FISHER INFORMATION ESTIMATES FOR

BOLTZMANN’S COLLISION OPERATOR
C. VILLANI
Abstract. We derive several estimates for Boltzmann’s collision
operator in terms of Fisher’s information. In particular, we prove
that Fisher’s information is decreasing along solutions of the Boltz-
mann equation with Maxwellian cross-section, in any dimension of
velocity space, thus generalizing results by G. Toscani, E. Carlen
and M. Carvalho.
Contents
1. Introduction 1
2. Main results 4
3. Arbitrary cross-sections 7
4. Maxwellian cross-sections 12
5. Related inequalities
and analogy between Q
+
and the rescaled convolution 17
References 20
1. Introduction
Let f be a probability density on R
N
, N ≥ 1. Fisher’s quantity of
information associated to f is defined as the (possibly infinite) nonneg-
ative number
(1) I(f) =
_
R
N
|∇f|
2
f
= 4
_
R
N
¸
¸
¸∇
_
f
¸
¸
¸
2
.
This formula defines a convex, isotropic functional I, which was first
used by Fisher [11] for statistical purposes, and plays a fundamental
role in information theory.
In 1959, Linnik [12] used this functional (therefore also called Lin-
nik’s functional) to give an information-theoretic proof of the central
limit theorem (see [1, 10] for recent improvements of Linnik’s methods).
Some years later, McKean [14], drawing an analogy between the
central limit theorem and the trend to equilibrium in kinetic theory,
1
2 C. VILLANI
adapted the work of Linnik to the kinetic theory of gases. In this way he
obtained the first explicit bound from below for the speed of approach
to equilibrium in Kac’s model, which is a one-dimensional caricature of
the Boltzmann equation (we note that the optimal bound, conjectured
by McKean, was recently derived by Carlen, Gabetta and Toscani [9],
using a completely different technique).
The key observation by McKean was that I, like the classical Boltz-
mann H-functional, is nonincreasing with time along solutions of Kac’s
model. This monotonicity property was extended by Toscani [18] to
the two-dimensional Boltzmann equation for Maxwellian molecules. It
is our purpose here to generalize this result to higher dimensions of
velocity space, and to give related estimates in a larger setting.
The fact that I is a Lyapunov functional for the Boltzmann equa-
tion with Maxwellian molecules has many applications. For instance,
Toscani [19] used it to derive strengthened limit theorems. Moreover, it
entails also a propagation of smoothness for the solution to the Boltz-
mann equation, some applications of which are given in [9].
Bobylev and Toscani on one hand, Carlen and Carvalho on the other,
noticed that the decreasing property of I can be seen as a consequence
of an inequality which is reminiscent of well-known inequalities in in-
formation theory. To understand this, let us go a little bit into the
details of the Boltzmann equation.
In the (spatially homogeneous) Boltzmann equation, the unknown is
a nonnegative integrable function f(t, v), standing for the probability
distribution at time t of the velocity v of the molecules in a gas. The
equation governing the evolution of f is
(2) ∂
t
f = Q(f, f),
where Boltzmann’s collision operator Q(f, f) is defined by
(3)
Q(f, f) =
_
dv

dσ B(v −v

, σ) (f

f


−ff

) ≡ Q
+
(f, f) −Q

(f, f),
with the usual conventions f

= f(v

), f

= f(v

), f


= f(v


), and
(4)
_
¸
¸
¸
_
¸
¸
¸
_
v

=
v + v

2
+
|v −v

|
2
σ
v


=
v + v

2

|v −v

|
2
σ.
The weight-function B : R
N
× S
N−1
→ R
+
is the so-called “cross-
section”, depending on the interaction between particles. On physical
grounds it is always assumed that B(z, σ) depends only on |z| and
FISHER INFORMATION ESTIMATES 3
(z/|z| · σ). In the case of the so-called Maxwellian molecules, i.e. re-
pelling each other with an inverse-power force law of exponent 2N −1,
B depends only upon (z/|z| · σ). More generally, we shall define a
Maxwellian cross-section as a weight-function B(z/|z| · σ).
Under very little assumptions on the initial datum f(t = 0) = f
0

L
1
(R
N
) and on the cross-section B, one can show that the equation (2)
admits a unique nonnegative solution, whose total mass is preserved
with time. Therefore, we shall always assume that f is a probability
density.
Let B be a Maxwellian cross-section. Then, by rotational invariance,
_
B(k · σ) dσ is independent of the unit vector k. Assuming its value
to be 1, Q

is simply
(5) Q

(f, f) = f
__
f
_
= f.
Therefore, due to the convexity of I, to prove that Fisher’s information
is decreasing with time along solutions of the Boltzmann equation with
Maxwellian cross-section, it is sufficient to prove that
(6) I
_
Q
+
(f, f)
_
≤ I(f).
Since Q
+
acts more or less like a (rescaled) convolution operator,
this inequality is strongly reminiscent of the well-known Blachman-
Stam inequality [2, 5, 16]. If f is a probability density, let us define
f
α
(v) = α
−N/2
f(α
−1/2
v); then, if f and g are any two probability
densities, the Blachman-Stam inequality reads
(7) I(f
α
∗ g
1−α
) ≤ αI(f) + (1 −α)I(g).
It was in fact proven by Bobylev and Toscani [4], using the Fourier-
transform representation of Boltzmann’s equation, that if the inequal-
ity (7) holds for an arbitrary functional I, then the inequality (6) also
holds in dimension 2 of velocity space (the general case is still open).
Carlen and Carvalho [7] also proved the inequality (6) for arbitrary
dimension N, but for a particular cross-section (which is constant in
another representation of (v

