Notes on the Boltzmann Equation

Alberto Bressan - Dept. of Mathematics, Penn State University

(Lecture notes for a summer course given at S.I.S.S.A. , Trieste, 2005)

We wish to describe the motion of a rarefied gas, consisting of a very large number of identical
particles, moving in a three-dimensional space. For (t, x, ) [0, [ IR3 IR3 , we consider a
function f (t, x, ) describing the density of particles at time t, at the point x, having speed .
In alternative, we may also think of f (t, x, ) as the probability of finding a particle with speed
near the point x, at time t.
If no collisions occur, the speed of each particle will remain constant in time. A particle with
speed and located at the point x at the initial time t = 0 will move to x + at a later time .
Therefore f (, x, ) = f (0, x , ). In this case, f provides a solution to the linear transport
equation
t f + x f = 0. (0.1)

The presence of collisions accounts for an additional quadratic term on the right hand side:

t f + x f = Q(f ). (0.2)

The equation (0.2) is the famous Boltzmann equation. A specific form of the collision kernel Q
will be derived in the next section.

1 - The collision kernel.

In the following, the particles of the gas will be modelled as hard spheres, all with the same
radius, that hit each other with perfectly elastic collisions.
We first observe that if two particles move along the same straight line and hit each other
with a perfectly elastic collision, their velocities will be exchanged after the impact (fig. 1).

figure 1

1
 More generally, if two spherical particles move in IR3 , by their angle of collision we mean the
unit vector n parallel to the segment joining the centers of the spheres at the instant of collision.
If the particles hit each other at an angle n, then the components of their velocities along n
will be exchanged, while the components perpendicular to n will remain the same (fig. 2). Calling
and the velocities of the particles before the collision, and , their respective velocities after
the collision, we thus have
(

= n ( ) n,

(1.1)
= + n ( ) n.

The exchange in the normal velocities is reflected by the identities

n = n , n = n .

figure 2

Notice that a change in the sign of n does not affect the linear transformation (1.1). Setting

. . + . | |
V = , 0 = , = , (1.2)
2 2

an alternative construction of the incoming and outgoing speeds is illustrated in fig. 3. Consider
the sphere
. 
S = IR3 ; | 0 | =

2
 n

*

*
0

n
figure 3

i.e. the surface of the ball B 0 , ) having the segment as a diameter. If the collision occurs at
an angle n, the speeds , of the two outgoing particles are obtained as the intersections of the
sphere S with the lines through , , parallel to n.
.
Let now S 2 = n IR3 ; |n| = 1 be the surface of the unit ball. To find the probability
that the collision occurs at an angle n S 2 , it is convenient to take a frame of reference in which
one particle is at rest, say = 0, while the other has speed V = . Instead of thinking of
two rigid spheres both with radius r > 0, we can equivalently think of a point particle hitting a
sphere with radius 2r. In this case (fig. 4), we see that the rate of collisions at an angle n is given
by |n V | dn, where dn denotes the surface element on the unit sphere S 2 .

dn

V= n

figure 4

Following the probabilistic interpretation, we see that the collision kernel has the effect of
replacing the two particles with speeds , with a cloud of particles whose speeds are distributed
over the sphere S having the segment as diameter. Actually, an elementary computation
shows that the speed of the outgoing particles is uniformly distributed over the sphere S . To see
this, referring to fig. 5 we let (, ) [0, ] [0, 2] be the spherical coordinates of the angle at

3
which the collision takes place, and let ( , ) be the spherical coordinates of the outgoing velocity
on the sphere S . Choosing a oordinate frame so that the x-axis is parallel to the vector ,
we thus have
n = (cos , sin cos , sin sin ),
0 = (cos , sin cos , sin sin ).
Notice that (1.1) yields
= 2 , = .
We observe that the probability that the collision angle n lies in the infinitesimal region described
by dA = [, + d] [, + d] is

Prob.(dA) = cos dn = cos sin dd ,

where dn is the area of the infinitesimal region. Hence, the area covered by the corresponding
speed is

meas(dA ) = 2 d sin d = 2 2d 2 sin cos d = 42 Prob.(dA) . (1.3)

According to (1.3), the area of any portion dA of the spherical surface S is proportional to the
probability Prob.{ dA }. This proves the uniform distribution of the outgoing velocities.
We remark that, as the intensity of collisions grows as | | and the surface of the sphere
is 42 = | |2 , it is clear that the above probability density per unit area decreases as 1 .

n


*

2
S
S

figure 5

We now consider one particular speed IR3 . The rate at which particles with speed will
hit other particles and thus change their speed is
Z Z
 
Q (f )() = n ( ) f ()f () dn d
IR3 S 2
Z (1.4)
= 2 | | f ()f ( ) d ,
IR3

where is some constant factor and dn denotes the surface element on the unit sphere. This is a
loss term, responsible for the decrease in the number of particles with speed .

4
Remark 1. As shown in fig. 4, the integral in (1.4) should actually range only over values of n in
the half sphere
. 
S = n S 2 ; n ( ) < 0 .
However, extending the integral over the whole sphere S we obtain exactly the same result, multi-
plied by two. It is thus convenient to integrate over S, and incorporate the factor 1/2 within the
constant of proportionality .

