EFFECTIVE HAMILTONIANS AND AVERAGING

FOR HAMILTONIAN DYNAMICS I

by

L. C. Evans 1 and D. Gomes 2

Department of Mathematics
University of California
Berkeley, CA 94720

Abstract. This paper, building upon ideas of Mather, Moser, Fathi, E and others, applies
PDE methods to understand the structure of certain Hamiltonian flows. The main point is
that the “cell” or “corrector” PDE, introduced and solved in a weak sense by Lions, Papanico-
laou and Varadhan in their study of periodic homogenization for Hamilton–Jacobi equations,
formally induces a canonical change of variables, in terms of which the dynamics are trivial.
We investigate to what extent this observation can be made rigorous in the case that the
Hamiltonian is strictly convex in the momenta, given that the relevant PDE does not usually
in fact admit a smooth solution.

1. Introduction.

This is the first of a projected series of papers that develop PDE techniques to under-
stand certain aspects of Hamiltonian dynamics with many degrees of freedom.

1.1. Changing variables.

The basic issue is this. Given a smooth Hamiltonian H : Rn × Rn → R, H = H(p, x),

we wish to examine the Hamiltonian flow

ẋ = Dp H(p, x)
(1.1)
ṗ = −Dx H(p, x)

under a canonical change of variables

(1.2) (p, x) → (P, X),

1 Supported in part by NSF Grant DMS-9424342.
2 Supported by Praxis XXI-BD 5228/95.

1
where

p = Dx u(P, x)
(1.3)
X = DP u(P, x)
for a generating function u : Rn ×Rn → R, u = u(P, x). Here we write Dx = ( ∂x ∂
1
, . . . , ∂x∂n )
∂
and DP = ( ∂P 1
, . . . , ∂P∂ n ). Assuming that we can find a smooth function u to solve the
Hamilton–Jacobi type PDE
(1.4) H(Dx u(P, x), x) = H̄(P ) in Rn ,
and supposing as well that we can invert the relationships (1.3) to solve for P, X as smooth
function of p, x, a calculation shows that we thereby transform (1.1) into the trivial dy-
namics

Ẋ = DH̄(P)
(1.5)
Ṗ = 0.
In terms of mechanics, P is an “action” and X an “angle” or “rotation” variable, as for
instance in Goldstein [Gd].
Of course we cannot really carry out this classical procedure in general, since the PDE
(1.4) does not usually admit a smooth solution and, even if it does, the transformation
(1.2), (1.3) is not usually globally defined. Only very special Hamiltonians are integrable
in this sense.

1.2. Homogenization.

On the other hand, under some reasonable hypotheses we can in fact build appropri-
ate weak solutions of (1.4), as demonstrated within another context in the classic-but-
unpublished paper Lions–Papanicolaou–Varadhan [L-P-V]. These authors look at the ini-
tial value problem for the Hamilton–Jacobi PDE
 ε  
ut + H Dx uε , xε = 0 in Rn × (0, ∞)
(1.6)
uε = g on Rn × {t = 0},
under the primary assumption that the mapping x → H(p, x) is Tn -periodic, where Tn
denotes the flat torus, that is, the unit cube in Rn , with opposite faces identified. Con-
sequently as ε → 0, the nonlinearity in (1.6) is rapidly oscillating; and the problem is to
understand the limiting behavior of the solutions uε . Lions et al. show under some mild
additional hypotheses on the Hamiltonian that uε → u, the limit function u solving a
Hamilton–Jacobi PDE of the form

ut + H̄(Dx u) = 0 in Rn × (0, ∞)
(1.7)
u = g on Rn × {t = 0}.
Here H̄ : Rn → R, H̄ = H̄(P ), is the effective (or averaged) Hamiltonian, and is built
from H as follows.
2
 1.3. How to construct H̄.

First, consider for fixed P ∈ Rn the cell (or corrector) problem


H(P + Dx v, x) = λ in Rn ,
x → v is Tn -periodic.
As proved in Lions–Papanicolaou–Varadhan [L-P-V] (and recounted in [E2] and in Braides–
Defranceschi [B-D, §16.2]), there exists a unique real number λ for which there exists a
viscosity solution. We may then define
H̄(P ) := λ,
and so rewrite the foregoing as

H(P + Dx v, x) = H̄(P ) in Rn ,
(1.8)
x → v is Tn -periodic.
Once we set
(1.9) u(P, x) := P · x + v(P, x),
the PDE in (1.8) is just (1.4).
Remark. We pause here to draw attention to some simple observations relating the
cell problem (1.8) and semiclassical approximations in quantum mechanics for periodic
potentials. These comments are intended as further motivation.
Consider the time-independent Schrödinger equation
2
(1.10) −∆ψ + V ψ = Eψ in Rn ,
2
where  is Planck’s constant, V : Rn → R is a Tn -periodic potential, and E is the energy
corresponding to the eigenstate ψ : Rn → C. A standard textbook procedure is to look for
a solution having the Bloch wave form
P ·x
(1.11) ψ = ei
φ,
where φ : Rn → C is Tn -periodic. We further suppose φ to have the WKB–structure

v
(1.12) φ = aei

for periodic a, v : Rn → R. Our substituting (1.11), (1.12) into (1.10) and taking real parts
yields
1
(1.13) |P + Dx v|2 + V (x) = E,
2
up to terms formally of size O().
Thus in the semiclassical limit  → 0, we heuristically obtain the cell problem (1.8) for
the Hamiltonian H(p, x) = 12 |p|2 + V (x) and H̄(P ) = E.

3
 1.4. Questions, absolute minimizers.

The procedure outlined in §1.3 provides us with at least a theoretical construction of

H̄ and of a generating function u. Returning then to the comments in §1.1, we can now
formulate these
Basic Questions. To what extent can we employ H̄ and u to understand the solutions
of the Hamiltonian flow (1.1)? In particular, how is information about the dynamics
“encoded” into H̄?
These are really hard issues, and to make at least a little progress we will need some
additional hypotheses on both the Hamiltonian and the particular trajectories of the ODE
we examine. Let us henceforth suppose that the mapping p → H(p, x) is uniformly convex,
in which case we can associate with H the Lagrangian

L(q, x) := max(p · q − H(p, x)).

p

Consider then a Lipschitz curve x(·) which minimizes the associated action integral, mean-
ing that
 T  T
(1.14) L(ẋ, x) dt ≤ L(ẏ, y) dt
0 0

for each time T > 0 and each Lipschitz curve y(·) with x(0) = y(0), x(T ) = y(T ). We call
x(·) a (one-sided) absolute minimizer. If we as usual define the momentum

p := Dq L(ẋ, x),

then (x(·), p(·)) satisfy Hamilton’s ODE (1.1).

A discovery of Aubry [A], Mather [Mt 1-4], Fathi [F1–F3], Moser [Mo], E [EW2], etc., is
that solutions of (1.1) corresponding to absolute minimizers are in a strong sense “better”
than other solutions. Indeed, these authors have shown that the Hamiltonian dynamics
are in some sense “integrable” for such special trajectories. The main goal of our work is to
continue this analysis, with particular emphasis upon PDE methods (based upon viscosity
solutions of (1.8)), applied to problems with many degrees of freedom.

1.5. Outline.

In §2 below we review the definition of the effective Hamiltonian H, introduce the corre-
sponding effective Lagrangian L̄, and recall the connections with the large time asymptotics
of absolute minimizers x(·).
4
 We then rescale in time x(·) and p(·) in §3, and introduce certain Young measures
{νt }t≥0 on phase space, which record the oscillations of the rescaled functions in asymptotic
limits. These measures contain information about the Hamiltonian flow, and so our goal
in subsequent sections is understanding their structure. In §4 we show that each ν = νt
is supported on the graph of the mapping p = Dx u(P, x) and furthermore “stays away”
from the discontinuities in Dx u.
In §5 we prove that u is well behaved on the support of σ, the projection of ν onto
x-space. For this, we firstly derive the formal L2 -bound

(1.15) |Dx2 u|2 dσ ≤ C
Tn

and then the L∞ -estimate

(1.16) |Dx2 u| ≤ C σ-a.e..