, v


)). Here we shall prove this inequality
in full generality, by a direct and entirely elementary computation,
relying on apparently new representations for ∇Q
+
(f, f), that have
interest on their own. In fact, one can draw a slightly better parallel
with the Blachman-Stam inequality : let f and g be two probability
densities, and define
Q
+
(f, g) =
1
4
_
Q
+
(f + g, f + g) −Q
+
(f −g, f −g)
_
.
4 C. VILLANI
=
1
2
_
R
N
×S
N−1
dv

dσ B(v −v

, σ)(f

g


+ g

f


).
Then, we shall prove that
(8) I
_
Q
+
(f, g)
_

1
2
_
I(f) + I(g)
_
.
We also investigate briefly the case of arbitrary potentials, and show
precisely why the Maxwellian case seems to depart from the others.
Some of our methods are directly inspired from [10] and [7].
To conclude this introduction, we want to emphasize the fact that
even though the decreasing property of I can apparently be shown only
for Maxwellian cross-sections, the consequences that can be drawn from
this property go far beyond this setting. This can be seen for instance
in the work by Carlen and Carvalho [7, 8]. In a forthcoming joint work
with G. Toscani, we shall show that the decreasing property of I can
be related to the properties of the Ornstein-Uhlenbeck regularization
of the entropy dissipation functional in the Boltzmann equation, and
give new estimates for the speed of trend towards equilibrium in a very
general setting.
Acknowlegement : The author thanks Eric Carlen for several fruit-
ful discussions.
2. Main results
Our first estimate concerns arbitrary cross-sections. It relies on the
following elementary identity.
Proposition 1. Let f and g be two smooth probability densities with
fast decay at infinity, and let Q be a Boltzmann collision operator asso-
ciated with a cross-section B ∈ L
1
loc
(R
N
×S
N−1
) (at most polynomially
increasing at infinity). Then
(9) ∇Q
±
(f, g) = Q
±
(f, ∇g) + Q
±
(∇f, g),
where, of course, [Q
±
(f, ∇g)]
i
= [Q
±
(∇g, f)]
i
= Q
±
(f, ∂
i
g).
Remark. Under suitable assumptions on B, formula (9) can be given
a distributional sense even if f is not as smooth as required above. We
shall however not try to do so because this proposition suffices to our
purposes.
Theorem 2. Let B(z, σ) ∈ L

(R
N
; L
1
(S
N−1
)) be an arbitrary cross-
section, and define
(10) A(z) =
_
S
N−1
dσ B(z, σ).
FISHER INFORMATION ESTIMATES 5
Let a = A
L

(R
N
)
. Then, for any two probability densities f and g,
(11) I
_
Q
±
(f, g)
_
≤ a [I(f) + I(g)] .
The noticeable point in this inequality is that it depends on A only
through L

bounds, and does not require any smoothness for the cross-
section. As a corollary, we give a simple (certainly not optimal) theo-
rem of (local in time) propagation of Fisher information bounds.
Corollary. Let B(z, σ) be an arbitrary cross-section, and let A be
defined by the formula (10). Assume that a ≡ A
L
∞ < ∞, and
b ≡ ∇A
L
∞ < ∞. Let f
0
be a probability density with finite variance
such that I(f
0
) < ∞, and let f(t, v) be the solution of the Boltzmann
equation with cross-section B and initial datum f
0
. Then, for all time
t, I(f(t)) < ∞ and there exists a constant C < ∞ depending only on
a and b such that
I(f(t)) ≤ e
4at
_
2I(f
0
) + C(1 + t
3
)
_
.
For a Maxwellian cross-section, we can impose that A(z) = 1. Then
Theorem 2 yields
I
_
Q
+
(f, g)
_
≤ I(f) + I(g).
But, due to the particular structure of the Maxwellian case, this in-
equality can be improved by a factor 2. To this purpose, we establish
the following representation.
Proposition 3. Let B(k·σ) be a Maxwellian cross-section such that for
any unit vector k,
_
dσ B(k · σ) = 1, and let f be a smooth probability
density with fast decay at infinity. Then
(12)
∇Q
+
(f, f) =
1
2
_
dv

dσ B(k·σ)
_
f


(I+P
σk
)(∇f)

+f

(I−P
σk
)(∇f)


_
where k = (v − v

)/|v − v

|, I : R
N
→ R
N
is the identity, and P
σk
:
R
N
→R
N
is the linear transformation defined by
P
σk
(x) = (k · σ)x + (σ · x)k −(k · x)σ.
A formula similar to (12) is obtained for Q
+
(f, g) by the usual dou-
bling procedure.
Theorem 4. Let B(k · σ) be a Maxwellian cross-section such that for
any unit vector k,
_
dσ B(k · σ) = 1. Then, for any probability density
f on R
N
,
(13) I
_
Q
+
(f, f)
_
≤ I(f).
6 C. VILLANI
Moreover if B is even, in the sense that B(k · σ) = B(−k · σ) for all
k, σ, then for any two probability densities f and g on R
N
,
(14) I
_
Q
+
(f, g)
_

