The Cauchy problem for the Boltzmann equation

Alexander Bastounis

Thomas Holding

Vittoria Silvestri

January 30, 2013

Introduction

New mathematical results do not emerge often from existing tools and techniques. The Boltzmann
equation is one of the core equations of mathematical physics, and its Cauchy problem was and still is
one of the most important open problems in the field. The new concepts that are key to the 1989 result
by DiPerna and Lions [1] are the notions of renormalised solutions of transport equations and velocity
averaging. This allows for a proof of the global existence of weak solutions via compactness arguments
without any a priori estimates on the derivatives. The regularity and uniqueness of these solutions is still
an open problem. As is common when a new method is developed, further results came quickly. In this
project we will present the proof of weak solutions to the Boltzmann equation and a later application
of the same tools by DiPerna and Lions[3] to ODEs to extend the Cauchy-Lipschitz theorem to some
non-Lipschitz forcefields in a slightly weaker sense.

Contents
1 Introduction

2 Linear transport equations

2.1 Basic overview . . . . . . . . . . . . . . . . . . . .
2.1.1 Characteristics and transportation . . . .
2.1.2 Mild solutions and Duhamels principle . .
2.1.3 Algebraic properties and conservation laws
2.1.4 Renormalised solutions . . . . . . . . . . . .
2.2 The setting . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Function Spaces and topologies . . . . . . .
2.2.2 Distributional solutions . . . . . . . . . . .
2.3 Existence and uniqueness . . . . . . . . . . . . . .
2.3.1 Smooth coefficients . . . . . . . . . . . . . .
2.3.2 Rough coefficients . . . . . . . . . . . . . .
2.4 Velocity averaging results . . . . . . . . . . . . . .
2.4.1 Characteristics . . . . . . . . . . . . . . . .
2.4.2 Velocity Averaging . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

2
. 2
. 2
. 3
. 4
. 4
. 5
. 5
. 6
. 7
. 8
. 8
. 9
. 9
. 10

3 The Cauchy Problem for the Boltzmann Equation

3.1 The equation . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Solution concepts for the Boltzmann equation
3.1.2 Main theorem . . . . . . . . . . . . . . . . . .
3.2 Short sketch of the proof . . . . . . . . . . . . . . . .
3.3 A priori estimates . . . . . . . . . . . . . . . . . . .
3.3.1 Entropy identity . . . . . . . . . . . . . . . .
3.4 Truncated problems . . . . . . . . . . . . . . . . . .
3.5 Weak compactness results . . . . . . . . . . . . . . .
3.5.1 Weak compactness of (f n ) . . . . . . . . . . .
3.5.2 Weak compactness of the collision term . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

13
13
14
15
15
16
18
20
24
24
25

3.6
3.7

3.5.3 Temporal regularity of f . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

Velocity averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Exponential formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4 Rough coefficients
4.1 Another use for renormalised solutions . . . . .
4.1.1 Cauchy-Lipschitz with rough coefficients
4.1.2 Strategy . . . . . . . . . . . . . . . . . .
4.1.3 Techniques and ideas used . . . . . . . .
4.2 A proof using the stability result . . . . . . . .
4.3 Proof of stability result . . . . . . . . . . . . .
4.3.1 Stability . . . . . . . . . . . . . . . . . .
4.3.2 Pointwise . . . . . . . . . . . . . . . . .
4.3.3 On compact time sets . . . . . . . . . .
4.3.4 Existence and uniqueness . . . . . . . .
4.4 Further remarks . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

5 Conclusions

40

A Appendix
A.1 Weak* compactness . . . .
A.2 Weak compactness . . . . .
A.3 Strong compactness . . . .
A.4 Continuity in Banach spaces

35
35
35
36
36
36
39
39
39
39
40
40

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

40
40
41
41
41

Linear transport equations

This section will give a refresher on basic linear transport theory, some notions of solution, and then
some existence results. To begin with the presentation will be unrigourous, with no spaces or regularity
specified. The existence and uniqueness results are sketched from [3], and the equivalence result is
combined from [3] and [2].

2.1

Basic overview

The equation we will look at is (with the Einstein summation convention),

u
u
+ bi (t, x)
+ c(t, x)u = f (t, x, u)
t
xi
u(0, x) = u0 (x)
u(t, x) : (0, T ) R

(1)
N

To begin with we will look at when c = 0 and f = f (t, x).

2.1.1

Characteristics and transportation

u
Locally we can view b as constant, and we see that u
t + bi xi is the directional derivative along (1, b).
The equation therefore states that u remains constant along this line. We can then solve the equation
by finding the value of u along this line backwards in time. Globally we need to stitch up all the locally
constant lines to form the characteristic curves.
The characteristic equations are

X(s; t, x) : R1+N RN
X = b(t, X)
X(s; s, x) = x

where X =
obeys

X
t .