Next, we observe that new particles with speed may emerge from interactions. Indeed, since
the collision dynamics is perfectly reversible, a new particle with speed will be produced by the
interaction of two particles with speeds , as in (1.1), at an angle n. Indeed, such collision
would exchange once again the components along n of the two speeds, thus yielding the original
two particles with speeds , . Notice that, for any given IR3 and any fixed angle n S 2 , the
equations (1.1) uniquely determine the speeds of two particles whose collision yields the desired
result. The rate at which such collisions occur, between particles of speeds , , is proportional
to    
n ( ) f ( )f ( ) = n ( ) f ( )f ( ). (1.5)
The above identity is an immediate consequence of (1.1). Integrating over all vectors IR3 and
all angles n S 2 , we find that new particles with speed are created at the rate
Z Z
 
Q+ (f )() = n ( ) f ( )f ( ) dn d . (1.6)

IR3 S2

This is a gain term, responsible for the increase in the density of particles with speed .

By a coordinate rescaling, we can assume that the proportionality constant in (1.4) and (1.6)
is = 1. The Boltzmann equation in (0.2) thus takes the form
Z Z
 
t f + x f = (f f f f ) n ( ) dn d . (1.7)
IR3 S2

It is here understood that

f = f (t, x, ), f = f (t, x, ),
(1.8)
f = f (t, x, ), f = f (t, x, ),

with , given at (1.1) in terms of , and n.

2 - Collision Invariants.
In this section we consider the homogeneous Boltzmann equation
Z Z
.  
t f = Q(f ) = (f f f f ) n ( ) dn d (2.1)
IR3 S2

where f = f (t, ) is independent of x. We seek functionals of the form

Z
.
(f ) = () f () d (2.2)
IR3

5
which remain constant in time, along every solution of (2.1). Of course, this holds provided that
Z
() Q(f ) d = 0 (2.3)
IR3
for every velocity distribution f . Observing that the collision kernel Q replaces two particles with
speed , with a cloud of particles uniformly distributed on the sphere S , we see that (2.3)
holds for every f Cc if and only if
Z
() + ( ) = ( ) d (2.4)
S

where d denotes the normalized surface area on the sphere S . From (2.4) we deduce
() + ( ) = ( ) + ( ) (2.5)
whenever the segments and are diameters of the same sphere, i.e. whenever (1.1) holds.

Lemma 2.1. Let : IR3 7 IR be a C 2 function that satisfies (2.5). Then

() = a + b + c||2 (2.6)
for suitable constants a, c IR and b IR3 .

Proof. Set
. . . 1
a = (0) , b = (0) , c = (0) .
6
We need to show that
() = () a b c ||2 0 .
Notice that is a collision invariant and satisfies
(0) = 0 , (0) = 0 , (0) = 0 .
Consider any and let n be any unit vector perpendicular to (see fig. 6a). Using (2.5) with
= sn, = 0 and = + sn, for every s we obtain
(s n) + ( ) = (0) + ( + s n) .


Letting s 0+ we compute the directional derivative ( ) n = 0. This implies that = ||
must be a radially symmetric function. Its Taylor expansion at the origin vanishes up to second
order.
Next, fix any unit vector n, set = sn, = rn, and choose , so that the segment is
perpendicular to (see fig. 6b). For 0 < s < r we obtain


(r) + (s) = 2 (r, s) , (2.7)
where s r
2  2
rs r+s r 2 + s2
(r, s) = + = .
2 2 2
We now keep r fixed and let s vary. Differentiating (2.7) twice w.r.t. s, at s = 0 we find

1
(r, s) = () q s,
s 2 r +s
2 2
2
  
2
 r 1

(r, s) 
= =0
s 2
s=0 2 2r

The computation of the radial derivative thus yields (r/ 2) 0, hence 0, as claimed.

6
 n

*

*
sn
*
0 0 s r n
*

figure 6a figure 6b

If now is a collision invariant, then the functional in (2.2) is constant in time. These
invariants have a clear physical meaning:

1. Taking () 1 one obtains the conservation of

Z
f () d (mass) (2.11)
IR3

.
2. Taking () = ei , i = 1, 2, 3 (with e1 , e2 , e3 the standard basis in IR3 ), we obtain the
conservation of the vector Z
f () d (momentum) (2.12)
IR3

.
3. The choice () = ||2 /2 yields the conservation of
Z
||2
f () d (energy) (2.13)
IR3 2

3 - Maxwellian distributions
We investigate here the existence of velocity distributions f = f () which yield a vanishing
collision integral:
Z Z
.  
Q(f )() = (f f f f ) n ( ) dn d = 0 IR3 . (3.1)
IR3 S2

7
 .
Toward this goal, using the identity (2.8) with f = g and () = log f (), we obtain the
Boltzmann inequality
Z Z Z Z
1

Q(f ) log f d = (f f f f ) log f + log f log f log f |n V | dn d d
IR3 2 IR3 IR3 S 2
Z Z Z
1
= (f f f f ) log(f f /f f ) |n V | dn d d
2 IR3 IR3 S 2
0
(3.2)
because of the elementary inequality

(z y) log(y/z) 0 for all y, z > 0 . (3.3)

Notice that the equality holds in (3.3) if and only if y = z. As a consequence, the left hand side of
(3.2) is zero if and only if
f f = f f . (3.4)
.
Taking the logarithms of both sides, we see that (3.4) holds if and only if = log f satisfies
(2.9). The set of functions f that yield a vanishing collision integral are thus the Maxwellian
distributions, having the form

f () = exp a + b + c| 2 |
 (3.5)
= A exp | v|2

for some A, > 0, v IR3 . These represent velocity distributions which are in statistical equilib-
rium.
Returning to the original Boltzmann equation

t f + x f = Q(f ), (3.6)

multiplying both sides by log f and integrating w.r.t. we obtain

Z Z
d
f ln f d = [ft ln f + ft ] d
dt
Z Z
= Q(f ) (ln f + 1) d ( x f ) (ln f + 1) d (3.7)
Z Z
= Q(f ) ln f d x f ln f ) d .