1
2
_
I(f) + I(g)
¸
.
From this theorem follows the decrease of Fisher’s information with
time along solutions of the Boltzmann equation. Moreover, thanks to
the uniqueness theorem given in [20] by G. Toscani and the author, we
obtain the
Corollary 4.1. Let B be a Maxwellian cross-section such that for any
unit vector k,
(15)
_
S
N−1
dσ B(k · σ)(1 −k · σ) < ∞.
Let f
0
be a probability density with finite variance and such that I(f
0
) <
∞, and let t −→f(t) be the unique solution to the Boltzmann equation
with cross-section B and initial datum f
0
. Then I(f(t)) is nonincreas-
ing with time.
The next two corollaries are conveniently proved by the use of “Boby-
lev’s lemma” [3] : if f and g are probability densities and M is a
Maxwellian distribution, then
Q
+
(f ∗ M, g ∗ M) = Q
+
(f, g) ∗ M.
Corollary 4.2. Let f be a smooth positive function, rapidly decreasing
at infinity, and B a Maxwellian cross-section satisfying the assump-
tion (15). Then
(16) −
1
8
I

(f) · Q(f, f) ≡
_
R
N
Q(f, f)


f

f
≥ 0.
Finally, let us recall the definition of Boltzmann’s H-functional, or
entropy :
H(f) =
_
f(v) log f(v) dv.
By adapting the arguments given in [10], we shall obtain easily the
Corollary 4.3. Let f and g be two probability densities, and let B(k·σ)
be a Maxwellian cross-section such that for any unit vector k,
_
dσ B(k·
σ) = 1. Then
(17) H
_
Q
+
(f, g)
_

1
2
[H(f) + H(g)] .
FISHER INFORMATION ESTIMATES 7
This last inequality is clearly reminiscent of the well-known Shannon-
Stam inequality [16, 10],
H(f
α
∗ g
1−α
) ≤ αH(f) + (1 −α)H(g).
We give an application to numerical simulations. Consider the ex-
plicit Euler scheme for the Boltzmann equation with Maxwellian cross-
section
(18)
_
¸
_
¸
_
f
0
= f
0
f
n+1
−f
n
= εQ(f
n
, f
n
), 0 ≤ ε ≤ 1, n ≥ 0.
Then
H(f
n+1
) = H
_
(1 −ε)f
n
+ εQ
+
(f
n
, f
n
)
_
≤ (1 −ε)H(f
n
) + εH
_
Q
+
(f
n
, f
n
)
_
≤ H(f
n
).
Thus, the entropy of the solution (f
n
) to this scheme is nonincreasing
with the time step. This phenomenon can be clearly seen in numerical
simulations [15]. We note that the decrease of the entropy always holds
for an implicit Euler scheme of the Boltzmann equation.
The plan of the paper is as follows. Arbitrary cross-sections are con-
sidered in section 3; then we turn to the Maxwellian case in section 4.
The use of Bobylev’s regularization and its consequences are left to
section 5.
3. Arbitrary cross-sections
We first establish Proposition 1. Let f and g be smooth probability
densities with fast decay at infinity, and let B be a smooth cross-section.
Then, by standard theorems of differentiation of integrals depending
on a parameter,
∇Q
+
(f, g) =
_
dv

dσ ∇
v
[B(v −v

, σ)]
_
f

g


+ g

f


2
_
+
_
dv

dσ B(v −v

, σ)∇
v
_
f

g


+ g

f


2
_
= −
_
dv

dσ ∇
v∗
[B(v −v

, σ)]
_
f

g


+ g

f


2
_
+
_
dv

dσ B(v −v

, σ)∇
v
_
f

g


+ g

f


2
_
.
8 C. VILLANI
Integrating the first term by parts, we find
(19) ∇Q
+
(f, g) =
_
dv

dσ B(v −v

, σ)(∇+∇

)
_
f

g


+ g

f


2
_
.
Now, from the formulas (4), one easily deduces that if g is an arbi-
trary (L
1
loc
) function, then, in distributional sense (or in classical sense
if g is smooth),
(20)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
∇(g

) =
1
2
(∇g)

+
1
2
v −v

|v −v

|
σ · (∇g)



(g

) =
1
2
(∇g)


1
2
v −v

|v −v

|
σ · (∇g)

∇(g


) =
1
2
(∇g)



1
2
v −v

|v −v

|
σ · (∇g)




(g


) =
1
2
(∇g)


+
1
2
v −v

|v −v

|
σ · (∇g)


.
In particular,
(21)
_
¸
_
¸
_
(∇+∇

)(g

) = (∇g)

(∇+∇

)(g


) = (∇g)


.
As a consequence,
∇Q
+
(f, g) =
_
dv

dσ B(v −v

, σ)
_
(∇f)

g


+ f

(∇g)


+ (∇g)

f


+ g

(∇f)


2
_
= Q
+
(f, ∇g) + Q
+
(g, ∇f).
The computation for Q

is exactly the same (in fact simpler). By
density, Proposition 1 extends to nonsmooth cross-sections.
From this we derive Theorem 2. Let f and g be smooth functions
with fast decay at infinity. Using the identity (∇f)/(2

f) = ∇

f, we
can write
∇Q
+
(f, g) =
_
dv

dσ B(v−v

, σ)
_
f

g


_
_

_
f
_
′ _
g


+ (∇

g)


_
f

_
+
_
dv

dσ B(v −v

, σ)
_
g

f


_
(∇

g)

_
f


+
_

_
f
_


_
g

_
.
FISHER INFORMATION ESTIMATES 9
From now on, we omit the argument of B for simplicity. Using first
the convexity of the square norm, then the Cauchy-Schwarz inequality,
|∇Q
+
(f, g)|
2

Q
+
(f, g)
1
2
_
dv

dσ Bf

g


__
dv

dσ B
_
f

g


_
_

_
f
_
′ _
g


+ (∇

g)


_
f

__
2
+
Q
+
(f, g)
1
2
_
dv

dσ Bg

f


__
dv

dσ B
_
g

f


_
(∇

g)