If X(t, x) solves this and u solves (1), then we observe that for each fixed s, u(t, X(t, x))



u Xi
u 
u
u 
d
+
=0
[u(t, X(t, x))] =
=
+
b
i
dt
t (t,X(t,x)) Xi t
t
xi (t,X(t,x))

Noting that at t = s = 0 we have u(0, X(0; 0, x)) = u(0, x) = u0 (x), we see that the value of u
at (t, X(0; t, x)) is u0 (x). Thus by the group property (X(s; t + , x) = X(s; t, X(t; , x))) and that
each X(s; t, ) is a diffeomorphism, we have that if y = X(0; t, x) then x = X(t; t, y) and hence
u(t, y) = u0 (X(t; t, y)). This gives us the solution,
u(t, x) = u0 (X(t; t, x))

(2)

X(t; t, x) means we look back along the characteristic curve that starts at (t, x) to find where it
crosses the plane t = 0. In this way, the transport equation rearranges the function u(t, ) over time,
transporting mass along the characteristics.
2.1.2

Mild solutions and Duhamels principle

(2) has meaning for any measurable function u0 , so we can extend our notion of solution to a far more
general space. To do this we introduce the solution operator St .
Definition 1 (Solution Operator). The solution operator to (1) with c = f = 0 is an operator on
measurable functions RN RN defined by
[St g](x) = g(X(t; t, x))
If b depends only on x, then S obeys the group property St Ss = St+s .
From now on we will write g ] (t, x) to mean St g, moving our equation along the characteristics. In
general this is not a group or even a semigroup due to the time dependence of b.
Definition 2 (Mild solution). The mild solution to

u
t

u
+ bi x
= 0, u(0, x) = u0 with u0 measurable is
i

u]0 .
All classical solutions are mild solutions.
We now look at when c is not identically zero. Along the characteristics we have
d ]
u + c] u] = 0
dt
we can solve this via the integrating factor method to give
 Z t

]
]
]
u (t, x) = u0 (x) exp
c (s, x) ds
0

which we can use to define a new solution operator.

Definition 3 (Solution Operator). The solution operator to (1) with f = 0 is defined by
 Z t

]
]
[St g](x) = g (x) exp
c (s, x) ds
0
[

We will now define g using this. This is an sort of exponential form of the solution. Note that if
c = 0 then g [ = g ] .
For the complete transport equation (1), we can use Duhamels principle. Formally, along the characteristics we have
d [
u = f[
dt
so integrating we obtain
Z t
Z t
d [
[
[
u (t, x) u0 (x) =
u (s, x) dt =
f (s, x, u)[ dt
dt
0
0
and we can use this for the general rigourous definition
3

Definition 4 (Mild Solutions). Let u0 be measurable, then u is a exponentially mild solution to (1) iff
For almost all x, c] , f (s, x, u)[ L1 [0, T )
For all t [0, T ) and for almost all x,
u[ (t, x) = u[0 (x) +

f (s, x, u)[ ds

If c = 0 we simply call this a mild solution. We can always take c = 0 by merging the cu term into f ,
(which can depend on u).
2.1.3

Algebraic properties and conservation laws

We now look at the properties of solutions to the transport equation with f = c = 0. Suppose u is a
solution to the transport equation. A renormalisation of u is given by (u) where : R R. Formally,
using the chain rule,


(u)
(u)
u
u
+ bi
= 0 (u)
+ bi
=0
t
xi
t
xi
so that (u) also solves the transport equation with initial condition (u(0)) = (u0 ). This works
because the equation is linear, first order and has no feedback term as c = 0 and f = 0. Even when this
is not the case, we still obtain that (u) satisfies
(u)
(u)
+ bi
+ cu 0 (u) = f 0 (u)
t
xi
This makes the next property useful. We look at the evolution of the mass of the solution when c = f = 0.
Z
Z
Z
d
u
u
u dx =
dx = bi
dx
dt
t
xi
Z
Z
[bi u]
bi
=
u
dx = u div b dx
xi
xi
In particular if div b = 0 then the mass is conserved, and we can approximate the Lp norms to show
their conservation (for u0 0). As (u) is also a solution, we have that
Z
Z
d
(u) dx = (u) div b dx
dt
These conservation properties can be extended to the full equation where c, f are included. We obtain
Z
Z
d
(u) dx = (u) div b cu 0 (u) f 0 (u) dx
dt
2.1.4

Renormalised solutions

We will now transition the intuition discussed above into rigorous mathematical theorems. Again, we
restrict to when c = f = 0. When u is very irregular, i.e. not even L1loc , we can take C 1 (R) bounded
and then look at the equation for (u). Of course this equation cannot be solved in the classical sense.
Instead we will view it in the sense of distributions. This allows us to define a notion of another solution
for very general u.
Definition 5 (Renormalised Solution). u is a renormalised solution to (1) (with c = f = 0) if and only
if for all continuously differentiable : R R, (u) obeys
(u)
(u)
+ bi
=0
t
xi

in D0 ([0, T ) RN )

Depending on the space u lies in, we will require different additional restrictions on . For example,
we may demand that L . We will discuss this later.
For the general case, we use the derived equation for (u) and obtain the following definition:
4

Definition 6 (Renormalised Solution). We say that u is a renormalised solution to (1) if and only if
for all continuously differentiable : R R, with | 0 (z)|(1 + |z|) bounded, (u) obeys
(u)
(u)
+ bi
+ cu 0 (u) = f 0 (u)
t
xi

in D0 ([0, T ) RN )

The condition | 0 (z)|(1 + |z|) bounded means that the extra term u 0 (u) is bounded also. To define
this with a non-zero f , we require a restriction on to make f 0 (u) regular. This will depend heavily
on the particular f . Later we will look at the particular case of the Boltzmann equation.
The notion of renormalised solutions is particular to the form of the transport equation. It requires
the equation to be first order and quasilinear. It allows us to deal with growth of u and f by making
the terms in the renormalised equation of lower order.