This yields the macroscopic balance equation

t H(t, x) + x I(t, x) = S(t, x) , (3.8)

where Z Z Z
. . .
H= f ln f d , I= f ln f d , S= Q(f ) ln f d .
IR3 IR3 IR3

By (3.2), S 0. Hence, assuming suitable decay at |x| , we obtain

Z
d
H(t, x) dx 0 .
dt IR3

8
 In the homogeneous case where f is independent of x, (3.8) simply yields
d
H(t) 0 .
dt
This is Boltzmanns H-Theorem.

4 - Symmetries
We consider here various transformations that leave invariant the family of solutions to the
Boltzmann equation
Z Z
 
t f + x f = (f f f f ) n ( ) dn d . (4.1)
IR3 S2
Let f = f (t, x, ) be a solution of (4.1). Then, for every t IR, the function
.
f(t, x, ) = f (t t, x, ) (translation in time) (4.2)
provides another solution. The same is true of
.
f(t, x, ) = f (t, x x, ) (translation in space) (4.3)
3
for every x IR . Moreover, for every unitary 3 3 matrix U , the function
.
f(t, x, ) = f (t, U x, U ) (rotation in space) (4.4)
provides still another solution.

Next, consider a dilation of the form

f = f, t = t, x = x, = ,
so that
.
f(t, x, )
=
f ( t, x, ), (4.5)
with > 0. A direct computation yields
t f = t f ,
x f = + x f ,
Z Z
 
Q(f) = (f f ff ) n ( ) dn d = 2+4 Q(f ) .
IR3 S2

Therefore, if f = f (t, x, ) is a solution, the function

f(t, x, ) = f ( t, x, )
will be another solution of the Boltzmann equation (4.1) provided that
= = + 4 . (4.6)
We thus have a 2-parameter family of dilations, leaving invariant the solution set of (4.1). In
particular, taking = 1, = = 1, = 0 we find the new solutions
f(t, x, ) = f (t, x, ). (4.7)
Moreover, taking = 4, = 0, = = 1 we find
f(t, x, ) = 4 f (t, x, ). (4.8)

9
5 - The macroscopic balance equations
From the function f = f (t, x, ), one can obtain a macroscopic description of the gas, by
integration w.r.t. the variable . We thus define the density of mass of the gas as
Z
.
(t, x) = f (t, x, ) d , (5.1)

the density of momentum q = (q1 , q2 , q3 ),

Z
.
qi (t, x) = i f (t, x, ) d , (5.2)

and the energy density Z

. 1
w(t, x) = ||2 f (t, x, ) d . (5.3)
2
Moreover, we define the macroscopic velocity v = (v1 , v2 , v3 ) as
R
. qi i f d
vi = = R , (5.4)
f d

the momentum flow m = mij

Z
.
mi,j = i j f d (i, j = 1, 2, 3) , (5.5)

and the energy flow r = (r1 , r2 , r3 )

Z
.
ri = ||2 i f d . (5.6)

The microscopic velocity of a particle can now be expressed as a sum

= v +c,
.
where c = v represents the random deviation of the velocity of a single particle from the average
velocity v. Of course Z
c f d = 0 .

It is convenient to split the momentum flow as

mij = vi vj + pij (5.7)

with Z
pij = ci cj f d . (5.8)

One can write the energy density as a sum of a kinetic energy and an internal energy

1
w= |v|2 + e . (5.9)
2
10
Here e is the internal energy per unit mass, defined by
Z
. 1
e = |c|2 f d . (5.10)
2

Similarly, the energy flow can be written as

1  3
X
2
ri = vi |v| + e + i + vj pij , (5.11)
2 j=1

where = (1 , 2 , 3 ) is the heat-flow

Z
. 1
i = |c|2 ci f d . (5.12)
2

The evolution of the macroscopic quantities , q, w can be derived from the Boltzmann equation

t f + x f = Q(f ), (5.13)

multiplying both sides by the collision invariants

||2
0 () 1 , i () = ei i (i = 1, 2, 3), 4 () = .
2
Observing that Z
j () Q(f ) d = 0 j = 0, 1, 2, 3, 4,

from (5.13) we obtain

Z X3 Z

j f d + i j f d = 0 .
t i=1
xi

In the cases j = 0, j = 1, 2, 3 and j = 4 respectively, one obtains

conservation of mass:
X3

+ (vi ) = 0 , (5.14)
t i=1
x i

conservation of momentum:

X3

(vj ) + (vi vj + pij ) = 0 j = 1, 2, 3, (5.15)
t i=1
xi

conservation of energy:

 |v|2  X3
 |v|2  X3
+ e + vi + e + i + vj pij = 0 . (5.16)
t 2 i=1
xi 2 j=1

11
 The above equations are in conservation form, showing that the integral of mass, momentum
or energy over a given set can change in time only because of a flux across the boundary . In
particular, if this flux vanishes, then the total mass
Z
.
M= dx ,

the total momentum Z

.
Q= v dx ,

and the total energy Z 

|v|2 
.
E= + e dx
2
inside remain constant in time.

The equations (5.14)-(5.16) represent a system of 5 scalar conservation laws, which however
is not closed. Indeed, the flux functions involve the additional quantities pij and i which depend
on higher moments of the distribution f (t, x, ), according to (5.8), (5.12). In order to obtain a
hyperbolic system of conservation laws, one can formally proceed as follows. Consider the equation

1
t f + x f = Q(f , f ). (5.17)

When the relaxation parameter > 0 is very small, we expect that the function f will converge
pointwise to an equilibrium distribution f such that Q(f, f) = 0. According to our previous
analysis, this must be a Maxwellian distribution

f(t, x, ) = A exp | v|2 .

