Rezakhanlou 2003

Commun. Math. Phys.
232, 327–375 (2003) Communications in

Digital Object Identifier (DOI) 10.1007/s00220-002-0739-7
Mathematical
Physics
A Stochastic Model Associated with Enskog Equation

and Its Kinetic Limit
Fraydoun Rezakhanlou∗
Department of Mathematics, University of California at Berkeley, Berkeley, CA 94720–3840, USA
Received: 29 January 2002 / Accepted: 30 July 2002

Published online: 14 November 2002 – © Springer-Verlag 2002
Abstract: We examine a system of particles in which the particles travel deterministic-

ally in between stochastic collisions. The collisions are elastic and occur with probability
εd when two particles are at a distance σ . When the number of particles N goes to infinity
and N εd goes to a nonzero constant, we show that the particle density converges to a
solution of the Enskog Equation.
1. Introduction
The Boltzmann Equation provides a successful description for dilute gases and can be
derived from a Hamiltonian particle system in a suitable scaling limit. In the hard sphere
model, one starts with N balls of radius ε that travel according to their velocities and
collide elastically. In a Boltzmann-Grad limit, we send N → ∞, ε → 0 in such a way
that Nεd−1 → 1. If f (x, v, t) denotes the density of particles of velocity v, then f
satisfies the Boltzmann Equation

ft + v · fx = (n · (v − v∗ ))+ [f (x, v )f (x, v∗ ) − f (x, v)f (x, v∗ )]dn dv∗ ,
Rd S
(1.1)
where S denotes the unit sphere, dn denotes the d − 1 dimensional Hausdorff measure
on S, and
v = v − (n · (v − v∗ ))n,
v∗ = v∗ + (n · (v − v∗ ))n.
The Boltzmann Equation ceases to be valid when the density of the gas increases. In 1922,
Enskog proposed a modification of the Boltzmann Equation (1.1) that provided a very
useful description of moderately dense gases. In the Enskog Equation, the
∗ Research supported in part by NSF Grant DMS-0072666.

328 F. Rezakhanlou
expression in the brackets in (1.1) is replaced with

σ d−1 f (x, v )f (x − σ n, v∗ ) − f (x, v)f (x + σ n, v∗ ) . (1.2)
See for example Resibois and De Leener [RD] for more information on the Enskog
Equation.
Even though (1.2) suggests a model in which a collision occurs when particles are at
distance σ , it is not known whether the Enskog Equation can be derived from a Hamil-
tonian particle system. In spite of the original motivation behind the Enskog Equation,
we only regard it in this article as a mollification of the Boltzmann Equation. Techni-
cally speaking, the Enskog Equation is as singular as the one dimensional Boltzmann
Equation (see Sect. 3 for more details). Because of this, it seems plausible that some of
the known techniques for one–dimensional particle systems with kinetic behavior can
be utilized for particle systems associated with the Enskog Equation.
In Pulvirenti [P2], a stochastic particle system is introduced as a microscopic model
associated with the Enskog Equation. In this model we have N particles in Rd that travel
freely in between the collision times. When two particles are at distance σ , then with
probability εd they collide elastically, and with probability 1−ε d they go ahead with
no change in their velocities. This model should be regarded as a suitable smoothing of
the hard sphere model because the interaction length does not go zero as the number
of particles N goes to infinity. In this article we show that as N → ∞ and ε → 0
with Nεd → Z, the macroscopic particle density f (x, v, t) satisfies an Enskog type
equation.
The rigorous derivation of the Boltzmann Equation for models with deterministic
collision rules was carried out by Lanford [L] and King [K]. Lanford established the
kinetic limit for short times for the hard sphere model. King in his thesis utilizes Lan-
ford’s method to treat the case of Hamiltonian systems. Later Illner and Pulvirenti [IP,
P1] showed that Lanford’s restriction on time can be replaced by an assumption on the
smallness of the initial density.
Caprino and Pulvirenti [CP] derived a discrete Boltzmann Equation for stochastic
particle systems on a line with four velocities. Their derivation is valid globally in time
with no smallness condition on the initial data. Their approach, as in [L] and [K], is
based on a detailed analysis of the hierarchy equations for the correlation functions. In
Rezakhanlou [R1] and Rezakhanlou-Tarver [RT], a different approach for the derivation
of Boltzmann-type equations was proposed. This approach explores the entropy bound
and some microscopic bound on the total number of collisions. The key idea behind the
latter is that some well-established techniques of Tartar [T] and Bony [B] have indeed
microscopic counterparts that can be exploited for our purposes.
The Enskog Equation differs from the Boltzmann Equation because of the presence
of the interaction length σ , making the Cauchy problem associated with the Enskog
Equation much easier to handle. We refer to Arkeryd [Ar] and Arkeryd and Cercignani
[ArC] and the references therein for the existence and uniqueness of the initial value
problem. As we will see in Sect. 3, one can establish the existence of solutions using the
dispersive behavior of the free motion part of the equation and Tartar’s ideas [T]. This
will allow us to follow [RT] in studying the collision term by obtaining bounds on the
expected value of the total number of collisions.
The organization of the paper is as follows. In Sect. 2 the main result is stated. In
Sect. 3 the proof of the main result is sketched. Section 4 is devoted to an entropy bound.
In Sect. 5 we establish a bound on the total number of collisions. In Sect. 6, a variant of
A Stochastic Model Associated with Enskog Equation and its Kinetic Limit 329
the Stosszahlensatz (Boltzmann’s molecular chaos principle) will be established. Sec-

tion 7 is devoted to an improvement of the bounds of Sect. 5. In the last section we give
a proof of the kinetic limit.
2. Notation and Main Result

This section is devoted to the statement of our main result. We first describe the model
for which the kinetic limit will be established.
In our model, we have N particles in the d-dimensional Euclidean space Rd . Define
the state space E = (Rd × Rd )N ; q ∈ E is the N -tuple,
q = (x, v) = (q1 , . . . , qN ) , x = (x1 , . . . .xN ) , v = (v1 , . . . , vN ) ,
where qi = (xi , vi ). The process q(t) is a Markov process with the infinitesimal gener-
ator Aε , where ε denotes the length scale and
Aε = A0 + Ac . (2.1)
We have, for any smooth function g : E → R,
N
∂g
A0 g(q) = vi · (q) , (2.2)
∂xi
i=1

Ac g(q) = 1
2 V σ,ε (|xi − xj |)B(vi − vj , −nij )(g(Sij q) − g(q)) , (2.3)
i,j
where V σ,ε(z) = εd−1 V̂ (ε −1 (z − σ )) with V̂ : R → [0, ∞) a continuous even function

satisfying R V (z)dz = 1; B : Rd × S → R is defined by B(v, n) = (v · n)+ with S
x −x
denoting the unit sphere; nij = |xii −xjj | ; Sij q is the configuration obtained from q by
j
replacing vi and vj with vi and vji defined by
j
vi = vi − (nij · (vi − vj ))nij ,
vji = vj − (nj i · (vj − vi ))nj i = vj + (nij · (vi − vj ))nij . (2.4)
Given a positive number β, define the measure νβ by
Nd N
β
νβ (dq) = exp −β |xi |2 + |vi |2 dq .
π
i=1
Here and below we use the notations

dx = dx1 . . . dxN , dv = dv1 . . . dvN , dq = dx dv .
Let µε be a family of probability measures on E and f 0 : Rd × Rd → [0, ∞) be an
integrable function.
Notation 2.1. We will say that µε ∼ f 0 if the following conditions hold:
(i) For every bounded continuous function J : Rd × Rd → R,
330 F. Rezakhanlou
N

lim εd J (xi , vi ) − J (x, v)f 0 (x, v)dx dv µε (dq) = 0 . (2.5)
ε→0 Rd Rd
i=1
(ii) There exist a function F ε : E → [0, ∞) and constants p > 1 and b > 0, such
that
µε (dq) = F ε (q)νβ (dq),

−bε−d
sup e (F ε (q))p νβ (dq) ≤ 1 .
ε>0
Note that if we choose J ≡ 1 in (2.5), we deduce that

εd N → Z =: f 0 (x, v)dxdv . (2.6)
Also, if we define
N
1
Gε (q) = f 0 (xk , vk )
ZN
k=1
with Z as in (2.6), then µε (dq) = Gε (q)dq

satisfies µε ∼ f 0 if and only if

(f 0 (x, v))p exp β(p − 1) |x|2 + |v|2 dxdv < ∞ .
Rd Rd
We assume that q(0) is distributed according to µε and we denote the expectation

associated with q(·) by Eε . We are now ready to state the main result.
Theorem 2.2. Suppose µε ∼ f 0 . Then for every bounded continuous function J : Rd ×
Rd → R,
N

d
lim Eε ε J (xk (t), vk (t)) − J (x, v)f (x, v, t)dx dv = 0,
ε→0 Rd Rd
k=1
where f (x, v, t) is the unique solution to

ft + v · fx = Q(f, f )
, (2.7)
f (x, v, 0) = f 0 (x, v)
where Q = Q(f, f ) is defined by

d−1

Q=σ B(v − v∗ , n) f (x, v )f (x −σ n, v∗ )−f (x, v)f (x + σ n, v∗ ) dn dv∗ .
Rd S
A solution to (2.7) is understood in the following sense:

(i) f ∈ C([0, T ], L1 (Rd × Rd )) and supt∈[0,T ] f log+ f dxdv < ∞,
T
(ii) 0 S B(v−v∗ , n)f (x, v)f
t
(x + σ n, v∗ )dn dv dv∗ dx dt =: X(T ) < ∞,
(iii) f (x, v, t) = f 0 (x − vt, v) + 0 Q(x − (t −s)v, v, s)ds for all t ∈ [0, T ] and
almost all (x, v),
(iv) limT →0 X(T ) = 0,
where Q(x, v, t) := Q(f, f )(x, v, t). It turns out that any such a solution satisfies

F (x, v, T )|v|r dxdv < ∞ (2.8)
for every nonnegative integer r, where F (x, v, T ) denotes the essential supremum of
f (x + vt, v, t) over the t-interval [0, T ]. The proof of this follows the arguments of [Ar].
For example, a slight modification of Lemma 4.2 of [Ar] can be used to establish (2.8)
for small T . The proof of (2.8) for general T is achieved by a bootstrap. We can then
appeal to Theorem 4.1 of [Ar] to assert that the above solution is in fact unique.
Remark 2.3. Most of our arguments are valid for an arbitrary collision rate B that is
merely Lipschitz continuous. It is only in Sect. 4 that the special form of B(v, n) =
(v · n)+ is used to construct an invariant measure for the process q(t).
3. Sketch of Proofs
A well-known proof of the existence of solutions to the Boltzmann-type equation in
dimension one is due to Tartar [T]. Even though we are dealing with the Enskog Equa-
tion in arbitrary dimension, technically speaking, the Enskog Equation is as singular as
the one-dimensional Boltzmann Equation. The reason is that in the collision term of the
Enskog Equation we have something like

f (x, v, t)f (y, v∗ , t)γ (dx, dy), (3.1)
where y belongs to a sphere of diameter σ about the point x. Hence the support of the
measure γ in (3.1) is a set of codimension one. In the Boltzmann Equation the measure
γ (dx, dy) is the Dirac measure δ(x − y). Hence the support is of codimension d. As
a result the codimension of the support of γ in the case of the Enskog Equation and
one-dimensional Boltzmann Equation are equal and this is responsible for the fact that a
Tartar-type argument can be used to establish bounds on the collision term in both cases.
To explain our method of proof in this paper, we start with a bound on the colli-
sion term. Suppose f is a solution to (2.7). A straightforward calculation yields that the
expression

d
|x − vt|2 f (x, v, t)dxdv = |x − vt|2 Q(x, v, t)dxdv ,
dt
equals to

1 d−1
σ B(v − v∗ , n)f (x, v, t)f (x + σ n, v∗ , t)
2 S

· |x − v t|2 + |x + σ n − v∗ t|2 − |x − vt|2 − |x + σ n − v∗ t|2 dndxdv

d
= −tσ B(v − v∗ , n)2 f (x, v, t)f (x + σ n, v∗ , t)dn dxdv .
S
Since the right–hand side is nonpositive, we deduce

sup |x − vt|2 f (x, v, t)dxdv ≤ |x|2 f 0 (x, v)dxdv .
t
332 F. Rezakhanlou
This and the conservation of the energy

d
|v|2 f (x, v, t)dxdv = 0
dt
imply that

(|x|2 + |v|2 )f (x, v, t)dxdv ≤ 2|x|2 + (1 + 2t 2 )|v|2 f 0 (x, v)dxdv .
(3.2)
Again, it is straightforward to show

d
x · v f (x, v, t)dxdv
dt

σd
= B(v − v∗ , n)2 f (x, v, t)f (x + σ n, v∗ , t)dndxdv
2 S

+ |v|2 f (x, v, t)dxdv .
From this and (3.2) we deduce that

T
B(v − v∗ , n)2 f (x, v, t)f (x + σ n, v∗ , t)dn dxdv
S
0
(3.3)
≤ c0 (|x|2 + |v|2 )f 0 (x, v)dxdv ,
for a constant c0 that depends on T only. From now on we assume that the right–hand
side of (3.3) is finite.
Put B1 (v, n) = B(v, n)?(B(v, n) ≤ 1) and define
T
X1 (T ) = B1 (v−v∗ , n)f (x, v, t)f (x + σ n, v∗ , t)dn dv dv∗ dx dt .
0 S
From (3.3) we learn
c1
X∞ (T ) − X1 (T ) ≤ (3.4)
1
for a constant c1 that depends on T only. As a result, a bound on X1 for any finite 1
implies a bound on X∞ .
We certainly have
f (x, v, t)f (x + σ n, v∗ , t) = f 0 (x − vt, v)f 0 (x + σ n − v∗ t, v∗ )
t
d
+ f (x − v(t −s), v, s)
0 ds
× f (x + σ n − v∗ (t −s), v∗ , s) ds
= f 0 (x − vt, v)f 0 (x + σ n − v∗ t, v∗ )
t
+ Q(x − v(t −s), v, s)f (x +σ n−v∗ (t −s), v∗ , s)ds
0 t
+ f (x − v(t −s), v, s)Q(x +σ n−v∗ (t −s), v∗ , s)ds.
0
Hence
X1 = 21 + 22 + 23 ,
where 2i is obtained by multiplying the i th term on the right-hand side of (3.2) by

B1 (v−v∗ , n) and integrating with respect to n, v, v∗ , x and t.
We start with the term 22 . After a translation we obtain that 22 equals to
T
Q(x, v, s)
0
T
× B1 (v−v∗ , n)f (x + σ n + (v−v∗ )(t −s), v∗ , s)dtdndv∗ dvdxds .
S s
A straightforward calculation yields
dz = σ |w · n| dn dt
for z = x + σ n + (t − s)w. As a result 22 is bounded above by

