1 2 1 2
=0+
0 t
L ◦ dT ◦ γ + ˙ L ◦ dT ◦ γ + ˙
0
−
0
L◦γ+ ˙
1 2
=0+
− 0. v(z) } ∈ S− . i.e.e.
v) + 1 > 0 when v = −y + i0.
164
. v(z) if Re z < 0. (5. 0) ∈ Tx+iy H. then v(z) > 0 on 2 Re z > 0 and v(z) < 0 on Re z < 0.7)
The cut locus of w is Re z = 0 and the basin of w is Γ− (w) = A ∪ dT (A) where A is the set of vectors (y. Thus w(z) = 0 = u(z) if Re z > 0.6) L(x. examples
Since by (5.164
5.
Deﬁne u
∞
:=
k∈N
arctan( u 2k
k)
. Given u ∈ C ∞ (M. R) is dense. R). For every lagrangian L there exist a residual set O ⊂ C ∞ (M. We recall which topology is used in C ∞ (M. R). Ma˜´ introduced the concept of generic property of a lagrangian ne L. R) with the translationinvariant metric u − v ∞ . denote by uk its C k norm. In this chapter we shall prove 60. This metric is complete. Endow C ∞ (M.
In [38]. R) such that if ψ is in O then L + ψ has the property P . R) such that if ψ ∈ O 165
.Chapter 6
Generic Lagrangians.1 Theorem. A set is called residual if it contains a countable intersection of open and dense subsets. One of Ma˜´’s objectives was to show that Mather’s thene ory of minimizing measures becomes much more accurate and stronger if one restricts it to generic lagrangians. hence the Baire property holds: any residual subsets of C ∞ (M. A property P is said to be generic for the lagrangian L if there exists a residual set O on C ∞ (M.
Note that · ∞ is not a norm.
We will prove only the ﬁrst part and give only the ideas of the proofs of the last part. Denote by M(L) the set of minimizing measures of the lagrangian L: M(L) := { µ ∈ M(L) : AL (µ) + c(L) = 0 }.1 was proved by Ma˜´ in [34].
Ma˜´ conjectures in [35] that there exists a generic set O such that ne this unique minimizing measure is supported on a periodic orbit or an equilibrium point. When µ is supported on a periodic orbit or a ﬁxed point.
61
Generic Families of Lagrangians.166
6.
1. The lagrangian L + ψ has a unique minimizing measure µ. In this section we shall prove the following 166
. and this measure is uniquely ergodic. Th proof of item 1 presented here is extracted from a more general result in [6] about families of lagrangian systems.2). 3. this orbit (point) Γ is hyperbolic and if its stable and unstable manifolds intersect then they do it transversally. In general M(L) may be inﬁnite dimensional. supp(µ) = Σ(L + ψ) = Σ(L + ψ). Item 2 is proved ne in [16] and item 3 in [13]. Observe that M(L) is a simplex whose extremal points are the ergodic minimizing measures (see exercises 61. Item 1 of theorem 60. ˆ 2. generic lagrangians.
In this chapter the manifold M is compact.
· · · . Then there exists a residual subset O of C ∞ (M. generic families of lagrangians. 2.1 Theorem. A simplex is a subset Σ ⊂ V such that the set of its extremal points is aﬃnely independent. there is a Borel set A such that µ(A) = 1 and ν(A) = 0).
Fix B > 0 and let HB be the set of holonomic measures supported on [v ≤ B]: HB := { µ ∈ C(M ) : µ([ v > B ]) = 0 }.2 Exercises: 1. Prove any set of ergodic invariant Borel probabilities is aﬃnely independent. vn − v0 } is linearly independent. 3. then they are either equal or mutually singular (i.61. 4. v1 .
In other words. · · · .
167
61. every point in the convex hull of F is uniquely expressible as a convex combination λ0 v0 + · · · + λn vn . L∈A =⇒ dim M(L − u) ≤ dim A. R) such that u ∈ O. Show that if µ and ν are ergodic invariant Borel probabilities of a dynamical system. Let A be a ﬁnite dimensional convex family of lagrangians. 5. vn } ⊂ Y .
61. Prove that the set of minimizing measures M(L) of a lagrangian L is a simplex in the vector space of signed Borel measures on T M whose extremal points are the ergodic minimizing measures. Show that a ﬁnite set F is aﬃnely independent if and only if {v1 − v0 . Let K ⊂ V be a closed convex subset.e. 0 ≤ λi ≤ 1 of the elements of F . Let V be a real vector space. 167
. Prove that K is the convex closure of its extremal points. and in this case the dimension of the convex hull dim(conv(F )) = n. there exist at most 1 + dim A ergodic minimizing measures for L − u. A set Y ⊂ V is aﬃnely independent if for any ﬁnite subset F = {v0 . An extreme or extremal point of K is a point in K which is not in the (relative) interior of any segment contained in K.
under this topology the set P(M ) is compact Let π∗ : HB → P be the pushforward induced by the projection π : T M → M . R) and let d(µ. (6. B) is residual in C ∞ (M.1. Let P = P(M ) be the set of probability measures on M endowed with the weak* topology. Let MP (L.3 Proposition. B) for all B > B0 (L). A compatible metric (cf.6 implies that there is B0 (L) > 0 such that M(L) ⊂ HB for all B > B0 (L).1. ν) =
n
1 1 2n cn
fn dµ −
fn dν . 61. Let MHB (L) := arg min L
HB
be the set of measures µ ∈ HB which minimize LHB .
Since M is compact. B) := π∗ MHB (L) . B) =⇒ dim MP (L. Then inequality 1. R) such that L ∈ A.
These remarks. generic lagrangians.168
6.1 and 36. together with proposition 61.3. From theorem 24. Let A be a ﬁnite dimensional convex family of lagrangians on M and let B > 0.1)
Proof of theorem 61. By corollary 36. R). There is a residual subset O(A.
By the graph property 38.
Consider a lagrangian L as a functional L : HB → R given by L(µ) = AL (µ). B) ⊂ C ∞ (M.2 M(L) = arg min L = MHB (L)
HB
for all
B > B0 (L). 168
. E(M(L)) = {c(L)}.1 dim M(L) = dim π∗ M(L) = dim MP (L. u ∈ O(A.3 prove theorem 61. B) ≤ dim A.1: The set O(A) := ∩B∈N O(A. on P(M ) is deﬁned as follows: ﬁx a countable dense set { fn  n ∈ N } in C ∞ (M.
B) =
ε>0
O(A. dim A) is open and dense for any compact subset D ⊂ A and any ε > 0.
Indeed. ε.a we prove that it is open and in 61.3: Deﬁne the εneighbourhood Wε of a subset W ⊂ P(M ) as the union of all the open balls in P(M ) which have radius ε and are centered in W . ε.
We prove that for all k ∈ N. B) then the inequality in (6. if ε is small enough. We shall prove that the proposition holds with O(A. We need a lemma: 61. denote by O(D. The set valued map (L. u) → MHB (L − u) is upper semicontinuous on A × C ∞ (M. the convex set MP (L. B) is residual.a
Open. This means that for any open subset U of HB . ε. u) ∈ A × C ∞ (M. In 61. k ∈ N.
169
Proof of proposition 61. B) is contained in the εneighbourhood of some kdimensional convex subset of P(M ). R) 169
. So we have to prove that O(A. the set O(D. Because otherwise for some L ∈ A the convex set MP (L.4 Lemma. such a ball is not contained in the εneighbourhood of any convex set of dimension dim A. ε > 0 and any compact D ⊂ A. Given D ⊂ A. R) such that for all L ∈ D. ε. ε > 0. if u ∈ O(A.61. the set {(L. k) the set of potentials u ∈ C ∞ (M. and. generic families of lagrangians.b that it is dense. k) is open. dim A). R) : MHB (L − u) ⊂ U } ⊂ A × C ∞ (M.
61.1) holds. It is enough to prove that O(A. R). B) would contain a ball of dimension dim A + 1.
the setvalued map (L.
We prove the density of O(D.
(6. Then the open sets in D × C ∞ (M. ε. R) as a linear functional w : P(M ) → R as w(ν) = w dν or w : HB → R as w(µ) = (w ◦ π) dµ. . . µ) −→ (L − u)(µ) = L(µ) − u dµ. k). This implies that there is a ﬁnite subcover of D × {u0 } by open sets of the form Ωi × Ui . . cover the compact subset D × {u0 }. u) ∈ D × C ∞ (M. R)  MP (L − u. u) → MP (L − u.2)
170
. with wi ∈ C ∞ (M.
61. Proof: The lemma is a consequence of the continuity of the map A × C ∞ (M. u. 61. R) containing u0 . There is an integer m and a continuous map Tm = (w1 . is open in A × C ∞ (M. . R). B) ⊂ Vε }. R) for any ε > 0. Consequently.5 Lemma. R). dim A) in C ∞ (M.b
Dense. Let w ∈ C ∞ (M. ε. ε.170
6.
Now let u0 ∈ O(D. wm ) : P(M ) → Rm . R) such that ∀x ∈ Rm
−1 diam Tm {x} < ε. For each L ∈ D there is a kdimensional convex subset V ⊂ P(M ) such that MP (L − u0 . R) of the form { (L. dim A). k) and contains u0 . where V is a kdimensional convex subset of P(M ). We conclude that the open set ∩Ui is contained in O(D. B) is also upper semicontinuous. B) ⊂ Vε . generic lagrangians. k) is an open set in C ∞ (M. R) × HB (L. We want to prove that w is in the closure of O(D. ε. ε. where Ωi is an open set in A and Ui ⊂ O(Ωi . We consider a function w ∈ C ∞ (M.
By extracting a subsequence. We have that
−1 MP (L − w − y1 w1 − · · · − ym wm .