_
f


+
_

_
f
_


_
g

__
2
≤ 2Q
+
(f, g)
_
dv

dσ B
¸
¸
¸
¸
_

_
f
_
′ _
g


+ (∇

g)


_
f

¸
¸
¸
¸
2
+ 2Q
+
(f, g)
_
dv

dσ B
¸
¸
¸
¸
(∇

g)

_
f


+
_

_
f
_


_
g

¸
¸
¸
¸
2
.
Dividing by Q
+
(f, g) and integrating with respect to v, we get
I
_
Q
+
(f, g)
_
=
_
dv
|∇Q
+
(f, g)|
2
Q
+
(f, g)
≤ 2
_
dv dv

dσ B
¸
¸
¸
¸
_

_
f
_
′ _
g


+ (∇

g)


_
f

¸
¸
¸
¸
2
+ 2
_
dv dv

dσ B
¸
¸
¸
¸
(∇

g)

_
f


+
_

_
f
_


_
g

¸
¸
¸
¸
2
.
By the involutive change of variables with unit Jacobian (v, v

) ←→
(v

, v


), the right-hand side of this last expression is
2
__
dv dv

dσ B
¸
¸
¸
_

_
f
_

g

+ (∇

g)

_
f
¸
¸
¸
2
+
_
dv dv

dσ B
¸
¸
¸(∇

g)
_
f

+
_

_
f
_


g
¸
¸
¸
2
_
.
By the exchange of variables v and v

, both integrals above are equal.
Since by assumption,
_
dσ B(v −v

, σ) ≤ a, we obtain, expanding the
square norms,
I
_
Q
+
(f, g)
_
≤ 4a
_
dv dv

_
|∇
_
f|
2
g

+
_
|∇

g|
2
_

f
_
+ a
_
dv dv

_
∇f · (∇g)

+∇g · (∇f)

_
.
Since
_
dv ∇f =
_
dv ∇g = 0,
10 C. VILLANI
the last integral in the previous expression vanishes. Moreover, since
_
dv f =
_
dv g = 1,
we obtain by Fubini’s theorem
(22) I
_
Q
+
(f, g)
_
≤ a (I(f) + I(g)) ,
and this is the desired result in the case when f and g are smooth and
rapidly decreasing.
In the general case, one can find sequences of smooth, rapidly de-
creasing densities (f
n
) and (g
n
) such that
f
n
L
1
−→f, g
n
L
1
−→g, I(f
n
) −→I(f), I(g
n
) −→I(g).
Then, it is easy to check that Q
+
(f
n
, g
n
) converges towards Q
+
(f, g),
at least weakly in L
1
, and by the weak lower semicontinuity of I the
conclusion follows.
Remarks.
(1) The use of the Cauchy-Schwarz inequality above can be seen
as a variant of the proof given by Carlen and Carvalho for a
particular cross-section; see [7, Lemma 3.3].
(2) In establishing Proposition 1, we have used the structure of Q
+
to report the derivatives of B onto f and g. One could wonder
whether the inverse manipulation is possible, thus obtaining
an estimate of I(Q
+
(f, g)) depending only on the smoothness
of B, like I(Q
+
(f, f)) ≤ Cf
2
L
1
. But such a result is clearly
false : if δ denotes any Dirac measure, then for a Maxwellian
cross-section, Q
+
(δ, δ) = δ.
(3) We recall that the main result in this respect is the estimate
that was first obtained by Lions [13] :
Q
+
(f, f)
H
N−1
2
≤ Cf
L
2f
L
1,
if B(z, σ) is smooth, compactly supported and subject to certain
additional technical assumptions.
Now, we deduce from Theorem 2 a result of propagation of Fisher
information bounds. Let B be a cross-section such that A, defined
by (10), belongs to W
1,∞
(R
N
). If we set
Lf = A ∗ f,
the solution of the Boltzmann equation with initial datum f
0
can be
written as
(23) f(t, v) =
_
t
0
e

R
t
s
Lf(τ,v) dτ
Q
+
(f(s, v)) ds + f
0
(v)e

R
t
0
Lf(τ,v) dτ
.
FISHER INFORMATION ESTIMATES 11
Therefore, by convexity and homogeneity of I,
(24)
I(f(t)) ≤ I
__
t
0
e

R
t
s
Lf(τ) dτ
Q
+
(f(s), f(s)) ds
_
+ I
_
f
0
e

R
t
0
Lf(τ) dτ
_

_
t
0
ds I
_
e

R
t
s
Lf(τ) dτ
Q
+
(f(s), f(s))
_
+ I
_
f
0
e

R
t
0
Lf(τ) dτ
_
.
Now, for any two nonnegative functions g and h,
I(gh) =
_
|∇(gh)|
2
gh
=
_
|g∇h + h∇g|
2
gh
≤ 2
__
g
2
|∇h|
2
gh
+
_
h
2
|∇g|
2
gh
_
= 2
__
g
|∇h|
2
h
+
_
h
|∇g|
2
g
_
≤ 2g
L
∞I(h) + 2
_
_
_
_
|∇g|
2
g
_
_
_
_
L

h
L
1.
(25)
We estimate the different terms appearing in the right-hand side
of (23). First,
(26) e