For each (t, x), the coefficients A = A(t, x), = (t, x) and v = v(t, x) can be uniquely determined
by the condition that the densities of mass, momentum and energy in f(t, x, ) should be the same
as for f (t, x, ), namely Z Z

f (t, x, ) d = f (t, x, ) d ,
IR3 IR3
Z Z
i f(t, x, ) d = i f (t, x, ) d i = 1, 2, 3 ,
IR3 IR3
Z Z
||2 ||2
f (t, x, ) d = f (t, x, ) d .
IR3 2 IR3 2
A direct computation yields
 3/2
3 4
= , A= e ,
4e 3

while v is given by (5.4). For the Maxwellian distribution f, the stress tensor is diagonal
2
pij = 3 e if i = j,
(5.18)
0 if i 6= j,

12
while the heat flow vector vanishes:

i = 0 i = 1, 2, 3. (5.19)

Using (5.18) and (5.19) we can close the system of equations (5.14)-(5.16) and obtain

X3

+ (vi ) = 0 , (5.20)
t i=1
x i

 2 
X3

(vj ) + vi vj + e = 0 j = 1, 2, 3, (5.21)
t i=1
xi 3

  2 
 |v|2  X 5 
3
|v|
+ e + vi + e = 0. (5.22)
t 2 i=1
xi 2 3

The equations (5.20)(5.22) represent a hyperbolic system of five conservation laws in three space
dimensions. For this system, an outstanding open problem is to prove the global existence of
entropy admissible weak solutions. One conjectures that these solutions can be obtained from
solutions to the Boltzmann equation (5.17), letting the relaxation parameter tend to zero. This
would provide a rigorous justification to the formal arguments outlined in this section.

6 - The spatially homogeneous case

In this section we look at the space homogeneous case, where the velocity distribution f does
not depend on x. The Cauchy problem thus takes the simpler form

t f = Q(f ) , (6.1)

f (0, ) = f0 () . (6.2)
We will show that the above problem can be reformulated as a continuous differential equation in
a suitable Banach space. Global existence and uniqueness of solutions will thus follow from general
O.D.E. theory.
It will convenient to work in a space a weighted integrable functions, with norm
Z
.  
kf k1,s = (1 + 2 )s/2 f () d . (6.3)

For s 2 we thus define the space of velocity distributions having moments up to order s:
. 
L1s = f : IR3 7 IR , kf k1,s < .
.
Here and in the sequel, with slight abuse of notation we write 2 = ||2 . For future use, we record
the interpolation inequality q
khk1,3 khk1,4 khk1,2 . (6.4)

13
To prove (6.4), we can assume h 0 and consider the absolutely continuous measure such that
d = h()d. Using Cauchys inequality one obtains
Z Z
khk1,3 = (1 + 2 )2/3 h() d = (1 + 2 ) (1 + 2 )1/2 d

k1 + 2 kL2 () (1 + 2 )1/2 L2 ()
Z 1/2 Z 1/2
2 2 2
= (1 + ) h() d (1 + ) h() d
1/2 1/2
= khk1,4 khk1,2 .

Theorem 6.1. Assume that f0 L14 and f0 () 0 for a.e. . Then the Cauchy problem (6.1)-
(6.2) has a unique solution, defined for all t 0. The map t 7 f (t) is continuously differentiable
as a map with values in L12 . Moreover, for every t 0 there holds
Z Z Z Z
2



f (t, ) d = f0 () d , 1 + f (t, ) d = 1 + 2 f0 () d . (6.5)

Toward a proof we observe that, by a possible rescaling of the time variable, it is not restrictive
to assume the normalization Z
f () d = 1 . (6.6)

On the space L12 we consider a closed convex subset of the form

 Z
= f L12 ; f () 0 for a.e. , f () d = 1 ,
Z Z  (6.7)
2



2 2
1 + || f () d 2 , 1 + || f () d 4

for some constants 2 , 4 In order to apply a general theorem on O.D.Es in Banach spaces, we
need to show that the collision operator f 7 Q(f ), as a map from into L12 , satisfies the following::

Holder continuity.

One-sided Lipschitz condition.

Tangency condition.

We begin by proving some estimates on the collision kernel, in the norms k k1,s .

Lemma 6.2. The map f 7 Q(f ) is uniformly Holder continuous on L12 , when restricted to the
domain .

Proof. We shall consider the loss term Q and the gain term Q+ separately. Let f, g .
Observing that
 

n ( ) | | || + | | = ||2 + | |2 + 2||| | 1/2 (1 + 2 )1/2 (1 + 2 )1/2 , (6.8)

14
and using the interpolation inequality (6.4) we compute

Z Z Z

 
Q (f ) Q (g, g) = 1+ 2 n ( ) |f f gg | dn d d
1,2
IR3 IR3 S2
ZZ    
2 (1 + 2 )3/2 (1 + 2 )1/2 f () g() f ( ) + g()f ( ) g( ) dd
 
= 2 f g 1,3 kf k1,2 + f g 1,2 kgk1,3 (6.9)
 1/2 
1/2
2 f g 1,4 f gk1,2 kf k1,2 + f g 1,2 kgk1,3
 
1/2 1/2 1/2
2 (22 )1/2 2 kf gk1,2 + 4 2 kf gk1,2 .

The estimate for the gain terms is entirely similar:

Z Z Z

 
Q+ (f ) Q+ (g, g) = 1+ 2 n ( ) |f f g g | dn d d
1,2
IR3 IR3 S2
Z Z Z
 2 2  
(1 + ) + (1 + ) n ( ) |f f g g | dn d d
IR3
Z Z ZIR
3 S2 (6.10)

 
2 1 + n ( ) |f f gg | dn d d
2

 
1/2 1/2 1/2
4 (22 )1/2 2 kf gk1,2 + 4 2 kf gk1,2 .