2.2
2.2.1

The setting
Function Spaces and topologies

We saw before that the transport equation (1) can be viewed as rearranging u under the forcefield b.
This means that we can make our notions of solution and spaces very general, even as general as all
measurable functions. To do this we need to say what topology we want on these spaces.
Definition 7. We define L as the space consisting of measurable functions from RN to R that are finite
almost everywhere. We give it the topology of local convergence in measure. This has a metric
d(, ) =

X 1
k| | 1kL1 (Bn )
2n

n1

This topology is equivalent to the topology induced by the notion of convergence given by un u if and
only if (un ) (u) in L1loc for all C(R) L (R).
We have two variables for most of our functions, so we need to define our regularity on each and the
way they depend on each other.
R
Definition 8. We define L1 ([0, T ]; L (RN )) by the norm kf k = [0,T ] ku(t, )k dt and similarly for
any other composed spaces with norms.
For spaces without norms, we have to be more careful.
Definition 9. The space C(R; L) consists of all continuous functions from R to L where we use the
topological definition of continuity. Similarly for other spaces.
Definition 10. f L (0, T ; L) means that for all C(R) L (R), (f ) L (0, T ; L1loc ).
Finally we need to define what we mean by the sum of spaces.
Definition 11. By f L1 + L we merely mean that f = f1 + f2 for some f1 L1 and f2 L . This
sum will not be unique.
This notation is long winded and distracting. For this reason we will abbreviate. Unless otherwise
specified
Our time interval will be [0, T ], and our spacial domain RN .
We will assume that the coefficients b, c, div b will be L1 in time, so that when we say b L we
mean b L1 (0, T ; L (RN )). Similarly for convergence of these.

2.2.2

Distributional solutions

Earlier we assumed that it was obvious what is meant for u to satisfy the transport equation in the
distributional sense. We will now make this rigorous.
Definition 12 (Test functions). We define D = D([0, T ) RN ) as the space of all C ([0, T ] RN )
functions with compact support in [0, T ) RN , where by saying that the support is compact in [0, T ) we
mean that the support can include 0, but must not contain T .
Definition 13 (Distributional Solution). By saying that u solves (1) in D0 with f = f (t, x) we mean
that for all D,


Z T
Z TZ
Z TZ

f (t, x) dx dt
div(b) + c dx dt
u0 (0, x) dx =
u
t
0
0
RN
0
RN
This makes sense for u Lp if we have the conditions on the coefficients
c div b Lqloc ,

b Lqloc ,

f (t, x) L1loc

where p, q are Holder conjugates.

In particular for u L , we require that c div b, b and f are in L1loc .
In a broad sense these notions of solutions are equivalent. In particular:
Theorem 1 (Equivalence). Let the ] operator be invertable and measure preserving (for this to hold it
is sufficient that div b = 0, b is independent of t and satisfies the upper bound |b(z)| C(1 + |z|)). Then:
If u, f L1loc ([0, T ) RN ), then mild and distributional solutions are equivalent.
If f 0 (u) L1loc ([0, T ) RN ) then mild solutions are renormalised solutions.
Renormalised solutions are mild solutions, and we need only check (z) = log(1 + z).
If in addition c 0 and for almost all x, (cu)] L1 [0, T ), then exponentially mild solutions are
mild solutions.
Proof. We start with the equivalence of mild and distributional solutions. Here we will merge c into f ,
so wlog we may assume that c = 0.
Let u be a distributional solution, and fix a test function D(RN ). Let n D([0, T )) satisfy
n 1[t1 ,t2 ] in D0 ([0, T )). By measure preservation and invertability the adjoint of ] is given by ](t, x)
and both operate on all measurable functions, so that when we test the equation with ] (t, x)n (t) we
obtain:


u
u
]
0=
+ bi
f, (t, x)n (t)
t
xi
 



]
= u,
+ bi
( (t, x)n (t)) f, ] (t, x)n (t)
t
xi





]
]
0
= u, n (t)
+ bi
(t, x) + (t, x)n (t) f, ] (t, x)n (t)
t
xi


= u, 0 + ] (t, x)0n (t) f, ] (t, x)n (t)


= u] , 0n f ] , n
Where
the inner
product is the action of distributions in D0 ([0, T ) RN ) and we note that the equality
h
i

]
t + bi xi (t, x) = 0 follows from the chain rule (as is smooth) and the definition of ]. As was
RT
arbitrary, we have 0 0n (t)f ] n (t)u] dt = 0 for almost all x. We then let n to obtain that for
almost all x,
Z
t2

u] (t2 , x) u] (t1 , x) =

f ] (s, x) ds

t1

So u is a mild solution. The converse comes from reversing this argument.

