Aspects of Applied Analysis, Functional Inequalities and

Applied Probability in Many Particle Models.

Amit Einav, TU Wien

DK School, Weissensee

September 2017
 What is Kinetic Theory?

• Kinetic Theory is the field in Mathematics that deals with systems of

many objects such as galaxies, gases, cells, and neurons.
• Amongst the phenomena it describes one can find behaviour or rarefied
gases, cell migration, swarming behaviour and neuron networking.
• Kinetic Theory thrives on taking threads from Analysis, Nonlinear PDEs,
Dynamical Systems and Probability, and weaving them together in order
to solve its prominent problems.
 Microscopic, Macroscopic and Mesocopic Approaches

When one deals with a system of many particles a few viewpoints are available:
• Microscopic. In this viewpoint one considers the system as a collection of
particles and tries to solve the associated equations of all the particles.
• Macroscopic. In this viewpoint one considers a small part of the system
as a unit. While the unit is small in comparison to the system, it is large
enough to contain a large amount of particles so it can be treated
statistically. A prime example for this approach is fluid dynamics.
• Mesoscopic. This viewpoint is purely Kinetic. It is the middle ground
between the Micro and Macro - In this approach we consider the behaviour
of an single average particle, bringing the statistical study to the problem.
While the Mesoscopic approach lies in the origin of Kinetic Theory, the modern
reach of the field is far greater:
1. Many particle models that describe average behaviour of systems.
2. Mean field limits - passing from many particle models to the evolution
equation of an average particle (the Mesoscopic equation).
3. Hydrodynamical limits - passing from the Mesoscopic equation to the
Macroscopic setting (under appropriate scaling).
 The Boltzmann Equation

• One of the most fundamental equations in the theory of rarefied gases is

the so-called Boltzmann equation. This equation describes the time
evolution of the density function, f (t, x , v ), of a rarefied gas under the
additional assumption of molecular chaos (Stosszahlansatz).
• In its full generality the equation reads as

∂t f (t, x , v ) + v · ∇f (t, x , v ) = QB (f )(t, x , v ), (1)

| {z } | {z }
Transport Part Collision Part

where the collision part is given by

Z
QB (f )(x , v ) = B (|v − v∗ | , σ)
Rd ×SN−1

f x , v 0 f x , v∗0 − f (x , v ) f (x , v∗ ) dv∗ dσ




 The Boltzmann Equation Cont.

• The function B, the collision kernel, contains all the physical information
about the interaction between the particles. dσ is the uniform measure on
the sphere and
v + v∗ |v − v∗ | σ v + v∗ |v − v∗ | σ
v0 = + , v∗0 = −
2 2 2 2
corresponds to the general scattering of velocities after a collision that
preserves momentum and energy.
• When the interaction is through a power low potential the collision kernel
is given by
B (|v − v∗ | , σ) = |v − v∗ |γ b (cos θ) ,
where γ ∈ (−d, ∞) measures the hardness of the potential, and
cos θ = hv|v−v∗ ,σi
−v∗ |
. The angular kernel, b, usually has a singularity near
θ = 0 (the so-called grazing collision regime), yet in
many cases one
removes this singularity and considers b ∈ L1 SN−1 (the so-called Grad’s
cut off assumption).
 The Boltzmann Equation Cont.

• Boltzmann’s equation can be shown to have several important

hydrodynamical limits such as the compressible Euler equation and
compressible Nvier-Stokes equation (at least formally).
• In the spatially homogeneous case, which we will focus on from this point
onwards, the equation gave rise to the famous Entropy Functional and
H-Theorem.
• Investigating the properties of the equation: existence, uniqueness,
regularity and etc., proved to be a difficult task that had seen giant
fundamental and innovative ideas in nonlinear PDEs.
 Kac’s Model
• To this day, there is no rigorous proof to the derivation of the Boltzmann
equation from Newtonian mechanics for all time. Lanford has made
significant progress towards resolving this issue in 19751 , yet the proof is
valid only for a time that is of order of magnitude of the average time for
the first collision in the model.
• In 1956 Marc Kac introduced an N−particle linear model from which a
caricature of the Boltzmann equation arose as a mean field limit2 .
• Kac considered a model of N indistinguishable particles with one
dimensional velocities that undergo random collision in the following
manner: Suppose that a moment before the collision the velocity vector of
the ensemble was (v1 , . . . , vN ). When the collision occurs, we pick a pair
of particles at random, say (vi , vj ), with equal probability for any i, j, and
collide them. After their collision the new velocity vector is given by
(v1 , . . . , vi (ϑ), . . . , vj (ϑ), . . . , vN ) where
   
vi (ϑ) vi cos ϑ + vj sin ϑ
=
vj (ϑ) −vi sin ϑ + vj cos ϑ
1
O. E. Lanford III, Time Evolution of Large Classical Systems. Dynamical Systems, Theory and
Applications (Recontres, Battelle Res. Inst., Seattle, Wash., 1974), pp. 1–111. Lecture Notes in
Phys., Vol. 38, Springer, Berlin, 1975.
2
M. Kac, Foundations of kinetic theory, Proceedings of the Third Berkeley Symposium on
Mathematical Statistics and Probability, vol. III (1956), 171–197.
 Kac’s Model Cont.
• Formally, in the language of Probability, for given a test function ψ the
expectation to find ψ in velocity v after the collision is
X 1 Z 2π
1
E(ψ)(v) =   ψ (Ri,j,ϑ v) dϑ = Qψ(v),
N 2π 0
i<j
2
where Ri,j,ϑ v = (v1 , . . . , vi (ϑ), . . . , vj (ϑ), . . . , vN ).
• If F (0) (v) is the density function of the particles before the collision, and
F (1) (v) is the density function after the collision, we conclude that
Z Z
Qψ(v)F (0) (v)dµ(v) = ψ(v)F (1) (v)dµ(v),
M M
where dµ is the appropriate measure of the relevant set M, which must
be symmetric under rotation (so that Q will make sense).
• For any measure dµ that is invariant to rotations, it is easy to see that Q
is self adjoint, and as such we can conclude that
F (1) (v) = QF (0) (v).
• Similarly, assuming independence of collisions, we can show that the
density function of the ensemble after exactly k collisions is given by
F (k) (v) = Q k F (0) (v).
 Kac’s Model Cont.

• In the above we only described the change that occurs when a collision
happens. However, in order to get a time evolution, we need to prescribe a
random process that tells us when such collision occurred.
• Motivated by the Boltzmann equation, we would like the time of the
collisions to be memoryless. This is usually done with by a choice of a
Poisson stream: The waiting time of the k−th particle between any two
consecutive collisions is exponentially distributed
t
P(Tk > t) = e − τ ,

where τ is the average time between collisions.

• From the above, and assuming independence of {Tk }k∈1,...,N we can
conclude that waiting time for a collision to occur, T , satisfies
Nt
P(T > t) = P(min(Tk ) > t) = e − τ ,
k

as we assume that the waiting times of the particles are independent.

 Kac’s Model Cont.
• Without loss of generality we can assume that τ = 1.
• If our initial data is given by F0 (v) then for a given t > 0, the density
function of the ensemble at time t and velocity v is given by
∞
X
F (v, t) = F (k) (v)P (Ck,t ) (2)
k=0

where Ck,t = {there have been exactly k collision until t}.

• Assuming that the times between two collisions are independent, and using
the fact that T has a density

fT (t) = Ne −Nt

we find that
X
P(Ck,t ) = P (T = t1 ) P (T = t2 − t1 ) . . .
| {z } | {z }
0≤t1 ≤t2 ≤···≤tk ≤t
first collision at t1 second collision at t2

P (T = tk − tk−1 ) P (T > t − tk )
| {z }| {z }
k-th collision at tk no more collisions
 Kac’s Model Cont.

Z tZ tk Z t3 Z t2
Ne −Nt1 Ne −N(t2 −t1 ) . . .



= ···
0 0 0 0

−N(tk −tk−1 )

−N(t−tk ) N k t k −Nt
Ne e dt1 . . . dtk = e .
k!

• From the above we see that (2) can be explicitly expressed as

∞
X Nk tk
F (v, t) = e −Nt Q k F (0) (v) = e −Nt(I−Q) F (0) (v).
k!
k=0

• This can be reformulated by the following equation

∂FN
(v, t) = N(Q − I)FN (v, t) . (3)
∂t
which we call Kac’s Master equation.
 Kac’s Master eq.
• Kac’s Master equation can be considered on RN ,√
but a more realistic
N−1
approach would be to restrict it to the set S ( N) (which we will call
’Kac’s sphere’) with dµ = dσN , the uniform probability measure on the
sphere. In this setting Q is a bounded, self adjoint operator satisfying
Q ≤ I and
Ker (I − Q) = span {1} . (4)
The boundedness, self adjointness and control by I are true on any
rotation-invariant set. However, (4) is only valid on spheres. Indeed,
integrating over all variables and using the symmetry of the density
function (since the particles are indistinguishable) together with the
invariance of dµ under rotation, we can see that
1 1 X Z Z 2π
hF , (I − Q)F iL2 (M,dµ) =  
4π N M 0
i<j
2

(F (v1 , . . . , vN ) − F ◦ Ri,j,ϑ (v1 , . . . , vN ))2 dϑdµ.