1 T +
Q (x, v, s) f (z, v∗ , s)dz dv∗ dvdxds, (3.5)
σ 0 A
where

+ d−1
Q (x, v, t) = σ B(v − v∗ , n)f (x, v , t)f (x − σ n, v∗ , t)dn dv∗ ,
Rd S
and
A = {z : z = x + (v − v∗ )τ + σ n for some (n, τ ) ∈ S × [0, T ]

with B(v − v∗ , n) ∈ [0, 1]} .
Note that the volume of A can be estimated as

T
|A| ≤ B1 (v − v∗ , n)dndt ≤ 1T dn .
0 S S
As a result,
22 ≤ σ d−2 γ X∞ (T ), (3.6)

where γ = supv γ (v, v∗ ) dv∗ and

γ = sup f (z, v∗ , s)dzdv∗ : s ∈ [0, T ], |A| ≤ c2 1T ,
A
with c2 the d − 1 dimensional measure of the sphere S. The term 23 can be treated
likewise. Moreover, it is not hard to show
2
21 ≤ c3 f (x, v)dx dv ,
0
334 F. Rezakhanlou
for some constant c3 . From this, (3.4) and (3.6) we deduce

2
X1 (T ) ≤ c3 f (x, v)dx dv + 2σ d−2 γ X∞ (T ) ≤ c4 + c4 γ X1 (T ) .
0
(3.7)
A bound on the entropy

H (T ) = sup f log+ f (x, v, t)dx dv < ∞ , (3.8)
0≤t≤T
2 2 0
the conservation of the kinetic energy |v| f (x, v, t)dxdv = |v| f (x, v)dxdv,
and

f (z, v∗ , s)dzdv∗ ≤ f (z, v∗ , s)?(|v∗ | ≤ 1 )dzdv∗
A A

+ (1 )−2 |v∗ |2 f (z, v∗ , s)dzdv∗ ,
imply that there exists a positive constant c5 such that
γ ≤ c5 [log((1 )d 1T )]−1 + c5 (1 )−2 , (3.9)

for every positive 1 . Hence, if T is sufficiently small, then we can choose an appropriate
1 so that c4 γ < 1. This in turn implies a bound of the form
c4
X1 (T ) ≤ . (3.10)
1 − c4 γ
In Sect. 4 we establish a microscopic analog of the entropy bound (3.8) and its conse-
quence (3.9). The bound (3.9) will be used in Sect. 5 to establish a microscopic analog
of the collision bound (3.10).
For our main result Theorem 2.2, we need to show that the number of collisions
near a point (x, v, t) converges to the quadratic functional Q as N diverges to in-
finity. Formally, this is the Stosszahlensatz of Boltzmann. This principle asserts that
the particles before a collision are almost (stochastically) independent. We establish
a variant of this principle
√ in√Sect. 6 that macroscopically has the following flavor: If
ζ δ (z, w) = δ −d ζ (z/ δ, w/ δ) for a nonnegative smooth function ζ of compact sup-
port with ζ (z, w)dz dw = 1, and if J is a bounded continuous function of compact
support, then
T
lim (Q1 (f ∗ ζ δ , f ∗ ζ δ ) − Q1 (f, f ))J dx dv dt = 0, (3.11)
δ→0 0
where Q1 denotes the collision term when B is replaced with B1 . Recall that by (3.4),
the replacement of Q with Q1 causes an error of order O(1/1). Note that (3.11) would
follow from a stronger statement
T
lim sup (Q1 (τz,w f, f ) − Q1 (f, f ))J dx dv dt = 0, (3.12)
δ→0 |z|,|w|<δ 0
where
(τz,w f )(x, v, t) = f (x + z, v + w, t) .
The limit (3.12) can be established by starting from (3.2) but now (x, v) in f (x, v, t)
and f 0 (x − vt, v) is replaced with (x + z, v + w). We then multiply both sides by J ,
integrate, compare with the case z = w = 0, and repeat our previous arguments. The
corresponding set A in (3.5) now satisfies
|A| ≤ c6 | log δ|−1 .

Again the entropy inequality can be used to conclude
T
(Q1 (τz,w f, f ) − Q1 (f, f ))J dx dv dt ≤ c7 | log δ|−1 . (3.13)
0
A microscopic analog of (3.13) will be established in Sect. 6 and this will allow us to
replace the microscopic collision term with
Q(f ∗ ζ δ , f ∗ ζ δ ) + o(1),
where f represents the microscopic density. Here o(1) represents a random term that
goes to zero as ε and δ go to zero. From this we deduce that the macroscopic density
satisfies
ft + v · fx = Q(f ∗ ζ δ , f ∗ ζ δ ) + o(1) (3.14)
(after sending ε → 0) where o(1) now represents an error term that goes to zero as
δ → 0. To complete the proof, we now need to establish some kind of uniform integ-
rability of the collision term so that we can replace back f ∗ ζ δ with f . This will be
carried out in two steps. The first step is carried out before sending ε → 0, and is only a
uniform integrability in the space variable. More precisely, we establish a microscopic
version of the bound
T
+
; Q (f, f )(x + vt, v, t)dt dx dv < ∞, (3.15)
0
where ;(z) = z log+ log+ z and

Q+ (f, f ) = σ d−1 B(v − v∗ , n)f (x, v )f (x − σ n, v∗ ) dn dv∗ . (3.16)
Rd S
It is more convenient to establish (3.15) for Q− = Q+ − Q;

T
; Q− (f, f )(x + vt, v, t)dt dx dv < ∞. (3.17)
0
In fact (3.8) and (3.17) imply (3.15) because (2.7) implies

T
Q+ (f, f )(x + vt, v, t)dt = f (x + vT , v, T ) − f 0 (x, v)
0
T
+ Q− (f, f )(x + vt, v, t)dt . (3.18)
0
A microscopic analog of (3.17) will be established in Sect. 7. This will be used in

Sect. 8 to show that if f is the macroscopic density, then
336 F. Rezakhanlou
T
sup ;(Q+ (f ∗ ζ δ , f ∗ ζ δ ))dx dv dt < ∞ .
δ>0 0
This will be used to replace f ∗ ζ δ with f in (3.14).

As we will see in Sect. 7, (3.17) follows if we can show that for some constant c8 ,

U (x, v)?(U (x, v) > 1)dx dv ≤ c8 (log log 1)−1 , (3.19)
where U (x, v) is the argument of ; in (3.17). Put A = {(x, v) : U (x, v) > 1}. After a
translation we can rewrite (3.17) as
T
B(v − v∗ , n)f (x, v)f (x + σ n, v∗ )? ((x − vt, v) ∈ A) dndxdv dv∗ dt
0 S
≤ c8 σ d−1 (log log 1)−1 . (3.20)
It seems plausable that we can establish (3.18) using the same approach we utilized
for (3.10) and (3.13). Instead we prefer to appeal to a new trick that has the same flavor
as the method used in Sect. 6 of [R2]. Define Y (t) to be

f (x, v, t)f (y, v∗ , t)?(|x − y| ≤ σ )[(x − y) · (v∗ − v)]+
×H (x, y, t, v, v∗ )dx dy dv dv∗
for a suitable nonnegative function H that will be determined later. We have that
dY
= 21 + 22 + 23 + 24 ,
dt
where

21 = − v · fx (x, v, t)f (y, v∗ , t) + v∗ · fy (y, v∗ , t)f (x, v, t)
?(|x − y| ≤ σ )[(x − y) · (v∗ − v)]+ H (x, y, t, v, v∗ )dx dy dv dv∗ ,

22 = f (x, v, t)Q(y, v∗ , t)?(|x − y| ≤ σ )[(x − y) · (v∗ − v)]+
×H (x, y, t, v, v∗ )dx dy dv dv∗ ,

23 = Q(x, v, t)f (y, v∗ , t)?(|x − y| ≤ σ )[(x − y) · (v∗ − v)]+
× H (x, y, t, v, v∗ )dx dy dv dv∗ ,

24 = f (x, v, t)f (y, v∗ , t)?(|x − y| ≤ σ )[(x − y) · (v∗ − v)]+
× Ht (x, y, t, v, v∗ )dx dy dv dv∗ .
After an integration by parts,

21 + 2 4 = − |v − v∗ |2 ? ((x − y) · (v∗ − v) > 0) f (x, v, t)f (y, v∗ , t)
?(|x − y| ≤ σ )H (x, y, t, v, v∗ )dx dy dv dv∗

2
+σ d f (x, v, t)f (x + σ n, v∗ , t) ((v − v∗ ) · n)+
S
H (x, x − σ n, t, v, v∗ )dx dn dv dv∗

+ f (x, v, t)f (y, v∗ , t)(Ht + v · Hx + v∗ · Hy )(x, y, t, v, v∗ )
?(|x − y| ≤ σ ) [(x − y) · (v∗ − v)]+ dx dy dv dv∗

=: 211 + 212 + 213 .
In view of (3.20) we would like to choose
H (x, y, t, v, v∗ ) = ? ((x − vt, v) ∈ A) . (3.21)
Note that H satisfies
Ht + v · Hx + v∗ · Hy = 0 (3.22)
weakly, and as a result 213 = 0. In order to obtain
T
2
((v − v∗ ) · n)+ f (x, v)f (x + σ n, v∗ )
0 S
? ((x − vt, v) ∈ A) dndxdv dv∗ dt ≤ c9 (log 1)−1 , (3.23)
it suffices to verify
T T T
Y (T ) + 22 dt + 23 dt + 211 dt ≤ c10 (log |A|)−1 (3.24)
0 0 0
because by Chebyshev’s inequality,

1 T
|A| ≤ B(v − v∗ , n)f (x, v, t)f (x + σ n, v∗ , t)dn dx dv dv∗ dt .
1 0 S
In fact one can establish (3.24) in just the same way we obtained (3.9).
To complete the proof of (3.15), we need to explain why (3.23) implies (3.20).
Observe that the left-hand side of (3.20) is bounded above by 2̂1 + 2̂2 , where
T
2̂1 = B(v − v∗ , n) ? (|(v − v∗ ) · n| ≤ δ) f (x, v)f (x + σ n, v∗ )
0 S
? ((x − vt, v) ∈ A) dndxdv dv∗ dt,
T
2̂2 = B(v − v∗ , n) ? (|(v − v∗ ) · n| ≥ δ) f (x, v)f (x + σ n, v∗ )
0 S
? ((x − vt, v) ∈ A) dndxdv dv∗ dt .
We can show
2̂1 ≤ c11 | log δ|−1 , (3.25)
338 F. Rezakhanlou
in just the same way we obtained (3.9). On the other hand, (3.23) implies
2̂2 ≤ c12 (δ log 1)−1 . (3.26)
This and (3.25) imply (3.20) if we choose δ = (log 1)−1/2 .
4. Entropy
In this section we establish an exponential bound on an entropy-like functional of the
microscopic particle density. This estimate will play the same role as the bound (3.8)
played in the previous section, and will prove to be the key to demonstrating the bound-
edness of the number of collisions. Before discussing our entropy bound and some of its
consequences, we need to make some definitions. We partition Rd into sets of the form
d
[ar , br ) ,
r=1
each of which is of side length δ. This partition will be denoted by J δ . We also partition
the set Rd × Rd into sets of the form
d d
[ar , br ) × [ar , br ),
r=1 r=1
each of which is of side length at most δ. This partition will be denoted by Jˆ δ . We then
define
N

N (q; K) = N (x1 , v1 , . . . , xN , vN ; K) = ?((xi , vi ) ∈ K) (4.1)
i=1
for every set K ⊆ Rd × Rd and define

@ε (q) = ϕ(N (q; I × Rd )) + ϕ(N (q; Rd × I )) ,
I ∈J ε
(4.2)
ˆ ε (q) =
@ ϕ(N (q; I )) ,
√
I ∈Jˆ ε
where ϕ(z) = z log z. Recall the constant p that appeared in Notation 2.1. The main
result of this section is Theorem 4.1.
Theorem 4.1. There exists a constant C0 (T ) such that

p−1 ε
Eε sup exp @ (q(s)) ≤ exp(C0 (T )ε −d ) ,
0≤s≤T 2p
(4.3)
p−1 ε
Eε sup exp @ˆ (q(s)) ≤ exp(C0 (T )ε −d ) .
0≤s≤T 2p
As the first step, we drop the supremum in (4.3) and replace the expectation with the
integration with respect to the measure νβ .
Lemma 4.2. There exists a constant c0 such that

exp[@ε (q)]νβ (dq) ≤ exp(c0 ε −d ) , (4.4)

ˆ ε (q)]νβ (dq) ≤ exp(c0 ε −d ) .
exp[@ (4.5)
Proof. We only prove (4.5) because√the proof of (4.4) is identical. Let us write I1 , I2 , . . .
for the elements of the partition Jˆ ε and write a for N (q; I ). Given a collection of
r r
nonnegative integers k1 , k2 , . . . with r kr = N , put nj = k1 + . . . + kj . Note that only
finitely many kr ’s are nonzero. Given i ∈ {1, . . . , N}, we can find a unique r = r(i)
such that nr < i ≤ nr+1 . Given r, put 1(r) = min{|q| : q ∈ Ir }. For each nonnegative
integer n, define

Jn = {r : n ≤ 1(r) < n + 1}, Nn = kr .
r∈Jn
We also write γn for the cardinality of the set Jn . We have that νβ (a1 = k1 , a2 = k2 , . . . )
is equal to
N!
∞ νβ (q1 , . . . , qn1 ∈ I1 , qn1 +1 , . . . , qn2 ∈ I2 , . . . )
r=1 kr !
N!
= ∞ νβ (qi ∈ Ir(i) for i = 1, . . . , N )
r=1 kr !
N
N!
= ∞ νβ (qi ∈ Ir(i) )
r=1 kr ! i=1
dN N
N! β
d N
≤ ∞ (ε ) exp −β 1(r(i))2
r=1 k r ! π
i=1
dN ∞
N! β
≤ ∞ (ε d )N exp −β n 2 Nn , (4.6)
r=1 kr ! π n=0
because
d d d
β β β
exp −β|q|2 dq ≤ |Ir | exp(−β1(r)2 ) = ε d exp(−β1(r)2 ) .
Ir π π π
As a consequence of Stirling’s theorem, we have
1 k log k
ec1 k ≤ e ≤ ec2 k (4.7)
k!
for some positive constants c1 and c2 . Also, ε d ≤ c3 /N for some constant c3 . From this
and (4.6) we learn
340 F. Rezakhanlou

ˆ ε (q))νβ (dq)
exp(@
∞
= ear log ar νβ (dq)
r=1
∞
= ekr log kr νβ (a1 = k1 , a2 = k2 , . . . )

kr =N r=1
∞ dN ∞
N! β
kr log krdN
≤ ∞ e ε exp −β n2 N n
r=1 k n ! π
kr =N r=1 n=1
dN ∞
β
≤ c3N e−c1 N ec2 N exp −β n2 Nn
π
kr =N n=1
∞

= ec4 N exp −β n2 Nn

kr =N n=1
 
∞

= ec4 N exp −β n2 Nn ? kr = Nn for n = 0, 1, . . . 