Now let u be a renormalised solution. Define (z) = log(1 + z). By the definition of renormalised
solutions



f
+ bi
(u) = 0 (u)f =
in D0
t
xi
1+u
Next, define (z) = 1 ((ez 1)) for > 0, which, when 0, will approximate the identity function.
Note that for fixed each is Lipschitz. We then have





0
+ bi
((u)) = ((u))
+ bi
(u)
in D0
t
xi
t
xi
1+z
But we see that 0 ((z)) = e(z) 0 ((e(z) 1)) = (1 + z) 0 (z) = 1+z
, so that for all > 0,



f
+ bi
((u)) =
in D0
t
xi
1 + u

By computation we see that v = ((u)) = 1 log(1 + u), and by the equivalence of mild and distributional solutions proved above, we have that for almost all x, f ] L1 [0, T ). We now go back to u by
]
noting that u] = ev1 1, so that u] is also the time integral of some function, and hence f ] , (which may
depend on u), is in L1 [0, T ) as required. Then by the equivalence again we obtain that for all t > s 0,
> 0 and almost all x,
Z t
f]
d
v] (t, x) v] (s, x) =
]
s 1 + u
When we take 0 we see that u is a mild solution.
The converse follows from putting observing that if u is a mild solution then v1] obeys the equation
above, and we then apply the equivalence of mild and distributional solutions.
Finally we have to show that exponentially mild solutions are equivalent to mild solutions. Suppose
u is an exponentially mild solution and the given conditions on c are satisfied. Then we have
For almost all x, c] , f [ L1 [0, T ).
For 0 s < t < T and almost all x, u[ (t, x) u[ (s, x) =

Rt
s

f (, x, u)[ d .

c0
For almost all x, (cu)] L1 [0, T )
and we need to show that
For almost all x, (f cu)] L1 [0, T ).
Rt
For 0 s < t < T and almost all x, u] (t, x) u] (s, x) = s (f cu)] (, x) d
 R

t
Recall that f [ = f ] exp 0 c] (s, x) ds , and so because c 0 and f [ L1 [0, t) we see that f ]
L1 [0, T ), and hence (f cu)] L1 [0, T ). Under these conditions we can go from u[ to u] and obtain a
mild solution.

2.3

Existence and uniqueness

We will show existence and uniqueness with initial conditions

u0 Lp ,

p [1, ]

(3)

and with coefficients that satisfy

c, div b L1 (L )
1,q
b L1 (Wloc
)

(4)

b
L1 (L1 ) + L1 (L )
1 + |x|
(5)
7

2.3.1

Smooth coefficients

Lemma 1. Assume (3) and (4), and let b(t, x), c(t, x), f (t, x) C . Then there is a classical smooth
solution.
Proof. By Cauchy-Lipschitz there is a unique and smooth (in both x and t) solution to the characteristic
equations, and by the growth condition on b it is global. We can then define our solution by the solution
operators defined previously, all of which is valid for in the smooth case.
2.3.2

Rough coefficients

The estimates given in 2.1.3 together with the previous smooth result and standard mollification arguments yield:
Proposition 2 (Existence in Lp ). Assume (3) and (4) are satisfied. Then there is a unique distributional
solution in C(0, T ; Lp ) or if p = merely L (0, T ; L ).
Proof. Mollification allows the application of the smooth result. Using the ideas of 2.1.3 and approximation we obtain the estimates
Z t
ku(t)k ku0 k +
kcuk ds C ku0 k
0
Z

Z
d
p
p
|u(t)| dx
|u(t)| dx kpc + div bk
dt
ku(t)kp C 0 ku0 kp a.e. on (0, T )

by Gronwall

We can now apply relative weak compactness criterion to see that our sequence of solutions (
u) is
relatively weakly compact if 1 < p < and relatively weak* compact for p = . A computation that
we skip then shows that a limit of a convergent subsequence is a solution.
This leaves p = 1, for which we need to approximate u0 with elements of D(RN ) converging in L1 ,
giving similar results.
Finally we must show uniqueness. To do this we will make the conservation laws developed above
rigorous. This requires a technical lemma:
Lemma 2. Under the conditions of the theorem, the mollification u
=u
(; t, x) of the actual solution
(not the solution to the equation with mollified coefficients) satisfies
u

+ bi
+ c
u = r
t
xi
0

where as  0 r 0 in L1 (Lp ) if p < or in L1 (Lp ) all p0 < for p = .

The only hard part of the proof of this is showing that the error of mollifying before taking x
i
instead of after in the term involving b tends to 0. We omit this.
We wish to deduce the equation for (u). Let u
be the mollification of the solution. By the lemma
u
is the smooth solution to an equation. Put (u) in this for C 1 , 0 bounded. Then we get

(
u)
(
u)
+ bi
+ c
u 0 (
u) = r 0 (
u)
t
xi
Taking  0 the right hand side tends to zero, and we get
(u)
(u)
+ bi
+ cu 0 (u) = 0
t
xi

in D0

For p < , approximation arguments using now give that if u0 = 0 then u = 0 which by linearity
gives uniqueness. We will skip most of this, noting only the method.
Approximate the Lipschitz function (z) = max(|z|, M )p , M (0, ) using C 1 functions.