R
t
s
Lf(τ) dτ
≤ 1, I
_
Q
+
(f, f)
_
≤ 2aI(f).
Next, since

_
e

R
t
s
Lf(τ) dτ
_
= −e

R
t
s
Lf(τ) dτ
_
t
s
∇(Lf)(v, τ) dτ,
we obtain the bound
(27)
¸
¸
¸∇
_
e

R
t
s
Lf(τ) dτ

¸
¸
2
e

R
t
s
Lf(τ) dτ
= e

R
t
s
Lf(τ) dτ
__
t
s
dτ∇A ∗ f(τ)
_
2

__
t
s
dτ∇A
L
∞f
L
1
_
2
≤ (t −s)
2
b
2
.
Finally, we note that
(28)
_
dv Q
+
(f, f) =
_
dv Q

(f, f) ≤ a.
Putting together (24), (25), (26), (27) and (28), it follows that
(29) I(f(t)) ≤ 4a
_
t
0
I(f(s)) ds + 2a
_
t
0
(t −s)
2
b
2
+ 2I(f
0
) + 2t
2
b
2
.
Our estimate of I(f(t)) follows by Gronwall’s lemma.
12 C. VILLANI
4. Maxwellian cross-sections
Let B(k · σ) be a smooth Maxwellian cross-section, such that for any
unit vector k,
_
dσ B(k · σ) = 1. We shall refine the computation done
in the proof of Theorem 2, taking advantage of the structure of B :
instead of writing

v
_
B(v −v

, σ)
_
= −∇

_
B(v −v

, σ)
_
,
we shall report the derivatives on the variable σ. We begin with an
elementary lemma in differential calculus, whose proof we omit.
Lemma 1. ∇
v
_
B
_
v −v

|v −v

|
· σ
__
=
1
|v −v

|
B

_
v −v

|v −v

|
· σ
_
Π
(
v−v∗
|v−v∗|
)
⊥σ,
where Π
k
⊥ is the orthogonal projection upon k

, that is (if k = 1),
Π
k
⊥σ = σ −(σ · k)k.
Our second lemma can be considered as a particular integration by
parts on the unit sphere.
Lemma 2. Let k be a fixed unit vector, and F a smooth function on
R
N
. Then
(30)
_
S
N−1
dσ B

(k · σ)F(σ)Π
k
⊥σ =
_
S
N−1
dσ B(k · σ)M
σk
∇F(σ),
where the linear operator M
σk
: R
N
→R
N
is defined by
M
σk
(x) = (k · σ)x −(k · x) σ.
Proof of the lemma. Let u be a smooth function defined on R
N
; we
define its spherical gradient at point σ ∈ S
N−1
by
(31) ∇
σ
u(σ) = Π
σ
⊥∇u(σ).
We note first that
(32)
_
S
N−1
dσ ∇
σ
u(σ) = (N −1)
_
S
N−1
dσ u(σ)σ.
To prove (32), we introduce a smooth function q on R
+
, identically
vanishing near 0 and ∞. Then we set
w(x) = u
_
x
|x|
_
q(|x|),
so that
∇w(x) =
1
|x|

σ
u
_
x
|x|
_
q(|x|) + u
_
x
|x|
_
q

(|x|)
x
|x|
.
FISHER INFORMATION ESTIMATES 13
Hence, by integration and spherical change of variables,
0 =
_
R
N
1
|x|

σ
u
_
x
|x|
_
q(|x|) +
_
R
N
u
_
x
|x|
_
q

(|x|)
x
|x|
=
__
S
N−1

σ
u(σ) dσ
___

0
q(r)r
N−2
dr
_
+
__
S
N−1
u(σ)σ dσ
___

0
q

(r)r
N−1
dr
_
.
Since
_

0
q

(r)r
N−1
dr = −(N − 1)
_
q(r)r
N−2
dr, and since q is arbi-
trary, the formula (32) follows.
As a consequence, if k denotes a fixed unit vector and u, v are smooth
functions on R
N
, the following formulas of integration by parts hold.
_
S
N−1
dσ u(σ)∇
σ
v(σ) = −
_
S
N−1
dσ ∇
σ
u(σ)v(σ) + (N −1)
_
dσ u(σ)v(σ)σ;
_
S
N−1
dσ u(σ)
_
k · ∇
σ
v(σ)
¸
= −
_
S
N−1

_
k · ∇
σ
u(σ)
¸
v(σ)
+ (N −1)
_
dσ (σ · k)u(σ)v(σ).
(33)
Now, let us be interested in the right-hand side of (30). Writing
∇F = ∇
σ
F + (σ · ∇F)σ, the terms involving σ · ∇F cancel out, and
this expression is
(34)
_
S
N−1
dσ B(k · σ)
_
(k · σ)∇
σ
F(σ) −(k · ∇
σ
F(σ)) σ
¸
≡ (a) −(b),
which depends only on the values of F on S
N−1
. Integrating by parts,
(35)
(a) =
_
S
N−1
dσ B(k·σ)(k·σ)∇
σ
F(σ) = −
_
S
N−1
dσ ∇
σ
_
B(k·σ)k·σ
¸
F(σ)
+ (N −1)
_
S
N−1
dσ B(k · σ)(k · σ)σF(σ);
On the other hand, choosing an orthonormal basis (e
i
), the i-th
component of (b) is
_
S
N−1
dσ B(k · σ)(k · ∇
σ
F(σ))σ
i
= −
_
S
N−1
dσ ∇
σ
_
σ
i
B(k · σ)
_
· kF(σ)
+ (N −1)
_
S
N−1
dσ B(k · σ)(k · σ)F(σ)σ
i
.
14 C. VILLANI
Therefore,
(36) (b) = −
_
S
N−1

i

σ
_
σ
i
B(k · σ)
_
· k e
i
F(σ)
+ (N −1)
_
S
N−1
dσ B(k · σ)(k · σ)σF(σ).
In view of (35) and (36), our lemma is proven provided that the fol-
lowing identity holds :
(37)