We rigorously establish some analogues of (1.15), (1.16), entailing difference quotients in

the x-variables. As an application, we provide in §6 a new proof of Mather’s theorem that
ν is supported on an n-dimensional Lipschitz continuous graph.
Section 7 extends the techniques from §5 to establish what amounts to an L2 -estimate
for the mixed second partial derivatives,

(1.17) |DxP
2
u|2 dσ ≤ CD2 H̄(P ).
Tn

More precisely, we prove a similar inequality involving difference quotients in the variable
P . An application of this bound appears in §8, where we demonstrate the strict convexity
of H̄ in certain directions.
In §9 we draw some further deductions under the assumptions that H̄ is differentiable
at P and the components of Q := DH̄(P ) are rationally independent.

A forthcoming companion paper [E-G2] addresses problems with time-dependent Hamil-

tonians. The primary new topics developed there include a weak interpretation of the
“adiabatic invariance of the action” and a discussion of the Berry–Hannay geometric phase
correction, computed in terms of effective Hamiltonians.
Our work is strongly related to some extremely interesting papers of Fathi [F1–F3],
which develop his “weak KAM theory”. We hope later to work out more clearly some of
the connections with Fathi’s discoveries.
Some other relevant papers include Mather [Mt1-4], Weinan E [EW1-2], Sobolevskii
[So1-2], Mañé [Mn2-3], Jauslin–Kreiss–Moser [J-K-M], Iturriaga [I], Dias Carneiro [DC],
Arisawa [Ar], etc. A good survey is Mather–Forni [M-F], and we have found Mañé’s book
[Mn1] to be very useful.
5
 See Concordel [C1,C2], Chou–Duffin [C-D], Nussbaum[N], etc. for connections with
nonlinear additive eigenvalue problems. Fathi [F4], Namah–Roquejoffre [N-F], Roquejoffre
[R], Barles–Souganidis [B-S] and Fathi–Mather [F-M] discuss some related questions about
large time asymptotics of solutions to Hamilton–Jacobi equations. Similar problems for
stochastic homogenization have been studied by Rezakhanlou [Rz] and Souganidis [S].
There is also a large literature for time–dependent Hamiltonians with one degree of
freedom. In this setting ordering properties for minimizing trajectories provide powerful
tools unavailable in higher dimensions. See Mather–Forni [M-F], Aubry [A], Bangert [B2],
etc. for more.
We are grateful to G. Barles, L. Barreira, M. Crandall, Weinan E, W. Oliva, D. Serre
and A. Weinstein for interesting suggestions and for references.

2. Effective Hamiltonians and Lagrangians.

2.1. The Hamiltonian and Lagrangian.

As in the introduction, Tn denotes the standard flat torus.

Hypotheses on the Hamiltonian. Assume now that the given, smooth Hamiltonian
H : Rn × Rn → R, H = H(p, x), satisfies these conditions:
(i) periodicity:

For each p ∈ Rn , the mapping
(2.1)
x → H(p, x) is Tn -periodic.

(ii) strict convexity:



 There exist constants Γ, γ > 0 such that

n 2
(2.2) γ|ξ|2 ≤ i,j=1 ∂p∂i ∂pH
ξi ξj ≤ Γ|ξ|2


j

for each p, x, ξ ∈ Rn .

The Lagrangian. We define the associated Lagrangian L : Rn × Rn → R, L = L(q, x),

by duality:

(2.3) L(q, x) := sup(p · q − H(p, x))

p

for q ∈ Rn . In view of (2.1), (2.2) we see that L is smooth,


for each q ∈ Rn , the mapping
(2.4)
x → L(q, x) is Tn -periodic,
6
 

 there exist constants Γ, γ > 0 such that

n 2
(2.5) γ|ξ| ≤ i,j=1 ∂q∂i ∂q
2 L
(q, x)ξi ξj ≤ Γ|ξ|2


j

for all q, x, ξ ∈ Rn .

We physically interpret x as position, p as momentum and q as velocity. The corre-

sponding capital letters X, P, Q will likewise respectively denote position, momentum and
velocity in new coordinates.

2.2. The effective Hamiltonian and Lagrangian.

As explained in the Introduction, we intend next to “average” H, following Lions,

Papanicolaou, Varadhan [L-P-V]:
Theorem 2.1. (i) For each P ∈ Rn there exists a unique real number, denoted H̄(P ),
such that the cell problem

(2.6) H(P + Dx v, x) = H̄(P ) in Rn

has a Tn -periodic, Lipschitz continuous solution v.

(ii) In addition, there exists a constant α such that

(2.7) Dx2 v ≤ αI in Rn

in the distribution sense.

We call the function

(2.8) H̄ : Rn → R

so defined the effective or averaged Hamiltonian.

Remarks. (i) We understand v to solve (2.6) in the sense of viscosity solutions. This
means that if φ = φ(x) is a smooth function and

 v − φ has a maximum (minimum) at

(2.9) a point x0 ∈ Rn , then


H(P + Dφ(x0 ), x0 ) ≤ H̄(P ) (≥ H̄(P )).

We will in fact mostly need only that v is differentiable a. e. with respect to n-dimensional
Lebesgue measure, and that v solves the PDE (2.6) at any point of differentiability.
(ii) The inequality (2.7) means that

the function v̄(x) := v(x) − α2 |x|2
(2.10)
is concave on Rn .
7
 (iii) If v is a solution of (2.6), we will hereafter often write

v = v(P, x) (P, x ∈ Rn )

to emphasize the dependence on P .

Given H̄ as above, we define also the effective Lagrangian

(2.11) L̄(Q) := sup(P · Q − H̄(P ))

P

for Q ∈ Rn .

2.3. Properties of H̄ and L̄.

Proposition 2.2. The mappings

P → H̄(P ), Q → L̄(Q)

are convex and real-valued. Furthermore, H̄ and L̄ are superlinear:

H̄(P ) L̄(Q)
(2.12) lim = lim = +∞.
|P |→∞ |P | |Q|→∞ |Q|

Proof. 1. See Lions, Papanicolaou, Varadhan [L-P-V] (or [E2]) for a proof that H̄ is convex.
The convexity of L̄ is immediate from (2.11).
2. In view of (2.2),

(2.13) H̄(P ) ≥ α|P + Dx v|2 − β ≥ α|P |2 + 2αP · Dx v − β a.e.

for appropriate constants α > 0, β ≥ 0. We integrate this inequality over Tn and recall v
is periodic, to deduce
H̄(P ) ≥ α|P |2 − β.
Thus H̄ is superlinear, and in particular L̄(Q) < ∞ for each Q. On the other hand, by
construction H̄(P ) < ∞ for each P ; whence the duality formula

H̄(P ) = sup(P · Q − L̄(Q))

Q

implies L̄ is superlinear.

In later sections we will relate H̄, L̄ to appropriately rescaled minimizers of the ac-
tion functionals, and for this will several times invoke the following results of Lions–
Papanicolaou–Varadhan [L-P-V, §IV]. (See also E [EW1], Braides–Defranceschi [B-D,
§16.2].)
8
Theorem 2.3. (i) If X : [0, T ] → Rn is a Lipschitz continuous curve and xε (·) → X(·)
uniformly, then
  
T T
xε 
(2.14) L̄(Ẋ) dt ≤ lim inf L ẋε , dt.
0 0 ε

(ii) Define
 
t
x
(2.15) Sε (x, y, t) := inf L ẋ, ds | x(t) = x, x(0) = y ,
0 ε

for x, y ∈ Rn , t > 0. Then

 
x−y
(2.16) Sε (x, y, t) → tL̄ as ε → 0,
t

uniformly on compact subsets of Rn × Rn × (0, ∞).

3. Young measures.

Next we study the asymptotic behavior as t → ∞ of certain curves that minimize the
action.

3.1. Hamilton’s ODE, rescalings.

Definition. A Lipschitz continuous curve x : [0, ∞) → Rn is called a (one-sided) absolute

minimizer if
 T  T
(3.1) L(ẋ, x) dt ≤ L(ẏ, y) dt
0 0

for each time T > 0 and each Lipschitz continuous curve y : [0, ∞) → Rn such that

x(0) = y(0), x(T ) = y(T ).