 Integrate the equation for (u) against a testR function R = (/R) where is 1 in |x| 1 and 0
d
(u)R dx and close by Gronwall.
outside |x| 2. The purpose is to look at dt
Z
Z
Z
d
R
(u)R dx = (cu 0 (u) div b(u))R dx + (u)bi
dx
dt
xi
The first integral is ok because of the boundedness of 0 and the conditions on the coefficients.
The only problem is the second integral, for which we apply the growth condition b/(1 + |x|)
L1 + L . The problem integral is only non-zero in |x| (R, 2R) and we get R1 decay from the
|b|
derivative on R . In this region we bound |b|
R 1{R|x|2R} 1+|x| 1{R|x|} which we split this into
L1 and L parts. The L1 part is integrable when we bound (u) M , while the L part just
adds a term to the bound for the first integral.
Sending R we can close by Gronwall in the case p < .
For p = a duality argument is required involving solving the backwards problem

+ bi
(c div b) =
t
xi
for
R (t = 0) = 0 and test functions . An argument we omit then uses this to show that
u dx dt = 0 and the uniqueness.
We have in fact shown that

Z
d
p
ku(t)kp + (pc div b)|u|p dx = 0
dt
which gives ku(t)kp C([0, T ]). If p > 1 then u C(0, T ; Lp ) follows from the form of the transport
equation. p = 1 is harder, using approximation arguments we omit here.

2.4

Velocity averaging results

The aim of this section is to establish results for the later existence proof for the Boltzmann equation,
which has the form
f
f
+ vi
=g
t
xi
f (t, x, v) : [0, T ) RN RN R
We will later see the exact form of the right hand side g = Q(f, f ), but for now it doesnt matter.

+ vi x
as T f = g. This is a
The equation is often written using the transport operator T = t
i
T
time independent inhomogeneous transport equation with b(x, v) = [v, 0] . This means that div b =
vi

xi + vj 0 = 0, and the homogeneous equation preserves phase-space. In this section we will look at the
properties of equations of the form
Tf = g
2.4.1

Characteristics

The characteristics split into two vectors and the equations have a simple form:
x, v) = V (t, x, v)
X(t,
V (t, x, v) = 0
X(0, x, v) = x
V (0, x, v) = v
This has a solution given by X(t, x, v) = x vt, V (t, x, v) = v. The characteristic map is then
f ] (t, x, v) = f (t, x + vt, v)
from which we gain our notion of mild solution.
9

2.4.2

Velocity Averaging

Here we will present one of a general class of results called velocity averaging lemmas. The transport
operator T gives no gain in regularity, but if we integrate through we obtain conservation of mass we get
a very large gain in regularity. What happens if we only partially integrate through?
R
Definition 14 (Velocity average). We define the velocity average of f (t, x, v) by F (t, x) = RN f (t, x, v) dv.
We will look at the Fourier transform of T f in the t and x y coordinates. In Fourier
space the transport operator takes the form T = i + ivj yj , which is very similar to the Sobolev norm
R
2
kf kH s = RRN (1 + | |2 + |y|2 )s |f(, y, v)|2 dy d . We can use this to show a gain in regularity from f
to F .
Lemma 3. Let f L2 (R RN RN ) with compact support, and T f L2 (R RN RN ), then
F (t, x) H 1/2 (R RN ).
Proof. We wish to bound kF kH 1/2 . We compute
F (, y) =

f(, y, v) dv

RN

R
So as f L2 (R RN RN ) it is sufficient to show that RRN (| |2 + |y|2 )1/2 |F (, y)|2 dy d is bounded.
2
R

Note that RRN |F (, y)|2 dy d = f 2 < . Let = (| |2 + |y|2 )1/2 . We will show that |F |2
L




c
C
f 2 + T f 2 and the result will follow as by Jensens inequality

Lv

Lv

kukL2 dx dt
v

RRN

|u|2 dv dx dt

RRN RN

Note that |Tcf |2 = | + vj yj |2 |fb|2 . We want to use this to obtain 1 . To do this, we split the integral
y
according to whether the velocity makes this small or large. Define 0 = and yj0 = j . Then
|Tb| = | + vj yj | = | 0 + vj yj0 |. For |Tb| 1, if K is the compact support of f in v, we have
!2
Z
2

|Flarge | =
f dv
|Tb|1

|T|

! Z

1
2


by Cauchy-Schwarz

RN

|Tb|1;vK

1
2

|Tb|2 |fb|2 dv

dv

!
0

| +

vj yj0 |2

dv

|Tb|1;vK

2

Tcf 2

Lv

| 0 +vj yj0 |1/;vK

| 0 + vj yj0 |2 dv

C
2
Tcf 2

Lv

2

Tcf 2

Lv

by direct integration

While for |Tb| < 1 we obtain:

|Fsmall |2 =

!2
f dv

|Tb|<1

! Z

1 dv
vK;|Tb|<1

|f|2 dv


by Cauchy-Schwarz

RN

C0
b 2
f 2

Lv

where in the last line we take the supremum over 0 , yj0 of the Lebesgue measure of {v K : |Tb| < 1} =
{v K : | 0 + vj yj0 | < 1/} which is a slice of width 2 in RN , intersected with K.
10

The compactness of H 1/2 in L2 allows the conversion of weak compactness into strong compactness.
The following lemma and corollary will later be used on the Boltzmann equation.
Lemma 4. Let the following assumptions hold:
(fn ) L1 ((0, T ) RN RN ) be weakly relatively compact.