i

σ
_
(σ · e
i
)B(k · σ)
¸
· k e
i
−∇
σ
_
B(k · σ)k · σ
¸
= B

(k · σ)Π
k
⊥σ.
Let us compute the left-hand side of (37) : this is

i

σ
⊥e
i
· k)B(k · σ)e
i
+

i
(σ · e
i
)B

(k · σ)(Π
σ
⊥k · k)e
i
−B

(k · σ)Π
σ
⊥k(k · σ) −B(k · σ)Π
σ
⊥k.
=

i
(e
i
· Π
σ
⊥k)e
i
B(k · σ) +

i
(σ · e
i
)e
i
B

(k · σ)(Π
σ
⊥k · k)
−B

(k · σ)Π
σ
⊥k(σ · k) −B(k · σ)Π
σ
⊥k
= Π
σ
⊥kB(k · σ) + σB

(k · σ)(Π
σ
⊥k · k) −B

(k · σ)Π
σ
⊥k(k · σ) −B(k · σ)Π
σ
⊥k.
Since the terms involving B(k·σ) cancel out, it only remains to compute
B

(k · σ)
_

σ
⊥k · k)σ −Π
σ
⊥k(k · σ)
_
=B

(k · σ)
_
σ −(σ · k)
2
σ −k(k · σ) + (σ · k)
2
σ)
=B

(k · σ)
_
σ −k(k · σ)
_
=B

(k · σ)Π
k
⊥σ.

Remark. If N = 2, let us denote by σ

the unit vector obtained from
σ by a counterclockwise rotation of angle π/2. Then
M
σk
(x) = (σ

· x)k

.
By the lemmas 1 and 2 above, if we set
k =
v −v

|v −v

|
,
F(x) = f
_
v + v

2
+
|v −v

|
2
x
_
f
_
v + v

2

|v −v

|
2
x
_
,
FISHER INFORMATION ESTIMATES 15
then we obtain
(38)
_
dv

dσ∇
v
_
B(k · σ)
¸
f

f


=
1
2
_
dv

dσB(k · σ)
_
f


M
σk
(∇f)

−f

M
σk
(∇f)


¸
,
Proposition 3 follows thanks to (20), with
P
σk
(x) = (σ · x)k + M
σk
(x).
We note that P
σk
is an odd function of both k and σ.
Let us be interested in the norm of the linear operator P
σk
.
Lemma 3. Let k, σ, x be vectors of R
2
, such that k = σ = 1. Then
P
σk
x = x.
Proof of the lemma. By the remark following the proof of lemma 2,
P
σk
x = (σ · x)k + M
σk
x = (σ · x)k + (σ

· x)k

.
Hence,
P
σk
x
2
= (σ · x)
2
+ (σ

· x)
2
= x
2
.

Lemma 4. For all σ, k, x ∈ R
N
such that k = σ = 1, one has
P
σk
x ≤ x,
with equality only if σ, k and x belong to the same plane.
Proof of the lemma. First note that if σ and k are colinear, then P
σk
=
±I, and equality holds. Therefore, let us assume that k and σ are free,
and let Π be the orthogonal projector onto the plane directed by σ and
k. We write
x = Πx + Π

x,
with Π

x · σ = 0, Π

x · k = 0. Then,
P
σk
x = (k · σ)Π

x +
_
(k · σ)Πx + (σ · Πx)k −(k · Πx)σ
¸
= (k · σ)Π

x + P
σk
(Πx).
It is obvious that P
σk
(Πx) lies in the plane directed by σ and k. In
view of lemma 3, P
σk
(Πx) = Πx, and therefore
P
σk
x
2
= (k · σ)
2
Π

x
2
+ Πx
2
≤ Π

x
2
+ Πx
2
= x
2
,
with equality only if Π

x = 0, since (k · σ)
2
< 1.
Now, we are ready to adapt the computation in section 3 to our new
representation (12) of Q
+
.
16 C. VILLANI
Proof of Theorem 4. First we note that the change of variables σ →−σ
exchanges f(v

) and f(v


), so Q
+
(f, f) is left invariant if we replace
B(k · σ) by
˜
B(k · σ) = (1/2)[B(k · σ) + B(−k · σ)]. Thus to prove the
theorem, it suffices to consider the case when B is even. Also we shall
only consider the case when f = g since the proof goes through for the
general case as well.
By Cauchy-Schwarz inequality, reasoning as in section 3,
(39) I
_
Q
+
(f, f)
_
=
_
dv
|∇Q
+
(f, f)|
2
Q
+
(f, f)

_
dv dv

dσ B
¸
¸
¸
_
f


_
I + P
σk
_
(∇
_
f)

+
_
f

_
I −P
σk
_
(∇
_
f)


¸
¸
¸
2
with k = (v −v

)/|v −v

|.
By the change of variables (v, v

) ←→ (v

, v


), which exchanges k
and σ,
(40)
I
_
Q
+
(f, f)
_

_
dv dv

dσ B
¸
¸
¸
_
f

_
I + P

_
(∇
_
f) +
_
f
_
I −P

_
(∇
_
f)

¸
¸
¸
2
=
_
dv dv

dσ B
¸
¸
¸
_
_
f


_
f +
_
f(∇
_
f)

_
+P

_
_
f


_
f −
_
f(∇
_
f)

_
¸
¸
¸
2
.
We expand the square norm, and notice that the cross product vanishes,
(41)
_
dv dv

dσ B
_
_
f


_
f+
_
f(∇
_
f)