Thus, F ∈ Ker (I − Q) if and only if F = F ◦ Ri,j,θ for all i, j and almost
everywhere in θ. This implies that F must be radial. The converse is
immediate since Q is just rotations and averaging.
• On a sphere, any radial function is constant, which brings us to (4).
 Kac’s Master eq.

• The Kernel of the operator is important to us as it determines the steady

states of the equations, which will turn out to be equilibriums.
√
• The reason the radius of the sphere is N is that it is the right
normalisation: since all the particles are indistinguishable the total kinetic
energy should be NE , where E is the kinetic√ energy of one particle.
Choosing E = 1 for convenience gives us N as the radius of the energy
sphere.
• Now that we have the model at hand, we would like to see how one can
get the Boltzmann equation from it. The idea is to find the equation that
one particle satisfies (it doesn’t matter which, since they are
indistinguishable) and take the number of particles to infinity. In order to
do that we need to find the equation for the first marginal of F .
• From this point onward we will adopt the notation of FN to indicate the
number of particles in play.
 The Connection to the Boltzmann eq.

• For any measure dµ on RN (not necessarily absolutely continuous with

respect to dv!) we can define the k−th marginal in the (i1 , . . . , ik ) −th
(i ,...,i )
variables as the measure Πk 1 k (µ) on Rk satisfying
(i ,...,ik )
(µ) (A1 × · · · × Ak ) = µ A(i1 ,...,ik ) ,


Πk 1 (5)

Al j = il , l = 1, . . . , k
where A(i1 ,...,ik ) = Ae1 × · · · × A
fN with Aej = .
R j 6= i1 , . . . , ik
• In our case dµ = FN dσN is a symmetric measure so we can choose
i1 , . . . , ik = 1, . . . , k and drop the superscript.
 The Connection to the Boltzmann eq. Cont.

• While dµ is not absolutely continuous with respect to dv, its marginal are.
This follows from the following Fubini-Tunelli type theorem:
Theorem
If dµr = FN
dσrN , with dσrN the uniform probability measure on SN−1 (r ), and
ϕ ∈ Cb Rk then
Z Z
ϕ(v1 , . . . , vk )dµr (v) = ϕ(v1 , . . . , vk )FN (v1 , . . . , vN )dσrN
RN SN−1 (r )

k
! N−k−2
2
SN−k−1
Z X
2
= ϕ(v1 , . . . , vk ) r − vi2
|SN−1 | r N−2 Pk vi2 ≤r 2 (6)
i=1 i=1
 
Z
N−k
 FN dσq  dv1 . . . dvk .
 q 
 Pk Pk
SN−k−1 r2− v2 r 2− v2
i=1 i i=1 i
 The Connection to the Boltzmann eq. Cont.

• This theorem implies that Πk (FN dσN ) is absolutely continuous with

respect to the Lebesgue measure and has a density function

k
! N−k−2
2
SN−k−1 X
Πk (FN ) (v1 , . . . , vk ) = N−2
N− vi2
|SN−1 | N 2
i=1 +
 
Z
N−k
 FN dσq .
 q 
 Pk Pk
SN−k−1 N− v2 N− vi2
i=1 i i=1

• Applying the master equation to the above, and using the symmetry of FN
again we find the following equation:
√
Z 2π Z N
1
∂t Π1 (FN )(v1 ) = √
(Π2 (FN )(v1 (ϑ), v2 (ϑ)) − Π2 (FN )(v1 , v2 )) dϑdv2
π 0 − N

which looks very similar to the Boltzmann equation!

 Chaoticity
• In order to have the above equation equal the Boltzmann equation we
would have needed Π2 (FN ) = Π1 (FN ) ⊗ Π1 (FN ). However, as this equation
is only supposed to approximate the Boltzmann equation and we are only
interested in the limit as N goes to infinity we can relax the equality. This
is what’s behind the notion of chaos:
Definition
Let X be a Polish space. A family of symmetric probability measures on X N ,
{µN }N∈N , is called µ−chaotic, where µ is a probability measure on X , if for any
k∈N
lim Πk (µN ) = µ⊗k ,
N→∞

where Πk (µN ) is the k−th marginal of µN , and the limit is taken in the weak
topology.
• Under the assumption of f −chaoticity on the family of density functions
{FN }N∈N (i.e. assuming that the family dµN = FN dσN is
dµ = f (v )dv −chaotic) the limit of the equation for the first marginal
turns into the Kac-Boltzmann equation:
Z
1
∂t f (v1 ) = (f (v1 (ϑ))f (v2 (ϑ)) − f (v1 )f (v2 )) dϑdv2 .
π
.
 Propagation of Chaos

• The problem with the notion of chaoticity is that it is a purely functional

one. Yet, in order to use it to obtain the Kac-Boltzmann equation we
needed to assume that FN (t) is f (t) chaotic for any t > 0, for some
function f (t).
• Kac’s Master equation is uniquely solvable for any given initial datum,
FN (0) (in fact, we wrote the solution as an infinite series), so the only
extra thing we can assume is that the initial datum is cahotic. Apriori, we
have no idea if this property propagates with the Master equation.
• That was the main theorem in Kac’s 1956 paper.
 Propagation of Chaos - Idea of the Proof
• We start by denoting
Z 2π
2 X 1
ΩFN = N (Q − I) FN = (F (Ri,j,θ ◦ v) − FN (v)) dθ
N −1 2π 0
i<j

and rewriting Kac’s Master equation as ∂t FN = ΩN FN .

• We are interested in the limit as N goes to infinity of
Z
hFN (t), gk iL2 (SN−1 (√N ),dσN ) = Πk (FN ) (t, v1 , . . . , vk )gk (v1 , . . . , vk )dv1 . . . dvk
Rk


for any gk ∈ Cb Rk , k ∈ N.
P∞ t m Ωm
• Since FN (t) = m=0 m!
N
FN (0),
where the series converges absolutely in
norm, and ΩN is bounded and self adjoint, one sees that
∞
X tm
hFN (t), gk iL2 (SN−1 (√N ),dσN ) = hFN (0), Ωm √
N gk iL2 (SN−1 ( N ),dσN ) .
m!
m=0

The latter expression involves FN (0), which we know to be chaotic.

Moreover, as we will see shortly, ΩN acting on gk gives us a sum of
functions of up to k + 1 variables, and the sum on the right converges
absolutely, and uniformly in N. This will imply that we can take the limit
of N inside the sum and use the chaoticity of FN (0).
 Propagation of Chaos - Idea of the Proof Cont.

• Let g1 be a function of one variable, v1 . Then

Z 2π
2 X 1
ΩN g1 (v) = (g1 (Ri,j,θ ◦ v) − g1 (v)) dθ
N −1 2π 0
i<j

Z 2π
2 X 1
= (g1 (v1 cos θ + vj sin θ) − g1 (v1 )) dθ,
N −1 2π 0
j>1

since if i is not 1 we see that g1 (Ri,j,θ ◦ v) = g1 (v1 ) = g1 (v).

• Moreover, due to the symmetry of FN (0)
Z Z 2π
1
hFN (0), ΩN g1 i = √
FN (0)
π SN−1 ( N) 0

(g1 (v1 cos θ + v2 sin θ) − g1 (v1 )) dθdσN

(we have dropped the subscript of the inner product for simplicity).
 Propagation of Chaos - Idea of the Proof Cont.
• What happens if our function depends on two variables, say v1 , v2 ?

2 X 1 Z 2π
ΩN g2 (v) = (g2 (v1 cos θ + vj sin θ, v2 ) − g2 (v1 , v2 )) dθ
N −1 2π 0
j>1,j6=2

X 1 Z 2π
2
+ (g2 (v1 , v2 cos θ + vj sin θ) − g2 (v1 , v2 )) dθ
N −1 2π 0
j>2
Z 2π
2 1
+ (g2 (v1 cos θ + v2 sin θ, −v1 sin θ + v2 cos θ) − g2 (v1 , v2 )) dθ
N − 1 2π 0
• While the last term is the only one that doesn’t add a new variable, when
using the symmetry of FN (0) we see that
Z Z 2π
1 N −2
hFN (0), ΩN g2 i = FN (0)
π N −1 √
SN−1 ( N ) 0

(g2 (v1 cos θ + v3 sin θ, v2 ) + g2 (v1 , v2 cos θ + v3 sin θ) − 2g2 (v1 , v2 )) dθdσN
Z Z 2π
1
+ FN (0)
π(N − 1) √
SN−1 ( N ) 0

(g2 (v1 cos θ + v2 sin θ, −v1 sin θ + v2 cos θ) − g2 (v1 , v2 )) dθdσN ,

and the last term will disappear as N goes to infinity.
 Propagation of Chaos - Idea of the Proof Cont.
• The above estimation also shows how Ω2N g1 will behave. In general, given
a function gk , for some k < N, and m ∈ N such that m + k ≤ N we define
k+m−1
1 X
gk+m (v1 , . . . , vk+m−1 , ω) =
2π
j=1
Z 2π 
gk+m−1 (v1 , . . . , vj cos (θ) + ω sin (θ) , . . . , vk+m−1 )
0

−gk+m−1 (v1 , . . . , vk+m−1 ) dθ.