171
The open sets Uw cover the complement of the diagonal in P(M )×P(M ). y)). Deﬁninig Tm = (w1 . µ) ∈ P(M ) × P(M ) w dν = w dµ . y) := arg min [F (L.61. . Take m large enough so that Tm (ν) = T (µ). For y = / (y1 . generic families of lagrangians. . µm ) ≥ ε. which satisfy d(ν. . We have that Tm (νk ) = Tm (µk ) for k ≥ m. This amounts to say that the sequence wk separates P(M ). R) we associate the open set Uw = (ν. . This is a contradiction. µ) ≥ ε. wm ).2) holds for m large enough. x) = min
µ∈HB Tm ◦π(µ)=x
(L − w)(µ). One can extract a countable subcover Uwk . Proof: To each function w ∈ C ∞ (M. ym ) ∈ Rm let Mm (L. we can assume that the sequences νm and µm have diﬀerent limits ν and µ. B) ⊂ Tm (Mm (L.
171
. x) − y · x. x) − y · x] ⊂ Rm
x∈Rm
be the set of points wich minimize the function x → F (L. Otherwise.
when x ∈ Tm (HB ) and Fm (L.
Deﬁne the convex function Fm : A × Rm → R ∪ {+∞} as Fm (L. . . . x) = +∞ if x ∈ Tm (HB ). k ∈ N. we would have some ε > 0 and two sequences νm and µm in P(M ) such that Tm (νm ) = Tm (µm ) and d(νm . we have to prove that (6. Such m exists because the sequence wk separates P(M ).
Hence at the limit Tm (ν) = Tm (µ). .
3) (6. y) = max y · x − Fm (L. Therefore. y) ≥ dim A + 1 } . the complement of Σ is dense in Rm . which follows from the next proposition.6 Proposition. generic lagrangians.5 it follows that y ∈ Om (w) =⇒ w + y1 w1 + · · · + ym wm ∈ O(A. dim A).172
6. From lemma 61.1 the Hausdorﬀ dimension HD(Σ) ≤ (m + dim A) − (dim A + 1) = m − 1. The set Om (w) is dense in Rm . x) m
x∈R
(6. So it is enough to prove that y∈Σ / =⇒ ∀L ∈ A : 172 dim Mm (L. ε. dim A) it is enough to prove that 0 is in the closure of Om (w).
Therefore. 61. y) ∈ A × Rm  dim ∂Gm (L.
. By proposition E. Let ∂Gm be the subdiﬀerential of Gm and let Σ = { (L. y) ≤ dim A }.4)
= max
µ∈HB
w + y1 w1 + · · · + ym wm − L dµ
It follows from the second expession that the function Gm is convex and ﬁnitevalued. ˜ Consequently. Proof: Consider the Legendre transform of Fm with respect to the second variable Gm (L. hence continuous on A × Rm .
Let Om (w) := { y ∈ Rm  ∀L ∈ A : dim Mm (L. in order to prove that w is in the closure of O(A. y) ≤ dim A. the projection Σ of the set Σ on the second factor Rm also has Hausdorﬀ dimension at most m − 1. ε.
/ it is enough to observe that dim Mm (L. y) is contained in the subdiﬀerential of the convex function Rm at the point y. z)
173
. dim ∂Gm (L.
173
Since we know by deﬁnition of Σ that if y ∈ Σ. z → Gm (L. y) ≤ dim ∂Gm (L. y). The last inequality follows from the fact that the set Mm (L.61. y) ≤ dim A. generic families of lagrangians.
174
.174
6. generic lagrangians.
On Ma˜´ [35]. ˆ (B)Moreover supp(µ) = Σ(L + ψ) = Σ(L + ψ). For every lagrangian L there exists a generic set O ⊆ C ∞ (M. Ma˜´ introduced the concept of generic property of a lagrangian ne L. A property P is said to be generic for the lagrangian L if there exists a generic set O (in the Baire sense) on the set C ∞ (M. this orbit (point) Γ is hyperbolic and its stable and unstable manifolds if intersect they do it transversally. (C) When µ is supported on a periodic orbit or a ﬁxed point. R) such that (A) If ψ is in O then L + ψ has a unique minimizing measure. One of Ma˜´’s objectives was to ne show that Mather’s theory of minimizing measures becomes much more accurate and stronger if we restrict ourselves to generic lagrangians. R) such that if ψ is in O then L + ψ has the property P .Chapter 7
Generic Lagrangians.
In [34].1 Theorem. The main purpose of this chapter is to proof the following 70. µ and this measure is uniquely ergodic. 175
. it is conjectured that there exists a generic set O such ne that this unique minimizing measure is supported on a periodic orbit or an equilibrium point.
The second ne statement is proved in [16] and the third one in [13]. From this follows that O is open. ψdν = m0 (ψ)}
M0 (ψ) = {ν ∈ K0 : 176
. It remains to prove that it is also dense. R) deﬁne
m(ψ) = min
ν∈C
L + ψdν. We will prove only the ﬁrst part and give only the ideas of the proofs of the last part. generic lagrangians.1)
So by theorem 24.
The ﬁrst statement of this was proved by Ma˜´ in [34]. Given a compact convex set K0 on C and potential ψ on C ∞ (M.176
7.
Proof of (A) Given a potential ψ on C ∞ (M. in fact if νn is in M (ψn ) then
L + ψdνn ≤ m(ψ) + 2ψ − ψn C 0
(7. L + ψdν = m(ψ)} > 0 let
M (ψ) = {ν ∈ C :
Where C is the set of holonomic measures. For O = {ψ : diam M (ψ) < } This set is open.
71
Generic Lagrangians. R) deﬁne
m0 (ψ) =
ν∈K0
min
ψdν.1 in chapter 2 if ψn → ψ then the sequence νn is precompact and the limit is in M (ψ).
ν) such that ψ(µ. Let K0 as before then if µ is an extremal point of K0 .ν) dµ = ψ(µ.71.ν) dν.
177
71. then f ◦ Tn =
i
λi 177
ψi
.2)
Using (7. generic lagrangians. M0 (ψ)) < Proof Denote by D the diagonal of K0 ×K0 for each pair (µ. ν) in K0 ×K0 − D take a potential ψ(µ. ν ) in U (µ. ν).1 Lemma.2) and the compactness of K0 × K0 it is easy to see that given there exist δ > 0 and n > 0 such that S ⊂ Rn ..ν) dµ = ψ(µ. R) such that diam M0 (ψ) < d(µ.. ψn dµ) ψn dν (7. νn )} of K0 × K0 − D and set ψn = ψ(µn . Deﬁne ψ = i λi ψi where f = i λi pi . then there is a neigbourhood U (µ.3)
Let B = Tn (k0 ) then B is a compact convex set.ν) dν for every (µ . ν) in K0 × K0 − D there exist n such that ψn dµ = Deﬁne Tn : C → Rn as Tn (µ) = ( ψ1 dµ. diam S < δ ⇒ diam T −1 S < (7.. Take a covering {U (µn .νn ) then if (µ. ν) contained on K0 × K0 such that ψ(µ. for all > 0there exists ψ on C ∞ (M. let f : Rn → R a linear function such that its minimum restricted to B is attained in only one point p. .
generic lagrangians. Set
m0 = m(ψ) f0 (ν) = f1 (ν) = Then f1 (ν) = f0 (ν) if ν ∈ S f1 (ν) ≥ 0 if ν ∈ M (ψ) ν ∈ M (ψ). f1 (ν) = 0 ⇒ ν ∈ S 178 (7. If ψ is on C ∞ (M. say m1 for all measures ν on K0 = M (ψ) on a set S = M0 (ψ1 ) such that d(µ. we can ﬁnd given a ψ1 such that ψ1 dν attains its minimum.5) (7.
and so
−1 M0 (ψ) = Tn (p)
Then by (7. 71.6) L + ψ − m0 dν ψ1 − m1 dν
. S) < .3) we get diam M0 (ψ) < The following lemma proves the density of O and hence the ﬁrst part of (A).2 Lemma.178
7.4) (7. R) and µ is an extremal point of M (ψ) then for every neighbourhood U of ψ and every > 0 there exists ψ1 on U such that diam M (ψ1 ) = Proof For K0 = M (ψ) applying the previous lemma.
generic lagrangians. Then by (7. νλn on M (λn ) = M (ψ + λn ψ1 ) such that d(µλn . νλn ) > . f0 (ν) = 0 ⇒ ν ∈ M (ψ) For λ > 0 deﬁne fλ = f0 + λf1 and set m(λ) = min fλ (ν).6) µ is in S. Similarly ν is in S.4) we have that m(λ) ≤ 0 0 ≥ m(λn ) = = f0 (µλn ) + λf1 (µλn ) ≥ λn f1 (µλn ) So f1 (µλn ) ≤ 0 and hence f1 (µ) ≤ 0. 179
. {µ}) ≤
(7.7)
M (λ) = {ν ∈ C : fλ (ν) = m(λ)}. This is a contradiction with the fact that the diameter of S is less than .8)
This proves the lemma since M (λ) = M (ψ + λψ1 ) Proof of the claim Suppose otherwise that there exist λn → 0 and µλn . Naturally d(µ. We claim that
λ→0
lim diam (M (λ). Now Because of (7.
ν∈C
(7. ν) ≥ .1) {µλn } and {νλn } are precompact and as in the proof of the open property we may assume that µλn → µ ∈ M (ψ) and νλn → ν ∈ M (ψ).5) f1 (µ) = 0 and then by (7.
179
ν ∈ C. since µ ∈ M (ψ) by (7. The fact that µ is uniquely ergodic follows from the fact that ergodic components of a minimizing measure are also minimizing.71. And the proof of (A) is complete.
180
. R) • (2) As in section 23 let C 0 be the set of continuous functions f on T M such that sup f (x. L : F → R linear maps satisfying • (a) The map E×F → R deﬁned by (w. K contained on F a metrizable convex subset and φ : E → F . Let K be C and F be the subspace of (C 0 )∗ spanned by C. we get the desired result. • (b) For any x = y in K there exists w in E such that φ(w)(x) = φ(w)(y).3 Proposition. If E is a Frechet space then there exists a residual set O contained on E such that if w is on E then M (w) has only one element. And M (w) = {x ∈ K : L(x) + φ(w)(x) = m(w) 71.