(6)

(n ) be bounded in L ((0, T ) R R ) and converge a.e. to

Additionally,
assume that (T fn ) be weakly relatively compact in L1loc ((0, T )RN RN ). Then the velocity
R
average RN fn n dv is strongly compact in L1 ((0, T ) RN ).
Corollary 1. Furthermore, ifR in addition fn f weakly in L1 ((0, T ) RN RN ) then the velocity
average converges strongly to RN f dv in L1 ((0, T ) RN ).
Proof. By Egorovs theorem have n uniformly except on a set of arbitrarily small measure, as
we only care about L1 convergence, we can take n uniformly everywhere and in fact we can take
n = . In a similar way, by uniform integrability (as (fn ) is relatively weakly compact), for any  > 0
we can take a fixed compact set K on whose complement all the integrals are <  in magnitude. For
this reason we may assume that (fn ) and f have fixed compact support K. We can also approximate
L by smooth functions k uniformly bounded in L , so that fn n satisfies the same conditions
as fn . This means we may take = 1, as the error is
Z

Z
Z
Z


 k fn dv fn dv  dx dt
|k ||fn | dv dx dt 0


RRN

RRN RN

because (gn ) is uniformly integrable. Finally by the compact support of (fn ), we have that T fn is weakly
relatively compact in the whole space L1 rather than simply L1loc .
In summary, we now have (fn ) compactly supported and weakly relatively compact in L1 , with T fn
weakly relatively compact in L1 , and we want to show (Fn ) strongly compact in L2 . The result would
then follow from the velocity averaging lemma as H 1/2 H 0 = L2 by Rellich-Kondrachov, and because
of the compact support, the compactness in L1 .
To do this we need to approximate with L2 functions. We define the large and small parts of fn by
the solutions to:
T fnsmall = (T fn )1{|T fn | M }
T fnlarge = (T fn )1{|T fn | > M }


fnlarge t=0 = fnsmall t=0 = 0
By linearity of T and uniqueness of solutions, fn = fnsmall + fnlarge . We can then use the explicit solution
operator to express fnlarge as
Z t
large]
fn
=
(T g)] 1{|T fn | > M }] ds
0

Then integrating and using measure preservation of the solution operator, we see that uniformly in n
large
fn
1 0 as M
L
However, by measure preservation again, both (fnsmall ) and (T fnsmall ) are bounded in L2 , which is what
we needed for the velocity averaging lemma.
Lemma 5. Assume that the following hold
fn f weakly in L1+ ((0, T ) RN RN )
For each fixed > 0, let (T (f n )) be weakly relatively compact in L1loc ((0, T ) RN RN )
where ( )>0 is a uniformly Lipschitz family in C(R; R) with (0) = 0
and (z) z as 0 uniformly on compact sets.
R
R
Additionally, assume that (6) holds. Then RN fn n dv RN f dv.
11

(7)

Proof. First note that:

n

kf (f )kL1 =

1{|f | < a}|f (f )| +

1{|f n | a}|f n (f n )|

The first term tends to 0 by the uniform convergence on compact sets. Using the bound (f nR) M |f n |
where M is the uniform Lipschitz constant, we see that the second integral is less than (1+M ) 1{|f n |
a}|f n | which tends to zero by the dominated convergence theorem. Hence we have that f n (f n ) 0
in L1 .
Now we apply lemma 4 to ( (f n )) for each fixed > 0, and obtain the result using the triangle
inequality, taking 0.
We can also deduce vector valued versions of these results, but first we need a technical lemma.
Lemma 6. Let
(E, ) be an arbitrary measure space.
U RN .
(un ) be weakly compact in L1 (K) for any compact K U with supp K E.
(k ) be uniformly bounded in L (U ; L1 (E)) and converge to 0 in L1 (U E).
Then supn kun k kL1 (U E) 0 as k .
Proof. First it should be noted that the reason (E, ) is so general is that we need nothing from it.
Let Km be a chain of compact sets in U that eventually surround any point. Then we note that by
the convergence of k we have the k-independent estimate
sup kk un kL1 (Km E) C sup kun kL1 (Km )
n

We then bound the large part of un ,

kk un 1{|un | M }kL1 (Km E) C kun 1{|un | M }kL1 (Km ) m (M )
where by the weak compactness of (un ), m (M ) 0 as M .
The small part of un can then be bounded by
kun k 1{|un | < M }kL1 (Km E) M kk kL1 (Km E)
We then take k and then M to obtain the result.
Corollary 2. Let (6) and (7) hold but with n , and their convergence in L ((0, T )RN RN ; L1 (RN )).
Then the velocity average converges,
Z
Z
fn n dv
f dv strongly in L1 ((0, T ) RN RN )
RN

RN

Proof. By arguments given in prior proofs we can reduce to when n = . Now (t, x, v) is valued in
L1 (RN ). To keep this space distinct we will label it E. If separates into a sum
(t, x, v) =

m
X

i (t, x, v)i where i L1 ((0, T ) RN RN ) and i L1 (E)

i=1

then the result is immediate by lemma 5. Otherwise we can approximate with functions of this form
and apply the technical lemma.