We have here used the symmetry of the collision kernel Q+ w.r.t. the variables , . Together,
(6.9) and (6.10) establish the lemma.
We remark that, taking g = 0, the above computations yield

1/2 3/2
Q(f ) 6 kf k1,3 kf k1,2 6 4 2 . (6.11)
1,2

Next, we show that the collision operator Q satisfies a one-sided Lipschitz estimate. Consider
the following semi-inner product on the space L12 :

.
hf, wi1,2 = kf k1,2 [f, w]1,2 ,

where
Z
.
[f, w]1,2 = (1 + 2 ) signf () w() d .

We recall that the collision operator Q replaces two particles with speeds , with a cloud of
particles with speeds , uniformly distributed on the sphere having as diameter. In the
following computation we use the symmetry of the integral expressing Q(f ) w.r.t. the variables
2 2
and . Moreover, we use and the identity 2 + 2 = + and the upper bound (6.8) for the

15
distance | |. This yields
 
f g , Q(f ) Q(g) 1,2
Z Z Z 
  2 2
 
 n ( ) (1 + ) + (1 + ) f ()f () g()g()

IR3IR 3 S2
h 
2

2

i

(1 + ) sign f () g() + (1 + ) sign f ( ) g( ) f ()f () g()g() dn d d
ZZ 

1  

= 2 | | (1 + 2 ) + (1 + 2 ) (f + g)()(f g)( ) + (f + g)( )(f g)()
2
h

i
(1 + 2 ) sign f () g() + (1 + 2 ) sign f ( ) g( )
1h i
(f + g)()(f g)( ) + (f + g)()(f g)() d d
2
ZZ   
 
= 2 | | (1 + 2 ) (f + g)()(f g)( ) + (f + g)( )(f g)()

h i
2
(1 + ) sign f () g() (f + g)()(f g)() + (f + g)( )(f g)() d d
ZZ h i
2
2 | | (1 + ) (f + g)()|f g|( ) + (f + g)( )|f g|()
h i
+ (f + g)()|f g|( ) (f + g)( )|f g|() d d
ZZ
4 (1 + 2 )1/2 (1 + 2 )1/2 (1 + 2 ) (f + g)()(f g)( ) d d

= 4 kf + gk1,3 kf gk1,2
84 kf gk1,2 .
(6.12)

We now consider the tangency condition. Since the collision operator preserves total mass and
energy, for every f and h > 0 we have
Z   Z   Z
f ()+h Q(f )() d = 1 , (1+ ) f ()+h Q(f )() d = (1+ 2 )f () d 2 .
2

.  (6.13)
Moreover, the vector Q(f ) is tangent to the positive cone + = f L12 ; f () 0 simply
because Q(f )() 0 whenever f () = 0.
The forthcoming analysis will establish the a priori bound kf k1,4 4 , provided that the
constant 4 is suitably large. This will require further estimates on the collision kernel.

Lemma 6.3 (Povzner). For every f L1s with f 0 and s 2, one has the estimate
Z

s/2
1 + ||2 Q(f ) d Cs kf k1,s kf k1,2 (6.14)

for some constant Cs , with Cs 0 as s 2.

Proof. We first observe that, for any p 1 there exists a constant Cp such that
ap + bp (a + b)p ap + bp + Cp (a1/2 bp(1/2) + ap(1/2) b1/2 ) (6.15)

16
for any numbers a, b 0. Moreover, Cp 0 as p 1.
If now , , , are related as in (1.1), then by (6.15)
 s/2  s/2
(1+ 2 )s/2 + (1 + 2 )s/2 1 + 2 + 1 + 2 = 1 + 2 + 1 + 2
 
(1 + 2 )s/2 + (1 + 2 )s/2 + Cs/2
(1 + 2 )1/2 (1 + 2 )(s1)/2 + (1 + 2 )(s1)/2 (1 + 2 )1/2
(6.16)
By the action of the collision kernel Q, a couple of particles with speeds , is replaced by
particles whose speeds are uniformly distributed on the sphere S . Using (6.8) and then (6.16)
to cancel the leading order terms, we thus obtain
Z Z ZZ


2 s/2
 
1 + || Q(f ) d n ( ) f ()f ()
 
 
(1 + 2 )s/2 + (1 + 2 )s/2 (1 + 2 )s/2 (1 + 2 )s/2  dn d d
ZZ
(1 + 2 )1/2 (1 + 2 )1/2 f ()f ()
h i

Cs/2 (1 + 2 )1/2 (1 + 2 )(s1)/2 + (1 + 2 )(s1)/2 (1 + 2 )1/2 d d
Cs kf k1,s kf k1,2 ,

with Cs = 2 Cs/2 .

The estimate (6.14) provides an a-priori bound on the L1s norm of a solution. Namely, by
mass and energy conservation there holds

f (t) 1 = f0 1 f (t) = f0 (6.17)
L L 1,2 1,2
for all t 0. In turn, (6.14) implies
d
f (t) Cs kf0 k1,2 f (t) ,

f (t)

1,s 1,s 1,s
exp Cs kf0 k1,2 t kf0 k1,s . (6.18)
dt

A sharper estimate is now given. First, consider the case of a particle with speed 0 colliding
with a particle with speed (each particle having unit mass). The result is a cloud of particles
(with total mass = 2) uniformly distributed along the sphere S0 , having a diameter joining the
points 0 with , and radius R = ||/2. Setting
r() = 2R cos , x() = R (1 + cos 2) ,
we now compute the momentum of order s of this cloud of particles (fig. 7)
Z
. 2
Is = | |s d
4R2 S0
Z /2
1  s  
= 2
r() 2R dx()
2R 0
Z /2
1 (6.19)
= (2R cos )s 2R 2R sin 2 d
2R2 0
Z /2
= (2R)s (4 cos )s+1 sin d
0
4
= ||s .
s+2

17
 r ()

2
0
R x ()

figure 7

Notice that when s = 2 we have Is = ||s . Otherwise

Is < ||s s > 2.