Then
2m X
Ωm
N gk (v1 , . . . , vN ) = gk+m (v1 , . . . , vk , vl1 , . . . , vlm )
(N − 1)m
l1 6∈ {1, . . . , k}
.
.
.

lm 6∈ 1, . . . , lm−1

m
2 X (l)
+ ηm+k−1 (v1 , . . . , vN ),
(N − 1)m
l
 Propagation of Chaos - Idea of the Proof Cont.

(l)
• One has that ηm+k−1 is a function of no more than m + k − 1 variables and

#l ≤ k(k + 1) . . . (m + k − 1)
((N − 1)m − (N − k)(N − k − 1) . . . (N − m − k + 1)) .

• A careful calculation shows that

2m X
Ωm
N gk − gk+m (v1 , . . . , vk , vl1 , . . . , vlm )
(N − 1)m
l1 ,...,lm ∞
k +j −2
  
m
≤ 4 k . . . (m + k − 1) 1 − Πm
j=1 1− kgk k∞
N −1
Using the latter inequality, together with the fact that FN is a probability
density, we conclude that for a fixed k and m

2m X
lim hFN (0), Ωm
N gk i − hFN (0), gk+m i = 0,
N→∞ (N − 1)m
l1 ,...,lk
 Propagation of Chaos - Idea of the Proof Cont.

• Another estimation shows that for any k, m ∈ N

kΩm m
N gk k∞ ≤ 4 k . . . (m + k − 1) kgk k∞ ,
P∞ |t|m
from which we conclude that m=0 m!
kΩm
N gk k∞ converges uniformly for
t < 41 (uniform time!).
• The symmetry of FN (0), and the f −chaoticity of it, assures us that

lim hFN (0), Ωm

N gk i
N→∞
Z
= 2m f ⊗(k+m) (v1 , . . . , vm+k ) gk+m (v1 , . . . , vm+k ) dv1 . . . dvm+k .
Rm+k

and we conclude that

∞
X (2t)m
lim hFN (t), gk i =
N→∞ m!
m=0
Z
f ⊗(k+m) (v1 , . . . , vm+k ) gk+m (v1 , . . . , vm+k ) dv1 . . . dvm+k
Rm+k
 Propagation of Chaos - Idea of the Proof Cont.
• The previous inequality, together with some compactness results from
probability, implies

that Πk (FN (t)) converges weakly to a density
fk (t) ∈ P R k and that
∞
(2t)m
Z X
fk (t, v1 , . . . , vk )gk (v1 , . . . , vk )dv1 . . . dvk =
Rk
m!
m=0 (7)
Z
⊗(k+m)
f (v1 , . . . , vm+k ) gk+m (v1 , . . . , vm+k ) dv1 . . . dvm+k .
Rm+k

• We want to show that fk = f1⊗k , or equivalently

Z
fk (t, v1 , . . . , vk )gk (v1 , . . . , vk )dv1 . . . dvk
Rk
Z
= f1 (t, v1 ) . . . f1 (t, vk )gk (v1 , . . . , vk )dv1 . . . dvk
Rk

• Choosing gk (v1 , . . . , vk ) = ψ (1) (v1 ) . . . ψ (k) (vk ) and computing directly

(with the definition of gk+m ) shows that the above is valid for such
functions, which is enough.
 Existence of Chaotic States

• Kac’s proof showed that if one starts with a chaotic state then his Master
equation propagates that property. However, are there chaotiic states?
• It is quite easy to show that FN = 1 is indeed M−chaotic, where M is
the standard Maxwellian
1 v2
M(v ) = √ e − 2 ,
2π
but can we find more non-trivial states? Moreover, if we are given a
function f ∈ P(R), can we find an f −chaotic family {FN }N∈N ?
√

• Had we considered our Master equation on RN instead of SN−1 N the
answer to the above would have been Yes. Indeed, chaoticity is a form of
’finite dimensional independence in the limit’ - on RN we can just start
with independent families that are constructed from f :

FN = f ⊗N .

This, however, doesn’t work on the sphere - the velocities of the particles,
given by the random variables {Vi }i=1,...,N whose density function is
PN
represented by (the same) f , are dependent (since i=1
Vi2 = N).
 Existence of Chaotic States Cont.

• Can the process we described for RN work on the sphere? Can we

somehow ’restrict’ a given tensorisation from RN to the sphere?
• Intuitively speaking, since every particle should have a unit
R P energy weN
know that R v 2 f (v )dv = 1. As such the expectation of i=1 Vi2 is indeed
√

N and restriction of f ⊗N to SN−1 N makes sense - at least in the
energy level - according to the law of large numbers.
• For a given probability density function f on R, with unit energy, we define
the family of conditioned tensorisation of f , FN , as

f ⊗ (v)
FN (v) = √
, (8)
ZN f , N

where, for any r > 0,

Z
ZN (f , r ) = ΠN N
i=1 f (vi )dσr ,
SN−1 (r )

with dσrN the uniform probability measure on SN−1 (r ).

 Existence of Chaotic States Cont.

• Our intuition that conditioned tensorisations are acceptable and natural

boils down to understanding the normalisation function ZN . We expect
√
that if ZN will exhibit good concentration properties around r = N, we
will get similar properties for FN as we would have to the independent
state on RN . While this is indeed the case, as we will see soon, we need to
understand what this ’concentration’ means.
Lemma
Let f ∈ P(R) be a density function with unit mass, representing the velocity
variable V . Let h be the density function of the variable V 2 . Then

2h∗N (r 2 )
ZN (f , r ) = ,
|SN−1 | r N−2

where h∗N is the N−fold convolution of h.

• The above lemma tells us that if we want to understand how ZN behaves
with respect to r , we need to understand the behaviour of h∗N . THis
expression shouldn’t surprise us - indeed, as we started with
PNa velocity
variable V with density f , h∗N is the density function of V 2 , where
i=1 i
{Vi }i=1,...,N are i.i.d. copies of V !
 Local CLTs
• Understanding the behaviour of the sum of i.i.d. variables is the realm of
Central Limit Theorems (CLT). However, as we want a pointwise
estimation, we need a more restrictive type of CLT - one that deals with
the density function. This type of CLT is called local CLT.
• Before we delve into this, let us explain how one proves Lemma 3: Let
ϕ((v )) = ϕ(|v |) be a radial function. Then by using radial coordinates we
have that
Z Z ∞ Z
E(ϕ) = ϕ(v)f ⊗N (v)dv = r N−1 ϕ(r )f ⊗N (r Ω)dΩdr
RN 0 SN−1 (r )
Z ∞
= SN−1 ϕ(r )r N−1 ZN (f , r ) dr
0

On the other
PNhand,2 since ϕ is radial, its expectation can be measured with
respect to V . As such
i=1 i
Z ∞ Z ∞
√
E(ϕ) = ϕ( r )h∗N (r )dr = 2r ϕ(r )h∗N (r 2 )dr .
0 0

Comparing between these two expressions yields the desired result.

 Local CLTs Cont.

√
• The fact that h∗N should be concentrated on the sphere of radius N is
something we established. However, if we want to an explicit
concentration estimation we need to somehow control the fluctuation of it.
This is usually done by requiring some higher moments. This is also the
case for local CLTs.
Theorem (Carlen et. al. 20103 )
R
Let f ∈ P(R) be such that f ∈ Lp (R) for some p > 1, R
v 2 f (v )dv = 1 and
R 4
R
v f (v )dv < ∞. Then
 
(r 2 −N)2
−
2 e 2NΣ2
ZN (f , r ) = √ √ + λN (r ) ,
NΣ |SN−1 | r N−2 2π
R
where Σ2 = R
v 4 f (v )dv − 1 and supr |λN (r )| −→ 0.
N→∞
√
• Note how strongly concentrated ZN is around r = N!

3
E. A. Carlen, M. C. Carvalho, J. Le Roux, M. Loss and C. Villani, Entropy and Chaos in the
Kac Model. Kinet. Relat. Models, 3 (2010), no. 1, 85–122.
 Local CLTs Cont.