Nn =N n=1 r∈Jn
∞ ∞
γn + N n − 1
c4 N
=e exp −β n2 Nn .
Nn
Nn =N n=1 n=0
It is not hard to see that γn ≤ c5 n2d−1 N for some constant c5 . Also

N
γn + Nn − 1 γn (γn + 1) . . . (γn + Nn − 1) (c5 n2d−1 + 1)N n
= ≤ .
Nn Nn ! Nn !
As a result,
∞ ∞
γn + N n − 1 N exp(c6 Nn log n)
≤N ∞n=1 .
Nn n=0 Nn !
n=0
From this and (4.7) we deduce

∞ ∞
γn + N n − 1
c2 N
≤e exp c6 Nn log n + N log N − Nn log Nn .
Nn
n=0 n=1 n
−2 .
Let us write α for nn Recall that ϕ(z) = z log z. We certainly have

N 1 1
ϕ(N ) = αϕ + N log α = αϕ Nn n2 + N log α
α α n n2
1
≤ ϕ(Nn n2 ) + N log α = Nn log Nn + 2 Nn log n + N log α .
n
n 2
n
Hence
∞ ∞
γn + N n − 1
c7 N
≤e exp c7 Nn log n .
Nn
n=0 n=1
As a result
∞ ∞

ˆ ε (q))ν β (dq) ≤ ec8 N
exp(@ exp −β n2 N n + c 7 Nn log n

Nn =N n=1 n=1
∞
β
c9 N
≤e exp − n2 Nn .
2
Nn =N n=1
A1 denote the set of N̄ = (N0 , N2 , . . . ) such that each Nn is a nonnegative integer,

Let
n Nn = N , N1 > 0, and Nn = 0 for every n > 1. Using (4.7), it is not hard to show
that

1+N −1
≤ ec10 (N+1) ,
N
for some constant c10 . Hence

∞

∞
β 2
ˆ (q))ν (dq) ≤ e
exp(@ ε β c11 N
exp − n Nn
2
1=1 N̄∈A1 n=1
∞
β
≤ ec11 N exp − 12
2
1=1 N̄∈A1
∞
c11 N 1+N −1 β 2
≤e exp − 1
N 2
1=1
∞
β
≤ ec11 N ec10 (N+1) exp − 12 ≤ ec12 N ,
2
1=1
for some constant c12 and this completes the proof.

Recall that initially the configuration q is distributed according to the measure
µε (dq) = F ε (q)νβ (dq) = Gε (q)dq ,

where Gε (q) = F ε (q) exp(−βK(q)) with K(q) = i |xi |2 + |vi |2 . Since dq is not an
invariant measure for the process q(t), let us first find an invariant measure to replace dq.
To this end, choose a positive constant c0 such that the support of V̂ is contained in the
interval (−c0 , c0 ). We take Ŵ : R → R to be a function such that Ŵ = −V̂ , Ŵ (z) = 0
for z > εc0 and Ŵ (z) = 1 for z < −εc0 . Define W σ,ε (z) = εd Ŵ (ε −1 (z − σ )), and
1 σ,ε
R(q) = W (|xi − xj |) , m(dq) = exp(−R(q))dq .
2
i,j
342 F. Rezakhanlou
Lemma 4.3. The adjoint of the operator Aε with respect to the measure m(dq) is −A0 +
Âc , where

Âc g(q) = 21 V σ,ε (|xi − xj |)B(vi − vj , nij )(g(Sij q) − g(q)) . (4.8)
i,j
Moreover, the measure m is invariant.

Proof. Let η1 (q) and η2 (q) be two continuously differentiable functions of compact
support. We have

1
A0 η1 η2 dm = − η1 (q)η2 (q) V σ,ε (|xi − xj |) (vi − vj ) · ni,j m(dq)
2
i,j

− η1 A0 η2 dm .
In other words,
1 σ,ε
A∗0 = −A0 − V (|xi − xj |) (vi − vj ) · ni,j . (4.9)
2
i,j
On the other hand, since R is independent of v and the Jacobian of the transformation
(v, v∗ ) → (v , v∗ ) is equal to one, we have

Ac η1 η2 dm

1 σ,ε
= V (|xi − xj |)B(vi − vj , −nij )η1 (Sij q))η2 (q)m(dq)
2
i,j
1 σ,ε
− V (|xi − xj |)B(vi − vj , −nij )η1 (q)η2 (q)m(dq)
2
i,j
1 σ,ε j
= V (|xi − xj |)B(vi − vji , nij )η1 (Sij q)η2 (q)m(dq)
2
i,j
1 σ,ε
− V (|xi − xj |)B(vi − vj , −nij )η1 (q)η2 (q)m(dq)
2
i,j
1 σ,ε
= V (|xi − xj |)η1 (q) B(vi − vj , nij )η2 (Sij q)
2
i,j

− B(v i − v j, −n ij )η 2 (q) m(dq)
1
= η1 Âc η2 dm + η1 (q)η2 (q) V σ,ε (|xi − xj |) (vi − vj ) · nij m(dq) .
2
i,j
This and (4.9) imply (4.8).

An immediate consequence of (4.8) is Aε ηdm = 0 for every continuously differ-
entiable function of compact support η. This implies the invariance of m.

As our next step, we choose m̂(dq) = exp(−βK(q))dm(q) for the initial distribu-
tion where K(q) = i |xi |2 + |vi |2 . Let us write m̂(t, dq) = Z(t, q)m(dq) for the

distribution of q(·) at time t. Define K(t, q) = i |xi − tvi |2 + |vi |2 .
Lemma 4.4. We have K(t, q(t)) ≤ K(0, q(0)) and
Z(t, q) ≤ exp(−βK(t, q))
for every t.
Proof. The proof of K(t, q(t)) ≤ K(0, q(0)) follows from two straightforward facts:
K(t, q(t)) does not change between the collision times and K(t, q(t)) decreases each
time a collision occurs. The former is obvious and the latter follows from
j
|xi − vi t|2 + |xj − vji t|2 − |xi − vi t|2 − |xj − vj t|2
j
= 2t (xi · vi + xj · vj − xi · vi − xj · vji )
j
= 2t (xi − xj ) · (vi − vi )
= 2t (vi − vj ) · (xi − xj ) ≤ 0 ,
q
whenever B(vi − vj , −nij ) = 0. Let us write Eε for the expectation when the process
q(t) is q at t = 0. If η is a nonnegative test function, then

η(q)Z(t, q) m(dq) = Eεq η(q(t)) exp(−βK(q)) m(dq)

≤ Eεq η(q(t)) exp(−βK(t, q(t))) m(dq)

= η(q) exp(−βK(t, q)) m(dq) ,
where for the last equality we used the fact that m is an invariant measure. This evidently
completes the proof.
We now write
µε (dq) = H0 (q)m(dq) ,
where H0 (q) = F ε (q) exp(R(q) − βK(q)). This implies that the configuration q(t) is
distributed according to the measure
µε (t, dq) = H (t, q)m(dq) ,
where H satisfies H (0, q) = H0 (q) and the backward equation

∂H
= Aε,∗ H ,
∂t
where Aε,∗ = −A0 + Âc is the adjoint of the operator Aε with respect to the measure
m.
Lemma 4.5. There exists a constant c such that

H (t, q) p −d
Z(t, q) m(dq) ≤ ecε ,
Z(t, q)
for every t.
344 F. Rezakhanlou
p
Proof. Define ψ(a, b) = ab b. It is not hard to show that ψ is a convex function. We
now show that the convexity of ψ implies

d
ψ(H (t, q), Z(t, q)) m(dq) ≤ 0 . (4.10)
dt
To see this, first observe that the left–hand side of (4.10) equals to

ψa (H, Z)Aε,∗ H + ψb (H, Z)Aε,∗ Z m(dq) ,
where ψa and ψb are the partial derivatives of ψ. To ease the notation, let us write B for
Aε,∗ . Note that B is an infinitesimal generator for a Markov process. We now claim that
the Maximum Principle and the convexity of ψ imply
Bψ(H, Z) ≥ ψa (H, Z)BH + ψb (H, Z)BZ . (4.11)
To show this, define
k(q̄) = ψ(H (t, q̄), Z(t, q̄)) − ψ(H (t, q), Z(t, q))
− ψa (H (t, q), Z(t, q))(H (t, q̄) − H (t, q))
− ψb (H (t, q), Z(t, q))(Z(t, q̄) − Z(t, q)) .
This function
attains its minimum value at q. By Maximum Principle, (Bk)(q) ≥ 0.
This and Bψ(H (t, ·), Z(t, ·)) dm = 0 imply (4.11). This in turn implies (4.10). As a
consequence,

ψ(H (t, q), Z(t, q)) m(dq) ≤ ψ(H (0, q), Z(0, q)) m(dq)

ε p −d
= F (q) exp(R(q)) exp(−R(q)) νβ (dq) ≤ ecε ,
because of our condition on the initial distribution and the fact that R(q) = O(ε −d ).

We are now ready to prove Theorem 4.1.
Proof of Theorem 4.1. We only prove the first bound in (4.3) because the proof of the
second bound is identical. By Hölder inequality and Lemma 4.5,

p−1 ε p−1 ε
Eε exp @ (q(t)) ≤ exp @ (q(t)) µε (t, dq)
p p
p−1 ε H (t, q)
= exp @ (q(t)) Z(t, q)m(dq)
p Z(t, q)
p−1
ε p −d
≤ exp @ (q(t)) Z(t, q)m(dq) ecε /p .
From this, Lemma 4.4 and Lemma 4.2 we deduce

p−1 ε −d
Eε exp @ (q(t)) ≤ ec1 ε (4.12)
p
for a constant c1 . We can now follow the proof of Theorem 2.1 of [RT] to deduce (4.3)
from (4.12). (See Steps 2 and 3 of that proof.)
Corollary 4.6 is the main consequence of Theorem 4.1 that is used in this paper. The
straightforward proof of this corollary is omitted.
Corollary 4.6. There exists a constant C1 (T , r) such that
r
Eε sup ε d @ε (q(s)) ≤ C1 (T , r) ,
0≤s≤T
r
Eε sup ε d @ ˆ ε (q(s)) ≤ C1 (T , r) .
0≤s≤T
We end this section with four lemmas that will prepare us to use Corollary 4.4 in the
proceeding sections. Let

|1 + log δ|−1 if δ < 1 ,
h(δ) =
1 otherwise .
We also fix a continuous function η : Rd × Rd → [0, ∞) of compact support. Given a

measurable function ρ, we set

ε
ρ (x, v) = ηε (x − z, v − w)ρ(z, w)dzdw ,
Rd Rd
√ √
where ηε (x, v) = ε−d η x/ ε, v/ ε .
Lemma 4.7. There exists a constant C1 (η) such that for every nonnegative ρ,
N

ε d ˆ ε (q)) ,
ρ ε (xi , vi ) ≤ C1 (η)ρL∞ h(ρL1 )(1 + ε d @
i=1
where · Lp denotes the Lp norm with respect to the set Rd × Rd .

The proof of Lemma 4.7 follows Lemma 5.2 of [RT] and is omitted.
Given a measurable set K ⊆ Rd × Rd , we choose ρ(x, v) = ? ((x, v) ∈ Bε K),
where
√
Bε K = K + ε[0, 1]2d .

We also choose η(x) = ? (x, v) ∈ [0, 1]2d . It is not hard to see that for such a pair
(ρ, η), we have
ρ ε (x, v) ≥ ? ((x, v) ∈ K) .
This and Lemma 4.7 yields
Lemma 4.8. There exists a constant Ĉ1 such that for every measurable set K,
ˆ ε (q)) .
εd N (q; K) ≤ Ĉ1 h(|Bε K|)(1 + ε d @
Similarly we define
B̂ε K = K + ε[0, 1]d ,
for every set K ⊆ Rd . In the same way, we can show
Lemma 4.9. There exists a constant C̃1 such that for every measurable set K ⊆ Rd ,

εd N q; K × Rd ≤ C̃1 h(|B̂ε K|)(1 + ε d @ε (q)),

εd N q; Rd × K ≤ C̃1 h(|B̂ε K|)(1 + ε d @ε (q)) .
346 F. Rezakhanlou
5. Collision Bound
In this section we establish a microscopic analog of (3.10). Define

Aε1 (q; z) = εd V σ,ε (|xi − xj + z|)B1 (vi − vj , −n̂ij (z)) ,
i,j
where B1 = B?(B ≤ 1) and

xi − xj + z
n̂ij (z) = .
|xi − xj + z|
We write Aε1 (q) for Aε1 (q; z) when z = 0. We also write Aε (q; z) (respectively Aε (q))
for Aε1 (q; z) when 1 = ∞ (respectively z = 0 and 1 = ∞).
T
Eε Aε (q(t))dt ≤ C2 (T ) . (5.1)
0
As in Sect. 3, we first find a bound on the number of collisions when the collision
rate is not small.
Lemma 5.2. There exists a constant C̄2 (T ) such that
T
εd V σ,ε (|xi (s) − xj (s)|)B(vi (s) − vj (s), −nij (s))2 ds ≤ C̄2 (T )
0 i,j
T C̄2 (T )
Eε (Aε (q(t)) − Aε1 (q(t)))dt ≤ .
0 1
Proof. By Lemma 4.4 and the conservation of the kinetic energy,

|xi (t) − vi (t)t|2 ≤ |xi (0)|2 , |vi (t)|2 = |vi (0)|2 .
i i i i
From this and our condition on the initial distribution µε we deduce

Eε sup |vi (t)|2 + sup ε d (xi (t) · vi (t)) ≤ c0 (5.2)
0≤t≤T i 0≤t≤T i
depends on T only.
for a constant c0 that
Set F (q) = εd i (xi · vi ). By the semigroup property,
T
Eε F (q(T )) = Eε F (q(0)) + Aε F (q(s))ds .
0
Using this and
j j
xi · vi + xj · vji − xi · vi − xj · vj = (xi − xj ) · (vi − vi ) = (xi − xj ) · (vi − vj ) ,
we deduce
T
Eε F (q(T )) = Eε F (q(0)) + εd |vi (s)|2 ds + εd V σ,ε (|xi (s)
0 i 0 i,j
− xj (s)|)B(vi (s) − vj (s), −nij (s)) |xi (s) − xj (s)| ds . 2
This, (5.2), and the trivial inequality B − B1 ≤ B 2 /1 complete the proof of the lemma.