Given as above an absolutely minimizing curve x(·), define the corresponding momen-
tum

(3.2) p(t) := Dq L(ẋ(t), x(t))

9
for t ≥ 0. Then

ẋ(t) = Dp H(p(t), x(t))
(3.3)
ṗ(t) = −Dx H(p(t), x(t))
for t ≥ 0.
We wish to understand the pair (x(·), p(·)) for large times, and to this end introduce
the rescaled dynamics 
xε (t) := εx(t/ε), pε (t) := p(t/ε)
xε (0) = εx(0), pε (0) = p(0).
It follows from (3.3) that
  
 ẋε (t) = Dp H pε (t), xεε(t)
(3.4)  
 ṗ (t) = − 1 D H p (t), xε (t)
ε ε x ε ε

for t ≥ 0.    
Remark. Since d
dt H pε (t), xεε(t) = 0, we have supt≥0 H pε (t), xεε(t) ≤ C for some
constant C, independent of ε. But H(p, x) ≥ γ2 |p|2 − C, and so
(3.5) sup{|pε (t)|, |ẋε (t)|} < ∞.
t≥0

3.2. Recording oscillations.

We expect the functions pε (·) and xεε(·) (mod Tn ) to oscillate as ε → 0, and so introduce
measures on phase space to record these motions. Invoking for instance the methods from
§1.E of [E1], we have
Proposition 3.1. There exists a sequence εk → 0 and for a.e. t > 0 a Radon probability
measure νt on Rn × Tn such that
   
xεk (t)
(3.6) Φ pεk (t), + Φ̄(t) := Φ(p, x) dνt (p, x)
εk Rn Tn
for each bounded, continuous function
Φ : Rn × Rn → R, Φ = Φ(p, x),
such that x → Φ(p, x) is Tn -periodic.
We call {νt }t≥0 Young measures associated with the dynamics (3.4).
Remark. The limit (3.6) means
 T    T
xεk
(3.7) Φ pεk , ζ dt → Φ̄ζ dt
0 εk 0
for each T > 0 and each smooth function ζ : [0, T ] → R.

10
Lemma 3.2. The support of the measure νt is bounded, uniformly in t.

This is clear from (3.5).

Lemma 3.3. For each C 1 function Φ as above,

 
(3.8) {H, Φ} dνt = 0
Rn Tn

for a.e. t ≥ 0, where

(3.9) {H, Φ} := Dp H · Dx Φ − Dx H · Dp Φ

is the Poisson bracket.

The identity (3.8) means that the measure νt is invariant under the Hamiltonian flow
(3.3).

Proof. We have

d  xε  ẋε
Φ pε , = Dp Φ · ṗε + Dx Φ ·
dt ε ε
1
= {H, Φ}
ε

according to (3.4). Take ζ : [0, T ] → R to be smooth, with compact support. Then

  
T
xε  T  xε 
{H, Φ} pε , ζ dt = − εζ̇Φ pε , dt.
0 ε 0 ε

Sending ε = εk → 0, we deduce (3.8).

3.3. Convergence of trajectories, the action vector.

From (3.5), we conclude that the curves {xε (·)}ε>0 are uniformly Lipschitz continuous.
Hence we may assume (passing if necessary to a further subsequence) that

(3.10) xεk → X

uniformly on compact subsets of [0, ∞), where X : [0, ∞) → Rn is Lipschitz continuous,

X(0) = 0.
11
Lemma 3.4. We have

(3.11) Ẋ(t) = Q(t) for a.e. t ≥ 0,

for
 
(3.12) Q(t) := Dp H(p, x) dνt .
Rn Tn

Proof. The limit (3.10) implies

ẋεk + Ẋ;

whence (3.11), (3.12) follow from (3.4).

Theorem 3.5. (i) For a.e. time t ≥ 0

 
(3.13) L̄(Q(t)) = L(Dp H(p, x), x) dνt .
Rn Tn

(ii) Furthermore, there exists P ∈ Rn such that

(3.14) P ∈ ∂ L̄(Q(t)), Q(t) ∈ ∂ H̄(P )

for a.e. t ≥ 0.

Recall that if Φ : Rn → R is convex, we write y ∈ ∂Φ(x) to mean

Φ(x) + y · (z − x) ≤ Φ(z) for all z ∈ Rn .

Remarks. (i) The point is that P does not depend on t. We call P an action vector
for the rescaled trajectories {xε (·)}ε>0 .
(ii) The second assertion above can be restated

Ẋ ∈ ∂ H̄(P)
for a.e. t ≥ 0,
Ṗ = 0

and this formulation should be compared with (1.5).

(iii) The existence of P is also a consequence of the Pontryagin Maximum Principle; cf.
Clarke [Cl].

Proof. 1. Let yε := xε (0) = εx(0) → 0. According to Theorem 2.3

x
(3.15) Sεk (x, yεk , t) → tL̄ (x ∈ Rn , t > 0),
t
12
uniformly on compact subsets. But
 
t
x
Sε (x, yε , t) = inf L ẋ, ds | x(t) = x, x(0) = yε ,
0 ε

and so
 
t
xε 
(3.16) Sε (xε (t), yε , t) = L ẋε , ds,
0 ε

since the curve xε (·) is an absolute minimizer.

2. From (3.10), (3.15) we see that
 
X(t)
(3.17) Sεk (xεk (t), yεk , t) → tL̄ .
t

But then (3.16) implies that

    
xε d X
(3.18) L ẋεk , k + tL̄ .
εk dt t

Now
       
d X X X X
(3.19) tL̄ ∈ L̄ + ∂ L̄ Ẋ − ≤ L̄(Ẋ),
dt t t t t
 
by convexity. Consequently, since ẋε = Dp H pε , xεε , we deduce from (3.18) that
 
(3.20) L(Dp H(p, x), x) dνt ≤ L̄(Ẋ(t))
Rn Tn

for a.e. t > 0.

Conversely, Theorem 2.3 implies
     
b b
xε  b
L̄(Ẋ(t)) dt ≤ lim L ẋε , dt = L(Dp H, x) dνt dt
a ε→0 a ε a Rn Tn

for all 0 ≤ a < b < ∞ and so

 
L̄(Ẋ(t)) ≤ L(Dp H(p, x), x) dνt
Rn Tn

for a.e. t. This and (3.20) establish (3.13).

3. In particular,
  
d X(t)
tL̄ = L̄(Ẋ(t)) = L̄(Q(t)) a.e.;
dt t
13
and so
 T   
T 
1 1 d X(t)
L̄(Q(t)) dt = tL̄ dt
T 0 T 0 dt t
 
X(T )
(3.21) = L̄
T
  
1 T
= L̄ Q(t) dt .
T 0

This identity, valid for each time T > 0, implies that {Q(t)}t≥0 lies a supporting domain
of L̄. This means that

(3.22) P ∈ ∂ L̄(Q(t)) for a.e. time t ≥ 0

for some vector P ∈ Rn . Equivalently, Q(t) ∈ ∂ H̄(P ).

T
To confirm (3.22), fix a time T > 0, write Q̄ := T1 0 Q(t) dt, and take any P ∈ ∂ L̄(Q̄).
Then owing to (3.21) we have

L̄(Q(t)) = L̄(Q̄) + P · (Q(t) − Q̄)

for a.e. time 0 ≤ t ≤ T . Thus Q(t) is a minimizer of the convex function L̄(Q) − L̄(Q̄) −
P · (Q − Q̄), and so P ∈ ∂ L̄(Q(t)), for a.e. time 0 ≤ t ≤ T . Taking a sequence of times
Tk → ∞ and passing if necessary to a subsequence, we obtain a vector P satisfying (3.22).

4. Structure of minimizing measures.

We next fix one of the Young measures νt and hereafter write ν = νt . Our goal is to
understand the form of this measure, and in particular to describe its support.
Our further deductions will be based entirely upon certain conclusions reached above.
These are firstly that ν is a compactly supported Radon probability measure on Rn × Tn ,
for which we define  
Q := Dp H(p, x) dν,
Rn Tn

as in (3.12) above. In addition, we have

 
(4.1) {H, Φ} dν = 0
Rn Tn

for each C 1 function Φ that is Tn -periodic, and furthermore

 
(4.2) L̄(Q) = L(Dp H(p, x), x) dν.
Rn Tn
14
These are, respectively, assertions (3.8) and (3.13) above.
Remarks. Our ν is therefore a minimal measure in the sense of Mather [Mt1], except
that we work in phase space. The advantage is that the flow invariance condition (4.1) is
fairly simple, and very useful, in the (p, x) variables.