It is worth to remark that the proof presented here is a particular case of Ma˜´’s original [38] more general setting: ne Let E. generic lagrangians.180
7. Denote by m(w) = min L(x) + φ(w)(x)
x∈K
which exists by (c). x) → φ(w)(x) is continuous.v) ≤ ∞. F be real convex spaces. and C the set of holonomic 1+v probabilities. • (1) Let E be the Banach space C ∞ (M. The reader can verify that with the following choices. • (c) For all w in E and c in R the set {x ∈ K : L(x) + φ(w)(x) ≤ c} is compact.
71. lim SL+φk (µk ) = SL+ψ (ν) > SL+ψ (µγ ) . C D in [35] and also [11] Proof of (C) Let O be the residual given by (A). R) φ(ψ) is the restriction to F of the linear map on (C 0 )∗ such that w →< w.h which are the continuation of γ. This general setting has some other applications see theorems A. Then if ν = µγ . We prove that A1 is relatively open on A. the velocities in the support of the minimizing measures µk := µηk are bounded (cf. since L is superlinear. We claim that if φk ∈ A. Indeed. ψ >. For energy levels h near to c(L + ψ) and potentials φ near to ψ. Let A be the subset of O of potentials ψ for which the measure on M(L + ψ) is supported on a periodic orbit. let ψ ∈ A1 and M(L + ψ) = {µγ } where µγ is the invariant probability measure supported on the hyperbolic periodic orbit γ for the ﬂow of L + ψ. on a small 181
. there exist hyperbolic periodic orbits γφ. and hence there exists a subsequence µk → ν converging weakly* to a some invariant measure ν for L + ψ. Therefore ν = µγ . corollary 36.9)
Thus if δk is the analytic continuation of the hyperbolic periodic orbit γ to the ﬂow of L + φk in the original energy level c(L + ψ). which contradicts the choice of ηk . SL+φk (µδk ) < SL+φk (µηk ) . φk → ψ and M(L + φk ) = {µηk }. generic lagrangians. since limk SL+ψk (µδk ) = SL+ψ (µγ ). then ηk → γ.6). and for ψ in C ∞ (M.
k
(7. Now. Let B := O \ A and let A1 be the subset of A on which the minimizing periodic orbit is hyperbolic.
181
• (3) Finally let L : F → R is the linear map such that if µ is in C then L(µ) = Ldµ.3 and inequality 1. For. for k large we have that.
0) = 0. Since O = A ∪ B is generic and (U ∪ ψ) ∩ (A ∪ B) = (U ∩ A) ∪ (U ∪ ψ) ∩ B ⊆ A1 ∪ B . then U ∪ ψ is open and dense in C ∞ (M. v >x −U (x). Lxx (x0 . Hence φk ∈ A1 and A1 contains a neighbourhood of φ in A. This implies that A1 ∪ B is generic. The perturbation to achieve hyperbolicity in a ﬁxed point is easy. 0) is a singularity of the energy level c(L).
neighbourhood of a hyperbolic orbit there exists a unique invariant measure supported on it. For. which is zero if and only if x is on π(Γ) that is nondegenerate in the transversal direction. We add a C ∞ small non negative potential ψ. And it is hyperbolic if and only if it is positive deﬁnite. since ηk → γ. then ηk is hyperbolic. 0) is positive semideﬁnite in linear coordinates in Tx0 M . ψ ∩ A = because A ⊆ A1 ⊆ U and ψ ∩ A ⊆ A \ U = . In fact. The reader 2 can verify that if x0 = max U then the Dirac measure supported on the point (x0 . then A1 ∪ B is generic. R) such that A1 = U ∩ A. Diﬀerentiating the energy function (1. So to achieve hyperbolicity we must just add a small quadratic form. It follows that Γ is also a minimizing solution of the perturbed lagrangian L + ψ. R). let ψ := int (C ∞ (M. We shall prove below that A1 is dense in A. Moreover. from the EulerLagrange equation (EL) we get that Lx (x0 . the minimizing property of µ implies that x0 is a minimum of the function x → Lxx (x. Let U be an open subset of C ∞ (M.182
7. In particular. R) \ U). Is very much as the mechanic case: L = 1 < v. 182
. And it is well known that this critical point is hyperbolic if and only if the maximum has non degenerate quadratic form.3) we see that (x0 . generic lagrangians. 0) is minimizing. Moreover. Thus. π(Γ) is a simple closed curve. Because of the graph property the projection of the orbit Γ. The perturbation needed in the case of a periodic orbit the same spirit. 0). and it is in fact supported in the periodic orbit.
Let H be the associated hamiltonian by the Legendre transformation on T ∗ M and ψ its ﬂow. θ2 ∈ T ∗ M are said to be conjugate if θ2 = ψτ (θ1 ) for some τ = 0 and dψτ (ψ(θ1 )) ∩ ψ(θ2 ) = {0}. Fix a riemannian metric on M and the corresponding induced metric on T ∗ M . Denote by π : T ∗ M → M be the canonical projection and deﬁne the vertical subspace on θ ∈ T ∗ M by ψ(θ) = ker(dπ).4 Proposition. E(θ)∩ψ(θ) = F(θ)∩ψ(θ) = {0}.
183
To prove that it actually is hyperbolic is much more diﬃcult. These bundles where constructed for disconjugate geodesics of riemannian metrics by Green [27] and of Finsler metrics by Foulon [24]. To explain the idea of the proof we need some deﬁnitions. In the general case where constructed in [13] We will only sketch the proof. Suppose that the orbit of θ ∈ T ∗ M does not contain conjugate points and H(θ) = e is a regular value of H. where X(θ) = (Hp . Two points θ1 . F ⊂ T (T ∗ M ) along the orbit of θ given by
E(θ) = lim dψ−t ψ(ψt (θ)) .
t→+∞
F(θ) = lim dψt ψ(ψ−t (θ)) . −Hq ) is the hamiltonian vector ﬁeld and Σ = H −1 {e}. E(θ)∪F(θ) ⊂ Tθ Σ. X(θ) ⊆ E(θ)∩F(θ) and dim E(θ) = dim F(θ) = dim M .71. Then Tθ T ∗ M splits as a direct sum of two lagrangian subspaces: the vertical subspace ψ(θ) = ker (dπ(θ)) and the horizontal sub183
. The reason is that the linearization of the ﬂow (the Jacobi equation) is along the periodic orbit and hence non autonomous as in the case of a singularity. generic lagrangians.
t→+∞
Moreover. Then there exist two ϕinvariant lagrangian subbundles E. A basic property of orbits without conjugate points is given by the following 71.
ε[ we have that dψt (E) ∩ ψ(ψt (θ)) = {0}. It can be checked that E is lagrangian if and only if in symplectic coordinates S is symmetric. Then we can write dψt (E) = graph S(t). ξ → (dπ(θ) ξ . that in local coordinates coincide with the matrices of partial derivatives
∂2H ∂2H ∂2H ∂qi ∂pj ∂2H ∂qi ∂qj
. 184
. If we
∂ choose local coordinates along t → π ψt (θ) such that t → ∂qi (π ψt (θ)) are parallel vector ﬁelds. and Hqq . since the hamiltonian H is convex.
. dp(ξ)). Moreover. Hpp and Hqq are linear operators on Tπ(θ) M . Let E be a lagrangian subspace of Tθ T ∗ M . (7. generic lagrangians. v) ∈ Tθ T ∗ M = H(θ) ⊕ ψ(θ) ≈ Tπ(θ) M ⊕ Tπ(θ) M .
space H(θ) given by the kernel of the connection map. Hqp . we obtain the hamiltonian Jacobi equations ˙ h= Hpq h + Hpp v . S(t) h(t) . Suppose that for t in some interval ] − ε. Consider a variation
αs (t) = qs (t). Using the isomor∗ phism K : Tθ T ∗ M → Tπ(θ) M × Tπ(θ) M . We derive now the Riccati equation. ˙
where the covariant derivatives are evaluated along π(αo (t)). if ξ ∈ E then dψt (ξ) = h(t). where S(t) : H(ψt θ) → ψ(ψt θ) is a symmetric map. then this identiﬁcation becomes ξ ↔ (dq(ξ). θ (π ξ)). That is. Writing dψt (ξ) = h(t). Then E is a graph of some linear map S : H(θ) → ψ(θ). αs is a solution of the hamiltonian H such d that α0 (0) = θ and ds αs (0)s=0 = ξ. ε[. ∂pi ∂pj and ∂pi ∂qj . ps (t) such that for each s ∈] − ε.184
7. then Hpp is positive deﬁnite. we can ∗ identify H(θ) ≈ Tπ(θ) M × {0} and ψ(θ) ≈ {0} × Tπ(θ) M ≈ Tπ(θ) M . Take θ ∈ T ∗ M and ξ = (h.10)
v = −Hqq h − Hqp v . v(t) . Let E ⊂ Tθ T ∗ M be an ndimensional subspace such that E∩ψ(θ) = {0}.
71. generic lagrangians.
185
Using equation (7.10) we have that ˙ Sh + S(Hpq h + Hpp Sh) = −Hqq h − Hqp Sh . Since this holds for all h ∈ H(ψt (θ)) we obtain the Riccati equation: ˙ S + SHpp S + SHpq + Hqp S + Hqq = 0 . (7.11)
Let Kc (θ) : H(θ) → ψ(θ) be the symmetric linear map such that graph(Kc (θ)) = dψ−c (ψ(ψc (θ))). Deﬁne a partial order on the symmetric isomorphisms of Tπ(θ) M by writing A B if A − B is positive deﬁnite. The following proposition based essentially on the convexity of H proves 71.4. 71.5 Proposition. For all ε > 0, (a) If d > c > 0 then K−ε (b) If d < c < 0 then Kε (c)
d→+∞
Kd Kd
Kc . Kc .
lim Kd = S, U.
d→−∞
lim Kd = U.
(d) S
(e) The graph of S is the stable green bundle E and the graph of U is the unstable green bundle F An example of the relationship between the transversality of the Green subspaces and hyperbolicity appears in the following 71.6 Proposition. Let Γ be a periodic orbit of ψt without conjugate points. Then Γ is hyperbolic (on its energy level) if and only if E(θ) ∩ F(θ) = X(θ) for some θ ∈ Γ, where X(θ) is the 1dimensional subspace generated by the hamiltonian vector ﬁeld X(θ). In this case E and F are its stable and unstable subspaces. 185
186
7. generic lagrangians.