_
·P

_
_
f


_
f−
_
f(∇
_
f)

_
= 0,
because P
σk
is an odd function of σ (and B is an even function of σ).
Therefore,
(42) I
_
Q
+
(f, f)
_

_
dv dv

dσ B
_
¸
¸
¸
_
f


_
f +
_
f(∇
_
f)

¸
¸
¸
2
+
¸
¸
¸P
σk
_
_
f


_
f −
_
f(∇
_
f)


¸
¸
2
_
.
(43) ≤
_
dv dv

dσ B
_
¸
¸
¸
_
f


_
f +
_
f(∇
_
f)

¸
¸
¸
2
+
¸
¸
¸
_
f


_
f −
_
f(∇
_
f)

¸
¸
¸
2
_
.
Using
_
dσ B(k · σ) = 1 and the symmetry in (v, v

), we obtain
I
_
Q
+
(f, f)
_
≤ 4
_
dv dv

f

¸
¸
¸∇
_
f
¸
¸
¸
2
= I(f).
FISHER INFORMATION ESTIMATES 17
Finally, this inequality extends to arbitrary probability distributions
by density.
Remark. By the “Cauchy-Schwarz equality”,
__
u
2
___
v
2
_

__
uv
_
2
=
_
dxdy
_
u(x)v(y) −u(y)v(x)
_
2
,
we can obtain an explicit lower bound for the dissipation of Fisher
information (see [8] for related arguments) : if we set
Tf(v, v

, σ) =
_
f


(I + P
σk
)(∇
_
f)

+
_
f

(I −P
σk
)(∇
_
f)


,
with k = (v −v

)/|v −v

| and
g

= g
_
v + v

2
+
|v −v

|
2
σ
_
≡ g

[σ],
g


= g
_
v + v

2

|v −v

|
2
σ
_
≡ g


[σ],
we obtain
(44)
I(f)−I
_
Q
+
(f, f)
_

_
dv dv

dσ dwdw


_
B
_
v −v

|v −v

|
· σ
_
_
f

f


[σ]Tf(w, w

, τ)
−B
_
w −w

|w −w

|
· τ
_
_
f

f


[τ]Tf(v, v

, σ)
_
2
,
with equality in dimension 2.
5. Related inequalities
and analogy between Q
+
and the rescaled convolution
In this section, we fix a collision operator Q with Maxwellian cross-
section B, such that for any unit vector k,
_
dσ B(k · σ) = 1. We begin
with a very useful proposition, which we shall call “Bobylev’s lemma”.
Lemma (Bobylev). Let f and g be two probability densities, and let
M be a Maxwellian density. Then
Q
+
(f ∗ M, g ∗ M) = Q
+
(f, g) ∗ M.
The proof can be found for instance in [3].
From this lemma we easily deduce Corollary 4.2. Let f
0
be a smooth
probability density, nonvanishing and rapidly decreasing at infinity.
Then there exists a unique solution f(t, v) to the Boltzmann equation

t
f = Q(f, f). Let M
δ
(v) = exp(−v
2
/2δ)/(2πδ)
N/2
. By Bobylev’s
lemma, f ∗M
δ
is also a solution of the Boltzmann equation. Moreover,
18 C. VILLANI
since ∂
t
(f ∗ M
δ
) = Q(f, f) ∗ M
δ
, f ∗ M
δ
lies in C

([0, ∞) × R
N
) and
does not vanish. This enables to justify the formula
d
dt
I(f(t) ∗ M
δ
) = −8
_


f ∗ M
δ

f ∗ M
δ
Q(f ∗ M
δ
, f ∗ M
δ
).
Passing δ to 0, we obtain in particular
_


f
0

f
0
Q(f
0
, f
0
) ≥ 0,
which proves Corollary 4.2.
We note that we are not aware of a direct proof for this last inequality.
Now, to prove Corollary 4.3, we recall another property related to
the rescaled convolution.
Lemma 5. Let f and g be two probability distributions, and let Q be
a Boltzmann operator with Maxwellian cross-section. Then, for any
λ > 0,
Q
+
(f
λ
, g
λ
) =
_
Q
+
(f, g)
¸
λ
Proof of the lemma. This is a straightforward computation :
2Q
+
(f, g)
λ
=
1

λ
N
_
dv

dσ B
_
v/

λ −v

|v/

λ −v

|
_
_
f
_
v/

λ + v

2
+
|v/

λ −v

|
2
σ
_
g
_
v/

λ + v

2

|v/

λ −v

|
2
σ
_
+g
_
v/

λ + v

2
+
|v/

λ −v

|
2
σ
_
f
_
v/

λ + v

2

|v/

λ −v

|
2
σ
__
.
By the change of variables w

=

λv

, the previous expression is equal
to
1
λ
N
_
dw

dσ B
_
v −w

|v −w

|
__
f
_
v + w

2

λ
+
|v −w

|
2

λ
σ
_
g
_
v + w

2

λ

|v −w

|
2

λ
σ
_
g
_
v + w

2

λ
+
|v −w

|
2

λ
σ
_
f
_
v + w

2

λ

|v −w

|
2

λ
σ
__
= 2Q
+
(f
λ
, g
λ
).