• The idea behind the proof of the theorem is to use the Fourier transform
since for any g ∈ L1 (R) we have that

gd g N (ξ),
∗N (ξ) = b kgk∞ ≤ kb
g kL 1 .
R R
Step 1: Due to the conditions on f has that R
h(v )dv = R
vh(v )dv = 1
and Z Z 2
v 2 h(v )dv = Σ2 + vh(v )dv < ∞.
R R
2p
In addition, for any q < p+1 h ∈ Lq (R). To simplify the proof we define
g(v ) = Σh(Σv + 1), which has a unit mass and energy and zero moment,
and deal with it from this√point onwards.
√

Step 2: Define gN (v ) = Ng ∗N Nv . It is simple to see that
 
N ξ
g
c N (ξ) = b
g √ .
N
 Local CLTs Cont.

Remembering that M
c = M we see that
Z
kgN (v ) − M(v )k∞ ≤ g
c N (ξ) − M(ξ) dξ
R
Z Z


= g
c N (ξ) − M(ξ) dξ + g
c N (ξ) + |M(ξ)| dξ,
√ √
|ξ|≤δ N |ξ|>δ N

where δ is an arbitrary small constant.

Step 3: Since g ∈ Lq (R) for some q > 1 and has a finite moment one can
show that
|ξ| ≥ η ⇒ |b g (ξ)| ≤ 1 − α(η), (9)
with α(η) > 0.
 Local CLTs Cont.
 N
Since M(ξ) = M √ξ we have that
N

√
Z Z  N
N


g
c N (ξ) + |M(ξ)| dξ = N |b
g (ξ)| + |M(ξ)| dξ
√
|ξ|>δ N |ξ|>δ

Nδ 2
√ e− 2
Z
N−q 0 q0
≤ N (1 − α(δ)) g (ξ)| dξ + c √
|b
R δ N
where q 0 is the Hölder conjugate of q and c is a uniform constant.
Step 4: The moment conditions of g imply that there exists (δ) > 0, that
goes to zero as δ goes to zero, such that

|ξ|2
 
|ξ| ≤ δ ⇒ g (ξ) −
b 1− ≤ (δ) |ξ|2 . (10)
2

Z Z    
ξ ξ
√
g
c N (ξ) − M(ξ) dξ =
√
gN
b √ − MN √ dξ
|ξ|≤δ N |ξ|≤δ N N N
 Local CLTs Cont.
As kb
g k∞ , kMk∞ ≤ 1 we have that
       
ξ ξ ξ ξ2

gN
b √ − MN √ ≤N g
b √ − 1−
N N N 2N
      2
ξ2 ξ ξ
+ 1− −M √ ≤ 2N(δ) √ = 2(δ) |ξ|2
2N N N
|ξ|
√
since √ N
≤ δ. This however, doesn’t give a small integral over |ξ| ≤ δ N. We
need to be more delicate when ξ is too large, which occurs as N goes to infinity.
Step 5: Combining the estimation for b g for small and large ξ, (9) and (10), we
see that there exists 0 < α0 < 1 such that for all ξ
 
ξ2
|b
g (ξ)| ≤ max 1 − , (1 − α0 ) .
4

N
Since 1 − u
N
≤ e −u we find that
 N
ξ2
g
cN (ξ) − M(ξ) ≤ max 1− , (1 − α0 ) + |M(ξ)|
4N
 
ξ2
≤ 2 max e− 4 , (1 − α0 )N .
 Local CLTs Cont.

Together with step 4 we have that

q 2
g
c N (ξ) − M(ξ) = g
c N (ξ) − M(ξ)

 
p ξ2 N
≤ 2(δ) |ξ| e − 8 + (1 − α0 ) 2

so
Z Z 
p ξ2 N

√
g
c N (ξ) − M(ξ) dξ ≤ 2(δ) |ξ| e − 8 2
2
dξ + (1 − α0 (δ)) β N .
|ξ|≤δ N R

Step 6: Combining the two estimation for the splitting of the integral of
g
c N − M yields the desired result via the connection between g and h.
 Conditioned Tensorisation and Chaos
• Now that we have the local CLT at hand, we shall see how our family of
conditioned tensorisation are automatically chaotic.
• Recall that

k
! N−k−2
2
SN−k−1 X
Πk (FN ) (v1 , . . . , vk ) = N−2
N− vi2
|SN−1 | N 2
i=1 +
 
Z
N−k
 FN dσq .
 q 
 Pk Pk
SN−k−1 N− v2 N− v2
i=1 i i=1 i

For conditioned tensorisation we get that

 Pk  N−k−2
2
SN−k−1 N− i=1
vi2
+
Πk (FN ) (v1 , . . . , vk ) = N−2
|SN−1 | N 2
 q 
Pk
ZN−k f, N− v2
i=1 i
√
f ⊗k (v1 , . . . , vk ).
ZN f, N
 Conditioned Tensorisation and Chaos Cont.

• The local CLT implies that

Πk (FN ) (v1 , . . . , vk )
Pk
v 2 +k
− i=1 i √  Pk 
e 2(N−k)Σ + 2πλN−k N − v2
r
N i=1 i
= √ √ f ⊗k (v1 , . . . , vk ).
N −k 1 + 2πλN ( N)
• The above is actually much stronger than chaoticity (it implies
convergence even in L1 ) and is extremely useful when one tries to do some
computations.
 Lévy Type Local Central Limit Theorem

• It is important to note that there are other types of local CLTs. The
fourth moment requirement in the theorem assures us that the function h,
for which we actually show the local CLT theorem, has a finite second
moment. As such, if there exists a CLT we expect it to be associated to
the gaussian. However, there are other possible attractors in the realm of
CLTs - those arising from Lévy processes.
 Lévy Type Local Central Limit Theorem Cont.
Theorem (Carrapatosso and E. 4 )
Let f be the probability density function of a random real variable V such that
f ∈ Lp (R) for some p > 1. Assume in addition that
√
Z x
νf (x ) = y 4 f (y )dy ∼ CS x 2−α ,
√ x →∞
− x

for some CS > 0 and 1 < α < 2 and that

Z
x 2 f (x )dx = E < ∞.
R

Then
   
√
2 1 r − NE
ZN f , r = N−2 1 γσ,α,1 1 + λN (r ) ,
|SN−1 | r 2 Nα Nα

where γσ,α,1 is the appropriate Lévy distribution, σ is given by the parameters

of the problems and supu |λN (u)| −→ 0. Moreover, an explicit estimation on
N→∞
the error term λN can be given.
4
K. Carapattos and A. Einav Chaos and entropic chaos in Kac’s model without high moments.
Electron. j. Probab. 18 (2013) no. 78.
 On the Convergence to Equilibrium
• The study of the rate of convergence to equilibrium (when an equilibrium
exists) is of immense importance in most kinetic equations. Kac’s Master
equation, with its connection to the Boltzmann equation, is no different.
• Remembering that we can write the solution to the equation as

FN (t) = e ΩN t FN (0)
with ΩN = N(Q − I) ≤ 0, we see that since ΩN is bounded and self
adjoint, with KerΩN = span {1}, one has (at least formally) that
Z
lim FN (t) = PKerΩN FN (0) = √
FN (0)dσN = 1,
SN−1 ( N )

where PKerΩN is the appropriate orthogonal projection.

• The above shows the existence of a (natural) equilibrium. The question
we now face is - can we estimate how quickly we go to that state? This
question is strongly related to the spectral properties of ΩN . If the kernel
of ΩN is isolated, and the next part of the spectrum of it has a distance,
∆N > 0, from 0 we say that ΩN has a spectral gap of size ∆N . In that
case we will be able to show that
kFN (t) − 1kL2 ≤ e −∆N t kFN (0) − 1kL2 . (11)
 On the Convergence to Equilibrium Cont.

• There might be a strong dependency on N for ∆N . However, to be able to

’push’ this rate of convergence to equilibrium from Kac’s model to the
Kac-Boltzmann equation (the limit equation for N going to infinity) we
need to remove this N dependency.
• Kac conjectured that lim inf N→∞ ∆N > 0.
• Kac’s conjecture remained open until 2000, when Janvresse showed it to
be true. Later on that year Carlen, Carvalho and Loss5 managed to
compute the spectral gap explicitly and showed that
N +2
∆N = .
2(N − 1)
• While the above might fool us to think that we are able to get exponential
convergence to equilibrium even as we take N to infinity, this is not the
case as the initial datum can, and sometimes does, depend strongly on N.

5
E. A. Carlen, M. C. Carvalho and M. Loss, Many body aspects of approach to equilibrium,
Journées ”Équations aux dérivées partielles”, Exp. No, XI (2000).
 On the Convergence to Equilibrium Cont.