This reflects the fact that collisions conserve the energy norm, but dissipate higher norms L1s with
s > 2. By a simple rescaling argument we obtain
 
Lemma 6.4. For each s > 2 and any 4/(s
+ 2) , 1 , one can find a constant large enough
so that the following holds. If || 1 + | | , then
Z
2  
(1 + | |2 )s/2 d (1 + 2 )s/2 + (1 + 2 )s/2 , (6.20)
| |2 S

where d denotes the surface area on S .

Proof. If the conclusion of the lemma fails, then there exists a sequence of couples , with

| |
| | 1 , ,
| |

such that (6.20) fails for every . Set

. 2 . 2
= , = .
| | | |

As , by possibly taking a subsequence we can assume that converges to some vector .

Clearly 0. Hence the sphere whose diameter is the segment joining with approaches

18
a sphere with unit radius, whose diameter is the segment with endpoints 0, . Using (6.9) with
R = 1 we thus find
Z
1 2

2 s/2
lim 1 + | | d
1 + | |2 )s/2 | |2
S

Z
1 2 4
= s 2
| |s d = > .
|| || S0 s+2

This contradiction establishes the lemma.

Lemma
6.5. Let t 7 f (t) be a solution of the homogeneous Boltzmann equation (6.1), with
f (t) 2 and f (t) 1,s locally bounded, for some s > 2. Then there exist a constant s
1,2
depending only on s and 2 such that
d
f (t) 0
1,s
whenever kf (t)k1,s s . (6.21)
dt

0 *

figure 8

The situation is illustrated in fig. 8. If the norm kf k1,2 is small, most particles have small
speed, contained in a neighborhood V of the origin. However, if the norm kf k1,s is very large,
there must be a few particles having very large speed. By the interactions between these high
speed particles and other slow particles, the L1s norm is dissipated, as shown in Lemma 6.4.

Proof of Lemma 6.5. If t 7 f (t) is a solution, its L12 norm is clearly constant in time. Assume
kf k1,2 2 . For a given s > 2, choose constants < 1 and according to Lemma 6.4. Define the
radius
.
= 22 .
R
Recalling the normalization f ()d = 1, this implies
Z
1
f ( ) d > .
| | 2

19
In the integral expression for Q(f ), we now split the domain IR3 IR3 = , setting
.  .
= (, ) ; , (1 + ) , = IR3 IR3 \ .

Using the two previous lemmas we find

ZZ Z
d f (t)k1,s =
 
n ( ) f ()f ()
dt 2
h S i
(1 + ) + (1 + 2 )s/2 (1 + 2 )s/2 (1 + 2 )s/2 dn d d
2 s/2

ZZ (6.22)


Cs f (t) 1.2 f (t) 1,s (1 ) | |(1 + 2 )s/2 f ()f () dd

Z
1 (1 + 2 )1/2
Cs f (t) 1.2 f (t) 1,s (1 + 2 )s/2 f () d .
2 ||(1+) 2

We now apply Jensens inequality

Z  Z
u d (u) d
X X

to the convex function (y) = y (s+1)/s and to the probability measure d = f ()d, with u() =
(1 + 2 )s/2 . This yields
Z (s+1)/s Z
2 s/2
(1 + ) f () d (1 + 2 )(s+1)/2 f () d
Z

(s+1)/2
(1 + 2 )(s+1)/2 f () d + 1 + 2 (1 + )2 .
||(1+)
(6.23)
Using (6.23) to provide a lower bound on the negative term on the right hand side of (6.22), we
obtain
d 1 (s+1)/s 1
(s+1)/2
kf k1,s Cs kf k1,2 kf k1,s kf k1,s + 1 + 22 (1 + 2 ) . (6.24)
dt 4 4
The left hand side of (6.24) is 0 provided that kf k1,s is large enough.

Recalling that < 1 and s > 2, the equation (6.24) can be compared with the O.D.E.

z = a z (s+1)/s + bz + c (6.25)

with
1
(s+1)/2
a= , b = Cs kf k1,2 , c = 1 + 22 (1 + 22 )2 .
4
For z large, the right hand side of (6.25) is < (a/2)z (s+1)/s . By a comparison with the O.D.E.
a
z = z (s+1)/s ,
2
we conclude that every solution will satisfy

f (t) K (1 + ts ) s > 2, t > 0, (6.26)
1,s

20


where the constant K = K s, kf0 k1,2 depends only on s and on the L12 norm of the initial data.

Corollary 6.6. For any given 2 , there exists a constant 4 such that the following holds. If

kf k1,2 2 , kf k1,4 4 ,

then Z
(1 + 2 )2 Q(f )() d 0 . (6.27)

Thanks to the above estimates, we can now establish Theorem 2 by applying a general existence
theorem for O.D.Es in a Banach space. Indeed, we have shown that the map f 7 Q(f ) is uniformly
bounded and Holder continuous on the closed convex set L12 . Moreover, it satisfies the one-
sided Lipschitz condition

 .  
f g , Q(f ) Q(g) 1,2 = kf gk1,2 f g , Q(f ) Q(g) 1,2 84 kf gk21,2 ,

together with the tangency condition

1

lim dist. f + h Q(f ) ; = 0 f .
h0+ h

By applying Theorem A1 in the Appendix, we now obtain the global existence of a unique solution
to the Cauchy problem (6.1)-(6.2).