q
Let us write Eε for the expectation when the process q(·) satisfies q(0) = q. To
prepare for the proof of Theorem 5.1, we start with an elementary identity that is a
consequence of the semigroup property. For any continuous function F : E × R → R,
t
Eεq F (q(t)) = F (St q) + Eεq Ac (F oSt−s )(q(s))ds , (5.3)
0
where for q = (x1 , v1 , . . . , xN , vN ) = (x, v),
Sθ q = (x1 + v1 θ, v1 , . . . , xN + vN θ, vN ) = (x + vθ, v),

F oSθ (q) = F (Sθ q) .
To show (5.3), fix t and define G(q, s) = F (St−s q). By the semigroup property,
t
∂
Eεq G(q(t), t) = G(q, 0) + Eεq + Aε G(q(s), s)ds .
0 ∂s
∂
This implies (5.3) because ∂s + Aε G = Ac F .
For Theorem 5.1 and Theorem 6.1 of the next section, we apply (5.2) with

F (q) = F ε (q; J, z, 1) = ε d V σ,ε (|xi − xj + z|)B1 (vi − vj , −n̂ij (z))J (xi , vi ) .
i,j
j
Let us write vi (s) and vji (s) for the outgoing velocities where the ingoing velocities
are vi (s) and vj (s). Define M(a) = a/|a| and
nij (s, θ ; z) = M(xi (s) − xj (s) + (vi (s) − vj (s))θ + z) ,

j
nij (s, θ ; z) = M(xi (s) − xj (s) + (vi (s) − vji (s))θ + z) ,
(5.4)
nkij (s, θ ; z) = M(xi (s) − xj (s) + (vik (s) − vj (s))θ + z) ,
ñkij (s, θ ; z) = M(xi (s) − xj (s) + (vi (s) − vjk (s))θ + z) .
We simply write nij (s, θ ), nij (s, θ ), nkij (s, θ ), and ñij (s, θ ) if z is zero in (5.4). We also
write nij (s) for nij (s, 0). We have,
T
Eεq F (q(t); J, z)dt = 21 (z, J ) + 22 (z, J ) + 23 (z, J ) + 24 (z, J ) , (5.5)
0
where
T
21 (z, J ) = εd V σ,ε (|xi − xj + (vi − vj )t + z|)B1 (vi − vj , −nij (0, t; z))
0 i,j
·J (xi + vi t, vi ) dt , (5.6)
348 F. Rezakhanlou
T t
1 q
22 (z, J ) = E V σ,ε (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))
2 ε 0 0 i,j,k
· V σ,ε (|xi (s) − xj (s) + (vik (s) − vj (s))(t − s) + z|)

× B1 (vik (s) − vj (s), −nkij (s, t − s; z))J (xi (s) + vik (s)(t − s), vik (s))
− V σ,ε (|xi (s) − xj (s) + (vi (s) − vj (s))(t − s) + z|)

× B1 (vi (s) − vj (s), −nij (s, t − s; z))J (xi (s) + vi (s)(t − s), vi (s)) ds dt ,
(5.7)
T t
1 q
23 (z, J ) = E εd V σ,ε (|xj (s) − xk (s)|)B(vj (s) − vk (s), −nj k (s))
2 ε 0 0 i,j,k
· V σ,ε (|xi (s)−xj (s) + (vi (s)−vjk (s))(t −s) + z|)

× B1 (vi (s) − vjk (s), −ñkij (s, t − s; z))J (xi (s)+vi (s)(t −s), vi (s))
− V σ,ε (|xi (s) − xj (s) + (vi (s) − vj (s))(t − s) + z|)

× B1 (vi (s) − vj (s), −nij (s, t − s; z))J (xi (s)+vi (s)(t −s), vi (s)) ds dt ,
(5.8)
T t
1 q
24 (z, J ) = E εd V σ,ε (|xi (s)− xj (s)|)B(vi (s) − vj (s), −nij (s))
2 ε 0 0 i,j
j
× V σ,ε (|xi (s)−xj (s) + (vi (s)−vji (s))(t −s) + z|)
j j j
× B1 (vi (s) − vji (s), −nij (s, t − s; z))J (xi (s) + vi (s)(t − s), vi (s))
− V σ,ε (|xi (s)−xj (s) + (vi (s)−vj (s))(t −s) + z|)

× B1 (vi (s) − vj (s), −nij (s, t − s; z))J (xi (s) + vi (s)(t − s), vi (s)) .
(5.9)
Remark 4.2. By the Optimal Sampling Theorem, we have that (5.5) holds when T is
replaced by a stopping time τ . By averaging over all configurations with the weights
given by the initial distribution µε , we also have (5.5) when Eε is replaced with Eε .
q
Proof of Theorem 5.1. Step 1. We apply (5.4) with J ≡ 1, z = 0, and denote the corre-
sponding terms by 21 , 22 , 23 , 24 . We start with 21 . We have
T
21 = εd V σ,ε (|xi − xj + (vi − vj )t|)B1 (vi − vj , −nij (0, t))dt . (5.10)
0 i,j
Define
H ε (v, T ) = {x : (x, v) ∈ Gε (T )} , (5.11)

where
T
x + θv
Gε (T ) = (x, v) : V σ,ε (|x + θ v|)B1 v, − dθ = 0 . (5.12)
0 |x + θv|
Recall B̂ε K = K + ε[0, 1]d . Since V is of compact support, there exists a constant
c0 such that V σ,ε (|z|) = 0 implies σ − c0 ε ≤ |z| ≤ σ + c0 ε. Hence, there exists a
constant c1 such that if x ∈ B̂ε H ε (v, T ), then σ − c1 ε ≤ |x + θv| ≤ σ + c1 ε for
some θ ∈ [0, T ]. Note that if σ − c1 ε ≤ |x + θv| ≤ σ + c1 ε for some θ ∈ [0, T ] and
|x| ∈
/ [σ − c1 ε, σ + c1 ε], then the line segment {x + θv : θ ∈ [0, T ]} intersects either
the sphere {z : |z| = σ + c1 ε} or the sphere {z : |z| = σ − c1 ε}. Moreover, the function
a
M(a) = |a| is uniformly Lipshitz away from the origin a = 0. As a result, there exists
a constant c2 such that
B̂ε H ε (v, T ) ⊆ A0 ∪ A− ∪ A+ ,
where A0 = {x : σ − c1 ε ≤ |x| ≤ σ + c1 ε} and

x + θv
A± = x : |x + θ v| = σ ± c1 ε, v · ∈ [−1 − c2 ε, c2 ε] for some θ ∈ [0, T ] .
|x + θ v|
Observe that if x ∈ A± , then x = −θ v + (σ ± c1 ε)n for some (θ, n) with θ ∈ [0, T ]
and v · n ∈ [−1 − c2 ε, c2 ε]. From this and the fact that dx = |v · n|dθ dn we deduce
T
|A± | = dx = |v · n|?(v · n ∈ [−1 − c2 ε, c2 ε])dn dθ ≤ αT (1 + c2 ε) ,
A± 0 S
(5.13)
where α denotes the d − 1-dimensional measure of the sphere S. As a result,
|B̂ε H ε (v, T )| ≤ c3 T (1 + ε) = c4 T (5.14)
for some constants c3 and c4 . Moreover,

a + θv
V σ,ε (|a + θ v|)B v, − dθ
R |a + θv|
a + θv −

d−1 −1
= ε V̂ ε (|a + θ v| − σ ) v · dθ
|a + θv|
R
= ε d−1 V̂ ε−1 (η − σ ) dη = ε d . (5.15)
R
From (5.10–15) we deduce

21 ≤ ε 2d ? (xi − xj , vi − vj ) ∈ Gε (T )
i,j

≤ε 2d
? xi ∈ H ε (vi − vj , T ) + xj
i,j
≤ε N 2d
sup N (q; K × Rd ) . (5.16)
|B̂ε K|≤c4 T
350 F. Rezakhanlou
From this and Lemma 4.9 we deduce that

21 ≤ c5 h (c4 T ) (1 + ε d @ε (q)) , (5.17)
for some constant c5 .
Step 2. We now consider 22 . First observe that 22 is bounded above by

1 q T t σ,ε
Eε V (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))
2 0 0 i,k

· V σ,ε (|xi (s) − xj (s) + (vik (s) − vj (s))(t − s)|)
j
× B1 (vik (s) − vj (s), −nkij (s, t − s; z))ds dt

1 q T σ,ε
= Eε V (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))
2 0 i,k
T
s j
× B1 (vik (s) − vj (s), −nkij (s, t − s; z))dt ds .

We concentrate upon the t-integral. As in Step 1, we have that the t-integral is bounded
by

c6 ε d ? (xi (s) − xj (s), vik (s) − vj (s)) ∈ Gε (T − s) ,
j
for some constant c6 . This, a repetition of (5.16) and Lemma 4.9 imply
T
|22 | ≤ c6 Eεq
Aε (q(s)) ε d sup N (q(s); K) ds
0 |B̂ε K|≤c4 T
T
≤ c7 Eεq Aε1 (q(s))ds h (c4 T ) sup (1 + ε d @ε (q(s)))
0 0≤s≤T
T
+c7 Eεq (Aε (q(s)) − Aε1 (q(s)))ds . (5.18)
0
The same estimate holds for 23 because 23 = 22 .
Step 3. As for 24 , note that |24 | is bounded above by
T
Eεq εd V σ,ε (|xi (s) − xj (s)|)B(vi (s) − vj (s), −nij (s))
0 i,j
T
j
· V σ,ε (|xi (s) − xj (s) + (vi (s) − vji (s))(t − s)|)
s
j
× B1 (vi (s) − vji (s), −nij (s, t − s))dt ds
T
d q
≤ ε Eε Aε (q(s))ds ,
0
where for the last inequality we used (5.15).
Step 4. After putting all the pieces together, we see that for some constant c8 ,
T
Eεq Aε1 (q(s))ds ≤ c8 h(c4 T )(1 + ε d @ε (q))
0 T
+ c 8 Eε
q
Aε1 (q(s))ds h (c4 T ) sup (1 + ε d @ε (q(s)))
0 0≤s≤T
T
+ c8 Eεq (Aε (q(s)) − Aε1 (q(s)))ds .
0
Since (5.5) holds whenever T is replaced by a stopping time, a similar argument
shows that
τ
Eε Aε1 (q(s))ds ≤ c9 Eε h(c4 τ ) + c9 /1
0
τ
+c8 Eε Aε1 (q(s))ds h (c4 τ ) sup (1 + ε d @ε (q(s)))ds ,
0 0≤s≤τ
(5.19)
where τ is any stopping time with τ ≤ T . Here we are using Lemma 5.2 and the fact
that εd Eε @ε (q(0)) is uniformly bounded in ε. We now choose

τ = τ ε := T ∧ inf t : c8 h(c4 t) sup (1 + ε d @ε (q(s))) ≤ 1
2 .
0≤s≤t
For such a stopping time τ we deduce

τ
Eε Aε (q(s))ds ≤ 2c9 h(c4 T ) + 2c9 /1 .
0
From this and Lemma 5.2 we deduce
τ
Eε Aε (q(s))ds ≤ 2c9 h(c4 T ) + c10 /1 . (5.20)
0
Final Step. From (5.2), it is not hard to show
T
Eε Aε (q(s))ds ≤ c11 ε −1 .
0
As a result,
T τ T
ε ε
Eε A (q(s))ds ≤ Eε A (q(s))ds + Eε Aε (q(s))?(τ < T )ds
0 0 τ 0
≤ Eε Aε (q(s))ds + c12 ε −1
Pε (τ < T ) .
0
We can now repeat the proof of Theorem 3.1 of [RT] and use Theorem 4.1 to argue that
for T sufficiently small,
Pε (τ < T ) ≤ exp(−c13 ε −d ) .
This and (5.20) imply (5.1) for sufficiently small T . The proof of general T is done by
a bootstrap.
As in [RT], we can readily deduce
352 F. Rezakhanlou
Corollary 5.3. There exists a function C2 (T , k) such that limT →0 C2 (T , k) = 0 and for
every integer k,
T k
ε
Eε A (q(s))ds ≤ C2 (T , k) ,
0
T +τ
sup Eε Aε (q(s))ds ≤ C2 (T , 1) ,
τ τ
where the supremum is over all stopping times.
We continue with two more corollaries that will be needed in Sects. 6 and 7.
Corollary 5.4. There exists a constant C̃2 (T ) such that for any 1 ∈ (0, 1],
T
εd V σ,ε (|xi (s) − xj (s)|)B1 (vi (s) − vj (s), −nij (s))ds
0 i,j
≤ C̃2 (T )| log(1 + ε)|−1 . (5.21)
Proof. The proof is similar to the proof of Theorem 5.1 and we only sketch it. As in the
proof of Theorem 5.1, the right-hand side of (5.21) can be written as 21 + . . . + 24 and
it suffices to show that
|2r | ≤ const.| log(1 + ε)|−1 , (5.22)
for each r. For example, we can readily show

T
|22 | ≤ c6 Eε Aε (q(s)) ε d sup N (q(s); K) ds
0 |B̂ε K|≤c3 T (1+ε)
T
≤ c7 Eε Aε (q(s))ds h (c3 T (1 + ε)) sup (1 + ε d @ε (q(s)))
0 0≤s≤T
in just the same way we showed (5.18). See also (5.14). We can now apply Schwartz’s
inequality, Corollary 5.3 with k = 2, and Corollary 4.6 with r = 2 to deduce (5.22) for
r = 2. The other cases are treated likewise.
Define

Âε (q) = εd V σ,ε (|xi − xj |)B(vi − vj , −nij )|vi | .
i,j
Corollary 5.5. There exists a constant Ĉ2 (T , k) such that,

T k
Eε Âε (q(s))ds ≤ Ĉ2 (T , k) , (5.23)
0
for every positive integer k.