This proposition follows ideas of Eberlein [18] and Freire, [25]. It is known that minimizing orbits do not have conjugate points. So by proposition 71.4 and 71.6 to prove the density of hyperbolicity it is enough to perturb to make the Green bundles transverse. This is done using two formulas for the index. One in the lagrangian setting and another one in the hamiltonian setting. Let ΩT be the set of continuous piecewise C 2 vector ﬁelds ξ along γ[0,T ] . Deﬁne the index form on ΩT by
T
I(ξ, η) =
0
˙ ˙ ξ Lvv η + ξ Lvx η + ξ Lxv η + ξ Lxx η dt , ˙ ˙
(7.12)
which is the second variation of the action functional for variations f (s, t) with ∂f ∈ ΩT . For general results on this form see Duistermaat [17]. ∂s From this formula it is easy to compare the index of the original and the perturbed lagrangian along the same solution Γ. Finally we use the following transformation of the index form. It is taken from Hartman [29] and originally due to Clebsch [10] see also [13]. Let θ ∈ T ∗ M and suppose that the orbit of θ, ψt (θ), 0 ≤ t ≤ T does not have conjugate points. Let E ⊂ Tθ T ∗ M be a lagrangian subspace such that dψt (E) ∩ ψ(ψt (θ)) = {0} for all 0 ≤ t ≤ T . Let E(t) := dψt (E) and let H(t), ψ(t) be a matrix solution of the hamiltonian Jacobi equation (7.10) such that det H(t) = 0 and E(t) = Image (H(t), ψ(t)) ⊂ Tψt (θ) (T ∗ M ) is a lagrangian subspace. For ξ = Hζ ∈ ΩT , η = Hρ ∈ ΩT , we obtain (see [13])
T
I(ξ, η) =
0
(Hζ )∗ (Hpp )−1 (Hρ ) dt + (Hζ)∗ (V ρ)T. 0
(7.13)
Deﬁne N(θ) := { w ∈ Tπ(θ) M  w, γ = 0 }. Then N(θ) is the sub˙ ∂ ∂ space of Tπ(θ) M generated by the vectors ∂x2 ,. . ., ∂xn . Let vo ∈ N(θ), ˜ vo  = 1 and let ξ T (t) := ZT (t) vo . Denote by IT and IT be the index forms on [0, T ] for L and L respectively. Using the solution (ZT , VT ) on 186
71. generic lagrangians.
187
formula (7.13), we obtain that ˜ ˜ IT (ξ T , ξ T ) = −(ZT (0) vo )∗ (VT (0) vo ) = −vo ∗ KT (0) vo . (7.14)
∂ ∂ Moreover, in the coordinates (x1 , . . . , xn ; ∂x1 , . . . , ∂xn ) on T U we have that
˜ IT (ξ T , ξ T ) =
0
T
˙ ˙ ˙ ξ T Lvv ξ T + 2 ξ T Lxv ξ T + ξ T Lxx ξ T dt ˙ ˙ ˙ ξ T Lvv ξ T + 2 ξ T Lxv ξ T + ξ T Lxx ξ T dt
T n
T
=
0
(7.15)
2
+
0
ε
i=2
T ξi
dt .
We have that ZT (0) = I and for all t > 0, limt→∞ ZT (t) = h(t) ˜ with h(t) the solution of the Jacobi equation for H corresponding to the stable Green bundle. Writing πN (ξ) = (ξ2 , ξ3 , . . . ξn ) then πN h(0) v0 = v0  = 1 because v0 ∈ N(θ). Hence there exists λ > 0 and T0 > 0 such that πN ξ T (t) = πN ZT (t) v0 > 1 for all 0 ≤ t ≤ λ and T > T0 . 2 Therefore ελ ˜ IT (ξ T , ξ T ) ≥ IT (ξ T , ξ T ) + . (7.16) 4 Let (h(t), v(t)) = dψt ◦ dπE(θ) be the solution of the Jacobi equation for H corresponding to the stable Green subspace E and let S(ψt (θ)) = v(t) h(t)−1 be the corresponding solution of the Riccati equation, with graph[S(ψt (θ)] = E(ψt (θ)). Using formula (7.13), and writing ξ T (t) = h(t) ζ(t), we have that
T −1
IT (ξ T , ξ T ) =
0
−1 ˙ ˙ (h ζ)∗ Hpp (h ζ) dt + 0 − h(0) ζ(0)
∗
v(0) ζ(0) (7.17)
∗ IT (ξ T , ξ T ) ≥ −vo S(θ) vo .
From (7.14), (7.16) and (7.17), we get that ελ ∗ ∗ ˜ . vo S(θ) vo ≥ vo KT vo + 4 187
188
7. generic lagrangians.
˜ From proposition 71.5, we have that limT →+∞ KT (0) = S(θ), where ˜ Therefore graph(S(θ)) = E(θ), the stable Green bundle for H.
∗ ∗ vo S(θ) vo ≥ v0 S(θ) vo +
ελ . 4
(7.18)
Similarly, for the unstable Green bundles we obtain that
∗ ∗ vo U(θ) vo + λ2 ≤ vo U(θ) vo
for vo ∈ N(θ), vo  = 1.
(7.19)
for some λ2 > 0 independent of vo . From proposition 71.5 we have that U(θ) S(θ). From (7.18) and (7.19) we get that UN UN SN SN . Since E(θ) = graph(S(θ)) and F(θ) = graph(U(θ)), we get that E(θ) ∩ F(θ) ⊆ X(θ) . Then proposition B shows that Γ is a hyperbolic periodic orbit for L + φ. This proves that A1 is dense in A. Let A2 be the subset of A1 of potentials ψ for which the minimizing hyperbolic periodic orbit Γ has transversal intersections. The proof that A2 is open and dense in A1 is similar to the previous proof, see [13].
72
Homoclinic Orbits.
Assume in this section that Σ contains only one static class. By theorem 70.1.(A), this is true for generic lagrangians. By proposition 311.4, the static classes are always connected, thus if we assume that there is only one static class, Σ must be connected. Given ε > 0, let Uε be the εneighbourhood of π(Σ). Since Σ is connected, the open set Uε is connected for ε suﬃciently small. Let H1 (M, Uε , R) denote the ﬁrst relative singular homology group of the pair (M, Uε ) with real coeﬃcients. We shall say that an orbit of L is homoclinic to a closed invariant set K ⊂ T M if its α and ωlimit sets are contained in K. Observe that to each homoclinic orbit x : R → M to the set of static orbits Σ we can associate a homology class in H1 (M, Uε , R). Indeed, 188
they are semistatic for lifts of L to suitable ﬁnite covers. R) homoclinic orbits to the set of static orbits Σ. R). We note that the homoclinic orbits that we obtain in theorem 72. Uε . R) such that if ψ ∈ O the lagrangian L+ψ has a unique minimizing measure 189
. Uε . we conclude that the number of static classes of L0 must be ﬁnite. R). theorem 70. Let us denote by H the subset of H1 (M. Let Uε be an εneighbourhood of supp(µ). For any ε suﬃciently small the set H of homoclinic orbits to supp(µ) generates over R the relative homology H1 (M.72. Given a lagrangian L there exists a generic set O ⊂ C ∞ (M. Then for any ε suﬃciently small the set H generates over R the relative homology H1 (M. In particular.
189
since there exists t0 > 0 such that for all t with t ≥ t0 . Uε . Combining corollary 72. R) homoclinic orbits to supp(µ). Uε . In particular. However. 72. From theorems 72. we obtain 72. Given a lagrangian L there exists a generic set O ⊂ C ∞ (M.2 Corollary. R) such that if ψ ∈ O the lagrangian L + ψ has a unique minimizing measure µ in M0 (L + ψ) and this measure is uniquely ergodic.3 Corollary.4.2. R) given by all the classes corresponding to homoclinic orbits to Σ.1 to L0 to deduce that the group generated by the homoclinic orbits to the set of static orbits of L coincides with H1 (M. there exist at least dim H1 (M.1 we obtain: 72. R). Z) by a ﬁnite index subgroup. the class of x[−t0 . Uε .1 and corollary 72.1 we consider ﬁnite coverings M0 of M whose group of deck transformations is given by the quotient of the torsion free part of H1 (M. there exist at least dim H1 (M. homoclinic orbits. Using that the lifted lagrangian L0 has the same critical value as L. Uε . Uε . Hence we can apply theorem 311.1 and lemma 72. Uε . Let M be a closed manifold with ﬁrst Betti number ≥ 2. x(t) ∈ Uε .t0 ] deﬁnes an element in H1 (M.2 have energy c but they are not semistatic orbits of L. R).1 Theorem. To prove theorem 72. Suppose that Σ contains only one static class.1 and 70.