7 - Pointwise bounds
In this chapter we consider the case of bounded initial data. Our goal is to establish uniform
L bounds valid for all times. We always assume that the density is normalized as in (6.6).

Theorem 7.1. In the same setting as Theorem 6.1, assume that the initial data is bounded and
has finite entropy, so that
Z
.
kf0 kL < , H0 = f0 () ln f0 () d < . (7.1)

Moreover, assume that Z

sup f () d < ,
E E
3
the supremum being taken over all two-dimensional planes E IR . Then the solution remains
uniformly bounded, i.e. f (t, ) L C for all t 0.

Proof. We proceed in several steps.

1. We claim that there exists a constant c0 > 0 depending only on the entropy H0 , such that
Z
| | f ( ) d c0 (7.2)

21
for all IR3 , t 0. Indeed, we can find r1 > 0 sufficiently small such that r [0, r1 ] and
Z
1
f ( ) d
| |r 2

together imply Z
1
f ( ) d . (7.3)
| |r, ,f ( )>r 1 4
If (7.3) holds, then
Z Z
| ln r|
H0 f ( ) ln f ( ) d f ( ) | ln r| d .
| |r, f ( )>r 1 4
. 
Therefore, setting r0 = min r1 , e4H0 , for every IR3 the above estimates yield
Z
1
f ( ) d ,
| |r0 2
Z Z
r0
| | f ( ) d | | f ( ) d ,
| |>r0 2
proving (7.2).

E

figure 9

2. We shall need an alternative expression for the gain term (fig. 9). In order to produce a particle
with speed , one can first choose a particle with arbitrary speed . This has to interact with
some other particle whose speed lies in the plane E through perpendicular to . This
motivates the formula
Z Z !
1
Q+ (f )() = f ( ) d f ( ) d .

(7.4)
| | E

22
The factor | |1 comes from (1.6) through a change of variables. A rigorous justification of
(7.4) requires some care, because one cannot regard IR3 and IR2 as independent variables.
Indeed, the range of E itself depends on .
To establish (7.4) we thus proceed as follows. Fix IR3 and let range in the fixed

plane E passing through and perpendicular to (see fig. 10a). For each triple (s, n, )
+ 2
IR S E we consider the two speeds
.
= s n , = E , (7.5)

where E denotes the perpendicular projection on the plane E . When = , the determinant
7 ( , ) defined at (7.5) is computed as (see fig. 10b)
of the map (s, n, )

  !  
( , ))
(s n , E ) (sn)
det = det = det = s2 .
(s, n, ) (s, n, ) (s, n)

Therefore, from (1.6) (with = 1) it follows

Z Z
1
Q+ (f )() = s f ( ) f ( ) d d ,
IR3 E s2

proving (7.4).

~


*
= s n
d

dn
0 1 s

figure 10a figure 10b

We now consider the quantities

Z Z
. .
J () = f ( ) d , I(E) = f () d .
| | E

From (7.4) it follows

Q+ (f )() J () sup I(E) , (7.6)
E

23
where the supremum is taken over all planes E IR3 .

3. For any plane E and > 0, denote by E the set of points whose distance from E is . As
usual, S is the sphere having the segment as diameter. Using (7.2) we compute
Z
d  
I(E) = Q+ (f )() Q (f )() d
dt
ZEZ   Z Z
f ()f () 1

lim meas S E dd 2 f () | | f ( ) d d
| | 0 2 E
ZZ Z
f ()f () dd 2c0 f ()d
E


(1 2c0 I(E) .
(7.7)
Indeed, the area of the portion of the sphere S which is contained inside E satisfies

. | |
meas S E 2 2R , R= .
2

.
4. Setting 0 = (1 + 2 )/2 and writing d for the element of area on the sphere S1 2 , recalling
(7.2) we compute
Z
d Q+ (f )( ) Q (f )( )
J () = d
dt | |
ZZ Z !
1 1

d( ) f (1 )f (2) d1 d2
|1 2 | S1 2 | |
Z Z
f ( ) (7.8)
2 d | | f ( )d
| |
ZZ Z !
1 1

d( ) f (1 )f (2 ) d1 d2 2 J () c0
|1 2 | S1 2 | 0 |

= 4 2c0 J () .

5. By (7.7) and (7.8), the quantities

. .
Jb = sup J () , Ib = sup I(E) ,
E

remain uniformly bounded for all times t 0. By (7.6) and (7.2) we have
d
f (t, ) = Q+ (f )() Q (f )() Jb Ib 2c0 f (t, ) . (7.9)
dt
This differential inequality implies
( )
Jb Ib
f (t, ) max , kf0 kL
2c0

for evry t 0 and a.e. IR3 .

24
Appendix: O.D.E. theory in Banach spaces
We review here some existence and uniqueness results for abstract O.D.Es, which were used
in Section 6.
Let E be a Banach space with norm k k, and let g : E 7 E be a Lipschitz continuous
mapping, so that
g(u) g(v) L ku vk u, v E . (A.1)
It is well known that in this case the Cauchy problem

u = g(u) , u(0) = u0 (A.2)

has a unique solution, which can be obtained as the fixed point of the Picard integral operator
u 7 Pu, with Z t


Pu(t) = u0 + g u(s) ds . (A.3)
0

Moreover, solutions depend continuously on the initial data. Indeed

d


u(t) v(t)
g u(t) g v(t) L u(t) v(t) . (A.4)
dt
Hence, by Gronwalls lemma,

u(t) v(t) eLt ku0 v0 k t 0. (A.5)

In order to achieve (A.5), the assumption of Lipschitz continuity can be greatly relaxed. For
example, if E is a Hilbert space with inner product h, i, assume that g is a continuous mapping
such that

u v , g(u) g(v) L ku vk2 u, v E . (A.6)
Then the computations at (A.4) can be replaced by

d D

E
u(t) v(t) 2 2 u(t) v(t) , g u(t) g v(t) 2L u(t) v(t) 2 . (A.7)
dt
Hence, for t 0,

u(t) v(t) 2 e2Lt ku0 v0 k2

and (A.5) still holds.