Proof. The proof of this corollary is also similar to the proof of Theorem 5.1 and we
only sketch it. First assume that k = 1. As in the proof of Theorem 5.1, we apply (5.5)
with z = 0 and J (x, v) = |v| to say that the left-hand side of (5.23) can be written as
21 + . . . + 24 . For example, we can readily show
T
1
22 ≤ Eεq V σ,ε (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))|vik (s)|
2 0 i,k
T
s j
× B(vik (s) − vj (s), −nkij (s, t − s; z))dt ds .
Using |vik | ≤ |vi | + |(vi − vk ) · nik | and Lemma 5.2 we deduce that for some constant c,
T
22 ≤ c6 Eε Âε (q(s)) ε d sup N (q(s); K) ds + c
0 |B̂ε K|≤c4 T
T
≤ c7 Eε Â (q(s))ds h(c4 T ) sup (1 + ε d @ε (q(s))) + c ,
ε
0 0≤s≤T
in just the same way we showed (5.18). We now replace T with the stopping time τ as
in the proof of Theorem 5.1 to deduce that for some constant c ,
τ
Eε Âε (q(s))ds ≤ c .
0
A repetition of the final step of the proof of Theorem 5.1 would complete the proof of
(5.23) when k = 1. The proof for the general k is achieved as in [RT].
6. Spatial Regularity of Collision

As we mentioned in the introduction, the main hypotheses behind a kinetic derivation is
the Stosszahlensatz of Boltzmann that asserts that a pair of particles before a collision
are uncorrelated. In our version of this principle, we show that shifting particles around
somewhat does not dramatically alter the value of the collision term. In the last section
we show that such a regularity of the collision term implies a weak form of Stosszahlen-
satz. See also Sect. 3 where we argued that a statement like (3.12) implies (3.11). The
next theorem is our regularity claim. Define

F ε (q; J, z) = εd V σ,ε (|xi − xj + z|)B(vi − vj , −n̂ij (z))J (xi , vi ) ,
i,j
j
(6.1)
F̃ ε (q; J, z) = εd V σ,ε (|xi − xj + z|)B(vi − vj , −n̂ij (z))J (xi , vi ) .
i,j
Theorem 6.1. For every continuously differentiable function J of compact support, there
exists a constant C3 (T , J ) such that
T
ε
sup sup Eε F (q(t); J, z) − F ε (q(t); J, 0) dt ≤ C3 (T , J ) h(δ + ε) , (6.2)
ε>0 |z|<δ 0
T
sup sup Eε F̃ ε (q(t); J, z) − F̃ ε (q(t); J, 0) dt ≤ C3 (T , J ) h(δ + ε) . (6.3)
ε>0 |z|<δ 0
354 F. Rezakhanlou
To prepare for the proof of Theorem 6.1, we start with some definitions and prelim-
inary lemmas. Define
G = {(x, v) ∈ Rd × Rd : |x + vt| = σ for some t ∈ R}

= (x, v) ∈ Rd × Rd : (x · v)2 − |x|2 |v|2 + σ 2 |v|2 ≥ 0 . (6.4)
Given (x, v) ∈ G, the line {x + vt : t ∈ R} intersects the sphere Sσ = {σ n :

n ∈ S} at at most two points. For such an (x, v), we define t − (x, v), t + (x, v) ∈ R,
n− (x, v), n+ (x, v) ∈ S so that
{x + vt : t ∈ R} ∩ Sσ = {σ n− (x, v), σ n+ (x, v)}

(6.5)
t − (x, v) ≤ t + (x, v) .
More explicitly,
x·v 1 1/2
t ± (x, v) = − ± (x · v) 2
− |x| 2
|v| 2
+ σ 2
|v| 2
|v|2 |v|2
± (6.6)
x + vt (x, v) 1
n± (x, v) = ±
= (x + vt ± (x, v)) .
|x + vt (x, v)| σ
We define
Gδ = {(σ n − vt, v) ∈ G : t ∈ R, |n · v| ≤ δ}
= {(x, v) ∈ G : |(n+ (x, v) · v| or |(n− (x, v) · v| ≤ δ} . (6.7)
We continue with an elementary lemma.
Lemma 6.2. We have,

∂t ± ∂n± 1 + |v|
(x, v) + (x, v) ≤ . (6.8)
∂x ∂x |v · n± (x, v)|
Moreover,
1 1
v · n± (x, v) = ±
2
(x · v)2 − |x|2 |v|2 + σ 2 |v|2 . (6.9)
σ
Proof. Observe that if we put g(x, v, t) = |x + vt|2 , then g(x, v, t ± (x, v)) = σ 2 . By
an implicit differentiation we obtain
∂t ± x + vt ± (x, v) n± (x, v)
(x, v) = − = − . (6.10)
∂x (x + vt ± (x, v)) · v n± (x, v) · v
Moreover, n± (x, v) = (x + t ± (x, v)v)/σ . From this and (6.10) we deduce

∂n± 1 v ⊗ n±
(x, v) = I− ± ,
∂x σ n ·v
where I denotes the identity matrix and a ⊗ b means the tensor product of vectors a, b.
From this and (6.10) we deduce (6.8).
The proof of (6.9) is straightforward and follows from (6.6).
Given a continuously differentiable function J of compact support, define fε (x, v,

w; T , J ) = fε (x, v, w) by
T
−d σ,ε x + vt −
fε (x, y, v, w) := ε V (|x + vt|) v · J (y + wt, w)dt .
0 |x + vt|
We also define
;(α) = {(x, v) ∈ G : t − (x, v) ∈ [−α, α] ∪ [T − α, T + α]} .
Lemma 6.3. There exist constants C4 (J ), C5 and Ĉ5 such that if (x, v) ∈ G − Gδ ,
/ ;(C5 ε/δ), and |v| ≤ Ĉ5 δε −1/2 , then
(x, v) ∈
ε
|fε (x, y, v, w) − J (y + wt − (x, v), w)?(t − (x, v) ∈ [0, T ])| ≤ C4 (J ) . (6.11)
δ
Proof. Choose a number c0 so that if V σ,ε (|a|) = 0, then σ − c0 ε ≤ |a| ≤ σ + c0 ε. We
then choose a constant C̃5 such that
(x · v)2 − |x|2 |v|2 + η2 |v|2 = σ 2 (v · n+ )2 + (η2 − σ 2 )|v|2

1
≥ σ 2 δ 2 − |v|2 |η2 − σ 2 | ≥ σ 2 δ 2 ,
2
whenever η ∈ [σ − c0 ε, σ + c0 ε] and |v| ≤ C̃5 δε −1/2 . As a result, if |x + vt| = η ∈
[σ − c0 ε, σ + c0 ε], then t = tˆ+ or tˆ− , where
x·v 1 1/2
tˆ± = − 2 ± 2 (x · v)2 − |x|2 |v|2 + η2 |v|2 . (6.12)
|v| |v|
For t ∈ Iε we
√ have η =
√ σ + cε for
√ some c ∈ [−c0 , c0 ]. From this and the elementary
inequality | a + b − a| ≤ |b|/ a we deduce
x·v 1 1/2
tˆ± + ∓ (x · v) 2
− |x| 2
|v| 2
+ σ 2
|v| 2
|v|2 |v|2
−1/2
≤ 2|c|σ ε + c2 ε 2 (x · v)2 − |x|2 |v|2 + σ 2 |v|2 .
This and (6.9) imply

C5 ε C5 ε
|tˆ± − t ± (x, v)| ≤ ≤ , (6.13)
|v · n± (x, v)| δ
√
for some constant C5 . Furthermore, if |v| ≤ Ĉ5 δε −1/2 with Ĉ5 = min(C̃5 , σ/C5 ),
and (x, v) ∈
/ Gδ , then we can use (6.6) and (6.9) to assert
2σ 2σ ε C5 ε
t + (x, v) − t − (x, v) = v · n+ (x, v) ≥ ≥2 .
|v|2 Ĉ5
2 δ δ
This and (6.13) imply that the set Iε =: {t : V σ,ε (|x + vt|) = 0} can be written as a
union of two disjoint intervals (aε− (x, v), bε− (x, v)) and (aε+ (x, v), bε+ (x, v)) such that
t ± (x, v) ∈ (aε± (x, v), bε± (x, v)) .

356 F. Rezakhanlou
We now write
fε (x, y, v, w) = fε− (x, y, v, w) + fε+ (x, y, v, w), (6.14)
where
bε± (x,v)
fε± (x, y, v, w) = ε −d V σ,ε (|x + vt|)J (y + wt, w)
aε± (x,v)

x + vt −
× ?(t ∈ [0, T ]) v · dt .
|x + vt|
As in (6.9) we have that if t ∈ [aε± , bε± ] and |x + vt| = η, then
x + vt 1 1
2
v· = ± (x · v)2 − |x|2 |v|2 + η2 |v|2 .
|x + vt| η
As a result fε+ = 0. Define
bε− (x,v)
x + vt −
fˆε− (x, y, v, w) := ε −d σ,ε
V (|x + vt|) v · ?(t ∈ [0, T ])dt
aε− (x,v) |x + vt|
· J (y + wt − (x, v), w),
bε− (x,v)
x + vt −
f˜ε− (x, y, v, w) := ε −d V σ,ε (|x + vt|) v · dt
aε− (x,v) |x + vt|
· ?(t − (x, t) ∈ [0, T ])J (y + wt − (x, v), w)
= ?(t − (x, t) ∈ [0, T ])J (y + wt − (x, v), w) .
From (6.13) and the Lipschitzness of the function J we learn
c1 ε
|fε− (x, y, v, w) − fˆε− (x, y, v, w)| ≤ (6.15)
δ
for some constant c1 . Once more we use (6.13) to claim that if (x, v) ∈/ ;(C5 ε/δ), then
the term ?(t ∈ [0, T ]) in fˆε can be replaced with the term ?(t − (x, t) ∈ [0, T ]). This
means fˆε− = f˜ε− . From this, (6.14) and (6.15) we conclude (6.11).
+
Lemma 6.4. There exists a constant
− √ C6 = C6 (T ) such that if (x, v) ∈ G, |t (x, v)| >
C6 , |t (x, v)| > C6 and |v| ≥ ε, then fε (x, y, v, w; T , J ) = 0 for every y and w.
Moreover, fε (x, y, v, w; T , J ) = 0 whenever (x, v) ∈
/ G.
Proof. Recall that Iε = {t : V √ σ,ε (|x + vt|) = 0}. Given (x, v) ∈ G, we can use (6.12)
√ √
and the elementary inequality a + b − a ≤ b to deduce that if t ∈ Iε , then either √
|t − t − | or |t − t + | is bounded above by |v|−1 ((σ + c0 ε)2 − σ 2 )1/2 . For v with |v| ≥ ε
we deduce
min(|t − t − |, |t − t + |) ≤ c1 ,
for some constant c1 . From this and the definition of fε we deduce the lemma.
Finally, it is not hard to show that if Iε = ∅, then (x, v) ∈ G. This implies that
fε (x, y, v, w; T , J ) = 0 if (x, v) ∈
/ G.

Define K(q) = i |vi |2 .
Lemma 6.5. There exists a constant C7 (J ) such that
N

εd (fε (x − xj , y, v − vj , w) − fε (x + z − xj , y, v − vj , w))
j =1

ˆ ε (q) + ε d @ε (q) + ε d K(q) + |v| , (6.16)
≤ C7 (J ) h(|z| + ε) 1 + ε d @
for every ε, |z| ∈ (0, 1].
Proof. Step 1. Using Lemma 6.4, we have that the left-hand side of (6.16) is bounded

above by 8r=1 2r , where
N

2r = εd (fε (x − xj , y, v − vj , w) − fε (x − xj + z, y, v − vj , w)) ξjr ,
j =1
where

ξj1 = ? |v − vj | ≤ Ĉ5 δε −1/2 , (x − xj , v − vj ),

× (x − xj + z, v − vj ) ∈ G − (Gδ ∪ ;(C5 ε/δ)),
ξj2 = ? (x − xj , v − vj ) ∈ ;(C5 ε/δ), |v − vj | ≤ ε−1/2 ,

ξj3 = ? (x − xj + z, v − vj ) ∈ ;(C5 ε/δ)), |v − vj | ≤ ε−1/2 ,
√
ξj4 = ? (x − xj , v − vj ) ∈ Gδ , ε ≤ |v − vj | ≤ δ −1/2 ,
√
ξj5 = ? (x − xj + z, v − vj ) ∈ Gδ , ε ≤ |v − vj | ≤ δ −1/2 ,
√
ξj6 = ? (x − xj + z, v − vj ) ∈ G, (x − xj , v − vj ) ∈ / G, ε ≤ |v − vj | ≤ δ −1/2 ,
√
ξj7 = ? (x − xj + z, v − vj ) ∈ / G, (x − xj , v − vj ) ∈ G, ε ≤ |v − vj | ≤ δ −1/2 ,

ξj8 = ? δ −1/2 or ε −1/2 or Ĉ5 δε −1/2 ≤ |v − vj | ,
√
ξj9 = ? |v − vj | ≤ ε .
Step 2. By Lemmas 6.2 and 6.3, (6.8) and the Lipschitzness of J we learn that if
(x, v), (x + z, v) ∈ G, (x, v), (x + z, v) ∈ Gδ ∪ ;(C5 ε/δ), |v| ≤ Ĉ5 δε −1/2 ,
then

ε |z|
|fε (x, y, v, w) − fε (x + z, y, v, w)| ≤ c0 + ,
δ δ
for some constant c0 . This implies

ε |z|
21 ≤ c1 + , (6.17)
δ δ

358 F. Rezakhanlou
Step 3. Note that by (5.15), the function fε is uniformly bounded. By Lemma 4.9,
29 ≤ c2 h(ε)(1 + ε d @ε (q)) (6.18)
for some constant c2 . Also, from ?(|v − vj | ≥ 1) ≤ |v − vj |/1 ≤ (|v| + 1 + |vj |2 )/1
we deduce
√ √
28 ≤ c3 ( δ + ε δ −1 )(|v| + 1 + K(q)) , (6.19)

It is not hard to show
√
|Bε (;(C5 ε/δ))| ≤ c4 ε/δ ,
for some constant c4 . We then use (5.15) and Lemma 4.8 to deduce
√
22 + 23 ≤ c5 h( ε/δ)(1 + ε d @ˆ ε (q)) . (6.20)
Step 4. Define Rδ,ε = {x : (x, v) ∈ Kδ,ε for some v}, where

√
Kδ,ε = (x, v) ∈ Gδ : |t + (x, v)| or |t − (x, v)| ≤ C6 , and ε ≤ |v| ≤ δ −1/2 .
We certainly have
B̂ε Rδ,ε ⊆ R 0 ∪ Rδ,ε

1
, (6.21)
where R 0 is the set of points x + a such that (x, v) ∈ G, |t + (x, v)| or |t − (x, v)| ≤ C6 ,
|v| ≤ δ −1/2 , but (x + a, v) ∈
/ G for some a ∈ ε[0, 1]d , and
Rδ,ε
1
= {x + a : (x, v) ∈ Kδ,ε and (x + a, v) ∈ G for some (a, v) ∈ ε[0, 1]d × Rd } .
It is not hard to show that
|R 0 | ≤ c6 εδ (1−d)/2 . (6.22)
Also, we can find a constant c7 such that if (x, v) ∈ Kδ,ε , then |x| ≤ c7 δ −1/2 . We use
1 with (x, v) ∈ K , (x + a, v) ∈ G, a ∈ ε[0, 1]d ,
this and (6.9) to show that if x ∈ Rδ,ε δ,ε
then
√
|v · n± (x + a, v) − v · n± (x, v)| ≤ c7 ε δ −3/4 .
This implies that

√
|v · n± (x + a, v)| ≤ δ + c7 ε δ −3/4 .
Moreover, from |x + vt ± (x, v)| = σ we deduce that if η = |x + a + vt ± (x, v)|,

then |η − σ | ≤ c8 ε. We can now repeat the proof of Lemma 4.6 and use |t + (x, v)| or
|t − (x, v)| ≤ C6 to deduce that |t + (x, v)| or |t − (x, v)| ≤ c9 for some constant c9 . In
summary, if z ∈ Rδ,ε1 , then
√
|v · n± (z, v)| ≤ δ + c7 ε δ −3/4 , |t + (z, t)| or |t − (z, v)| ≤ c9 . (6.23)
Recall that dz = |v · n|dvdt if z = σ n − vt. This and (6.23) imply that |Rδ,ε
1 | ≤
√ −3/4
c10 (δ + ε δ ). From this, (6.22) and (6.21) we deduce
√
|B̂ε (Rδ )| ≤ c11 (δ + ε δ −3/4 + εδ (1−d)/2 ) ,
for some constant c11 . This and Lemma 4.9 imply

√
24 + 25 ≤ c12 h δ + ε δ −3/4 + εδ (1−d)/2 (1 + ε d @ε (q)) . (6.24)
Moreover, if Aδ (z) is the set of points (x, v) ∈ G such that (x + z, v) ∈/ G for some
v with |t − (x, v)| or |t + (x, v)| ≤ C6 and |v| ≤ δ −1/2 , then we can show that for some
constant c13 ,
|B̂ε Aδ (z)| ≤ c13 |z|δ (1−d)/2 ,
in just the same way we established (6.22). This and Lemma 4.9 imply

27 + 28 ≤ c13 h |z|δ (1−d)/2 (1 + ε d @ε (q)) . (6.25)
Final Step. Now (6.16) follows from (6.17–19), (6.21), (6.24–25) if we choose δ =
(|z| + ε)α with α any positive number strictly less than min(1/2, 2/(d − 1)).