Notation. (i) We write M := spt(ν) and call M the Aubry–Mather set.
(ii) We denote by σ the projection of ν onto the x-variables. That is,

σ(E) := ν(Rn × E)

for each Borel subset E of Tn .

Take now any P ∈ ∂ L̄(Q) and let v = v(P, x) be any viscosity solution of the corre-
sponding cell problem

H(P + Dx v, x) = H̄(P ) in Rn
(4.3)
x → v(P, x) is Tn -periodic,

satisfying the semiconcavity condition (2.7). We hereafter set

u(P, x) := P · x + v(P, x).

4.1. Differentiability on the support of ν.

Theorem 4.1. (i) The function u is differentiable in the variable x σ-a.e., and σ-a.e.
point is a Lebesgue point for Dx u.
(ii) We have
p = Dx u(P, x) ν-a.e.

(iii) Furthermore,
  
(4.4) H(p, x) dν = H(Dx u, x) dσ = H̄(P );
Rn Tn Tn

and if H̄ is differentiable at P ,
  
Dp H(p, x) dν = Dp H(Dx u, x) dσ = DH̄(P ).
Rn Tn Tn

Thus ν is supported on the graph p = Dx u(P, x) = P + Dx v(P, x), which is single-valued

σ-a.e. Also, the PDE (4.3) holds pointwise, σ-a.e.
Remarks. Formula (4.4) explicitly displays H̄ as an average of H; but for this to be
useful, we need to know more about the measure σ. We will later, in §9, discover a bit
more about the structure of σ.
15
 Observe also that from (4.4) we deduce

(4.5) H̄(P ) = H̄(P̃ ) if P, P̃ ∈ ∂ L̄(Q).

Finally, compare assertion (ii) with the canonical change of variables (1.3).

Proof. 1. To ease notation, we do not display the dependence of u on the variable P , and
also write Du for Dx u.
Take ηε to be a smooth, nonnegative, radial convolution kernel, supported in the ball
B(0, ε). Then set
uε := ηε ∗ u.

The strict convexity of H implies for all p, q ∈ Rn that

γ
H(q, x) ≥ H(p, x) + Dp H(p, x) · (q − p) + |q − p|2 .
2

Take q = Du(y), p = Duε (x) = Rn ηε (x − y)Du(y) dy in this expression, multiply by
ηε (x − y), and then integrate with respect to y:

H(Du (x), x) ≤
ε
ηε (x − y)H(Du(y), x) dy
Rn

γ
− ηε (x − y)|Du(y) − Duε (x)|2 dy.
2 Rn

Since the PDE H(Dx u, x) = H̄(P ) holds pointwise a.e., we conclude that

(4.6) βε (x) + H(Duε (x), x) ≤ H̄(P ) + Cε

for each x ∈ Tn , where


γ
(4.7) βε (x) := ηε (x − y)|Du(y) − Duε (x)|2 dy.
2 Rn

2. Recalling again the strict convexity of H with respect to the variable p, we deduce
 
γ
|Duε (x) − p|2 dν
2 Rn Tn
(4.8)  
≤ H(Duε (x), x) − H(p, x) − Dp H(p, x) · (Duε (x) − p) dν.
Rn Tn

Now Duε = P + Dv ε , where v ε = ηε ∗ v is periodic. Consequently

 
Dp H · Dv ε dν = 0,
Rn Tn
16
according to (4.1). This observation and (4.6) imply
  
γ
|Du − p| dν +
ε 2
βε dσ
2 Rn Tn Tn
(4.9)  
≤ H̄(P ) − H + Dp H · (P − p) dν + Cε.
Rn Tn

Next, P ∈ ∂ L̄(Q) implies

L̄(Q) + H̄(P ) = P · Q.
Furthermore
L(Dp H(p, x), x) + H(p, x) = Dp H(p, x) · p.

Recalling that Q = Rn Tn Dp H dν and substituting into (4.9), we find
  
γ
|Du − p| dν +
ε 2
βε dσ
2 Rn Tn Tn
(4.10)  
≤ −L̄(Q) + L(Dp H, x) dν + Cε = Cε,
Rn Tn

according to (4.2).
3. Now send ε → 0. Passing as necessary to a subsequence we deduce first from (4.10)
that
βε → 0 σ-a.e.
Thus σ-a.e. point x is a point of approximate continuity of Du, and Du is σ-measurable.
Since u = x · P + v and v is semiconcave as a function of x (Theorem 2.1,(ii)), it follows
that u is differentiable in x, σ-a.e. Thus

Duε → Du

pointwise, σ-a.e., and so (4.10) in turn forces

p = Du(x) = P + Dv(x) ν-a.e.

This proves assertion (ii), and (iii) follows then from the cell PDE.

Remark. As a consequence of the foregoing proof, we have the identity
  
(4.11) Dp H(p, x) · Dx v dν = Dp H(Dx u, x) · Dx v dσ = 0,
Rn Tn Tn

which we will need later. To confirm this, recall from above that
 
Dp H · Dx v ε dν = 0.
Rn Tn

Since Dx v ε → Dx v boundedly, ν-a.e., we can apply the Dominated Convergence Theorem.

17
5. Derivative estimates in the variable x.

We devote this section to showing that our solution u of the cell problem is “smoother”
on the support of σ than it may be at other points of Tn . This is a sort of “partial
regularity” assertion.

5.1. Formal L2 - and L∞ -estimates.

First of all, we provide for the reader some purely formal L2 and L∞ estimates for
Dx2 u on the support of σ, calculations which provide motivation for the rigorous bounds
obtained afterwards.
L2 -inequalities. We assume for this that the generating function u is smooth, then
differentiate the cell PDE twice with respect to xi , and finally sum for i = 1, . . . , n:

Hpk pl (Dx u, x)uxk xi uxl xi + Hpk (Dx u, x)uxk xi xi

+ 2Hpk xi (Dx u, x)uxk xi + Hxi xi (Dx u, x) = 0.

The first term on the left is greater than or equal to γ|Dx2 u|2 . Thus
  
γ |Dx2 u|2 dσ + Dp H · Dx (∆x u) dσ ≤ C + C |Dx2 u| dσ.
Tn Tn Tn

Since ∆x u = ∆x v is periodic, the second term on the left equals zero, according to (4.1).
We consequently conclude

(5.1) |Dx2 u|2 dσ ≤ C,
Tn

for some constant C depending only on H and P .

∞
L -inequalities. We can similarly differentiate the cell PDE twice in any unit direc-
tion ξ, to find

Hpk pl (Dx u, x)uxk ξ uxl ξ + Hpk (Dx u, x)uxk ξξ

+ 2Hpk ξ (Dx u, x)uxk ξ + Hξξ (Dx u, x) = 0,

n
for uξξ := i,j=1 uxi xj ξi ξj . Take a nondecreasing, function Φ : R → R, and write φ :=
Φ ≥ 0. Multiply the above identity by φ(uξξ ), and integrate with respect to σ. After some
simplifications, we find
  
γ
|Dx uξ | φ(uξξ ) dσ +
2
Dp H · Dx (Φ(uξξ )) dσ ≤ C φ(uξξ ) dσ.
2 Tn Tn Tn
18
Since uξξ = vξξ is periodic, the second term on the left is zero. We select

1 if z ≤ −µ
φ(z) =
0 if z > −µ,

for a constant µ > 0. Since |Dx uξ |2 ≥ u2ξξ , we conclude that σ({uξξ ≤ −µ}) = 0 if µ is
large enough. Because (2.10) provides the opposite estimate uξξ ≤ α, we thereby derive
the formal bound

(5.2) |uξξ | ≤ C σ-a.e.,

the constant C depending only upon known quantities.

Remark. As the interested reader may wish to confirm, the foregoing derivations are
especially transparent for the classical Hamiltonian
1 2
H(p, x) = |p| + V (x),
2
in which case the cell PDE (4.3) is the eikonal equation

1
|Dx u|2 + V (x) = H̄(P )
2
and (4.1) corresponds to the transport equation

div(σDx u) = 0.

A clear message is that these two PDE should be considered together as a pair, in accor-
dance with formal semiclassical limits. (See the Remark in §1.3.)