12

3
3.1

The Cauchy Problem for the Boltzmann Equation

The equation

In this section, we turn to the discussion of the Cauchy problem for the Boltzmann equation, namely

f
(t, x, v) + v x f (t, x, v) = Q(f, f )(t, x, v)
(B)
t
f (0, x, v) = f (x, v)
0
where (t, x, v) R+ Rd Rd . This equation was first formulated in 1872 by Ludwig Boltzmann to
model the time evolution of the molecular density of a rarefied gas. Note that if the gas is so rarified
that the molecules can be assumed not to interact with each other, then we expect their density to follow
the linear transport equation
f
+ v x f = 0 ,
t
and hence to remain constant along each molecular path. In this case the problem can be solved explicitly
just by evolving the initial datum along characteristics. If, on the other hand, the molecules interact
with each other, this conservation no more holds, and a term to describe the change of the molecular
density due to these interactions needs to be added to the equation. This corresponds to Q(f, f ) in (B).
In order to explain the structure of Q(f, f ) we need to choose a model for the molecular interaction. In
this notes, we focus on the so called hard spheres model, according to which the molecules are described
as perfectly elastic spheres moving independently (following the laws of classical mechanics) until they
collide. Call v and v the velocities of two molecules colliding at time t in position x, and v 0 , v0 their

Figure 1: Schematic representation of a collision between two molecules.

velocities after the collision respectively. Since we are assuming the collisions to be perfectly elastic, the
momentum and kinetic energy must be conserved, and we obtain the first relations
v + v = v 0 + v0
|v|2 + |v |2 = |v 0 |2 + |v0 |2

(8)

From this, if we let be the unit vector, directed along the line joining the centres of the spheres, we
deduce
(
v 0 = v [(v v ) ]
(9)
v0 = v + [(v v ) ]
and this easily implies the relations
v v = v 0 v0
|v v | = |v 0 v0 |
0

(v v ) = (v

v0 )

(10)

that will be useful later on. Equations (9) allow us to explicitly define the collision operator Q(f, f ) as
the average change in the molecular densities of colliding particles, weighted with the collision kernel.
Z Z
Q(f, f ) =
q(v v , )(f 0 f0 f f )dv d
Rd

S d1

13

where f = f (t, x, v), f = f (t, x, v ), f 0 = f (t, x, v 0 ), f0 = f (t, x, v0 ), and the hard spheres collision kernel
q is given by
q(v v , ) = |(v v ) | .
Observe that we can interpret q as the probability of a collision happening between two molecules having
velocities v, v and such that the unit vector joining their centres is . The collision operator Q is the
main problem with trying to solve the Boltzmann equation. It is non-linear and also non-local in the
sense that its value at a point depends on the values of f everywhere. Despite this, Q splits nicely into
positive and negative parts, and the negative part can be split into local and non-local parts.
Z Z
Q+ (f, f ) =
f0 f 0 q(v v , )dv d
Q (f, f ) = f L(f )
Z Z
Z
L(f ) =
q(v v , )f ddv = A(v v )f dv = A f
d1
Z S
A(z) =
q(z, )d .
S d1

Observe that Q has a local part f and a non-local part A f . The non-local part will be the main
angle of attack, once we have a bound on it, we can use an entropy inequality to bound Q+ . In this way
we control the nasty non-local part of Q+ with the nicer but still non-local part of Q .
We remark here that more general collision kernels can be considered, and in fact Assumptions 1 below
are enough to prove the main existence result, namely Theorem 5.
3.1.1

Solution concepts for the Boltzmann equation

The concepts of mild and renormalised solutions for the Boltzmann equation read as follows:
Definition 15 (Mild solution of the Boltzmann equation). A mild solution to the Boltzmann equation
(B) is a measurable function f (t, x, v) : [0, ) Rd Rd for which
Q (f, f )] L1loc [0, ) for almost all x, v, (so the integral below makes sense).
Rt
For all t 0 and for almost all x, v, we have f ] (t, x, v) = f0 (x, v) + 0 Q(f, f )] (s, x, v) ds.
Definition 16 (Renormalised solution of the Boltzmann equation). A renormalised solution to the
Boltzmann equation is a measurable function f (t, x, v) L1+ ([0, )loc Rd Rd ), for which

Q (f, f )
L1loc ([0, ) Rd Rd )
1+f

(11)

(so that it exists as a distribution).

For all Lipschitz continuous functions (z) : [0, ) R for which | 0 (z)|(1 + |z|) L , f satisfies
T [(f )] = 0 (f )Q(f, f )

in D0

Note how the bound on the derivative of combined with the condition on Q acts to put the right
hand side of the equation in L1loc .
Lemma 7 (Equivalence for Boltzmann equation). The two notions of solutions are equivalent in the
following sense.
If f is a renormalised solution, but for only (z) = log(1 + z), then f is a mild solution.
If f is a mild solution, and (11) holds, then f is a renormalised solution.
Proof. We just apply Theorem 1.