Proof of Theorem 6.1. Step 1. We only prove (6.1) because the proof of (6.2) is similar.
Let T0 ∈ [0, T ]. Using formula (5.5), we write
T0 T0
X(q) := Eεq R(q(t))dt := Eεq [F (q(t); J, z) − F (q(t); J, 0)]dt
0 0
= 2̂1 + 2̂2 + 2̂3 + 2̂4 ,
where 2̂r = 2r (z, J ) − 2r (0, J ) for r = 1, 2, 3, 4 and 2r (z, J ) are defined by

(5.6),. . .,(5.9). We first consider 2̂1 ,

2̂1 = εd gε (xi , xi , vi ; q),
i
where

gε (x, y, v; q) = εd (fε (x − xj , y, v − vj , v) − fε (x + z − xj , y, v − vj , v)) .
j
By Lemma 6.5,
ˆ ε (q) + ε d @ε (q) + ε d K(q) + |v|) ,

|gε (x, y, v; q)| ≤ c0 h(|z| + ε)(1 + ε d @
for some constant c0 . Hence,

ˆ ε (q) + ε d @ε (q) + ε d K(q)) ,
|2̂1 | ≤ c1 h(|z| + ε)(1 + ε d @ (6.26)
for some constant c1 . Here we are using |v| ≤ 1 + |v|2 .

360 F. Rezakhanlou
Step 2. We may write 2̂2 = 2̂21 − 2̂22 where

1 q T0 d σ,ε
2̂21 = Eε ε V (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))
2 0 i,k

· εd fε (xi (s) − xj (s) + z, xi (s), vik (s) − vj (s), vik (s); T0 − s, J )
j

− fε (xi (s) − xj (s), xi (s), vik (s) − vj (s), vik (s); T0 − s, J ) ds,
T0
1
2̂22 = Eεq V σ,ε (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))
2 0
i,k
d
·ε fε (xi (s) − xj (s) + z, xi (s), vi (s) − vj (s), vi (s); T0 − s, J )
j

− fε (xi (s) − xj (s), xi (s), vi (s) − vj (s), vi (s); T0 − s, J ) ds .
Note that by the conservation of the energy, K̂ε = εd K(q(s)) is independent of s. From
applying Lemma 6.5 we obtain
T
|2̂2 | ≤ c2 h(|z| + ε)Eεq ˆ ε (q(s)) + ε d @ε (q(s)) + K̂ε
Aε (q(s)) 1 + ε d @ ds
0
T
+ c2 h(|z| + ε)Eεq Âε (q(s)) ds .
0
(6.27)
The term |2̂3 | is treated likewise.
Step 3. As we saw in the previous section,

T
|2̂4 | ≤ c3 ε d Eεq Aε (q(s))ds .
0
(See Step 3 of the proof of Theorem 5.1.) From this, (6.26) and (6.27) we deduce
ˆ ε (q) + ε d @ε (q) + K̂ε )

|X(q)| ≤ c2 h(|z| + ε)(1 + ε d @
T
+ c4 h(|z| + ε)Eεq Aε (q(s)) 1 + ε d @ˆ ε (q(s)) + ε d @ε (q(s)) + K̂ε ds
0
T
+ c4 h(|z| + ε)Eεq Âε (q(s)) ds .
0
(6.28)
We certainly have
T 2 T T −s
Eε R(q(s))ds = 2Eε R(q(s)) Eεq(s) R(q(t))dt ds
0 0 0
T T −s
≤ 2Eε |R(q(s))| Eεq(s) R(q(t))dt ds . (6.29)
0 0
From (6.28) we deduce that the left-hand side of (6.29) is bounded above by
T
2c4 h(|z| + ε)Eε |R(q(s))| 1 + ε d @ ˆ ε (q(s)) + ε d @ε (q(s)) + K̂ε
0
T T −s
+ 2c4 h(|z| + ε)Eε Aε (q(t)) + Âε (q(t))
|R(q(s))| Eεq(s)
0 0

d ˆε d ε d
· 1 + ε @ (q(t)) + ε @ (q(t)) + ε K(q(t)) dt ds
T T
ε
≤ c5 h(|z| + ε)Eε |R(q(s))|ds · T + A(q(t)) + Â (q(t))dt
0 0

· sup 1 + ε d @ ˆ ε (q(t)) + ε d @ε (q(t)) + K̂ε
0≤t≤T
≤ c6 h(|z| + ε) ,
where for the last inequality, we used Hölder’s inequality, our conditions on the initial
distribution, and Corollaries 4.6, 5.3 and 5.5 for the last inequality. This completes the
proof of (6.1).
7. Uniform Integrability
Define
t
Xt (x, v) = εd V σ,ε (|xi (s) − xj (s)|)B(vi (s) − vj (s), −nij (s))
0 i,j
ζ ε (xi (s) − x − svi (s), vi (s) − v) ds , (7.1)

where ζ : Rd ×Rd → [0, ∞) is a smooth function of compact support with ζ dxdv =
√ √
1 and ζ ε (x, v) = ε−d ζ (x/ ε, v/ ε). Put
τ (x, v) = τ1 (x, v) := inf{t : Xt (x, v) ≥ 1} .

sup Eε ψ(XT (x, v))dxdv ≤ C8 (T ), (7.2)
ε>0
where

α(log log α)1/2 α ≥ ee
ψ(α) = .
ee α < ee
The next lemma holds the key to verifying (7.2).

Lemma 7.2. There exists a constant C9 (T ) such that

Eε XT (x, v) − XT ∧τ1 (x,v) (x, v) dxdv

≤ C9 (T ) (log log 1)−1 + ε d log 1 + | log ε|−1 . (7.3)
362 F. Rezakhanlou
We prove Lemma 7.2 below and omit the rest since the remainder of the proof is
essentially identical to the proof of Theorem 7.1 in [R1].
Proof of Lemma 7.2. Step 1. Suppose that the support of V̂ is contained in the interval
(−c0 , c0 ). We take Ŵ : R → R to be a function such that Ŵ = −V̂ , Ŵ (z) = 0 for
z > εc0 and Ŵ (z) = 1 for z < −εc0 . Define W σ,ε (z) = εd Ŵ (ε −1 (z − σ )) and
+
F ε (q, t) = ε d W σ,ε (|xi − xj |) (xi − xj ) · (vj − vi ) H (xi , vi , t) , (7.4)
i,j
where
H (x, v, t) = ? (τ1 (x − vt, v) < T , |v| ≤ 11 ) ,
for a given constant 11 . The randomness of τ1 (x, v) and the nonsmoothness of the in-
dicator function does not allow us to apply (5.3). Because of this we introduce two
approximation procedures. Let δ be a small positive number such that k0 = δ −1 is an
integer. We then divide the interval [0, T ) into smaller subintervals [λk , λk+1 ) of length
δT and put
g(a, v; k, δ) = ? (τ1 (a, v) ∈ [λk , λk+1 ) , |v| ≤ 11 ) ,

g̃(a, v, t) = ? (τ1 (a, v) < t < T , |v| ≤ 11 ) ,
g(a, v) = ? (τ1 (a, v) < T , |v| ≤ 11 ) .
Using this we define

ĝε (a, v; k, δ) = g(a − a , v − w; k, δ)ζ ε (a , w)da dw,

gε (a, v) = g(a − a , v − w)ζ ε (a , w)da dw,

g̃ε (a, v, t) = g̃(a − a , v − w, t)ζ ε (a , w)da dw .
We then define
Hkδ (x, v, t) = ĝε (x − vt, v; k, δ),
Hε (x, v, t) = gε (x − vt, v), (7.5)
Kε (x, v, t) = g̃ε (x − vt, v, t) .
We replace H in (7.4) with Hkδ . The resulting expression is denoted by Fkδ (q, t). We
finally put
k
0 −2
Gδ (q, t) = Fkδ (q, t)?(t ≥ λk+1 ) . (7.6)
k=0
To understand the motivation behind (7.6), observe that

0 −2
k
lim ?(τ1 ∈ [λk , λk+1 ), t ∈ [λk+1 , T )) = ?(τ1 < T , t ∈ (τ1 , T )) . (7.7)
δ→0
k=0
Also note that the time intergration in the expression XT (x, v) − XT ∧τ1 (x,v) (x, v) is
over the interval (T ∧ τ1 , T ). To be able to use the Markov property of the process
q(·), we would rather integrate over (λk+1 , T ) whenever τ1 (x, v) ∈ [λk , λk+1 ), and sum
over k. Because of (7.7), the resulting error is small. If Ft is the σ –field generated by
(q(s) : s ≤ t), then the event τ1 ∈ [λk .λk+1 ) is measurable with respect to the σ -field
Fλk+1 . Because of this, we can use the semigroup property (5.3) to have

T ∂Fkδ
Eε Fkδ (q(T ), T ) = Eε Fkδ (q(λk+1 ), λk+1 ) + Eε + Aε Fkδ (q(s), s)ds .
λk+1 ∂t
As a result,
k
0 −2
Eε Gδ (q(T ), T ) = Eε Fkδ (q(λk+1 ), λk+1 )
k=0

0 −2 T
k
∂Fkδ
+Eε + A0 Fkδ (q(s), s)ds
λk+1 ∂t
k=0
0 −2 T
k
+Eε (Ac Fkδ )(q(s), s)ds
k=0 λk+1
k
0 −2 k
0 −2
=: 2δ1 + 2δ2,k + 2δ3,k . (7.8)
k=0 k=0
Note that the function Fkδ is not differentiable and we can not apply (5.3) directly. But
now the term [(xi − xj ) · (vi − vj )]+ is the only nonsmooth component and after re-
placing this with a smooth approxiamtion, applying (5.3) and passing to the limit, we
d +
obtain (7.8) where ?(z ≥ 0) is used for dz z .
Step 2. A straightforward calculation yields,
∂Fkδ
+ A0 Fkδ
∂t

= εd V σ,ε (|xi − xj |)[nij · (vj − vi )][(xi − xj ) · (vj − vi )]+ Hkδ (xi , vi , t)
i,j

d
+ε W σ,ε (|xi − xj |)[(vi − vj ) · (vj − vi )] ? (xi − xj ) · (vj − vi ) ≥ 0
i,j
·Hkδ (xi , vi , t)

+ε d W σ,ε (|xi − xj |)[(xi − xj ) · (vj − vi )]+
i,j

∂ ∂
· + vi · Hkδ (xi , vi , t)
∂t ∂xi
=: X1 + X2 + X3 .
364 F. Rezakhanlou
We certainly have X3 = 0. Hence
k
0 −2
2δ2,k = X̂1 (δ) + X̂2 (δ),
k=0
where
0 −2 T
k 2
X̂1 (δ) = Eε εd V σ,ε (|xi (s) − xj (s)|) [nij (s) · (vj (s) − vi (s))]+
k=0 λk+1 i,j
|xi (s) − xj (s)|Hkδ (xi (s), vi (s), s)ds , (7.9)
0 −2 T
k
X̂2 (δ) = −Eε εd W σ,ε (|xi (s) − xj (s)|)|vi (s) − vj (s)|2
k=0 λk+1 i,j
Hkδ (xi (s), vi (s), s)ds . (7.10)
Since Fkδ ≥ 0, we deduce
X̂1 (δ) ≤ Eε Gδ (q(T ), T ) − X̂2 (δ) − X̂3 (δ) , (7.11)
for
k
0 −2
X̂3 (δ) = Eε 2δ3,k .
k=0
Note that
k
0 −2
?(τ1 ∈ [λk , λk+1 ), t ∈ [λk+1 , T )) ≤ ?(τ1 < T , t ∈ [0, T )) . (7.12)
k=0
From this, (7.7) and the dominated convergence theorem we deduce

T 2
lim X̂1 (δ) = Eε εd V σ,ε (|xi (s) − xj (s)|) [nij (s) · (vj (s) − vi (s))]+
δ→0 0 i,j
|xi (s) − xj (s)|Kε (xi (s), vi (s), s)ds . (7.13)

Step 3. We now turn to X̂2 (δ). Recall K(q) = i |vi |2 and K̂ε = εd K(q(s)). We have
0 −2 T
k
d
−X̂2 (δ) ≤ c1 Eε ε εd |vi (s) − vj (s)|2 Hkδ (xi (s), vi (s), s)ds
k=0 λk+1 i,j
k0 −2 T
≤ c2 Eε ε d εd Hkδ (xi (s), vi (s), s)(N 11 + K(q(s)))ds ,
k=0 λk+1 i,j
because |vi − vj |2 ≤ 2|vi |2 + 2|vj |2 and if Hkδ (x, v, t) = 0, then |v| ≤ 211 for small δ
and large 11 . This and (7.12) imply
T
−X̂2 (δ) ≤ c2 Eε ε 2d Hε (xi (s), xj (s), s, vi (s), vj (s))(N 11 + K(q(s))ds
0 i,j