Let p : M1 → M2 be a covering such that c(L1 ) = c(L2 ). Proof: Since A is diﬀeomorphic to a circle. y) ≥ Φ2 (px.20) and the fact that x2 is semistatic we have for s ≤ t. we have that H1 (M. Uε ) coincides with the singular homology of the pair (M. Uε . Then any lift of a semistatic curve of L2 is a semistatic curve of L1 . R) − 1 ≥ 1.20)
for all x and y in M1 . When this measure is supported on a periodic orbit. this orbit is hyperbolic and the stable and unstable manifolds have transverse homoclinic intersections. 72. py). c c (7. generic lagrangians. Suppose now that x2 : R → M2 is a semistatic curve of L2 and let x1 : R → M1 be any lift of x2 to M1 .4 Lemma. R) ≥ 2. py).
in M0 (L + ψ) and this measure is uniquely ergodic.190
7. y) ≥ Φ2 (px. A) and therefore the vector space H1 (M. R) is non zero for all ε suﬃciently small. Also the projection of a static curve of L1 is a static curve of L2 . p is a ﬁnite covering. 190
. k k for all x and y in M1 . Let M be a closed manifold with ﬁrst Betti number b1 (M.5 Lemma. R) must have dimension ≥ b1 (M.1 we shall need the following lemma: 72. Hence if we write c = c(L1 ) = c(L2 ) we have Φ1 (x. Uε . Proof: Observe ﬁrst that for any k ∈ R we have that Φ1 (x. For the proof of theorem 72. the singular homology of the pair (M. then any lift of a static curve of L2 is a static curve of L1 . If in addition. Then if A ⊂ M is a closed submanifold diﬀeomorphic to S 1 and Uε denotes the ε neighborhood of A. Using (7.
there exists a curve β joining x1 (t) to x1 (s) such that AL1 +c (x1 [s.t] ) = AL2 +c (x2 [s. The vector space H1 (M. Using (7. given s ≤ t and ε > 0. Since p is a ﬁnite covering. where Σ(L) is the set of static vectors of L. x1 (t)) = AL1 +c (x1 [s. R)).20) and the fact that x1 is static we have for s ≤ t.t] ) c = Φ2 (x2 (s). x1 (s)). Let x1 : R → M be any lift of x2 to M1 . x2 (s)) ≥ −Φ1 (x1 (t).t] ) + AL2 +c (α) ≤ ε.t] ) + AL1 +c (β) ≤ nε. Hence. the set U is also connected for small ε. c c Hence x1 is semistatic for L1 . 191
. there exists a positive integer n. Suppose now that p is a ﬁnite covering and let x2 : R → M2 be a static curve of L2 . −Φ1 (x1 (t). Let i : U → M be the inclusion map. Proof of theorem 72. x1 (s)) = Φ1 (x1 (s). homoclinic orbits. Suppose now that x1 : R → M1 is a static curve of L1 and let x2 : R → M2 be p ◦ x1 . T ] → M2 with α(0) = x2 (t).72. x1 (t)) ≤ AL1 +c (x1 [s. there exists a curve α : [0.1: def Let U = Uε denote the εneighborhood of π(Σ(L)). x1 (t)). Since x2 is static. c c c Hence x2 is static for L2 . x2 (t)) ≤ Φ1 (x1 (s). and thus x1 is static for L1 . R) is isomorphic to the quotient of H1 (M. R) by i∗ (H1 (U. α(T ) = x2 (s) such that AL2 +c (x2 [s. such that the nth iterate of x2 [s.t] ) = AL2 +c (x2 [s. U.
191
Φ1 (x1 (s). Since we are assuming that Σ(L) contains only one static class.t] ) c c ≥ Φ2 (x2 (s). x2 (t)) ≥ −Φ2 (x2 (t). bounded from above by the number of sheets of the covering.t] ∗ α lifts to M1 as a closed curve.
we obtain a surjective homomorphism π1 (M ) → G/J. c(L) = c(L0 ) and therefore by lemma 72.7 Lemma. For any J as above. Proof: Let L0 denote the lift of the lagrangian L to M0 .2. Indeed.5 we have π0 (Σ(L0 )) = p−1 (π(Σ(L))). Since J is a subgroup of G = H/K.
Let H denote the torsion free part of H1 (M. the image of H under j is precisely G/J.192
7. G/J acts transitively and freely on the set of connected components of p−1 (U ) which coincides with the set of connected components of p−1 (π(Σ(L))). Let us write G = H/K = Z ⊕ . Z) → H. (7. generic lagrangians. since there exists t0 > 0 such that for all t with t ≥ t0 . H → G and j : G → G/J.21) 192
def
. of G. Z)). 72. U. Z). k . the class of x[−t0 . Z). Let J be a ﬁnite index subgroup . If we take the Hurewicz map π1 (M ) → H1 (M. whose kernel will be the fundamental group of a ﬁnite covering M0 of M with covering projection map p : M0 → M and group of deck transformations given by the ﬁnite abelian group G/J. R).6 Lemma. x(t) ∈ U .t0 ] deﬁnes an element in H1 (M. and we compose it with the projections H1 (M. There is a surjective homomorphism j : G → G/J given by the projection. Observe ﬁrst that by proposition 27. U. Observe that to each homoclinic orbit x : R → M to Σ(L) we can associate a homology class in H/K. Let us denote by H the subset of H/K given by all the classes corresponding to homoclinic orbits to Σ(L). Therefore we have 72. Z) and let K denote the torsion free part of i∗ (H1 (U. ⊕ Z where k = dim H1 (M. There is a one to one correspondence between elements in G/J and connected components of p−1 (π(Σ(L))).
Since p : M0 → M is a ﬁnite covering there are lifts x1 of px and y1 of py such that d0 (x1 .
193
fig. Let us prove now that L0 satisﬁes the hypothesis of theorem 311.1. Using (7.21) and proposition 311. that is. In fact. the number of static classes of L0 is ﬁnite. y) = 0. homoclinic orbits. where π0 : T M0 → M0 is the canonical projection of the tangent bundle T M0 to M0 . py) = 0.4 we see that the projection of a static class of L0 to M0 must be contained in a single connected component of p−1 (π(Σ(L))). Since static classes are connected x1 and y1 must c belong to the same connected component of p−1 (π(Σ(L))) and thus there is a covering transformation taking x1 into x and y1 into y which implies that d0 (x. 1: Creating homoclinic connections with ﬁnite coverings. y) = 0 as desired. it suﬃces to show that if x and y belong to a connected component of p−1 (π(Σ(L))) then d0 (x. y1 ) = 0. we shall show that the projection to M0 of a static class of L0 coincides with a connected component of p−1 (π(Σ(L))). Hence. Since we are assuming that Σ(L) c contains only one static class we have that dc (px. c 193
.72.
But this contradicts (7. If A has the property that for all J as above j(A) = G/J. Let A be a subgroup of G. 72. • If the rank of A equals the rank of G. Therefore if we combine lemma 72.22) G/A = {0} and thus G = A. one can easily construct a subgroup J ⊂ G with ﬁnite index such that A ⊆ J and G/J = {0}.
Now theorem 311.8 Lemma. Given a ﬁnite index subgroup . generic lagrangians. We shall need the following algebraic lemma.1.1 and (7. Let G = Z ⊕ .21) imply that every covering transformation in G/J can be written as the composition of covering transformations that arise from elements in H. k . ⊕ Z.7 and lemma 72. then A has ﬁnite index in G and by (7.22) because A/A ∩ J = {0}.8 with H = A we deduce that the set H of classes corresponding to homoclinic orbits generates G and must have at least k elements thus concluding the proof of theorem 72.22)
• If the rank of A is strictly less than the rank of G.194
7. that is. then A = G. Proof: The hypothesis readily implies that A/A ∩ J is isomorphic to G/J (7. j( H ) = G/J. J ⊂ G let us denote by j : G → G/J the projection homomorphism.
194
.
Observe now that any set H of a free abelian group G of rank k such that the group generated by H is G must have at least k elements.
A Absolutely continuous functions. b]. (ii) f ∈ L1 ([a.
whenever ]s1 . b].
. if B is a Borel set 195
t af
(s) ds.e. b] → R is absolutely continuous.Appendix. We claim that µ deﬁnes a ﬁnite signed Borel measure on [a. The function µ can be extended to a σadditive function on A. b]). .2 Proposition. Moreover. (iii) f (t) = f (a) + Proof: Deﬁne µ([s. b] → R is absolutely continuous if and only if (i) The derivative f (t) exists for a. A function f : [a. A. t1 [. if ∀ ε > 0 ∃ δ > 0 such that
N N
ti − si  < δ
i=1
=⇒
i=1
f (ti ) − f (si ) < ε. b]. t ∈ [a.
A.1 Deﬁnition. tN [ are disjoint intervals in [a. The function f : [a. Indeed. . let A be the algebra of ﬁnite unions of intervals. ]sN . . . t]) := f (t) − f (s).
196
.
By the Lebesgue diﬀerentiation theorem 1 f (t + h) − f (t) = lim h→0 h h→0 h lim
t+h
g = g(t)
t
for a. f dm. where m is the Lebesgue measure.196
appendix. m(A) < δ =⇒ µ(A) < ε. t ∈ [a. let µ(A) =
Conversely. A. t]) =
a
g(s) ds. Then by (iii). t ∈ [a. b] → R is absolutely continuous if and only if t there exists g ∈ L1 ([a.
and {An }n∈N ⊂ A is a family with An ↓ B. A µ([s.
The Lebesgue diﬀerentiation theorem gives the following characterization. then µ
m is equivalent. t]) = f (t) − f (s) Then µ
for s. The function f : [a. using the BorelCantelli lemma. suppose that (i)(iii) hold. then µ(B) := limn µ(An ) exists because µ(Bn \ Bm ) → 0 when n.
1
If µ is ﬁnite. dµ Let g = dm be the RadonNikodym derivative.e. b]. Then g ∈ L1 and
t
f (t) − f (a) = µ([a. Using (ii). Observe that the properties of external measures and the absolute continuity of f imply that µ m. b].
m implies1 that f is absolutely continuous. to
∀ e > 0 ∃ δ > 0. m → +∞.3 Corollary. b]) such that f (t) = f (a) + a g(s) ds.
Then any positive linear functional I : P (X) → R deﬁnes a unique Borel measure µ on X such that I(f ) = f dµ for all f ∈ Cb (X).B.
197
. [28] Let X be a locally compact Hausdorﬀ topological space and let Cb (X) the vector space of continuous functions f : X → R with compact support.1 Riesz Theorem. measure theory
197
B
Measure Theory
B.
if the set { (x. (c) If ∂f (x) = {p} then f is diﬀerentiable at x and f (x) = p.
A function f : Rn → R is said convex if f λ x + (1 − λ) y ≤ λ f (x) + (1 − λ) f (y) for all 0 ≤ λ ≤ 1 and x. For the proof of the following proposition see Rockafellar [62]. y ∈ Rn . C. (a) ∂f (x) = for every x ∈ Dom(f ).1 Proposition. For x0 ∈ Rn the subdiﬀerential of f at x0 is the set ∂f (x0 ) := { p ∈ Rn  f (x) ≥ p (x − x0 ) + f (x0 ) }. r) ∈ Rn × R  r ≥ f (x) } is convex. r) ∈ Rn × R  r = p (x − x0 ) + f (x0 ) } are called supporting hyperplanes for f at x0 . and the planes { (x.