• Indeed, one can show (which we will skip here) that when considering
conditioned tensorisation of a function f , one usually gets something of
the form
kf kL2 ≥ C N
for some C > 1. This implies that the right-hand side of (11) behaves like
C N e −∆N t . Thus, the time we need to wait until a significant decay occurs
is of the order of N ∆logN C , which is proportional to N as ∆N ≈ 21 .
• We conclude that while the spectral method for Kac’s model is interesting,
it is not useful to fulfil Kac’s vision and getting new information about the
Kac-Boltzmann equation from his model. This is due to the fact that the
L2 distance is not the appropriate distance to look at in this setting. We
need is a different ’distance’.
• Taking a leaf from Boltzmann’s study, a natural quantity to explore in our
setting is the entropy.
 The Entropy

• For a given f ∈ P(Rd ) one can define the Boltzmann entropy as

Z
H(f ) = f (v) log f (v)dv.
Rd

• This functional has played an important role in Boltzmann’s study of his

equation, mainly via the famous H−theorem that states that H(f (t))
decreases along the flow of the equation.
• While the entropy is very useful (due to it being a Lyapunov functional)
the fact that it is decreasing, and as such converges as the time goes to
infinity, gives us no indication where it converges to. Also, if we are to
think of the entropy as our new ’distance’ the fact that it can become
negative poses some problems.
• What we really need in order to investigate Boltzmann’s or Kac’s
equations is the concept of the relative entropy - an entropy-like functional
that encodes in it both the ’good entropic’ properties, as well as the fact
that it is measured relative to something.
 The Relative Entropy
• Given two probability measures dµ and dν, with dµ being absolutely
continuous with respect to dν with a Radon-Nikodym derivative f , we
define the relative entropy of dµ with respect to dν as
Z Z
dµ dµ
 
H(µ|ν) = log dν = f log fdν.
dν dν
• Defining ψ(x ) = x log x − x + 1, for x ≥ 0, we see that since dµ and dν
have the same mass
Z Z Z
dµ
 
ψ dν = H(µ|ν) − dµ + dν = H(µ|ν).
dν
• This equality serves to show two things:
1. Since ψ(x ) ≥ 0 for all x > 0 we see that H(µ|ν) ≥ 0.
2. Since ψ(x ) = 0 if and only if x = 1 we conclude that H(µ|ν) = 0 if and
only if dµ = dν.
• Moreover, the famous Kullback-Pinsker inequality

kµ − νk2TV ≤ 2H(µ|ν),

gives us an explicit control of the total variation norm by the relative

entropy.
 The Relative Entropy Cont.

• If a kinetic equation has an equilibrium, that equilibrium is a natural

candidate to the measure dν in our definition of the relative entropy.
• For the Boltzmann equation we choose dµ = f (v )dv and dν = M(v )dv .
• For Kac’s model we choose dµ = FN dσN and dν = dσN . Thus we define
the entropy on Kac’s model as
Z
HN (FN ) = H (FN dσN |dσN ) = √
FN log FN dσN .
SN−1 ( N )

• Besides the relevance to the Boltzmann equation, our entropy respects the
chaoticity structure of our equation. Indeed, if one considers the family of
conditioned tensorisation of a function f , FN , we have that
N Z
X f ⊗N √


HN (FN ) = √
√
log f (vi )dσN − log ZN f , N
i=1 SN−1 ( N ) ZN f , N
 The Relative Entropy Cont.

√

!
f ⊗N log ZN f ,
Z
N
=N √
√
log f (v1 )dσN −
SN−1 ( N ) ZN f , N N
v 2 +1
r 1
− 2(N−1)Σ √

+ 2πλN−1 N − v12
Z
N e
=N √ √ f (v1 ) log f (v1 )dv1
N −1 √
|v1 |≤ N 1 + 2πλN ( N)
 N−2

log SN−1 N 2
log N
 
+ +O
N N

• From the above one can show that

Z
HN (FN ) 1 + log 2π
−→ f (v ) log f (v )dv + = H(f |M).
N N→∞
R
2

• Thus, intuitively we expect that for good chaotic states

HN (FN ) ≈ NH(f |M).
 The H−Theorem
• Now that we have our ’distance’, we would like to see that it is suitable for
our equation. In order to do that we need to reproduce Boltzmann’s
H−theorem.
• Differentiating HN (FN ) along Kac’s flow, and remembering that ΩN is self
adjoint with a kernel that is spanned by the function 1, we see that
Z Z
d
HN (FN (t)) = √
ΩN FN (t) log FN (t)dσN + √
ΩN FN dσN
dt SN−1 ( N ) SN−1 ( N )

XZ Z 2π
1 N
=   √
(FN ◦ Ri,j,θ − FN ) log FN dσN dθ
2π N SN−1 ( N ) 0
i<j
2
changing v to Ri,j,−θ v, and then −θ to θ gives us the same expression, up
to a sign, with Ri,j,θ in the logarithm. Thus, by simple averaging we get
XZ Z 2π
d 1 N
HN (FN (t)) = −   √
φ (FN ◦ Ri,j,θ , FN ) dσN dθ
dt 4π N SN−1 ( N ) 0
i<j
2
x


where φ(x , y ) = (x − y ) (log(x ) − log(y )) = (x − y ) log y
. Since φ ≥ 0
we conclude that HN decreases along Kac’s flow.
 The Entropy Method

• With a decreasing relative distance at hand, we can turn our attention to

trying to estimate how quickly the entropy converges to zero (we expect it
to go to zero as we know the solution converges to 1).
• The fact that we actually have an explicit expression to the dissipation of
the relative entropy is the key to what is known as the entropy method.
• The entropy method has a fairly simple idea: Once you have discovered a
Lyapunov functional H, you define its production, D, to be minus the
formal derivative of H under the flow of the equation. At this point you
forget the evolution equation and try to find a functional inequality of the
form
D(f ) ≥ Φ(H(f |M)),
where M is the appropriate equilibrium and Φ is a function that satisfies
some growth condition at infinity.
• Assuming that we managed to find such Φ we can recall the fact that D is
minus the derivative of H along the flow of the equation to find the
inequality
d
H(f (t)|M) ≤ −Φ (H(f (t)|M)) .
dt
 The Entropy Method Cont.

• The most desirable (and in many cases natural) inequality is of the form

D(f ) ≥ CH(f |M), (12)

where C > 0 doesn’t depend on f or the dimension. This inequality gives

an exponential rate of convergence for H(f (t)|M).
• Other explicit, and still good, inequalities are of the form

D(f ) ≥ CH(f |M)1+α , (13)

for C , α > 0 (algebraic rate of convergence) and

CH(f |M)
D(f ) ≥ , (14)
|log H(f |M)|
1
β
where C > 0 (stretched exponential, i.e. e −ct for β > 1, rate of
convergence).
 The Entropy Method Cont.

• If one chooses H(f |M) to be kf − Mk2L2 , (12) is equivalent to finding a

C
spectral gap (which will be 2
).
• The entropy method’s most famous success is with linear Fokker-Planck
equation
  
f (v )
∂t f (v ) = ∇·(∇f (v ) + vf (v )) = ∇· f (v )∇ log , t > 0, v ∈ Rd
M(v )

where M is the standard gaussian, and the Lyapunov functional

Z  
f (v )
H(f |M) = f (v ) log dv .
Rd
M(v )

In that case, the entropy production D is called the relative FIsher

information, is denoted by I, and is given by
Z   2 Z r 2
f (v ) f (v )
I(f |M) = f (v ) ∇ log dv = 4 ∇ M(v )dv
Rd
M(v ) Rd
M(v )
 The Entropy Method Cont.

• The functional inequality connecting between H and I, which is of the

form (12) is the famous log-Sobolev inequality.
• It is interesting to note that a very useful way to prove the log-Sobolev
inequality is via the so-called Bakry-Emery method. In this method one
differentiate I again under the flow of the equation, and finds a linear
functional inequality between it and its dissipation. A proper integration
on it yields the desired inequality.
• In Kac’s model’s settings, the existence of a constant C such that there is
a linear connection between HN and its production DN is a conjecture that
is called Cercignani’s Many Body conjecture (named after the appropriate
conjecture for the Boltzmann equation). We will say that the existence of
a power relation between HN and its production (like (13)) is called the
Almost Cercignani’s conjecture.
 Cercignani’s Conjecture
• Denoting by
DN (ξN )
ΓN = inf ,
ξN HN (ξN )
the entropy-entropy production ratio, Villani has shown in 20036 that
2
ΓN ≥ .
N −1
• Villani actually managed to show more than the above. Defining

XZ Z 2π
1 N
γ
DN,γ (FN ) =   √
1 + (vi2 + vj2 )
4π N SN−1 ( N ) 0
i<j
2

φ (FN , FN ◦ Ri,j,θ ) dσN dθ

for 0 ≤ γ ≤ 1 (γ = 0 fits our normal Kac model), Villani showed that
HN (FN )
DN.γ (FN ) ≥ Cγ
N 1−γ
6
C. Villani, Cercignani’s conjecture is sometimes true and always almost true, Comm. Math.
Phys., 234 (2003), 455–490.
 Cercignani’s Conjecture Cont.