≤ c3 Eε sup ε d Hε (xi (s), vi (s), s)(11 + K̂ε ) . (7.14)
0≤s≤T i
Observe

Hε (x, v, t) = g(x − vt − a , v − w)ζ ε (a , w)da dw

= g(x − (v − w)t − a , v − w)ζ ε (a + wt, w)da dw

= H (x − a , v − w, t)ζ ε (a + wt, w)da dw

=ζ̄ H (x − a , v − w, t)ηε (a , w)da dw,

where ζ̄ = ζ (a + wt, w)dadw and ηε (a, w) = ζ ε (a + wt, w)/ζ̄ . We now apply
Lemma 4.7 and (7.14) to assert

−X̂2 (δ) ≤ Eε c4 h sup H (x, v, s)dxdv sup 1 + ε d @ˆ ε (q(s)) (11 + K̂ε ) .
0≤s≤T 0≤s≤T
(7.15)
For this, we need to estimate the intergal in the argument of h. By the Chebyshev’s
inequality,

H (x, v, s)dxdv ≤ ? (τ1 (x − vs, v) < T ) dxdv = ?(τ1 (x, v) < T )dxdv

1 1 T ε
≤ XT (x, v)dxdv ≤ c4 A (q(t))dt .
1 1 0
From this and (7.15) we deduce

T
c4
−X̂2 (δ) ≤ c3 Eε h Aε (q(t))dt sup ˆ ε (q(t)) (11 + K̂ε ) . (7.16)
1 + εd @
1 0 0≤t≤T
In the same fashion we can show,

lim EGδ (q(T ), T ) ≤ c5 Eε ε 2d Hε (xi (T ), vi (T ), T )
δ→0
i,j
T
c4
≤ c6 Eε h ε
A (q(t))dt ˆ ε (q(T ) (11 + K̂ε ) .
1 + εd @
1 0
(7.17)
366 F. Rezakhanlou
Step 4. Once more we apply (7.12) to obtain,

T
lim (−X̂3 (δ)) ≤ Eε εd V σ,ε (|xi (s) − xk (s)|)B(vi (s) − vk (s), −nik (s))
δ→0 0 i,j,k
× W σ,ε (|xi (s) − xj (s)|)|xi (s) − xj (s)||vi (s) − vj (s)|
× Hε (xi (s), vi (s), s)ds
T
+ Eε εd V σ,ε (|xi (s) − xj (s)|)B(vi (s) − vj (s), −nij (s))
0 i,j
× W σ,ε (|xi (s) − xj (s)|)|xi (s) − xj (s)||vi (s) − vj (s)|
× Hε (xi (s), vi (s), s)ds
=: X31 + X32 .
We have
T
c4 T ε
ε
X31 ≤ c7 Eε A (q(s))ds h A (q(t))dt
0 1 0

× sup 1 + ε d @ ˆ ε (q(t)) (11 + K̂ε ) ,
0≤t≤T
in just the same way we obtained (7.14) and (7.15). Since W ε = O(ε d ), we have
T
d
X32 ≤ c8 ε Aε (q(s))ds .
0
Hence
T
ε c4 T ε
lim (−X̂3 (δ)) ≤ c7 Eε A (q(s))ds h A (q(t))dt
δ→0 1 0
0
T
· sup ˆ ε (q(t)) (11 + K̂ε ) + c8 ε d
1 + εd @ Aε (q(s))ds .
0≤t≤T 0
From this, Theorem 5.1, (7.11), (7.16) and (7.17) we deduce

T T
ε c4 ε
lim X̂1 (δ) ≤ c9 Eε A (q(s))ds + 1 h A (q(t))dt
δ→0 0 1 0

· sup ˆ ε (q(s)) (11 + K̂ε ) + c9 ε d
1 + εd @
0≤s≤T
 1
T 4 41  4 4
≤ c9 Eε Aε (q(s))ds + 1 ˆ ε (q(s))
Eε 1 + ε d sup @
0  0≤s≤T 
T 41 1
1
Aε (q(s))ds + c9 ε d
4
· E ε h4 Eε (11 + K̂ε )4
1 0
1 1 1 1
=: Y14 Y24 Y34 Y44 + c9 ε d . (7.18)
By Corollary 4.4 and Theorem 5.1 we have Y1 ≤ c10 , Y2 ≤ c10 , Y4 ≤ c10 11 for some
constant c10 . By the concavity of h4 and the Jensen’s inequality,
41
1 c4 T
ε c11
Y3 ≤ h
4 4
Eε A (q(s))ds ≤h ≤ c12 (log 1)−1 .
1 0 1
This, (7.13) and (7.18) imply
T
2
Eε εd V σ,ε (|xi (s) − xj (s)|) [nij (s) · (vi (s) − vj (s))]+
0 i,j
|xi (s) − xj (s)|Kε (xi (s), vi (s), s)ds ≤ c13 (log 1)−1 + c9 ε d .
Hence
T 2
Eε εd V σ,ε (|xi (s) − xj (s)|) [nij (s) · (vi (s) − vj (s))]+
0 i,j
Kε (xi (s), vi (s), s)ds ≤ c14 11 (log 1)−1 + c9 ε d . (7.19)

Step 5. By Corollary 5.4,
T
Eε εd V σ,ε (|xi (s) − xj (s)|)B(vi (s) − vj (s), −nij (s))
0 i,j

×? |nij (s) · (vi (s) − vj (s))| < δ Kε (xi (s), vi (s), s)ds
≤ c15 | log(δ + ε)|−1 .
From this with δ = (log 1)−1/2 and (7.19) we deduce
T
Eε εd V σ,ε (|xi (s) − xj (s)|)B(vi (s) − vj (s), −nij (s))Kε (xi (s), vi (s), s)ds
0 i,j

≤ c16 (log log 1)−1 + c16 ε d log 1 + c16 | log ε|−1 + c16 11 / log 1 , (7.20)
because
2
(v · n)− ≤ δ −1 (v · n)− + (v · n)− ?(|v · n| ≤ δ) .
The left–hand side of (7.20) equals to
T
0 i,j
× ? (τ1 (xi (s) − vi (s)s − x, vi (s) − v) < s, |vi (s) − v| ≤ 11 ) ζ ε (x, v)dsdxdv .
After a change of variables (x, v) → (x − xi + vi s, v − vi ), we have that the left-hand
side of (7.20) equals to
T
0 i,j
? (τ1 (x, v) < s, |vi (s) − v| ≤ 11 ) ζ ε (xi (s) − vi (s)s − x, vi (s) − v)dsdxdv .
368 F. Rezakhanlou
This implies

Eε XT (x, v) − XT ∧τ1 (x,v) (x, v) ?(|v| ≤ 11 )dxdv

≤ c17 (log log 1)−1 + ε d log 1 + | log ε|−1 + 11 / log 1 . (7.21)
Equation (7.3) now follows from this and Corollary 5.5 by choosing 11 = log log 1
because
T
2
Eε XT (x, v) − XT ∧τ1 (x,v) (x, v) ?(|v| ≥ 11 )dxdv ≤ Eε Âε (q(s)) ds .
11 0

8. The Kinetic Limit

Let ζ be a nonnegative smooth function of compact support such that ζ dxdv = 1,
√ √
and define ζ ε (x, v) = ε−d ζ (x/ ε, v/ ε). Define a process F ε to be
N

F ε (x, v, t; q) = F ε (x, v, t) := ε d ζ ε (xi (t) − x, vi (t) − v) .
i=1
The map q → F ε (·; q) induces a probability measure on the Skorohod space D([0, T ];
L1 (Rd × Rd ), which we refer to as Pε . We now restate Theorem 2.1 as follows:
Theorem 8.1. The sequence Pε converges to P, where P is concentrated upon the single
function f that solves (2.7).
The proof of Theorem 8.1 is omitted because it is a straightforward consequence of

Lemmas 8.2–8.4 below. (See Theorem 6.1 of [RT].) The proofs of Lemmas 8.3 and 8.4
are very similar to the proof of Lemmas 6.2 and 6.4 of [RT]. The proof of Lemma 8.3 is
only sketched.
Lemma 8.2. Suppose J is a smooth function of compact support. Then

∞
∂J ∂J
lim lim sup +v· f + J Q f ∗ ζ δ , f ∗ ζ δ dxdt
δ→0 ε→0 0 ∂t ∂x

+ J (x, v, 0)f 0 (x, v)dxdv Pε (df ) = 0. (8.1)
Lemma 8.3. The family Pε is precompact.
Lemma 8.4. Let P be any limit point of Pε . Then any P is concentrated on the set of f
for which
T
sup ψ(f ∗ ζ d (x, v, t)f ∗ ζ δ (x + σ n, v∗ , t))
δ>0 0 S Rd
× B(v − v∗ , n)dx dv dv∗ dn dt < ∞.
Proof of Lemma 8.2. Put

N

G1 (t, q) = ε d J (xi , vi , t) ,
i=1
G2 (t, q) = F ε (x, v, t)J (x, v, t)dxdv .
Since J is smooth, we clearly have

√
G2 (t, q) = G1 (t, q) + O( ε) . (8.2)
It is well known that the processes

s
∂G1 ε
M(s) = G1 (s, q(s)) − G1 (0, q(0)) − + A G1 (t, q(t))dt ,
s 0 ∂t
N(s) = (M(s)) −2
(A G1 − 2G1 Aε G1 )(t, q(t))dt
ε 2
0
are Martingales. Choose T large enough so that J (x, v, t) = 0 for t ≥ T . It is straight-

forward to show that A0 G21 − 2G1 A0 G1 = 0 and

|Ac G21 − 2G1 Ac G1 | ≤ c0 ε 2d V σ,ε (|xi − xj |)B(vi − vj , −nij ) , (8.3)
i,j
for some constant c0 . Using this, EN (T ) = 0, Doob’s inequality and Theorem 5.1 we
deduce,
T
d
Eε sup (M(s)) ≤ 4Eε (M(T )) ≤ c0 ε Eε
2 2
Aε (q(t))dt ≤ c1 ε d . (8.4)
0≤s≤T 0
It is not hard to show

T
∂G1
G1 (0, q(0)) + + A0 G1 (t, q(t))
0 ∂t
∞
∂J ∂J √
= J (x, v, 0)F ε (x, v, 0)dxdv + +v· F ε dxdt + O( ε) .
0 ∂t ∂x
(8.5)
Note that
T
Ac G1 (q(t))dt = X̂ε − Xε ,
0
where
T
Xε = εd V σ,ε (|xi (t) − xj (t)|)B(vi (t) − vj (t), −nij (t))J (xi (t), vi (t), t)dt,
0 i,j
T j
X̂ε = εd V σ,ε (|xi (t) − xj (t)|)B(vi (t) − vj (t), −nij (t))J (xi (t), vi (t), t)dt .
0 i,j
370 F. Rezakhanlou
On the other hand, Theorem 6.1 and the Lipschitzness of B and J imply

Eε |Xε − Yε | + |X̂ε − Ŷε | ≤ c2 h(δ + ε) , (8.6)
where
T
Yε = εd V σ,ε (|xi (t) − z1 − xj (t) + z2 |)
0 i,j
× B(vi (t) − vj (t) + w1 − w2 , −nij (t, z2 − z1 ))J (xi (t) − z1 , vi (t) − w1 , t)
× ζ δ (z1 , w1 )ζ δ (z2 , w2 )dz1 dw1 dz2 dw2 dt ,
T
Ŷε = εd V σ,ε (|xi (t) − z1 − xj (t) + z2 |)
0 i,j
j
× B(vi (t) − vj (t) + w1 − w2 , −nij (t, z2 − z1 ))J (xi (t) − z1 , vi (t) − w1 , t)
× ζ δ (z1 , w1 )ζ δ (z2 , w2 )dz1 dw1 dz2 dw2 dt .
After a change of variables we deduce,
T
z1 − z2
Yε = εd V σ,ε (|z1 − z2 |)J (z1 , w1 , t)B w1 − w2 , −
0 |z1 − z2 |
i,j
× ζ δ (xi (t) − z1 , vi (t) − w1 )ζ δ (xj (t) − z2 , vj (t) − w2 )dz1 dz2 dw1 dw2 dt
T
z1 − z2
= ε−d V σ,ε (|z1 − z2 |)J (z1 , w1 , t)B w1 − w2 , −
0 |z1 − z2 |
× (F ε ∗ ζ δ )(z1 , w1 , t)(F ε ∗ ζ δ )(z2 , w2 , t)dz1 dz2 dw1 dw2 dt .
If we replace (F ε ∗ζ δ )(z2 , w2 , t) with (F ε ∗ζ δ )(z1 +σ n, w2 , t) with V σ,ε (|z1 −z2 |) = 0

and n = |zz11 −z
−z2
2|
, then we cause an error of order εδ −d−1/2 . Hence
T
Yε = ε −d V σ,ε (|z1 − z2 |)J (z1 , w1 , t)B(w1 − w2 , −n)(F ε ∗ ζ δ )(z1 , w1 , t)
0
× (F ε ∗ ζ δ )(z1 + nσ, w2 , t)dz1 dz2 dw1 dw2 dt + O(εδ −d−1/2 ) .
In the same fashion we can show
T
Ŷε = ε −d V σ,ε (|z1 − z2 |)J (z1 , w1 , t)B(w1 − w2 , −n)
0
× (F ε ∗ ζ δ )(z1 , w1 , t)(F ε ∗ ζ δ )
× (z1 + nσ, w2 , t)dz1 dz2 dw1 dw2 dt + O(εδ −d−1/2 ) ,
where w1 = w1 − (n.(w1 − w2 ))n. The result follows after making a change of variables
z2 − z1 → z2 and writing the z2 –integration in polar coordinates.
Proof of Lemma 8.3. Put Ẑ = supε Z ε . Let M̂(Rd × Rd ) denote the space of nonneg-
ative measures µ(dx, dv) such that µ(Rd × R d d d
) ≤ Ẑ. Let L̂ (R × R ) be the space
1
of measures µ(dx, dv) = f (x, v)dxdv with f (x, v)dxdv ≤ Ẑ. Note that the space
M̂(Rd ×Rd ) is a complete separable metric space with respect to the weak topology. We
define D = D([0, T ]; M̂(Rd × Rd )) which is also a complete separable metric space.
We regard Pε as a family of probability measures on D. Since D is a Polish space, we
can appeal to the Prohorov’s theorem to assert that the family Pε is relatively compact
if it is tight.
Observe that Lemma 4.4 implies,

sup Eε sup ϕ(F ε (x, v, t))dxdv < ∞.
ε>0 0≤t≤T
Since the space of functions f with

sup ϕ(f (x, v, t))dxdv ≤ 1
0≤t≤T
is weakly closed in D for every 1, we deduce that any limit point of Pε is concentrated
on the space D([0, T ], L̂1 (Rd × Rd )).
As in Lemma 8.2, one can readily show that for every smooth function J of compact
support,

lim sup J (x, v)(f (x, v, t) − f (x, v, s))dxdv
ε→0 0≤t,s≤T
t t
∂J
− v· (x, v)f (x, v, θ )dθ + Y ε (θ )dθ Pε (df ) = 0, (8.7)
s ∂x s
where by Corollary 5.3 we have