5.2. An L2 -estimate of difference quotients in x.

We now establish an analogue of estimate (5.1), with difference quotients replacing some
of the derivatives.
Theorem 5.1. There exists a constant C, depending only on H and P , such that

(5.3) |Dx u(P, x + h) − Dx u(P, x)|2 dσ ≤ C|h|2
Tn

for h ∈ Rn .
Remark. If Dx u(P, x + h) is multivalued, we interpret (5.3) to mean

(5.4) |ξ − Dx u|2 dσ ≤ C|h|2
Tn
19
for some σ-measurable selection ξ ∈ Dx u(P, · + h).

Proof. 1. To simplify notation we do not display the dependence of u on P , and just write
Du for Dx u.
Fix h ∈ Rn and define the shifted function

ũ(·) := u(· + h).

Then
H(Dũ, x + h) = H̄(P ) a.e. in Rn .

Mollifying as in the proof of Theorem 4.1, we have

H(Dũε , x + h) ≤ H̄(P ) + Cε in Rn .

Therefore
H(Dũε , x) − H(Du, x) ≤ Cε + H(Dũε , x) − H(Dũε , x + h)

σ-a.e., and consequently

 
γ
|Dũ − Du| dσ +
ε 2
Dp H(Du, x) · (Dũε − Du) dσ
2 Tn T n

(5.5) ≤ Cε + H(Dũε , x) − H(Dũε , x + h) dσ
Tn

≤ C(ε + |h| ) −
2
Dx H(Dũε , x) · h dσ.
Tn

2. Since Dũε − Du = Dṽ ε − Dv, the second term on the left hand side of (5.5) vanishes,
in view of (4.1), (4.11). Therefore
 
γ
|Dũ − Du| dσ ≤ C(ε + |h| ) −
ε 2 2
Dx H(Du, x) · h dσ
2 Tn Tn

+C |Dũε − Du||h| dσ,
Tn

and thus   
γ
|Dũ − Du| dσ ≤ C(ε + |h| ) −
ε 2 2
Dx H · h dν.
4 Tn Rn Tn

However (4.1) implies the last term here is zero; whence


(5.6) |Dũε − Du|2 dσ ≤ C(ε + |h|2 ).
Tn

3. We send ε → 0. Passing as necessary to a subsequence we have

Dũε + ξ weakly in L2σ

20
and

(5.7) |ξ − Du|2 dσ ≤ C|h|2 .
Tn

4. To conclude, we must show

(5.8) ξ ∈ Dũ = Du(· + h) σ-a.e.,

which means that for σ-a.e. point x there exists a constant C such that

(5.9) ũ(y) ≤ ũ(x) + ξ · (y − x) + C|y − x|2

for all y. To confirm this, recall that ũ, and so also ũε , are semiconcave:

ũε (y) ≤ ũε (x) + Dũε (x) · (y − x) + C|y − x|2

for all x, y. Take g ∈ L2σ , g ≥ 0. Then fixing y and integrating the variable x with respect
to σ, we find

0≤ (−ũε (y) + ũε (x) + Dũε (x) · (y − x) + C|y − x|2 )g(x) dσ(x).
Tn

Let ε → 0 and note ũε → ũ uniformly. We conclude


0≤ (−ũ(y) + ũ(y) + ξ · (y − x) + C|y − x|2 )g(x) dσ(x).
Tn

This inequality is true for all g as above; whence (5.7) holds for σ-a.e. point x and all
y.

5.3. L∞ -estimates of difference quotients in x.

We next refine the integration arguments above, to derive an L∞ bound on second-order

difference quotients. This will be a variant of the formal estimate (5.2) above.
Theorem 5.2. There exists a constant C, depending only on H and P , such that

(5.10) |u(P, x + h) − 2u(P, x) + u(P, x − h)| ≤ C|h|2

for all h ∈ Rn and each point x ∈ spt(σ).

Proof. 1. Take h = 0, and write

ũ = u(· + h), û = u(· − h).

21
We as before consider the mollified functions ũε , ûε , where we take

(5.11) 0 < ε ≤ η|h|2 ,

for small η > 0. As in the earlier proofs, we have


H(Dũε , x + h) ≤ H̄(P ) + Cε,
H(Dûε , x − h) ≤ H̄(P ) + Cε.

Therefore for σ-a.e. point x,

H(Dũε , x) − 2H(Du, x) + H(Dûε , x)

≤ Cε + H(Dũε , x) − H(Dũε , x + h) + H(Dûε , x) − H(Dûε , x − h).

Hence
γ
(|Dũε − Du|2 + |Dûε − Du|2 ) + Dp H(Du, x) · (Dũε − 2Du + Dûε )
2
≤ C(ε + |h|2 ) + (Dx H(Dûε , x) − Dx H(Dũε , x)) · h,

and consequently
γ
(|Dũε − Du|2 + |Dûε − Du|2 )
4
+ Dp H(Du, x) · (Dũε − 2Du + Dûε ) ≤ C(ε + |h|2 ).


2. Fix now a smooth, nondecreasing,  function Φ : R → R, and write φ := Φ ≥ 0.
Multiply the last inequality above by φ ũ −2u+û
ε ε

|h|2 , and integrate with respect to σ:

  
γ ũε − 2u + ûε
(|Dũ − Du| + |Dû − Du| )φ
ε 2 ε 2
dσ
4 Tn |h|2

(5.12) + Dp H(Du, x) · (Dũε − 2Du + Dûε )φ(· · · ) dσ
Tn

≤ C(ε + |h| )
2
φ(· · · ) dσ.
Tn

Now the second term on the left hand side of (5.12) equals
   
ũε − 2u + ûε
(5.13) |h| 2
Dp H(p, x) · Dx Φ dν
Rn Tn |h|2

and thus is zero. (To see this, note from (4.1) that the expression (5.13) vanishes if we
replace u by a mollified function uδ . Let δ → 0, recalling the estimates in the proof of
Theorem 4.1.)
22
 So now dropping the above term from (5.12) and rewriting, we deduce
  
uε (x + h) − 2u(x) + uε (x − h)
|Du (x + h) − Du (x − h)| φ
ε ε 2
dσ
Tn |h|2
(5.14)   ε 
u (x + h) − 2u(x) + uε (x − h)
≤ C(ε + |h| )
2
φ dσ.
Tn |h|2

3. We confront now a technical problem, as (5.14) entails a mixture of first-order dif-

ference quotients for Duε and second-order difference quotients for u, uε . We can however
relate these expressions, since u is semiconcave.
To see this, first of all define

(5.15) Eε := {x ∈ spt(σ) | uε (x + h) − 2u(x) + uε (x − h) ≤ −µ|h|2 },

the large constant µ > 0 to be fixed below. Now according to (2.10), the functions

α 2 α
(5.16) ū(x) := u(x) − |x| , ūε (x) := uε (x) − |x|2
2 2

are concave. Also a point x ∈ spt(σ) belongs to Eε if and only if

(5.17) ūε (x + h) − 2ū(x) + ūε (x − h) ≤ −(µ + α)|h|2 .

Set
 
h
(5.18) ε
f (s) := ū ε
x+s (−|h| ≤ s ≤ |h|).
|h|

Then f is concave, and

ūε (x + h) − 2ūε (x) + ūε (x − h) = f ε (|h|) − 2f ε (0) + f ε (−|h|)

 |h|

= f ε (x)(|h| − |s|) ds
−|h|
 |h|
 
≥ |h| f ε (s) ds (since f ε ≤ 0)
−|h|
 
= |h|(f ε (|h|) − f ε (−|h|))
= (Dūε (x + h) − Dūε (x − h)) · h.

Consequently if x ∈ Eε , this inequality and (5.17) together imply

2|ūε (x) − ū(x)| + |Dūε (x + h) − Dūε (x − h)||h| ≥ (µ + α)|h|2 .

23
Now |ūε (x) − ū(x)| ≤ Cε on Tn , since u is Lipschitz continuous. We may therefore take η
in (5.11) small enough to deduce from the foregoing that

µ
(5.19) |Dūε (x + h) − Dūε (x − h)| ≥ ( + α)|h|
2

But then

µ
(5.20) |Duε (x + h) − Duε (x − h)| ≥ ( − α)|h|
2

4. Return now to (5.14). Taking µ > 2α and


1 if z ≤ −µ
φ(z) =
0 if z > −µ,

we discover from (5.14) that

µ
( − α)2 |h|2 σ(Eε ) ≤ C(ε + |h|2 )σ(Eε ).
2

We fix µ so large that
µ
( − α)2 ≥ C + 1,
2
to deduce
(|h|2 − Cε)σ(Eε ) ≤ 0.