We now examine how the one-sided Lipschitz estimate (A.6) can be reformulated in the context
of a general Banach space. Observe that, still in a Hilbert space,

hu, wi hu + sw , u + swi1/2 hu, ui1/2 ku + swk kuk

= lim = lim . (A.8)
kuk s0 s s0 s

Notice that the right hand side of (A.8) is well defined also in an arbitrary Banach space. Indeed,
for every fixed u, w E, the map s 7 ku + swk is convex, being the composition of a linear
function with a convex one. Moreover, it is globally Lipschitz continuous because

ku swk ku s wk |s s k kwk .

25
Therefore, there exists the (possibly distinct) limits

. ku + swk kuk ku + swk kuk .

[u, w] = lim lim = [u, w]+ . (A.9)
s0 s s0+ s

Motivated by (A.8), for u, w E we thus define the upper and lower semi-inner products
. .
hu, wi+ = kuk [u, w]+ , hu, wi = kuk [u, w] .

In a Hilbert space the norm is smooth and we clearly have hu, wi+ = hu, wi = hu, wi. In a general
Banach space there holds

kuk kwk hu, wi hu, wi+ kuk kwk .

We are now ready to formulate a general result on the well-posedness of the Cauchy problem
for an O.D.E. in a Banach space. In the following, the distance of a point v from the set is of
course defined as
.
dist. (v; ) = inf kv k .

Theorem A1 (Global existence and uniqueness of solutions). Consider:

A Banach space E, with norm k k and lower semi-inner product h, i.

A closed, convex subset E.

A bounded, continuous mapping g : 7 E satisfying the one-sided Lipschitz condition


u v , g(u) g(v) L ku vk2 u, v (A.10)

and the tangency condition

1

lim inf dist. u + hg(u) ; = 0 u . (A.11)
h0+ h

In the above setting, for every initial data u0 , the Cauchy problem

d
u = g(u) , u(0) = u0 (A.12)
dt

has a unique solution t 7 u(t) , defined for all t 0.

Proof. Fix an arbitrarily large time interval [0, T ]. Following [M], the solution will be constructed
as a limit of polygonal approximations.

1. By the boundedness assumption, there exists a constant M such that

g(u) M u . (A.13)

26
 g

u(t 2 )
u(t1)
u u(t)
u0
u(t )
u(T)

u
u

u+hg(u)
figure 11

Let > 0 be given. For any u , by the continuity of g and the tangency condition (A.11), we
can find u and h > 0 such that

u u h g(u)
, g(u) g(v) if ku vk (M + 1)h .
h 2 2
As a consequence, the linear map
s 7 v(s) = u + s(u u) s [0, h]
satisfies


u(s) , u(s) g u(s) (A.14)
for s [0, h].
2. Given > 0, we construct an -approximate solution u : [0, ] 7 by transfinite induction
(fig. 11). Namely, by Step 1 there exists t1 > 0 and affine map u : [0, t1 ] 7 such that (A.14)
holds for t [0, t1 ].
Next, assume that a piecewise affine map u : [0, [ 7 has been constructed, and satisfies
(A.14) for a.e. t < . Here = sup t . Since g is bounded, the map u is Lipschitz continuous,
hence the value u( ) = limt u(t) is certainly well defined. If = T we are done. Otherwise,
we have two cases:

(i) If is a limit ordinal, we simply define

t = sup t , u(t ) = lim u(t) .
tt

27
 (ii) If is a successor ordinal, then we apply step 1 with u = u(t1 ).

This allows us to extend the -approximate solution to a larger interval [0, t ]. Since t+1 > t
for every , we must have t T for some countable ordinal .

3. Finally, consider a sequence of approximate solutions u , with 0. We claim that this

sequence is Cauchy. Indeed, let u , v be two -approximate solutions. By our previous construction
we can find a decomposition of the form
!
[
[0, T ] = I N ,

where the I are countably many open intervals where u and v are both affine, and N has measure
zero. The distance function t 7 u (t) v (t) is Lipschitz continuous. Moreover, recalling (A.9),
the dissipativity assumption (A.10) implies

d h i
u (t) v (t) = u (t) v (t) , u (t) v (t)
dt
h

i
u (t) v (t) , g u (t) g v (t) + 2


L u (t) v (t) + 2 .

for a.e. t [0, T ]. In turn, this implies

Lt
u (t) v (t) 2 e .
L

As 0, we thus have the convergence u u uniformly on [0, T ]. The function u() is clearly a
solution to the Cauchy problem, because of the continuity of g.

Remark. As usual, the boundedness assumption on g can be replaced by the assumption of

sub-linear growth

g(u) C 1 + kuk .

Basic References
[C] C. Cercignani, The Boltzmann Equation and its Applications, Springer, New York, 1988.

[CIP] C. Cercignani, R. Illner and M. Pulvirenti, The Mathematical Theory of Dilute Gases, Springer-
Verlag, 1994.

[M] R. Martin Jr., Functional Analysis and Differential Equations in Banach Spaces, Wiley Inter-
science, 1975.

[V] C. Villani, A review of mathematical topics in collisional kinetic theory, in Handbook of Math-
ematical Fluid Mechanics, S. Friedlander and D. Serre Eds., Elsevier Science, 2002.

28