τ +δ
lim sup Eε YL (θ )dθ ≤ Ĉ1 (δ),
ε→0 0≤τ ≤T τ
where τ ranges over all stopping times taking values inthe interval [0, T ]. By the Aldous’
t
theorem [Al] we deduce the tightness of the process 0 YL (θ )dθ . This and (8.7) imply
the tightness of the family Pε .
Proof of Lemma 8.4. Define

Q−
ε (x, v, t) = ε
d
V σ,ε (|xi (t) − xj (t)|)B(vi (t) − vj (t), nij (t))
i,j
ε
×ζ (xi (s) − x, vi (s) − v) ,

Q+
ε (x, v, t) = εd V σ,ε (|xi (t) − xj (t)|)B(vi (t) − vj (t), nij (t))
i,j
ε j
×ζ (xi (s) − x, vi (s) − v) ,
Qε = Q + −
ε − Qε .
Let H : Rd × Rd → R, η : [0, ∞) → R be two smooth functions of compact support

and put J (x, v, t) = H (x − vt, v)η(t). As in Lemma 8.2, it is not hard to show
∞
∂J ∂J
lim Eε +v· F ε + J Qε dx dv dt
ε→0 0 ∂t ∂x

+ F ε0 (x, v)J (x, v, 0)dxdv = 0.
372 F. Rezakhanlou
We then make a change of variable (x, v, t) → (x + vt, v, t) and use Qε < Q+ ε to

deduce that if both H and η are nonnegative, then
∞
lim Eε H (x, v) (f (x + vt, v, t)η (t) + Q+
ε (x + vt, v, t)η(t))dt
ε→0 0
−
+f 0 (x, v)η(0) dx =0. (8.8)
Here {a}− = − min{a, 0}. Put

t
; ε (x, v, t) = Q+
ε (x + vs, v, s)ds .
0
To take advantage of Theorem 7.1, we extend Pε (df ) to a probability measure P̃ε (df, d;)
that is the distribution of the pair
(F ε , ; ε ) ∈ D([0, T ]; L̂1 (Rd × Rd )) × L1 (Rd × Rd )

with respect to Pε . Suppose η is nonincreasing and that the support of η is contained in
[0, T ]. Then
∞ ∞
+
Qε (x + vt, v, t)η(t)dt = − ; ε (x, v, t)η (t)dt
0 0 ∞
≤ −; ε (x, v, T ) η (t)dt = ; ε (x, v, T )η(0) .
0
Hence (8.7) implies

∞
lim H (x, v) f (x + vt, v, t)η (t)dt + ;(x, v, T )η(0)
ε→0 0
−
+f 0 (x, v)η(0) dx dv P̃ε (df, d;) = 0, (8.9)
provided H ≥ 0, η ≤ 0 and the support of η is contained in [0, T ]. Theorem 7.1 says

sup ψ (;(x, v, T )) P̃ε (df, d;) ≤ C8 (T ) . (8.10)
ε>0
Let us denote the integrand of (8.9) by Y(f, ;). The functional Y(f, ;) is a continuous
functional of the pair (f, ;). Although Y is not bounded, by (8.10) we can approximate
Y(f, ;) by a sequence of bounded continuous functionals. Therefore we can pass to the
limit ε → 0. If P̃ is any limit point of P̃ε , then we have

Y(f, ;)P̃(df, d;) = 0 .
This means
∞

H (x, v) f (x + vt, v, t)η (t)dt + ;(x, v, T )η(0) + f (x, v)η(0) dxdv ≥ 0
0
0
(8.11)
with probability one with respect to P̃. Since the space of nonnegative smooth functions
is separable, we may take a countable dense set of H in (8.11) to obtain

f (x + vt, v)η (t)dt + ;(x, v, T )η(0) + f 0 (x, v)η(0) ≥ 0 ,
for almost all x, and with probability one with respect to P̃. By taking another count-
able dense set of smooth η of compact support in [0, T ) with η ≥ 0, η ≤ 0, we can
approximate functions of the form ?[0,s] (t) for s ∈ [0, T ], to deduce
ˆ
f (x + vt, v, t) ≤ f 0 (x, v) + ;(x, v, T ) =: ;(x, v) (8.12)
for almost all (x, v, t) and with probability one with respect to P̃.
On the other hand, since ψ is convex, the functional ; → ψ(;)dx is lower semi-
continuous with respect to the weak topology. Therefore (8.10) implies

ψ(;(x, v, T ))dxdv P̃(df, d;) < ∞ .
This in turn implies,

ˆ
ψ(;(x, v))dxdv P̃(df, d;) < ∞ . (8.13)
We have
T
ψ(f (x, v, t)f (x + σ n + z, v∗ , t))B(v − v∗ , n)dx dn dv dv∗ dt
0 S Rd
T
≤ ˆ − vt, v);(x
ψ(;(x ˆ + σ n + z − v∗ t, v∗ ))
0 S Rd
× B(v − v∗ , n)dx dn dv dv∗ dt.
Furthermore,
T
ˆ − vt, v);(x
ψ(;(x ˆ + σ n + z − v∗ t, v∗ ))B(v − v∗ , n)dx dn dv dv∗ dt
0 S Rd
T
≤ ee ˆ − vt, v);(x
?(;(x ˆ + σ n + z − v∗ t, v∗ ) ≤ ee )
0 S Rd
× B(v − v∗ , n)dx dn dv dv∗ dt
T
ˆ − vt, v);(x ˆ − vt, v)| 2
ˆ + σ n + z − v∗ t, v∗ )| log log ;(x
1
+ ;(x
0 S Rd
×B(v − v∗ , n)dx dn dv dv∗ dt
T
+ ˆ − vt, v);(x
;(x ˆ + σ n + z − v∗ t, v∗ )|
0 S Rd
ˆ + σ n + z − v∗ t, v∗ )| 2 B(v − v∗ , n)dx dn dv dv∗ dt
1
× log log ;(x
:= 21 + 22 + 23 .
374 F. Rezakhanlou
We then make a change of variable (x − vt, x + σ n + z − v∗ t) → (a, b) to find that

22 + 23 bounded above by a constant multiple of

ˆ
ψ(;(a, ˆ
v));(b, ˆ
v∗ ) + ;(a, ˆ
v)ψ(;(b, v∗ )) da db dv dv∗ .
This implies
T
sup ψ(f (x, v, t)f (x + σ n + z, v∗ , t))B(v − v∗ , n)dv dv∗ dx dn dt < ∞
z 0
(8.14)
with probability one with respect to P. Finally by the convexity of ψ and Jensen’s
inequality, we deduce
T

ψ (f ∗ ζ δ )(x, v, t)(f ∗ ζ δ )(x + σ n, v∗ , t)
0 S Rd
× B(v − v∗ , n)dx dn dv dv∗ dt
T
≤ sup ψ(f (x, v, t)f (x + σ n + z, v∗ , t))dx dn dv dv∗ dt .
z 0 S Rd
This and (8.14) complete the proof.

Acknowledgements. I wish to thank Mario Pulvirenti for introducing me to the problem and many fruitful
discussions. A part of this work was done when I was visiting Universitá di Roma uno and Universitá di
Roma due during the Spring of 1998. I would like to thank both universities for their generous support.
Special thanks to Rossana Marra and Mario Pulvirenti for their warm hospitality.
References
[Al] Aldous, D.: Stopping times and tightness. Annals of Probability 6, 335–340 (1978)
[Ar] Arkeryd, L.: On the Enskog equation with large initial data. SIAM J. Math. Anal. 21, 631–646
(1990)
[ArC] Arkeryd, L., Cercignani, C.: Global existence in L1 for the Enskog equation and convergence of
the solutions to solutions of the Boltzmann equation. J. Statist. Phys. 59 no. 3–4, 845–867 (1990)
[B] Bony, J.: Solutions globales bornées pour les modèles discrets de l’equation de Boltzmann en
dimension 1 d’espace. Actes Journées E.D.P. St. Jean de Monts, XVI (1987)
[IP] Illner, R., Pulvirenti, M.: Global validity of the Boltzmann equation for a two- and three-dimen-
sional rare gas in vacuum. Commun. Math. Phys. 105, 189–203 (1986); Erratum and improved
results, Commun. Math. Phys. 121, 143–146 (1989)
[K] King, F.: BBGKY hierarchy for positive potentials. Ph.D. Thesis, Dept of Mathematics, Univer-
sity of California at Berkeley (1975)
[L] Lanford, O.E.: Time evolution of large classical systems. In: Lecture Notes in Physics, ed.
J. Moser, Vol. 38, Berlin, Heidelberg: Springer, pp. 1–111
[P1] Pulvirenti, M.: Global validity of the Boltzmann equation for a three-dimensional rare gas in
vacuum. Commun. Math. Phys. 113, 79–85 (1987)
[P2] Pulvirenti, M.: On the Enskog hierarchy: Analyticity, uniqueness and derivability by particle
systems. Rendiconti del Circolo Matematica di Palermo, Serie II, Suppl. 45, 529–542 (1996)
[R1] Rezakhanlou, F.: Kinetic limits for a class of interacting particle systems. Probab. Theory Related
Fields 104, 97–146 (1996)
[R2] Rezakhanlou, F.: Large deviation from a kinetic limit. Annals of Probability 26, 1259–1340
(1998)
[RT] Rezakhanlou, F., Tarver III, J.L.: Boltzmann-Grad limit for a particle system in continuum. Ann.
Inst. Henri Poincaré 33, 753–796 (1997)
[RD] Resibois, P., De Leener, M.: Classical Kinetic Theory of Fluids. New York: Wiley, 1977
[T] Tartar, L.: Some existence theorems for semilinear hyperbolic systems in one space variables.
University of Wisconsin, MRS Technical Summary Report (1980)
Communicated by H. Spohn

Rezakhanlou 2003

Uploaded by

Copyright:

Available Formats

Rezakhanlou 2003

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rezakhanlou 2003

Uploaded by

Copyright:

Available Formats

Commun. Math. Phys.

232, 327–375 (2003) Communications in

A Stochastic Model Associated with Enskog Equation

Received: 29 January 2002 / Accepted: 30 July 2002

Abstract: We examine a system of particles in which the particles travel deterministic-

∗ Research supported in part by NSF Grant DMS-0072666.

expression in the brackets in (1.1) is replaced with

the Stosszahlensatz (Boltzmann’s molecular chaos principle) will be established. Sec-

2. Notation and Main Result

where V σ,ε(z) = εd−1 V̂ (ε −1 (z − σ )) with V̂ : R → [0, ∞) a continuous even function

Here and below we use the notations

µε (dq) = F ε (q)νβ (dq),

Note that if we choose J ≡ 1 in (2.5), we deduce that

with Z as in (2.6), then µε (dq) = Gε (q)dq

We assume that q(0) is distributed according to µε and we denote the expectation

where f (x, v, t) is the unique solution to

where Q = Q(f, f ) is defined by

A solution to (2.7) is understood in the following sense:

This and the conservation of the energy

From this and (3.2) we deduce that

where 2i is obtained by multiplying the i th term on the right-hand side of (3.2) by

A straightforward calculation yields

for z = x + σ n + (t − s)w. As a result 22 is bounded above by

A = {z : z = x + (v − v∗ )τ + σ n for some (n, τ ) ∈ S × [0, T ]

Note that the volume of A can be estimated as

for some constant c3 . From this, (3.4) and (3.6) we deduce

A bound on the entropy

imply that there exists a positive constant c5 such that

γ ≤ c5 [log((1 )d 1T )]−1 + c5 (1 )−2 , (3.9)

|A| ≤ c6 | log δ|−1 .

where ;(z) = z log+ log+ z and

It is more convenient to establish (3.15) for Q− = Q+ − Q;

In fact (3.8) and (3.17) imply (3.15) because (2.7) implies

A microscopic analog of (3.17) will be established in Sect. 7. This will be used in

This will be used to replace f ∗ ζ δ with f in (3.14).

?(|x − y| ≤ σ )[(x − y) · (v∗ − v)]+ H (x, y, t, v, v∗ )dx dy dv dv∗ ,

After an integration by parts,

?(|x − y| ≤ σ ) [(x − y) · (v∗ − v)]+ dx dy dv dv∗

because by Chebyshev’s inequality,

2̂2 ≤ c12 (δ log 1)−1 . (3.26)

This and (3.25) imply (3.20) if we choose δ = (log 1)−1/2 .

for every set K ⊆ Rd × Rd and define

Lemma 4.2. There exists a constant c0 such that

As a consequence of Stirling’s theorem, we have

It is not hard to see that γn ≤ c5 n2d−1 N for some constant c5 . Also

From this and (4.7) we deduce

 A1 denote the set of N̄ = (N0 , N2 , . . . ) such that each Nn is a nonnegative integer,

for some constant c10 . Hence

for some constant c12 and this completes the proof. 

Recall that initially the configuration q is distributed according to the measure

µε (dq) = F ε (q)νβ (dq) = Gε (q)dq ,

Moreover, the measure m is invariant.

This and (4.9) imply (4.8). 

Lemma 4.4. We have K(t, q(t)) ≤ K(0, q(0)) and

Z(t, q) ≤ exp(−βK(t, q))

µε (t, dq) = H (t, q)m(dq) ,

where H satisfies H (0, q) = H0 (q) and the backward equation

Lemma 4.5. There exists a constant c such that

From this, Lemma 4.4 and Lemma 4.2 we deduce

where V σ,ε(z) = εd−1 V̂ (ε −1 (z − σ )) with V̂ : R → [0, ∞) a continuous even function

γ ≤ c5 [log((1 )d 1T )]−1 + c5 (1 )−2 , (3.9)

A1 denote the set of N̄ = (N0 , N2 , . . . ) such that each Nn is a nonnegative integer,

for some constant c12 and this completes the proof.

This and (4.9) imply (4.8).

where · Lp denotes the Lp norm with respect to the set Rd × Rd .

(x, v), (x + z, v) ∈ G, (x, v), (x + z, v) ∈ Gδ ∪ ;(C5 ε/δ), |v| ≤ Ĉ5 δε −1/2 ,