198
. The functional p ∈ Rn is called the slope of the hyperplane. Its elementst are called subderivatives or subgradient of f at x0 . Equivalently.198
appendix.
C
Convex functions.
∗ ∗
(b) A ﬁnite convex function is continuous and Lebesgue almost everywhere diﬀerentiable.
199
D
The Fenchel and Legendre Transforms. f (x) ≥ A x−B(A) . more explicitly.D.1 Proposition. the fenchel and legendre transforms. Proof: 1. If f and f ∗ are superlinear then f ∗∗ = f . From (D. 1. ∀x ∈ Rn ⇐⇒ f ∗ (p) ≤ B(p) . then f ∗ is ﬁnite on all Rn . p2 ∈ Rn we have that f ∗ λ p1 + (1 − λ) p2 = max λ p1 + (1 − λ) p2 x − f (x) n
x∈R
∗
≤ λ max [ p1 x − f (x) ] + (1 − λ) max [ p2 x − f (x) ] n n = λ f (p1 ) + (1 − λ) f (p2 ). D.
Given a convex function f : Rn → R the Fenchel Transform (or the convex dual of f is the function f : (Rn )∗ → R ∪ {+∞} deﬁned by f ∗ (p) = max [ p x − f (x) ] n
x∈R
(D. f is superlinear if and only if f ∗ is bounded on balls.1 is attained at some point x ∈ Rn .1) we get that f (x) ≥ p x − f ∗ (p) for all x ∈ Rn . If f is convex then f ∗ is convex. 2. 3. the maximum D. 2. If f is superlinear.1)
∗
The function f admits a supporting hyperplane with slope p ∈ Rn if and only if f ∗ (p) = +∞. Given 0 ≤ λ ≤ 1 and p1 . p ∈ (Rn )∗ 199
x∈R ∗ x∈R ∗
. If f is superlinear. ∀p ∈ Rn
∗
4.
suppose that fx (p) ≤ B(p). Given A ∈ R and x ∈ Rn n∗ such that p  = A and p v = p  x = A x. We have that
∗ fx (p) = max [ p v − fx (v) ] n v∈R
≤ max [ p v − p v ] + B(p) n
v∈R
= B(p). f ∗ is ﬁnite.
Hence. Then ∀ y ∈ Rn . Hence
f ∗ (px ) = max px y − f (y) = px x − f (x). n
y∈R
And f (x) = px x − f ∗ (px ) ≤ max p x − f (x) = f ∗∗ (x).
∗ Conversely. there exists px ∈ R x x x Then
f (x) = max [ p x − f ∗ (p) ] ∗
p∈Rn
≥ px v − B(px ) = A x − B(A).
f (y) ≥ f (x) + px (y − x). 4.
p∈Rn∗
Let px ∈ ∂f (x) =
. If b > 0 is such that f (x) > (p + 1) x − b.
max
200
. ∗
p∈Rn
3. Let p ∈ Rn . then p x − f (x) < b − x < f ∗ (p) − 1 Hence f ∗ (p) =
x≤b+1−f ∗ (p)
∗
for x > b + 1 − f ∗ (p). By item 3. [ p x − f (x) ]. f (x) ≥ sup [ p x − f ∗ (p) ] = f ∗∗ (x).200
appendix.
∗
then the Legendre transform L : Rn → Rn is a C 1 diﬀeomorphism given by L(x) = dx f . by L(x) = { p ∈ Rn  p x = f (x) + f ∗ (p) }. Proof: The function f is convex and it is superlinear because
1
f (x) = f (0) +
0 1
f (sx) ds
1
= f (0) +
0 0
sx f (ts x) x dt ds
≥ f (0) + 1 a x2 .
D.
∗
(D.
201
and the maximum is attained at some interior point xp in the closed ball x ≤ b + 1 − f ∗ (p). If f : Rn → R is C 2 and there is a > 0.
Observe that in this case we have f ∗ (0) = − min f (x) n
x∈R
and
f (0) = − min∗ f ∗ (p).3 Proposition.D. the fenchel and legendre transforms.
p∈Rn
If f : Rn → R is convex and superlinear we deﬁne the Legendre n∗ Transform L : Rn → 2R of f . such that y · f (x) · y ≥ a y2 for all x.2)
D. In this case f ∗∗ = f .2 Corollary. ∗ If f : Rn → R is convex and superlinear then so is f ∗ : Rn → R. y ∈ Rn . 2 201
.
This proves that L diﬀerentiable and singled valued. ∗
p∈Rn
Moreover.
∗
(D.3).1 the maximum f ∗ (p) = max [ p x − f (x) ] n
x∈R
is attained at some xp ∈ Rn . Moreover since dx L = f (x) is nonsingular. Then p ∈ L(xp ).2. By item 4 in proposition D. then L is a local C 1 diﬀeomorphism.2 implies that L(x) = arg max { p x − f ∗ (p) } = . from (D. Since
1
(y − x) · [ dy f − dx f ] =
0
f
sx + (1 − s)y ds > 0.1) we get that p x ≤ f (x) + f ∗ (p) for all x ∈ Rn .3)
By proposition D. We now prove that L is surjective.
From (D. Thus p = dx f = L(x).
202
. f ∗ is superlinear. Then item 4 in proposition D.
then x → dx f = L(x) is injective.202
appendix. if p ∈ L(x) then x = arg maxx∈Rn { p x − f (x) }. p ∈ Rn .
singular sets of convex funcions.6)
x − x ≥ ( − )(x − x ) ≥ x − x ≥ x − x . see [2. (E.5) . let Σk (f ) := { x ∈ Rn  dim ∂f (x) ≥ k }. ∀y ∈ Rn }. If f : Rn → R is a convex function then for all 0 ≤ k ≤ n the Hausdorﬀ dimension HD(Σk (f )) ≤ n − k.2 Lemma.E.
1 2
y − x2
∀x. We recall here an elegant proof due to Ambrosio and Alberti. (E. If k ∈ N.
(E.1 Proposition.
Since f is superlinear. y ∈ Rn . More can be said on the structure of Σk . x−x
2
2
. E.
203
E
Singular sets of convex funcions. 66] for example.4) x − x  ≤ − .
Let f : Rn → R be a convex function. see [3].
∈ ∂f (x).4) we have that f (x ) ≥ f (x) + (x − x) + f (x) ≥ f (x ) + (x − x ) + Then 0 ≥ ( − )(x − x ) + x − x − Therefore −
2 2 1 2 1 2
x −x . By adding x2 if necessary (which does not change Σk ) we can assume that f is superlinear and that f (y) ≥ f (x) + (y − x) + E. Then the sets ∂f (x) ⊂ Rn are convex.
∈ ∂f (x )
Proof: From inequality (E. Recall that its subdiﬀerential at x ∈ Rn is the set ∂f (x) := { : Rn → R linear  f (y) ≥ f (x) + (y − x) . the subdiﬀerential ∂f is surjective and we have: 203
. =⇒
∀ ∈ ∂f (x).
Since F is Lipschitz.
E.
204
. There exists a Lipschitz function F : Rn → Rn such that ∈ ∂f (x) =⇒ x = F ( ).204
appendix. For example Ak = {x ∈ Rn  x has at least k rational coordinates }. Observe that x ∈ Σk =⇒ ∂f (x) intersects Ak =⇒ x ∈ F (Ak ).
Therefore Σk ⊂ F (Ak ).3 Corollary.1: Let Ak be a set with HD(Ak ) = n − k such that Ak intersects any convex subset of dimension k. we have that HD(Σk ) ≤ HD(Ak ) = n − k.
Proof of Proposition E.
symplectic linear algebra.
205
.
205
F
Symplectic Linear Algebra.F.
206
appendix.
206
.
. e Non Lin´aire 16 (1999). Inst. Contreras. Lyapunov minimizing measures for expanding maps of the circle.. Thieullen. One dimensional variational problems whose minimizers do not satisfy the eulerlagrange equation. 55 (1958). J. Progress in Mathematics. [3] Giovanni Alberti and Luigi Ambrosio. Mikhael Gromov. Z.. 1077–1085. 325–388. no. 6. Marsden. Poincar´ Anal. Mech. Benjamin / Cummings Publishing Co. [7] C´sar Camacho and Alcides Lins Neto. Arch. [4] J. and Ph. A geometrical approach to monotone functions in n . 17–27. Math. Reine Angew. On minimizing measures of the action of autonomous Lagrangians. 1. Ergodic Theory Dynam. Math. no. e Birkhauser. 230 (1999). [6] Patrick Bernard and Gonzalo Contreras. Ration. 2007. [8] M.. Reading. 1978. e [9] Mario Jorge Dias Carneiro. 6. Systems 21 (2001). 61. ¨ J. Geometric theory of foliations. Foundations of Mechanics.Bibliography
[1] Ralph Abraham and Jerrold E. Nonlinearity 8 (1995). Mizel. 1379–1409. [11] G. 254–276. Lopes. Boston. Ball and V. Boston. O. H. Inc. 667–690. Var. Manifolds of nonpositive curvature. A generic property of families of lagrangian systems. & Partial Diﬀerential Equations 2 (1994). Ann. 2. Dias Carneiro and Arthur Lopes. Calc. Mass. no.
207
. 1985. no. Math. no. A. Anal. [10] Alfred Clebsch. 1985. On the structure of singular sets of convex functions. 259–316. [2] Giovanni Alberti. Birkhauser. On the minimal action function of autonomous Lagrangians associated to magnetic ﬁelds. and Viktor Schroeder. to appear in Ann. vol. uber die reduktion der zweiten variation auf ihre einfachste form. 5. 90 (1985). Advanced Book Program. [5] Werner Ballmann.
2. Geom. Hopf. 645–666. e
[24] Patrick Foulon. Ordinary Diﬀerential Equations. 4. 3. Gabriel P. 788–809. 1950. 155–196. J.