Thus σ(Eε ) = 0 if η in (5.11) is small enough, and this means

uε (x + h) − 2u(x) + uε (x − h) ≥ −µ|h|2

for σ-a.e. point x. Now let ε → 0:

u(x + h) − 2u(x) + u(x − h) ≥ −µ|h|2

σ-a.e. Since
u(x + h) − 2u(x) + u(x − h) ≤ α|h|2

owing to the semiconcavity, we have

|u(x + h) − 2u(x) + u(x − h)| ≤ C|h|2

for σ-a.e. point x. As u is continuous, the same inequality obtains for all x ∈ spt(σ).

24
6. Application: Lipschitz estimates for the support of ν.

We next improve the second derivative bounds from the previous section, and then show
as a simple consequence that spt(ν) lies on a Lipschitz continuous graph.
Theorem 6.1. (i) There exists a constant C, depending only on H and P , such that

(6.1) |u(P, y) − u(P, x) − Dx u(P, x) · (y − x)| ≤ C|x − y|2

for all y ∈ Tn and σ-a.e. point x ∈ Tn .

(ii) Furthermore,

(6.2) |Dx u(P, y) − Dx u(P, x)| ≤ C|x − y|

for all y ∈ Tn and for σ-a.e. point x ∈ Tn .

(iii) In fact, u is differentiable at each point x ∈ spt(σ), and estimates (6.1), (6.2) hold
for all y ∈ Tn , x ∈ spt(σ).
Remark. When Dx u(P, y) is multivalued, (6.2) asserts

|ξ − Dx u(P, x)| ≤ C|x − y|

for all ξ ∈ Dx u(P, y). In particular, for multivalued Dx u(P, y) we have the estimate

diam(Dx u(P, y)) ≤ C dist(y, spt(σ)),

providing a quantitative justification to the informal assertion that “spt(σ) misses the
shocks in Du”.

Proof. 1. Fix y ∈ Rn and take any point x ∈ spt(σ) at which u is differentiable.
According to Theorem 5.2 with h := y − x, we have

(6.3) |u(y) − 2u(x) + u(2x − y)| ≤ C|x − y|2 .

By semiconcavity, we have

(6.4) u(y) − u(x) − Du(x) · (y − x) ≤ C|x − y|2 ,

and also

(6.5) u(2x − y) − u(x) − Du(x) · (2x − y − x) ≤ C|x − y|2 .

Use (6.5) in (6.3):

u(y) − u(x) − Du(x) · (y − x) ≥ −C|x − y|2 .
25
This and (6.4) establish (6.1).
2. Estimate (6.2) follows from (6.1), as follows. Take x, y as above. Let z be a point to
be selected later, with |x − z| ≤ 2|x − y|.
The semiconcavity of u implies that

(6.6) u(z) ≤ u(y) + Du(y) · (z − y) + C|z − y|2 .

Also,

u(z) = u(x) + Du(x) · (z − x) + O(|x − z|2 ), u(y) = u(x) + Du(x) · (y − x) + O(|x − y|2 ),

according to (6.1). Insert these indentities into (6.6) and simplify:

(Du(x) − Du(y)) · (z − y) ≤ C|x − y|2 .

Now take
Du(x) − Du(y)
z := y + |x − y|
|Du(x) − Du(y)|
to deduce (6.2).
3. Now take any point x ∈ spt(σ), and fix y. There exist points xk ∈ spt(σ) (k = 1, . . . )
such that xk → x and u is differentiable at xk . According to estimate (6.1)

|u(y) − u(xk ) − Du(xk ) · (y − xk )| ≤ C|xk − y|2 (k = 1, . . . ).

The constant C does not depend on k or y. Now let k → ∞. Owing to (6.2) we see that
{Du(xk )} converges to some vector η, for which

|u(y) − u(x) − η · (y − x)| ≤ C|x − y|2 .

Consequently u is differentiable at x and Du(x) = η.

As an application of these bounds, we show next that the set M = spt(ν) lies on an
n-dimensional Lipschitz continuous graph. This theorem (in position-velocity variables) is
due originally to Mather [Mt2].
Theorem 6.2. There exists a constant C, depending only on P and H, such that

(6.7) |Dx u(P, x1 ) − Dx u(P, x2 )| ≤ C|x1 − x2 |

for σ-a.e. pair of points x1 , x2 .

Proof. In view of (6.2) we can extend the mapping x → Du(x) to a uniformly Lipschitz
function defined on all of Tn . The support of ν lies on the graph of this mapping.

26
7. Derivative estimates in the variable P .

We turn next to some bounds involving variations in P . These are rather subtle and
involve the smoothness properties of H̄. (See Pöschel [P, p. 656–657] for an explicit linear
example, showing that u can be less well behaved in P than in x.)

7.1. A formal L2 -estimate.

As in §5.1, we begin with a simple, but unjustified, calculation that suggests the later
proof. So for the moment suppose u and H̄ are smooth, differentiate the cell PDE twice
with respect to Pi , and sum on i:

(7.1) Hpk pl (Dx u, x)uxk Pi uxl Pi + Hpk (Dx u, x)uxk Pi Pi = H̄Pi Pi (P ).

2
The first term on the left is greater than or equal to γ|DxP u|2 . Consequently
 
γ |DxP u| dσ +
2 2
Dp H · Dx (∆P u) dσ ≤ ∆H̄(P ),
Tn Tn

where ∆H̄ = ∆P H̄ is the Laplacian of H̄ in P . Since ∆P u = ∆P v is periodic, the second

term on the left equals zero. Therefore

(7.2) |DxP
2
u|2 dσ ≤ C∆H̄(P ).
Tn

7.2. An L2 -estimate of difference quotients in P.

We next provide a rigorous version of the foregoing calculation, replacing derivatives by

difference quotients.
Theorem 7.1. There exists a positive constant C, depending only on H, such that

(7.3) |Dx u(P̃ , x) − Dx u(P, x)|2 dσ ≤ C(H̄(P̃ ) − H̄(P ) − Q · (P̃ − P ))
Tn

for all P̃ ∈ Rn .
  
Remark. Recall that Q = Rn Tn Dp H(p, x) dν = Tn Dp H(Dx u, x) dσ and that Q ∈
∂ H̄(P ). In (7.3), u(P̃ , x) = P̃ · x + v(P̃ , x) and v = v(P̃ , x) is any viscosity solution of the
cell problem

H(P̃ + Dx v, x) = H̄(P̃ ) in Rn
(7.4)
x → v(P̃ , x) is Tn -periodic.
27
If Dx u(P̃ , x) is multivalued, we interpret (7.3) to mean

|ξ̃ − Dx u(P, x)|2 dσ ≤ C(H̄(P̃ ) − H̄(P ) − Q · (P̃ − P ))
Tn
for some σ-measurable selection ξ̃ ∈ Dx u(P̃ , ·).

Proof. 1. Write ṽ(·) = v(P̃ , ·), ũ = x · P̃ + ṽ. Mollifying, we have
(7.5) H(Dũε , x) ≤ H̄(P̃ ) + Cε.
Therefore for σ almost every point
γ
|Dũε − Du|2 + Dp H(Du, x) · (Dũε − Du) ≤ H(Dũε , x) − H(Du, x)
(7.6) 2
≤ H̄(P̃ ) − H̄(P ) + Cε.
Observe that Dũε − Du = P̃ − P + (Dṽ ε − Dv) and
 
Dp H · (Dṽ ε − Dv) dν = 0.
Rn Tn
Consequently (7.6) yields
 
γ
|Dũ − Du| dσ ≤ H̄(P̃ ) − H̄(P ) −
ε 2
Dp H(Du, x) · (P̃ − P ) dσ + Cε
(7.7) 2 Tn Tn
= H̄(P̃ ) − H̄(P ) − Q · (P̃ − P ) + Cε.
Let ε → 0.