• The proof by Villani follows the idea of the Bakry-Emery method - but
with a twist. Villani has considered DN,1 and differentiated it again not
along Kac’s flow but along a related flow - the heat flow (playing the role
of the Fokker-Planck equation on the sphere). This, together with the fact
that γ = 1, allowed him to close an appropriate inequality, which yielded
the linear connection between DN,1 and HN after an appropriate
integration. All other inequalities are obtained by interpolation.
• Villani conjectured that ΓN = O (1/N). which was proved to be essentially
true in work of E. from 20117 . It is worth to note that since

HN (FN (t)) ≤ e −ΓN t HN (FN (0)) ,

the time for significant decay behaves like 1/ΓN ≈ N. It would seem that
the non-linear method is no better than the linear one!

7
A. Einav, On Villani’s conjecture concerning entropy production for the Kac master equation,
Kinet. Relat. Models, 4 (2011), 479–497.
 Villani’s Conjecture

• The initial step in showing Villani’s conjecture was done by Carlen et. al.
in their 2010 work, where they have proved the local CLT we mentioned
before.
v2
• Denoting by Ma (v ) = √ 1 e − 2a
, the authors followed similar ideas to
2πa
those considered for the Boltzmann equation and defined

fδ (v ) = δM δ (v ) + (1 − δ)M 1−δ (v ).
2 2

Noticing that the first part of fδ has mass δ but the same kinetic energy
(1/2) as the second part, the above function represents the situation that
a particle has very low probability to be in a very excited energetic state
with potentially high velocity, and a high probability to be at a relatively
stable state. Considering the conditioned tensorisation of fδ yields a state
where most of the particles in the system are stable, but a small amount
of them is extremely energetic. Since the collisions are relatively rare, it
will take very long for this state to reach equilibrium.
 Villani’s Conjecture Cont.
• Indeed, the authors showed that

DN (FN (δ))
ΓN ≤ ≤ −C δ log δ.
HN (FN (δ))
• While the above shows that lim supN→∞ ΓN = 0, it doesn’t give an explicit
bound in terms of N.
• The main idea to get such a bound, which is the crux of the work of E., is
to consider the same type of states as before, but allow δ to depend on N:

fδN (v ) = δN M δN (v ) + (1 − δN )M 1−δN (v ).
2 2

• The main problem with this approach is that the local CLT Carlen et. al.
have used didn’t allow the generating function f to change as the number
of variable increases with N. A new local CLT was developed to allow this
(with some restrictions on fN ), and with it E. has shown that
DN (FN (δN ))
ΓN ≤ ≤ −C (δN )δN log δN .
HN (FN (δN ))

• the choice of δN = N −η , for 0 < η < 1 is permitted in the theorem, and in

log N
that case C (δN ) = C (η), which gives us a convergence rate of order Nη
.
 The Almost Conjecture

• It seems that Cercignani’s conjecture is not valid in the general case. Can
we, however, say anything about the almost conjecture? For that we turn
our attentoin briefly to the Boltzmann equation.
• In his (previously

mentioned) 2003 paper, Villani has shown that if
f ∈ P Rd has a unit second moment, some lower bounds and enough
moments, then for 0 ≤ γ ≤ 1 there exists  > 0, depending on the above,
such that
Dγ (f ) ≥ Cf ,γ, H (f |M)1+ , (15)
where
Z
1
γ
1 + |v − v∗ |2 Ψ f (v )f (v∗ ), f (v 0 )f (v∗0 ) dvdv∗ dσ


Dγ (f ) =
2π R2d ×Sd−1

(Dγ is the production of H (f |M) under the Boltzmann equation).

• Moreover, in the case γ = 1 one can choose  = 0 and a fixed constant C
• The constant Cf ,γ, depends on the aforementioned properties of f . The
smaller  is, the more restrictive the conditions are.
 The Almost Conjecture Cont.

• Inequality (15), which can also be shown to hold in the Kac-Boltzmann

equation, gives us hope to get a similar inequality for the sphere.
• We have seen before that if FN is the conditioned tensorisation of a good
enough function f
HN (FN )
lim = H (f |M) .
N→∞ N
It is as simple to show that
DN,γ (FN ) Dγ (f )
lim =
N→∞ N 2
• This gives rise to the idea that we need to look for inequalities of the form
 1+
DN,γ (FN ) HN (FN )
≥ C . (16)
N N
 First Attempt - Log Scalability

• Taking lead from Villani’s work, we start by attempting to restrict our

discussion to ’bounded’ functions in some sense.
Definition √


Let FN ∈ P SN−1 N . We say that FN is log-scalable if there exists
CF > 0, independent of N, such that

sup √ |log FN (v)| ≤ CF N.

v∈SN−1 ( N)
 First Attempt - Log Scalability Cont.

Theorem (Carlen, Carvahlo, E. 20178 )

N−1
√


Let FN ∈ P S N be symmetric and log-scalable. Assume in addition
that there exists k > 1 such that
Z
M2k = sup M2k (Π1 (FN )) = √
|v1 |2k FN dσN < ∞.
N SN−1 ( N)

Then,
 k−1
1+ 1−γ
DN,γ (FN ) Cγ,k HN (FN )
≥ 1−γ ,
N N k−1 N
where Cγ,k depends on the log-scalability constant, γ and M2k .

• The log-scalability property, as well as finite moments, propagate under

Kac’s flow. While the above functional inequality is not the desired form,
γ−1
it still gives a algebraic rate of decay with significant decay after N k−1 -
an improved dependency on N!

8
E. A. Carlen, M. C. Carvalho and A. Einav, Entropy Inequalities for the Kac’s Walk. To
appear in Kinet. Rel. Mod.. https://arxiv.org/abs/1608.00223.
 Second Attempt - A Measure of Cancellation
• The previous theorem, while improving a known result, is still not good
enough to try and take the limit of N to infinity. Going back to the almost
result for the limit equation, one notices that taking a function f satisfying
that inequality will results in the desired almost conjecture on the sphere if
one considers its tensorisation.
• Why does it work so well for these functions? The answer lies in the
following strong cancellation phenomena: If FN ≈ f ⊗N then
 
f (v1 )f (v2 )
φ (FN , FN ◦ Ri,j,θ ) ≈ (FN − FN ◦ Ri,j,θ ) log .
f (v1 (θ)f (v2 (θ)))
• The first part in the above is simple to control, while the second is
independent in N. This is true not only for this prescribed φ.

Definition √


We say that a family of probability densities {FN }N∈N ∈ P SN−1 N is
log-power-chaotic of order β > 0 if there exists C > 0, independent of N, such
that Z 2π Z
1
√
φβ (FN , FN ◦ R1,2,θ ) dσN dθ ≤ C 1+β ,
2π 0 SN−1 ( N )

where φβ (x , y ) = |x − y | |log x /y |1+β .

 The Almost Conjecture Revisited
Theorem (Carlen, Carvahlo, E. 20179 )
N−1
√


Let FN ∈ P S N be a log-power-chaotic family of functions of order β
with associated constant CF > 0.
(i) If there exists k > 1 + 1/β such that M2k < ∞ then,
 1+ (1−γ)(1+β)
kβ−(1+β)
DN,γ (FN ) HN (FN ) (17)
≥ Ck,γ,β .
N N

(ii) If there exists a, µ > 0 such that

Z
µ
Mexp = sup √
e a|v1 | FN (v) dσN < ∞.
N SN−1 ( N )

Then,
HN (FN )
DN,γ (FN ) N
≥ !! 2(1−γ) .
N µ
21+µ Ck,γ,β
a
log
1+β
HN (FN ) β
6N

9
E. A. Carlen, M. C. Carvalho and A. Einav, Entropy Inequalities for the Kac’s Walk. To
appear in Kinet. Rel. Mod.. https://arxiv.org/abs/1608.00223.
 Idea of the Proof

• Splitting the sphere into the sets where 1 + v12 + v22 > λ and
1 + v12 + v22 ≤ λ we get that

DN,1 (FN ) ≤ λ1−γ DN,γ (FN )

Z Z 2π
N
1 + v12 + v22 φ (FN , FN ◦ R1,2,θ ) dσN dθ.