208
. Brasil. 901–952. Inst. no. no. Fluxos geod´sicos em variedades riemannianas compactas de e curvatura negativa. Paternain. Paris S´r. no. 9. 8 (1998). Evans and Ronald F. 327 (1998). Master’s thesis. [18] Patrick Eberlein. A theorem of E. Connecting orbits between static classes for generic Lagrangian systems.. Acad. 173–195. Estimation de l’entropie des syst`mes lagrangiens sans points e conjugu´s. [20] Albert Fathi. minimizing measures and Ma˜´’s critical values. I Math. IMPA. ne [13] Gonzalo Contreras and Renato Iturriaga. Acad. Paris S´r. New York. [28] Paul R. Renato Iturriaga. 4. Duistermaat. 2. C. C. [15] Gonzalo Contreras. 326 (1998). [27] L. Lagrangian ﬂows: the dynamics of globally minimizing orbits. 437–463. J. e e e e C.. 325 (1997). Sur la convergence du semigroupe de LaxOleinik. On the Morse index in variational calculus. C. 5 (1958). and Miguel Paternain. New YorkLondonSydney. and Renato Iturriaga. 2. Measure theory and ﬁne properties of functions. Th´or`me KAM faible et th´orie de Mather sur les syst`mes lagrangiens. Paris S´r. W. 3. 117–146. I Math. Convex Hamiltonians without conjugate points. e e [25] Alexandre Freire. e e Acad. no. Sci. Orbites h´t´roclines et ensemble de Peierls. CRC Press. . 21 (1976). Sci. Mat. 1992. Math. Acad. 1964. Renato Iturriaga. [26] Alexandre Freire and Ricardo Ma˜´. 1043–1046. 649–652. no. R. J. Ann. 324 (1997). e . no. 6. Sci. Solutions KAM faibles conjugu´es et barri`res de Peierls. Michigan Math. 31–34. Advances in Math. [14] Gonzalo Contreras. (N. Gariepy. Ergodic Theory Dynam.S. 267–270. Paris S´r. R.208
BIBLIOGRAPHY
[12] Gonzalo Contreras. 375–392. When is a geodesic ﬂow of anosov type? i.) 28 (1997). no. [16] Gonzalo Contreras and Gabriel P. Rio de Janeiro. Inc. R. C. 1213–1216. no. Invent. Bol. Halmos. no. Paternain. special issue in honor of Ricardo Ma˜´. [19] Laurence. ee e I Math. ne Funct. Measure theory. 5. Nonlinearity 11 (1998). Topology 41 (2002). On the entropy of the geodesic ﬂow in ne manifolds without conjugate points. Lagrangian graphs. 355–361. Diﬀerential Geometry 8 (1973). Jorge Delgado. and H´ctor S´nchezMorgado. Anal. [29] Philip Hartman. no. 1981. R. Henri Poincar´ 57 (1992). On the cree a ation of conjugate points for Hamiltonian systems. Green. John Wiley & Sons. 69 (1982). 10. e [21] [22] [23] . no. I Math. [17] J. Systems 19 (1999). SpringerVerlag. Soc. Sci. II.
1983. 19o Coloquio Bras. 466–468. Vol. 1978. Stable norms of surfaces: local structure of the unit ball of rational directions. 2 (Z¨rich. Rio de Janeiro. 1987. . Soc 128 (2000). . 230. . no. IXth International Congress on Mathematical Physics (Swansea.S. no. 6. Longman. and Varadhan S. [31] Wilhelm Klingenberg. pp. Minimal action measures for positivedeﬁnite Lagrangian systems. DordrechtBoston. On the minimizing measures of Lagrangian dynamical systems. Global Variational Methods in Conservative Dynamics. [32] P. 120–131. Homogenization of hamiltonjacobi equations. no. IMPA. 59–70. 221–224. Paris S´r. 996–1010. C. Mat. [41] .. 2. SpringerVerlag. unpublished. Modulus of continuity for Peierls’s barrier. 3. 1988). Papanicolau. Ergodic theory and diﬀerentiable dynamics. Z. . Helv. Grundlehren der Mathematischen Wissenschaften.. [33] Ricardo Ma˜´. no.R. Reidel. Mather. Basel. Amer. 1987. R. 169–207. u 1994). Lectures on closed geodesics. 7 (1997). Math. Funct. . 1. 2. 273–310. Translated from the Portuguese by Silvio Levy. 2.) 28 (1997). Geom. 1989. Nonlinearity 9 (1996). (N. G. Minimal measures. SpringerVerlag. 1995). Bol.) 21 (1990). Vol.L.S. MA. Proc. [43] . Diﬀerentiability of the minimal average action as a function of the rotation number. Comment. pp. I Math.
[44] [45] [46]
209
. 1986). 1996. Anal. 141–153. Bristol. 1993. Reprinted in Bol. . Action minimizing invariant measures for positive deﬁnite Lagrangian systems. Generic properties and problems of minimizing measures of Lagrangian systems. 2.BIBLIOGRAPHY
209
[30] Renato Iturriaga and H´ctor SanchezMorgado. a . 375–394. pp. Mat. Nonlinearity 5 (1992). 1216–1220. 64 (1989). pp. . Hilger.
[42] John N.S. 3. 3311–3316. Rio de Janeiro. [34] [35] [36] [37] . Proceedings of the International Congress of Mathematicians. Ergodic variational methods: new techniques and new problems. Math. Soc. preprint. BerlinNew York. Instituto de Matem´tica Pura e ne c˜ ` o a Aplicada. Brasil. Periodic solutions of Hamiltonian systems and related topics (Il Ciocco. 1987. Introdu¸ao a teoria erg´dica. 177–202. Soc. International Conference on Dynamical Systems (Montevideo. 623–638. Acad. (N. Birkh¨user. Lions. Sci. Normes stables des surfaces. Lagrangian ﬂows: the dynamics of globally minimizing orbits.
[38] [39]
[40] Daniel Massart. Finsler metrics and action poe tential. Harlow. Brasil. 1995. no. no. e 324 (1997). BerlinNew York. 207 (1991). Mat. Math. 1. no. no.
26 (1924). 1991. 1997. Helv. Lectures given at the Third C.E. no. J. and Degli Esposti. 3. M. 92–186.I. given at the Third C. July 6–13. 2. Acad. Roquejoﬀre. Palais. Edited by S. 207–263. Hyperbolic dynamics of EulerLagrange ﬂows on prescribed energy levels. Action minimizing orbits in Hamiltonian systems. e e e e e Grenoble I. preprint.
[52] G.
[60] [61]
210
. and Forni. [59] . 109–121. 1991. Ann. Paternain and Miguel Paternain. Variational construction of connecting orbits. 1. no. 127–151. Soc. no. SpringerVerlag.
[49] John N. R. Longman. Math. July 6–13. . Graﬃ. and Isola. 1991). Critical values of autonomous Lagrangian systems. 1994. Session held in Montecatini Terme. C. 15. pp.M. Sci. 25–60. . Ergodic Theory Dynamical Systems 12 (1992). 481–499. SaintMartind’H`res. N. S. Amer. No. 113 (1991). no. e
[57] Gabriel P. Transition to chaos in classical and quantum mechanics (Montecatini Terme. Edited by S. Topology 2 (1963).E. Trans. 324 (1997). Soc. 217 (1994). 367–376. pp. [58] 1994. Graﬃ. J. Math. 1367–1370. Paternain. no. 299– 340.M. On Anosov energy levels of convex Hamiltonian systems. Amer. Entropy and completely integrable Hamiltonian systems. Calculus of variations in the large. Harlow. Morse theory on Hilbert manifolds. Math. 871–873. 1995). no. no. S´minaire de Th´orie Spectrale et G´om´trie. 3. Lectures by Bellissard. Inst. 1349–1386. Session held in Montecatini Terme. First derivative of topological entropy for Anosov geodesic ﬂows in the presence of magnetic ﬁelds. On the topology of manifolds with completely integrable geodesic ﬂows. S. Springer. J. . Anosov geodesic ﬂows and twisted symplectic structures. Variational construction of orbits of twist diﬀeomorphisms. Comportement asymptotique des solutions d’une classe d’equations paraboliques et de hamiltonjacobi. and Graﬃ. A fundamental class of geodesics on any closed surface of genus greater than one. Serie I. pp. Fourier (Grenoble) 43 (1993). G. .210
BIBLIOGRAPHY
[47] [48]
. and Mather. International Conference on Dynamical Systems (Montevideo. 121–131. Soc. 4 (1991). Proc.M.. [55] [56] . 3. Comment. 1. Mather and Giovanni Forni. Math. Ann´e 1996–1997. 5. Berlin. [50] Marston Morse. Nonlinearity 10 (1997). Z. Univ. Math. Namah and J.I. On the topological entropy of autonomous lagrangian systems. 72 (1997). [53] Richard S. Amer. [54] Gabriel P. 132–145. 1996. [51] . Berlin. 1994. .
no. Z. Sci. Serie I. 41–55. Tyrrell Rockafellar. [64] Rafael Oswaldo Ruggiero. [66] Ludˇk Zajˇek. 270–284. Princeton University Press. On the creation of conjugate points. Ann. Comportement asymptotique des solutions d’equations de hamiltonjacobi monodimensionnelles. Asymptotic cycles. C.
211
. 185– 189. Carolinae 19 (1978). e c Math. 1970. Princeton Mathematical series. New Jersey. of Math. 21.326 (1998). (2) 66 (1957). Comment. 1. Univ. Roquejoﬀre. R. On the points of multiplicity of monotone operators. [63] J. 208 (1991). [65] Sol Schwartzman. 179–189.M. Acad. Math.BIBLIOGRAPHY
211
[62] R. Convex analysis.