Remark. For use later, we record the estimate

(7.8) lim sup βε dσ ≤ H̄(P̃ ) − H̄(P ) − Q · (P̃ − P ),
ε→0 Tn
for 
γ
(7.9) βε (x) := ηε (x − y)|Dx u(P̃ , y) − Dx uε (P̃ , x)|2 dy.
2 Tn
To see this, note that as in the proof of Theorem 4.1 we can replace (7.5) by the stronger
inequality
βε (x) + H(Dũε , x) ≤ H̄(P̃ ) + Cε.


Corollary 7.2. (i) For each P ∈ Rn , we have

|Dx u(P̃ , x) − Dx u(P, x)|2 dσ ≤ O(|P̃ − P |) as P̃ → P.
Tn
(ii) If H̄ is differentiable at P ,

|Dx u(P̃ , x) − Dx u(P, x)|2 dσ ≤ o(|P̃ − P |) as P̃ → P.
Tn
(iii) If H̄ is twice-differentiable at P ,

|Dx u(P̃ , x) − Dx u(P, x)|2 dσ ≤ O(|P̃ − P |2 ) as P̃ → P.
Tn

28
8. Application: strict convexity of H̄ in certain directions.

The next estimate allows us to deduce certain strict convexity properties of H̄.
Theorem 8.1. (i)There exists a positive constant C such that for each R ∈ Rn , we have
 1/2
H̄(P + tR) − 2H̄(P ) + H̄(P − tR)
(8.1) −R · Q̃, R · Q̂ ≤ C lim inf ,
t→0 + t2

where Q̃, Q̂ ∈ ∂ H̄(P ).

(ii) In particular, if H̄ is twice differentiable at P , then

(8.2) |DH̄(P ) · R| ≤ C(R · D2 H̄(P )R)1/2

for each R ∈ Rn .
Proof. 1. Fix R ∈ Rn , t > 0, and take

ũ = u(P + tR, ·), û = u(P − tR, ·).

Then for σ-a.e. point x:

H(Dũε , x) − 2H(Du, x) + H(Dûε , x) ≤ H̄(P + tR) − 2H̄(P ) + H̄(P − tR) + Cε.

Similarly to the proof in §5.2, we deduce


(8.3) |Dũε − Du|2 + |Dûε − Du|2 dσ ≤ C(H̄(P + tR) − 2H̄(P ) + H̄(P − tR)) + Cε.
Tn

2. Since H(Dũε , x) ≤ H̄(P + tR) + Cε, we have


H̄(P ) − H̄(P + tR) ≤ H(Du, x) − H(Dũε , x) dσ + Cε
Tn
(8.4)  1/2
≤C |Du − Dũ | dσ
ε 2
+ Cε.
Tn

Likewise,
 1/2
(8.5) H̄(P ) − H̄(P − tR) ≤ C |Du − Dû | dσ
ε 2
+ Cε.
Tn

Combining (8.3)–(8.5), sending ε → 0, and recalling the convexity of H̄, we find

−tQ̃(t) · R, tQ̂(t) · R ≤ C(H̄(P + tR) − 2H̄(P ) + H̄(P − tR))1/2

29
for any
Q̃(t) ∈ ∂ H̄(P + tR), Q̂(t) ∈ ∂ H̄(P − tR).

Taking any tk → 0, we may assume Q̃(tk ) → Q̃, Q̂(tk ) → Q̂ with Q̃, Q̂ ∈ ∂ H̄(P ). Estimate
(8.1) follows.

Remarks. (i) From (8.1) we deduce that H̄ is strictly convex in any direction R which
is not tangent to the level set {H̄ = H̄(P )}, provided H̄ is differentiable at P . (Compare
this assertion with Iturriaga [I].)
(ii) More generally, if H̄(P ) > minRn H̄, and so 0 ∈
/ ∂ H̄(P ), there exists an open convex
cone of directions R in which H̄ is strictly convex at P .
Therefore the graph of H̄ can contain an n-dimensional flat region only possibly at
its minimum value. This can in fact happen, even though H is uniformly convex in the
variable p: see Lions–Papanicolaou–Varadhan [L-P-V] or Braides–Defranceschi [B-D, p.
149]. Consult Concordel [C1,C2] for more. Physically, a flat region at the minimum of H̄
corresponds to “nonballistic” trajectories for the dynamics.
(iii) See also Bangert [B] and Weinan E [EW2] for an example showing that the level
sets of H̄ can have corners and/or flat parts.

9. Application: averaging in the variable X.

Assume for this section that H̄ is differentiable at P and furthermore that Q = DH̄(P )
satisfies the nonresonance condition:

(9.1) Q · k = 0 for each vector k ∈ Zn , k = 0.

Notation. For h > 0, we write the vector of difference quotients

 
h u(P + hel , x) − u(P, x)
(9.2) DP u(P, x) := . . . , ,... ,
h

for el := (0, . . . , 1, . . . , 0), the 1 in the lth -position.

Theorem 9.1. Suppose Q = DH̄(P ) satisfies (9.1). Then
 
h
(9.3) lim Φ(DP u(P, x)) dσ = Φ(X) dX
h→0 Tn Tn

for each continuous, Tn -periodic function Φ.

Proof. 1. Let ul (·) := u(P + hel , ·), and uεl := ηε ∗ ul , for l = 1, . . . , n.
Since H is smooth, we have for all p, q lying in a compact subset of Rn that

H(q, x) = H(p, x) + Dp H(p, x) · (q − p) + R, with |R| ≤ C|q − p|2 .

30
 
Take q = Dul (y), p = Duεl (x) = Rn ηε (x − y)Dul (y) dy, multiply by ηε (x − y), and
integrate with respect to y:
 
(9.4) ε
H(Dul (x), x) = ηε (x − y)H(Dul (y), x) dy − ηε (x − y)R dy.
Rn Rn

Furthermore the PDE H(Dul , x) = H̄(P +hel ) holds pointwise a.e., and so we can conclude
that

(9.5) H(Duεl , x) = H̄(P + hel ) + γεl ,

where the error term is estimated by

|γεl | ≤ C(ε + βεl )

for 
γ
βεl (x) := ηε (x − y)|Dul (y) − Duεl (x)|2 dy.
2 Rn

2. We introduce the partially smoothed vector of difference quotients

 
h ε uεl − u
(9.6) DP u (P, x) := . . . , ,... ,
h

and take then a vector of integers k = (k1 , . . . , kn ), k = 0.

Next, observe that the function
h ε h ε
e2πik·DP u = e2πik·x e2πik·DP v

is Tn -periodic, even though DPh uε is not periodic. Hence

  
h ε
0= Dp H(Du, x) · Dx e2πik·DP u dσ
Tn
(9.7)  
n
h ε uεl − u
= 2πi e2πik·DP u
kl Dp H(Du, x) · Dx ( ) dσ.
Tn h
l=1

3. Now (9.5) implies

H(Duεl , x) − H(Du, x) = H̄(P + hel ) − H̄(P ) + γεl .

Consequently
Dp H(Du, x) · D(uεl − u) = H̄(P + hel ) − H̄(P ) + Γlε ,

where

(9.8) |Γlε | ≤ C(ε + βεl + |Duεl − Du|2 )

31
for l = 1, . . . , n.
Therefore
 
uε − u H̄(P + hel ) − H̄(P ) 1 l
(9.9) Dp H(Du, x) · Dx ( l ) = Ql + − Ql + Γε .
h h h

4. Insert (9.9) into (9.7), and then estimate

   n  
  Cε  H̄(P + hel ) − H̄(P )
(Q · k) e2πik·DPh ε
u 
dσ  ≤ +C − Ql
 h h
Tn l=1
n 
C
+ βεl + |Duεl − Du|2 dσ by (9.8)
h
l=1 T
n

n  
(9.10)
Cε  H̄(P + hel ) − H̄(P )
≤ +C − Ql
h h
l=1
n 
C
+ βεl dσ,
h Tn
l=1

the last inequality following from (7.7) in the proof of Theorem 7.1.
Next, send ε → 0, and remember (7.8):
   n  
   H̄(P + he ) − H̄(P )
(Q · k) e2πik·D h
u
dσ  ≤ C
l
− Ql .

P

Tn h
l=1

Since Ql = H̄Pl (P ) and Q · k = 0, we conclude that


h
lim e2πik·DP u dσ = 0
h→0 Tn

for all k ∈ Zn , k = 0. Because any continuous, Tn -periodic funtion Φ can be uniformly

approximated by trigonometric polynomials, this implies assertion (9.3).