+ √
4π SN−1 ( N ),1+v12 +v22 >λ 0

• Using Hölder inequality we get that the above is bounded by

1
! 1+β
Z 2π Z
N 1
λ1−γ DN,γ (FN ) + √
φβ (FN , FN ◦ R1,2,θ ) dσN dθ
2 2π 0 SN−1 ( N )

β
! 1+β
Z

1+β
2 √
1+ v12 + v22 β
FN dσN
SN−1 ( N ),1+v12 +v22 >λ
 Idea of the Proof Cont.
• Thus
β kβ
DN,1 (FN ) DN,γ (FN ) 2 1+β 3 1+β CF 1 β
≤ λ1−γ + kβ
(1 + 2M2k ) 1+β .
N N 2 λ 1+β
−1

• Optimising over λ > 0 and using the fact that DN,1 ≥ CHN (FN ) yields the
result.
• The same approach gives the exponential case, and the log-scalable
theorem.
• Unfortunately, it is not clear that the log-power-chaoticity property
propagates with Kac’s flow. However, by using conditioned tensorisation
of a given function f , and a newly defined concentration properties, one
can fulfil Kac’s original vision and obtain information for the limit equation
from his many particle model. To do that one considers a solution to the
Kac-Boltzmann equation, f (t), and look at its conditioned tensorisation.
That family, FN (t), satisfies nice chaosticity properties that are uniform in
time, making it possible to take the limit of (17) to obtain a similar
inequality for f (t).
• Some restrictions on f are required, and some properties of concentration
- to generalise when one has a local CLT. The conditions presented in the
work of Carlen, Carvalho and E. are shown to propagate along the
Kac-Boltzmann equation, which is crucial for the above process.
 The Connection with Geometric Functional Inequalities
• A surprising connection was found by Carlen in 200710 between a geometric
functional inequality, and Cercignani’s conjecture on Kac’s sphere.
• On RN , the entropy functional enjoys

a special property called
subadditivity: Given FN ∈ P RN we have that
N  
X (i)
H Π1 (FN ) ≤ H (FN ) .
i=1

Basically, what the above says is that if we compute the ’disorder’ of each
individual particle in the system and sum all of them up - we will always
be below the total ’disorder’ of the system.
• The above can be generalised to the relative entropy with respect to the
standard gaussian: If, in addition, we assume that FN has a finite second
moment then:
N  
X (i)
H Π1 (FN ) |M ≤ H (FN |MN ) , (18)
i=1

|v|2
−
e 2
where Mk (v1 , . . . , vk ) = k and M = M1 .
(2π) 2
10
E. A. Carlen, The rate of local equilibration in kinetic theory, Prospects in mathematical
physics, 71–88, Contemp. Math., 437, Amer. Math. Soc., Providence, RI, 2007
The Connection with Geometric Functional Inequalities Cont.

• The proof of (18) follows immediately from the first inequality. To show
the first inequality we consider the probability measure
 
(i)
GN = Π N
i=1 Π1 (FN )

we notice that
N  
X (i)
0 ≤ H (FN |GN ) = H(FN ) − H Π1 (FN ) .
i=1

• inequality (18) is more appropriate when one considers Kac’s model as the
gaussian√ measure

on RN and the uniform probability measure on
SN−1 N are intimately related and in many cases behave the same
when N is very large. This is known as equivalence of ensembles in
statistical physics.
• In this setting, the uniform measure dσN corresponds to the
micro-canonical ensemble, representing a fixed number of particles with a
fixed total energy while the gaussian measure corresponds to the canonical
ensemble, representing a fixed number of particle in thermal equilibrium.
The Connection with Geometric Functional Inequalities Cont.

• For simple systems, the equivalence of ensembles means that for any
observable function ψ of finitely many particles, v1 , . . . , vk , one has that
Z Z !
lim √
ψ (v1 , . . . , vk ) dσN − ψ (v1 , . . . , vk ) MN (v)dv = 0,
N→∞
SN−1 ( N ) RN

which is simple to show. A problem occurs when one tries to take into
account all the particles.
• In order to see an analogue to (18) we must first understand what
marginals on the sphere mean. Remembering that marginal can be defined
by integration against a function of finitely many variable, we define the
first marginal on the sphere in the i−th variable of the density function
(N)
FN , Fi , as the unique density function (on the sphere) such that
Z Z
(N)
√
φ(vi )Fi (vi ) dσN = √
φ(vi )FN (v) dσN .
SN−1 ( N) SN−1 ( N )
The Connection with Geometric Functional Inequalities Cont.

• One can show

Z
(N) N−1
Fi (vi ) = p
FN dσpN−v 2 ,
SN−2 N−vi2 i

where dσrk is the uniform probability measure on Sk−1 (r ).

• The connection between the ’spherical marginal’ and the regular marginals
is given by
 N−3
SN−2

(i) 1 v2 2
(N)
Π1 (FN ) (v ) = N−1 √ 1− Fi (v ).
|S | N N +
 The Breaking of the Subadditivity on the Sphere

• In 2004 Carlen, Lieb and Loss managed to show that in general (18) does
not hold on the sphere:
Theorem (Carlen, Lieb, Loss (2004)11 )
N−1
√


Given FN ∈ P S N , dσN then
N  
X (N)
HN F i ≤ 2HN (FN ) . (19)
i=1

Moreover, the constant 2 is sharp.

• Note that the proof of the (18) relied heavily on our ability to define a
new probability by multiplying the first marginals of FN . This is not
possible on the sphere!
• This deviation from the principle of equivalence of ensembles speaks much
about the geometric nature of the sphere and is the key to Carlen’s proof
of Villani’s lower bound on ΓN without using the Bakry-Emery method.

11
E. A. Carlen, E. H. Lieb and M. Loss A sharp analog of Young’s inequality on S N and related
entropy inequalities, J. Geom. Anal. 14 (2004), no. 3, 487–520.
 Subadditivity and Cercignani’s Conjecture

• The first thing that Carlen noticed was that due to its form, if one defines
(r )
DN (ξ)
ΓN (r ) = inf (r )
ξ∈P (SN−1 (r )) HN (ξ)

with the entropy and dissipation on the sphere SN−1 (r ), then ΓN (r ) is

independent of the radius r .
• Another√observation
√ one can easily make is that for a fixed
vi ∈ (− N, N)
 
FN (v)
p 
N−2 N−1
(N)
∈P S N− vi2 , dσp .
Fi (vi ) N−vi2

• Carlen’s proof relies on the above two observation, (19), and an induction.
Since DN (F ) = N h(I − Q)F , log F iL2 we can rewrite
!
hQξN , log ξN i
ΓN = N 1− sup √ = N (1 − ΛN )
ξN ∈P (S N−1 ( N )) HN (ξN )
 Subadditivity and Cercignani’s Conjecture Cont.

• The operator Q can be written as an average of similar operators that

leave one axis untouched. Indeed, for 1 ≤ k ≤ N we can define
Z
1 X 1
Qk (F )(v) =   √
F ◦ Ri,j,θ (v)dθ,
N −1 2π SN−1 ( N )
i<j,i,j6=k
2

1
PN
and a simple counting shows that Q = N k=1
Qk .
• We have that
N Z  
1 X (N) FN (v)
hQFN , log FN i = √
Fk (vk )Qk (N)
N S N−1 ( N) Fk (vk )
k=1
   
FN (v)

(N)
log (N)
+ log Fk (vk ) dσN .
Fk (vk )
 Subadditivity and Cercignani’s Conjecture Cont.
N−1
• Remembering that we can write dσN = dσp dµ(vk ) and that ΛN
N−v 2
k
doesn’t depend on the radius (since ΓN doesn’t), we see that
N Z  
1 X FN (v) FN (v)

(N)
hQFN , log FN i ≤ √
Fk (vk )ΛN−1 (N) log (N)
dσN .
N SN−1 ( N ) F (vk ) Fk (vk )
k=1 k
Z  
FN (v)
 
(N) (N)
+ √
Fk (vk )Qk (N)
log Fk (vk ) dσN
SN−1 ( N ) Fk (vk )
Additionally, since Qk is self adjoint and leaves function that depend on vk
unchanged
Z  
FN (v)
 
(N) (N)
√
Fk (vk )Qk (N)
log Fk (vk ) dσN
SN−1 ( N ) Fk (vk )
Z   Z  
(N) (N) (N)
√
FN (v) log Fk (vk ) dσN = √
Fk (vk ) log Fk (vk ) dσN
SN−1 ( N ) SN−1 ( N )
 