212
BIBLIOGRAPHY
212
.
t). 45. 127. 23 C 0 . y). y. R). 99.Index
AL (γ). 44 Γ+ . 56 dk (x. 51. 38 S + (γ). y. 134 LipK (M. 71 Σ− (L). 71 P = π(Σ(L)). 198 213
. 24 e0 . 44 M(h). 89 ρ(µ). 30. 84 Γ+ . 79 S± (k). 71. 72 α(ω) = β ∗ (ω). 44. 39 M. 71 0 . 9 AL (µ). 50 A. 44 C(x. 141 β(h). 52 ca (L). 30. 88 f (x). q2 . 141 Σ(L). 9 CT (x. y). 52 cu (L). T ). 62 Σ+ (L). 198 . 53 d1 (γ1 . 10. 165 µγ . 50 M . 47 c(L). 38 [h]Lip . 23. 12. 141 . 44 ω(v). 38. T ). 143 ess sup . 28 N = Σ(L). 134 δ(s. 71 H. k). 23 Φk (x. y). 50 H(u). 28 · ∞ . 72 ∂f (x). 50 c0 (L). 31. 55 S(x. 71 Σ± (k). 84 0 Γ± (u. 28 C k (q1 . 134 α(v). 40 C(M ). y). 129 · . γ2 ). 137 Φk (x. T ). 31. 51 M(L).
167 alfa limit. T ). 90 class static. 7 superlinearity.214
INDEX
h(x. 27 of the homology function ρ. 24 of the critical value. 130
C
chain recurrent. 38. 15 Busemann weak KAM. 15 convexity. 44 boundedness. 31. 50 alpha limit. 79 hT (x. 79 xv . 71 AubryMather theory. 57 action of a curve. 23 of ﬁnite time. 24 aﬃnely independent. 9 of a measure. 79–81 basin of a static class. 13 for the Legendre transform. 90 transitive. 62. 84
. 62 action potential Lipschitz property. 8
A
abelian cover. 118 asymptotic cycle. 121 beta function. 52 real –. 86 properties. 47–51
of a weak KAM. y). 72 Anosov energy level. 73. 8. 47. 50 bound a priori – for speed. 7 conﬁguration space. 53 absolutely continuous functions. 48 continuity of the action potential. 188 of a semistatic. 56 absolutely equicontinuous. 47 Aubry set. 16 for the Peierls barrier. 24 triangle inequality. 86–87 condition boundedness. 84 closed form. y. 98 coboundary Lipschitz. 44 potential. 137 214
B
barrier Peierls’. 23. 79 lower – for a lagrangian. 14 weak. 63 for Lv . 90 alpha function. 16 for the energy. 47 contraction by a vector ﬁeld. y) = Φc (x. 8. 195–196 metric. 8.
45. 134 real abelian. 24. 13 regular. 118 compactness. 88 strict (c0 ). 52 of the universal cover. 106. 141 covering. 12–13 kinetic. 166 ess sup. 115 function. 118 of a minimizing measure. 17 ﬂow. 18 215
D
deck transformation. 191. 189 deformation retract. 122. 127 dominated ( ). 50 of the β function. 134 dual convex –.INDEX
215
convex dual. 50. 100 of a static class. 189. 84 of a weak KAM. 171. 8 magnetic lagrangian. 71 of a timefree minimizer. 204 convexity. 52. 53 strict. 15. 50 e0 ≤ cu ≤ ca ≤ c0 ≤ c(L). 38. 70 cut locus of a point. 48 distance
. 199
E
energy Finsler. 122. 52 curve semistatic. 52 ﬁnite. 17 equicontinuous absolutely. 50. 121
absolutely continuous. 17 level Anosov. 76 of a semistatic curve. 141 EulerLagrange equation. 44. 51. 198–204 set. 50 convolution. 118 critical value. 45. 105 abelian. 65 potential –. 23. 52 of the abelian cover (ca ). 19 mechanic lagrangian. 99 dominated function. 56 dominated. 199 function. 52–53. 8 reversible. 53 universal. 50. 70 energy of. 7 of the α function. 53. 88 of a covering. 99–102 domination properties. 71 static. 192 properties. 17 riemannian lagrangian. 88. 57 ergodic uniquely.
17 twisted geodesic. 198 216
H
HamiltonJacobi equation weak solution. 171. 14 function convex.216
INDEX
example c(L) > c0 (L) > e0 . 122. 30 homoclinic orbit homology of –. 23. 139 ﬂow embedded. 189. 199
F
Fenchel transform. 199 ﬁnite – time potential. 160
. 53 c0 (L) > cu (L). 98 properties. 115 lagrangian. 17. 47. 50. 53 extremal point. 192 Finsler energy. 97 hamiltonian. 28 superlinear. 14 riemannian. 114 of a Finsler metric. 44. 17. 21 EulerLagrange. 191 ﬁnite covering. 188 of a measure. 62 covering. 17 metric. 17 twisted –. 47 Horocycle ﬂow. 8. 98 Liouville’s 1form. 14 holonomic measure. 114 reversible. 17 vector ﬁeld. 188 to a static class. 18 graph lagrangian. 82–85. 98 growth linear. 199. 14 closed. 137. 103–104 theorem. 19 gradient. 15. 113–117 ﬁxed point LaxOleinik semigroup. 167
G
generic point. 14 ﬂow. 48 geodesic ﬂow Finsler. 188 homology of a homoclinic orbit. 50. 7. 134 submanifold. 167 extreme point. 8 Finsler geodesic –. 19 form canonical symplectic –. 18 riemannian geodesic. 17 riemannian. 204 (.
72 exact. 71 submanifold. 79 lagrangian weak KAM solution’s. 139 supporting. 18 217
. 167 Liouville’s 1form. 137 inverse. 40 reversible.INDEX
217
subspace. 88 coboundary. 137. 14 aﬃnely. 8. 18. 44 graph. 82 isotropic subspace. 100. 44 magnetic. 97 images of LaxOleinik. 14 inequality Lipschitz c(L) ≥ e0 . 17 Rademacher’s theorem. 198 ﬁxed point. 137 I Legendre transform. 17 action potential’s. 7. 122 solution. 8 smallest – constant. 23.. 18 mechanic. 140 slope of –. 199 Lipschitz images. 106 ﬂow. 15 weak contraction prop. 44 for a lagrangian. 137 hyperplane convergence. 86 invariant measures M(L). 18 subsolutions of HJ. 98 static curve. 97 magnetic lagrangian. 97 Hurewicz homomorphism. 134 Finsler. 31. 109 graph. 192 symmetric. 52. 121 property kinetic energy. 198. 17 M natural. 15. 23. 84 K properties KAM weak KAM solution’s. 53 LaxOleinik semigroup. 67 L Peierls barrier’s. 98 local cohomology class of. 53. 137 hypothesis on L. 98 lower bound lower bound. 17 Ma˜´ ne riemannian. 201 independent Lie derivative. 24 ﬁnitetime potential’s. 17 set.
47–51 Mather set. 48 potential action –.218
INDEX
EL equation. 158 is weak KAM. 188 of a semistatic. 127 Lipschitz property. 79. 106. 88. 137
P
partial order. 122. 72. 158. 137 weak contraction. 130. 73. 134
N
natural lagrangian. 86–87 covering –. 17 a priori bound for speed. 82 chain recurrency. 31 mechanic lagrangian. 17 minimizing measure. 198. 139 218
. 47 invariant. 31. 90 operator LaxOleinik. 31 support of. 76 invariance of –. 167 generic. 199 point extremal. 51 minimizing. 82 minimizing measures. 48 plane supporting. 167 extreme. 90 empty. 62. 44 properties coboundary –. 24 energy. 79 Peierls set P. 17
O
omega limit. 19 Mather beta function. 51. 38. 23 properties. 89 Peierls barrier. 76 phase space. 51 set. 17 EL equation. 75 energy of –. 75 invariance of –. 47 Mather minimizing. 140 ﬁxed point. 141
Lipschitz images. 44. 137 convergence. 63 probability holonomic. 160 nonemptyness. 75 molliﬁcation. 56 Finsler. 81 with KAM. 17 metric absolutely continuous. 71 theory. 44. 79–81 characterization with action potential. 129 inﬁnite. 71. 71 measure holonomic. 50 crossing lemma.
73. 134 graph –. 103– 104 static class. 198. 51. 70 α and ωlimits. 137 convergence. 134 weak contraction –. 167 slope of a hyperplane. 82–85. 90 simplex. 71 convex. 172. 134 Lipschitz –. 15 residual. 165 reversible ﬂow. 137 weak contraction. 79. 139 Lipschitz images.INDEX
219
domination –. 72 orbit. 71 lift of. 198 submanifold graph. 137
semistatic curve. 89–95 regular energy level. 122. 53 recurrence properties. 71. 17 riemannian lagrangian EL equation. 84. 32. 47
S
Schwartzman. 50 semigroup LaxOleinik. 199 smoothness properties. 18 riemannian hamiltonian. 97 219
. 134 solution weak – of HJ equation. 17 lagrangian. 106 ray. 98 lagrangian. 90 energy of. 71 existence. 204 Ma˜´. 73. 17 Riesz theorem. 143 real abelian cover. 190 strict critical value (c0 ). 90 curve local. 122. 71 ne Mather. 82 chain recurrency. 76 lift of. 71 Peierls. 52 subderivative. 89–95 smoothness –. 137. 47 semicontinuous. 73. 190 set Aubry. 122. 198 subdiﬀerential. 203–204 subgradient. 122. 197 rotation of a measure. 198. 134 recurrence –. 122. 137
R
Rademacher’s theorem. 70 energy of. 51. 140 ﬁxed point.
14
W
weak contraction. 18. 109
T
theorem HamiltonJacobi. 38. 50 supporting hyperplane. 199 superlinearity. 12 symmetric lagrangian. 199 of the α function. 19. 97 superlinear. 79 twisted geodesic ﬂow. 105 subspace isotropic. 14 canonical. 199 Legendre. 198. 23. 65 Tonelli minimizer. 97 lagrangian. 103–104 subsolution of HJ. 44 of convex duals. 130 solution of HJ. 160 uniquely ergodic. 55 timefree minimizer energy of. 100. 166 vector ﬁeld hamiltonian. 50 of the β function. 53 transform Fenchel. 201 triangle inequality for the action potential. 24 220
. 49 of the projection of an energy level. 121 Busemann –. 55 theorem. 97
for the Peierls barrier. 7. 15.220
INDEX
subsolution of the HamiltonJacobi equation. 53 symplectic form. 137 KAM solution. 55 torsion. 199 surjectivity of the homology function ρ. 97 Tonelli’s. 15.