Remarks. (i) Recalling the formal change of variables (1.3), we interpret (9.3) to assert

(9.11) “ dσ = | det DxP

2
u| dx ”

in some weak sense, provided (9.1) holds. See [E-G2, §5.1] for related formal computations.
(ii) Theorem 9.1 provides a partial, but rigorous, interpretation of the following heuris-
tics.
Suppose that our generating function u is smooth, and induces the global change of
variables (p, x) → (P, X) by (1.3). The the dynamics (1.1) become (1.5); that is,

Ẋ = DH̄(P)
Ṗ = 0.
32
Consequently X(t) = Qt + X0 , P(t) ≡ P . In view therefore of the nonresonance condition
(9.1), we have
 λT 
1
lim Φ(X(t)) dt = Φ(X) dX.
T →∞ λT 0 Tn

for each λ > 0. However

 λT  λT
1 1
Φ(X(t)) dt = Φ(DP u(P, x(t))) dt
λT 0 λT 0

1 λ xε (t) 1
= Φ(DP u(P, )) dt for ε =
λ 0 ε T
 λ
1
→ Φ(DP u(P, x)) dσt dt.
λ 0 Tn

Consequently  
Φ(DP u(P, x)) dσt = Φ(X) dX
Tn Tn

for all t ≥ 0.

33
 References
[Ar] M. Arisawa, Multiscale homogenization for first-order Hamilton–Jacobi–Bellman equations, Ad-
vances in Differential Equations (to appear).
[A] S. Aubry, The twist map, the extended Frenkel–Kantorova model and the devil’s straircase, Phys-
ica D 7 (1983), 240–258.
[B1] V. Bangert, Minimal geodesics, Ergodic Theory and Dyn. Systems 10 (1989), 263–286.
[B2] V. Bangert, Geodesic rays, Busemann functions and monotone twist maps, Calculus of Variations
2 (1994), 49–63.
[B-S] G. Barles and P. E. Souganidis, On the long time behavior of solutions of Hamilton–Jacobi
equations, J. Math Analysis (to appear).
[B-D] A. Braides and A. Defranceschi, Homogenization of Multiple Integrals, Oxford Univ. Press, 1998.
[C-D] W. Chou and R. J. Duffin, An additive eigenvalue problem of physics related to linear program-
ming, Advances in Appl. Math 8 (1987), 486–498.
[Cl] F. Clarke, Optimization and Nonsmooth Analysis, Wiley-Interscience, 1983.
[C1] M. Concordel, Periodic homogenization of Hamilton-Jacobi equations I: additive eigenvalues and
variational formula, Indiana Univ. Math. J. 45 (1996), 1095–1117.
[C2] M. Concordel, Periodic homogenisation of Hamilton-Jacobi equations II: eikonal equations, Proc.
Roy. Soc. Edinburgh 127 (1997), 665–689.
[DC] M. J. Dias Carneiro, On minimizing measure of the action of autonomous Lagrangians, Nonlin-
earity 8 (1995), 1077–1085.
[EW1] Weinan E, A class of homogenization problems in the calculus of variations, Comm Pure and
Appl Math 44 (1991), 733–754.
[EW2] Weinan E, Aubry–Mather theory and periodic solutions of the forced Burgers equation, Comm
Pure and Appl Math 52 (1999), 811–828.
[E1] L. C. Evans, Weak Convergence Methods for Nonlinear Partial Differential Equations, American
Math Soc, 1990.
[E2] L. C. Evans, Periodic homogenization of certain fully nonlinear PDE, Proc Royal Society Edin-
burgh 120 (1992), 245–265.
[E-G2] L. C. Evans and D. Gomes, Effective Hamiltonians and Averaging for Hamiltonian Dynamics II,
to appear.
[F1] A. Fathi, Théorème KAM faible et théorie de Mather sur les systèmes lagrangiens, C. R. Acad.
Sci. Paris Sr. I Math. 324 (1997), 1043–1046.
[F2] A. Fathi, Solutions KAM faibles conjuguées et barrières de Peierls, C. R. Acad. Sci. Paris Sr. I
Math. 325 (1997), 649–652.
[F3] A. Fathi, Orbites hétéroclines et ensemble de Peierls, C. R. Acad. Sci. Paris Sr. I Math. 326
(1998), 1213–1216.
[F4] A. Fathi, Sur la convergence du semi-groupe de Lax-Oleinik, C. R. Acad. Sci. Paris Sr. I Math.
327 (1998), 267–270.
[F-M] A. Fathi and J. Mather, Failure of convergence of the Lax–Oleinik semigroup in the time–periodic
case, preprint (2000).
[Gd] H. Goldstein, Classical mechanics (2nd ed.), Addison-Wesley, 1980.
[G] D. Gomes, Hamilton–Jacobi Equations, Viscosity Solutions and Asymptotics of Hamiltonian
Systems, Ph.D. Thesis, University of California, Berkeley (2000).
[J-K-M] H. R. Jauslin, H. O. Kreiss and J. Moser, On the forced Burgers equation with periodic boundary
conditions, preprint (1998).
[L-P-V] P.-L. Lions, G. Papanicolaou, and S. R. S. Varadhan, Homogenization of Hamilton–Jacobi equa-
tions, unpublished, circa 1988.
[I] R. Iturriaga, Minimizing measures for time-dependent Lagrangians, Proc. London Math Society
73 (1996), 216–240.
[Mn1] R. Mañé, Global Variational Methods in Conservative Dynamics, Instituto de Matemática Pura
e Aplicada, Rio de Janeiro.
[Mn2] R. Mañé, On the minimizing measures of Lagrangian dynamical systems, Nonlinearity 5 (1992),
623–638.
34
[Mn3] R. Mañé, Generic properties and problems of minimizing measures of Lagrangian systems, Non-
linearity 9 (1996), 273–310.
[Mt1] J. Mather, Minimal measures, Comment. Math Helvetici 64 (1989), 375–394.
[Mt2] J. Mather, Action minimizing invariant measures for positive definite Lagrangian systems, Math.
Zeitschrift 207 (1991), 169–207.
[Mt3] J. Mather, Differentiability of the minimal average action as a function of the rotation number,
Bol. Soc. Bras. Math 21 (1990), 59–70.
[Mt4] J. Mather, Variational construction of connecting orbits, Ann. Inst. Fourier, Grenoble 43 (1993),
1347–1386.
[M-F] J. Mather and G. Forni, Action minimizing orbits in Hamiltonian systems, Transition to Chaos
in Classical and Quantum Mechanics, Lecture Notes in Math 1589 (S. Graffi, ed.), Sringer, 1994.
[Ms] J. Moser, Recent developments in the theory of Hamiltonian systems, SIAM Review 28 (1986),
459–486.
[N-F] G. Namah and J.-M. Roquejoffre, Comportement asymptotique des solutions d’une classe d’equations
paraboliques and de Hamilton–Jacobi, C. R. Acad. Sci. Paris Sr. I Math. 324 (1997), 1367–1370.
[N] R. Nussbaum, Convergence of iterates of a nonlinear operator arising in statistical mechanics,
Nonlinearity 4 (1991), 1223–1239.
[P] J. Pöschel, Integrability of Hamiltonian systems on Cantor sets, Comm. Pure Appl. Math 35
(1982), 653–696.
[Rz] F. Rezakhanlou, Central limit theorem for stochastic Hamilton–Jacobi equations, preprint (1998).
[R] J.-M. Roquejoffre, Comportement asymptotique des solutions d’equations de de Hamilton–Jacobi
monodimensionnelles, C. R. Acad. Sci. Paris Sr. I Math. 326 (1998), 185–189.
[So1] A. N. Sobolevskii, Periodic solutions of the Hamilton–Jacobi equation with periodic forcing term,
Russian Math Surveys 53 (1998), 1375–1376.
[So2] A. N. Sobolevskii, Periodic solutions of the Hamilton–Jacobi equation with a periodic nonhomo-
geneous term and Aubry–Mather theory, Sbornik: Mathematics 190 (1999), 1487–1504.
[S] P. E. Souganidis, Stochastic homogenization of Hamilton–Jacobi equations and some applications,
Asymptotic Analysis 20 (1999), 1-11.

35