Corollary. The semigroup associated with the Boltzmann equation with
Maxwellian cross-section commutes with the adjoint Ornstein-Uhlenbeck
semigroup.
FISHER INFORMATION ESTIMATES 19
This corollary (due to Bobylev and Carlen [6]) is a direct conse-
quence of the last two lemmas, since the adjoint Ornstein-Uhlenbeck
semigroup is obtained by rescaling and convolution with a Maxwellian
distribution. It was first used by Carlen and Carvalho for a particular
cross-section.
Now, corollary 4.3 follows easily. Let f and g be any two fixed
probability densities. Let us denote by M the Maxwellian distribution
with 0 mean and unit variance, and by S
t
h the solution of the Fokker-
Planck (or Ornstein-Uhlenbeck) equation with initial datum h, i.e.

t
(S
t
h) = ∇·
_
∇(S
t
h) + (S
t
h)v
_
.
As t goes to infinity, S
t
h −→M in all Sobolev norms (Cf. [17]). By an
easy computation, and the same arguments as in [17],

d
dt
H
_
S
t
[Q
+
(f, g)]
_
= I
_
S
t
[Q
+
(f, g)]
_
−I(M).
Integrating this inequality in time from 0 to ∞ and using the last
Corollary, we get
H
_
Q
+
(f, g)
_
−H(M) =
_

0
dt
_
I
_
Q
+
(S
t
f, S
t
g)
¸
−I(M)
_
.
By inequality (8),
H
_
Q
+
(f, g)
_
−H(M) ≤
1
2
_

0
dt (I(S
t
f) −I(M))+
1
2
_

0
dt (I(S
t
g) −I(M))
=
1
2
_
H(f) + H(g)
¸
−H(M),
and the proof is complete. Rigorous justification for such manipulations
can be found in full detail in [10, 17].
Note added after publication: In the published version of this
paper (J. Math. Pures Appl., 1998), I mistakenly stated (14) without
assuming B to be even. But if B is not even, the symmetry argument
used in (41) does not work. (At least one should assume for instance
that B has zero first moment.) If f = g then one can always sym-
metrize the collision kernel B without changing Q
+
(f, f), which is why
this assumption is not needed at the level of (13). I don’t know if the
property of B being even is important to get (14), or if it is just an ar-
tifact of the proof. It is Eric Carlen, Maria Carvalho and Jos´e Antonio
Carrillo who kindly drew my attention on these issues (June 2007).
20 C. VILLANI
References
[1] A. R. Barron. Entropy and the central limit theorem. Ann. Probab., 14 :
336–342, 1986.
[2] N.M. Blachman. The convolution inequality for entropy powers. IEEE Trans.
Info., 2 : 267–271, 1965.
[3] A.V. Bobylev. The theory of the nonlinear, spatially uniform Boltzmann equa-
tion for Maxwellian molecules. Sov. Sci. Rev. C. Math. Phys., 7 : 111–233,
1988.
[4] A.V. Bobylev and G. Toscani. On the generalization of the Boltzmann H-
theorem for a spatially homogeneous Maxwell gas. J. Math. Phys., 33 (7) :
2578–2586, 1992.
[5] E.A. Carlen. Superadditivity of Fisher’s information and logarithmic Sobolev
inequalities. J. Funct. Anal., 101 (1) : 194–211, 1991.
[6] E.A. Carlen. Personal communication.
[7] E.A. Carlen and M.C. Carvalho. Strict entropy production bounds and stability
of the rate of convergence to equilibrium for the Boltzmann equation. J. Stat.
Phys., 67 (3–4) : 575–608, 1992.
[8] E.A. Carlen and M.C. Carvalho. Entropy production estimates for Boltzmann
equations with physically realistic collision kernels. J. Stat. Phys., 74 (3–4) :
743–782, 1994.
[9] E.A. Carlen, E. Gabetta, and G. Toscani. Propagation of smoothness in ve-
locities and strong exponential convergence for maxwellian molecules. Georgia
Tech. and Pavia Preprint. To appear in Comm. Math. Phys.
[10] E.A. Carlen and A. Soffer. Entropy production by block variable summation
and central limit theorems. Comm. Math. Phys., 140 : 339–371, 1991.
[11] R.A. Fisher. Theory of statistical estimation. Proc. Cambridge. Philos. Soc.,
22 : 700–725, 1925.
[12] Yu. V. Linnik. An information-theoretic proof of the central limit theorem with
the Lindenberg condition. Th. Prob. Appl., 4 : 288–299, 1959.
[13] P.L. Lions. Compactness in Boltzmann’s equation via Fourier integral opera-
tors and applications. J. Math. Kyoto Univ., 34 (2) : 391–427, 1994.
[14] H.P. McKean, Jr. Speed of approach to equilibrium for Kac’s caricature of a
Maxwellian gas. Arch. Rat. Mech. Anal., 21 : 343–367, 1966.
[15] S. Mischler. Personal communication.
[16] A. Stam. Some inequalities satisfied by the quantities of information of Fisher
and Shannon. Inform. Control, 2 : 101–112, 1959.
[17] G. Toscani. Entropy production and the rate of convergence to equilibrium for
the Fokker-Planck equation. Preprint, 1997.
[18] G. Toscani. New a priori estimates for the spatially homogeneous Boltzmann
equation. Cont. Mech. Thermodyn., 4 : 81–93, 1992.
[19] G. Toscani. Strong Convergence in L
p
for a spatially homogeneous Maxwell
gas with cut-off. Transp. Theory Stat. Phys., 24 (1–3) : 319–328, 1995.
[20] G. Toscani and C. Villani. Probability metrics and uniqueness of the solution
to the Boltzmann equation for a Maxwell gas. Preprint, 1997.
´
Ecole Normale Suprieure, DMI, 45 rue d’Ulm, 75230 Paris Cedex 05,
FRANCE. e-mail villani@dmi.ens.fr

Sign up to vote on this title
UsefulNot useful