(N)
= HN F k
 Subadditivity and Cercignani’s Conjecture Cont.
• Since Z  
(N) FN (v) FN (v)
√
Fk (vk ) (N)
log (N)
dσN
SN−1 ( N ) Fk (vk ) Fk (vk )
Z    
(N) (N)
= HN (FN ) − √
FN (v) log Fk (vk ) dσN = HN (FN ) − HN Fk
SN−1 ( N )

we get that
N 
1 X
    
(N) (N)
hQFN , log FN i ≤ ΛN−1 HN (FN ) − HN Fk + HN Fk .
N
k=1

Thus  
PN (N)
H Fk
k=1 N
hQFN , log FN i 1 − ΛN−1
≤ ΛN−1 +
HN (FN ) N HN (FN )
which, using inequality (19) yields
hQFN , log FN i 2 (1 − ΛN−1 )
≤ ΛN−1 + .
HN (FN ) N
 Subadditivity and Cercignani’s Conjecture Cont.
• By taking supremum over FN and using the fact that ΓN = N(1 − ΛN ) we
see that
N −2
ΓN ≥ ΓN−1 ,
N −1
from which we conclude that
N −2 N −3 1
ΓN ≥ ΓN−1 ≥ ΓN−2 ≥ · · · ≥ Γ2 .
N −1 N −1 N −1
A direct computations shows that Γ2 = 2 which completes the proof.
• Notice the following observation: if we had had the subadditivity from RN ,
the same process would have shown that ΓN ≥ ΓN−1 , giving a uniform
lower bound on ΓN !
• Even if we didn’t have exact subadditivity, but an approximate one of the
form
N  
X (N)
HN Fk ≤ (1 + N ) HN (FN )
k=1
the induction process would have yielded
  
j
ΓN ≥ ΠN
j=3 1− Γ2 .
j −1
P∞ j
• The above can give a lower bound that is uniform in N if j=3 j−1
< ∞,
for instance if j is any negative power of j.
 An Almost Subadditivity for the Entropy on the Sphere
• With what we have seen in mind, the last thing we would like to do in this
mini course is to try and correct our deviation from the subadditivity on
RN .
• While we are searching for a general geometric functional inequality, we
are motivated by Kac’s model and as such we will use tools and quantities
that are natural to it.
• On of these tools, which we briefly mentioned earlier, is the Fisher
information functional - this time on Kac’s sphere:
Definition √


Let FN ∈ P SN−1 N , dσN . The Fisher Information (on the sphere) of FN
is defined as
Z
IN (FN ) = IN (FN dσN |dσN ) = √
|∇S log FN |2 FN dσN ,
SN−1 ( N )

where ∇S is the gradient on the sphere.

• The Fisher information can be obtained from the our entropy functional by
differentiating along the appropriate Fokker-Planck flow - which is the
heat flow.
 An Almost Subadditivity for the Entropy on the Sphere Cont.
Theorem (E. 201612 )
√


Let FN ∈ P SN−1 N , dσN be such that:
PN (i)


Mk Π1 (FN )
i=1
(i) Ak = supN N
< ∞ for some k > 2 where
  Z Z
(i) k (i)
Mk Π1 (FN ) = |v | Π1 (FN ) (v )dv = √
|vi |k FN dσN .
R SN−1 ( N )

(ii) inf N HN N(FN ) ≥ CH > 0.

and either PN

(i)
I Π1 (FN )
i=1
(iii) (Non-intrinsic condition) AI = supN < ∞, or N
(iii’) (Intrinsic conditions) supN IN (F
N
N)
≤ CI < ∞ and there exists 2 < q < k
such that PN (i)
Pq (FN )
APq = sup i=1 <∞
N N
(i)
(i)
R Π1 (FN )(v ) R FN
where Pq (FN ) = R
q dv =

q−2 √
SN−1 ( N )  q dσN
 q−2
1− vN
2 v2
1− Ni

12
A. Einav, On the subadditivity of the entropy on the sphere, J. Geom. Anal. 26 (2016), no. 4,
3098–3128.
 An Almost Subadditivity for the Entropy on the Sphere Cont.
Then there exists N depending on Ak , CH and AI or CI and APq that goes to
zero like a negative power of N, depending on k and possibly q, such that
N  
X (N)
HN F i ≤ (1 + N ) HN (FN ) . (20)
i=1

• The idea behind the proof is to try and utilise the subadditivity on RN by a
natural extension of FN to RN , and then try to relate it back to the sphere
by using a new probabilistic distance, the Wasserstein distance, that
together with the Fisher information can control HN .

Definition
Let X be a Polish space and let µ, ν ∈ P (X ). For any 1 ≤ q we define the
Wasserstein distance of order q between µ and ν as
 Z  q1
q
Wq (µ, ν) = inf d (x , y ) dπ(x , y ) ,
π∈Π(µ,ν)
X ×X

where Π (µ, ν), the space of coupling, is the space of all probability measures
on X × X with marginals µ and ν respectively.
 Ideas of the Proof - Step 1: Extension

N
• The first step is to try and √
utilize

the
subadditivity in R . Towards that
N−1
goal, given FN ∈ P S N , dσN we define its euclidean extension
N


as the probability density F N ∈ P R
f

√
 
v
F
f N (v) = FN N MN (v) ,
|v|

for v ∈ RN \ {0}.
• Due to the fact that MN is radial and FN depends only on the spherical
variable it is simple to see that
 
H F N |MN = HN (FN ) .
f
Ideas of the Proof - Step 2: Connection between the Marginal and the
Marginals of the Extension

• A short computation yields the following relationship between the first

marginal of FN and that of F
f N:

N Z √
(j)
  SN−1 N 2 sgn(v ) N
(j) v N−1 − Nv22
Π1 F
f N (v ) = N
Π1 (FN ) (x ) e 2x dx . (21)
(2π) 2 0
xN

• Using the 1−Wasserstein distance, which can be shown to have a nice

formula (Kantorovich-Rubinstein identity), one finds that
  21
(j)
   M2 Π1 (FN )
(j) (j)
W1 Π1 (FN ) , Π1 F
f N ≤ √ (1 + τN ) , (22)
2N
where τN , which is given explicitly and independently on FN , satisfies
τN −→ 0 as N goes to infinity.
N→∞
Ideas of the Proof - Step 2: Connection between the Marginal and the
Marginals of the Extension Cont

• The Wasserstein distance of order 1 is not ideal for dealing with the
entropy. However, with additional control of moments one can estimate
’higher orders’ of the Wasserstein distance. Using (22) one can show that
(i)
if Mk Π1 (FN ) < ∞ for some k > 2, then for any 2 ≤ q < k

 k1
Ck
   3+1

(i) (i)
Wq Π1 (FN ) , Π1 F
f N ≤ 22 q 1+
2
  2q1 − 2k1
 k1 M2 Π(i)
1 (FN ) 1
    
(i)
1 + Mk Π1 (FN ) 1 1 1+O ,
(2N) 2q
− 2k N

where Ck is a constant that is independent of N and FN .

Ideas of the Proof - Step 3: HWI and the Entropy Connection
• Now that a distance that converges to zero as N goes to infinity has been
found, we would like to connect it to the entropy of the appropriate
densities. To do that we recall the HWI inequality with respect to the
gaussian measure on R:
p
H(f |M) ≤ H(g|M) + I (f |M)W2 (f , g) .

as well as the following ’distorted’ version of it: for any 1 < p < ∞
Z   p  p1
d f (v )
H(f |M) ≤ H(g|M) + log f (v )dv Wq (f , g) .
R
dv M(v )

where q is the Hölder conjugate of p.

• Using the appropriate HWI inequality with respect to the conditions (iii)
or (iii 0 ) gives
     
(i) (i) (i)
H Π1 (FN ) |M ≤ H Π1 F
f N |M + e
N ,

(i) (i)
where eN depends on the relevant quantities of Π1 (FN ) and goes to zero
as a negative power of N.
 (i)
Ideas of the Proof - Step 4: Connection between Entropies of Π1 (FN ) and
(N)
Fi

(i) (N)
• Using the formula relating Π1 (FN ) and Fi , as well as the conditions
specified in the main theorem yields the following:
   
(N) (i) (i)
HN F i ≤ H Π1 (FN ) |M + b
N ,

(i) (i)
where bN depends on the relevant quantities of Π1 (FN ) and goes to zero
as a negative power of N.
• Combining all our steps, along with the fact that

N      
X (N)
H Π1 F N |γ ≤H FN |γN = HN (FN ) ,
f f
i=1

HN (FN )
and the fact that N ≤ CH
, yields the proof of the theorem.
 Will the Almost Subadditivity Work?

• The second version of the theorem is far more intrinsic to the sphere. The
(i)
appearance of the quantity Pq is due to complications that may arise near
the poles of the sphere. Pq is a quantitative measure of how
’concentrated’ the state around the pole is - corresponding the to situation
where the ensemble has one particle that is very energetic while the rest
are almost stationary.
HN (FN )
• The requirement that N
be ’far’ from zero is actually not intuitive to
Kac’s model as HN (FN ) will decrease to zero.
• The above requirement can also be removed, resulting in an inequality of
the form
N  
X (N)
HN F i ≤ HN (FN ) + ΥN ,
i=1

where ΥN is given explicitly.

• While this subadditivity correction hasn’t resolved Cercignani’s conjecture,
it has given us some inklings to new and interesting quantities, as well as
ideas, to explore in the near future.
Here Endeth the Mini Course

Thank you for your attention!