14

3.1.2

Main theorem

In order to state the main existence result for (B) we need to make the following assumptions.
Assumptions 1. Let z = v v . Assume for the collision kernel q(z, ) the following properties:
q depends only on |z| and |z |

1
1 + |v|2

Z
A(v v )dv 0

as |v|

(12)

|v |R

d
A L
loc (R ).

Note that the hard spheres kernel q(v v , ) = |(v v )| trivially satisfies all the above properties.
We also make the following assumptions on the initial datum.
Assumptions 2. Assume the initial datum f0 of (B) belongs to L1+ (Rd Rd ) and is such that
Z Z
f0 (1 + |x|2 + |v|2 )dxdv <
Z Z
f0 | log f0 |dxdv <
Provided these assumptions are satisfied, the following result holds
Theorem 3 (DiPerna and Lions, 1989). Suppose the collision kernel q and the initial datum f0 satisfy
Assumptions 1 and 2 respectively. Then there exists a renormalized solution f of the Boltzmann equation
(B) such that f C(R+ , L1 (Rd Rd )) and the following bounds hold:
Z Z
Z Z
f (t, x, v)(1 + |x|2 + |v|2 )dxdv
f0 (x, v)(1 + 2|x|2 + (2t2 + 1)|v|2 )dxdv
Z Z

Z Z

Z Z
e(f )(s, x, v)dxdvds

f (t, x, v)| log f (t, x, v)|dxdv +

f0 (| log f0 | + 2|x|2 + 2|v|2 ) + Cd

where Cd is a constant depending only on the dimension d.

We outline below the strategy of the proof.

3.2

Short sketch of the proof

The proof is based on compactness arguments. In particular, a renormalized solution f will be obtained
as the limit of (a subsequence of) solutions (f n ) to nicer problems (Bn).
We start by proving some a priori estimates which follow directly from the structure of (B).
Define a sequence of truncated problems (Bn) corresponding to nice initial data f0n and truncated
collision operators Qn (f, f ), and show (using a contraction argument) that they have unique global
solutions (f n ) satisfying the the a priori bounds uniformly in n.
Observe that these bounds imply (by Dunford-Pettis criterion, Theorem 15) that the sequence
(f n ) is weakly compact in L1 ((0, T ) Rd Rd ), so there exists a function f in the same space
such that f n * f up to the extraction of a subsequence. We aim to show that f is the claimed
renormalized solution to (B). To this end, we turn to discuss the convergence of the sequence of
truncated collision operators.
As we are interested in a renormalised solution we turn our attention to

Qn (f n , f n )
. We first
1 + fn

Qn (f n , f n )
is contained in a weakly compact subset of L1 ((0, T ) Rd BR ) for all
1 + fn
T, R < : this is a consequence of DP criterion, the a priori estimates and the careful definition
of the collision operators in (Bn).
show that

15

 From this, we obtain the compactness of

n
n
Qn
+ (f ,f )
1+f n

via the entropy inequality

Qn+ (f n , f n ) KQn (f n , f n ) +
L1

n
n
Qn
(f ,f )
is weakly compact. The
n
1+f
Q (f,f )
1+f , because of the non linearity.

Hence, for now, f n * f , and

whether

n
n
Qn
(f ,f )
1+f n

goes to

1
en (f n ) .
log K
problem is that we dont know

On the other hand, with a careful use of the velocity averaging results, we can show that the above
convergence holds, even strongly, for the respective velocity averages, i.e.
R n n n
R
Q (f , f )dv
Q (f, f )dv
R
R

n
1 + f dv
1 + f dv
in L1 ((0, T ) Rd ), for each compactly supported function L ((0, T ) Rd Rd ).
To conclude, we introduce an equivalent form of (B), and show that it is satisfied by f . This,
(f,f )
together with the right integrability conditions on Q1+f
, implies that f is a renormalised solution
to (B), as claimed.

3.3

A priori estimates

Let f be a solution to (B) with initial datum f0 satisfying Assumptions 2. Take C (Rd ) and assume
we have all the necessary decay for the calculations below to make sense.
We start by integrating against the collision operator:
Z
Z Z Z
(v)Q(f, f )dv =
q(v v , )(v)(f0 f 0 f f )dv dvd
(13)
Recall that q(v v, ) = q(v v , ) and exchange v and v , to get
Z
Z Z Z
(v)Q(f, f )dv =
q(v v , )(v )(f0 f 0 f f )dv dvd

(14)

Now operate in (13) the change of variables (v, v ) 7 (v, v0 ), or more explicitly
(
v 0 = v [ (v v )]
v0 = v + [ (v v )]
and note that the Jacobian associated with this transformation is 1. We get
Z
Z Z Z
(v)Q(f, f )dv =
q(v 0 v0 , )(v 0 )(f f f0 f 0 )dv dvd
Z Z Z
=
q(v v , )(v 0 )(f0 f 0 f f )dv dvd

(15)

where we have used the second relation in (10) for the second equality. Finally we once again exchange
v and v in (15) to get
Z
Z Z Z
(v)Q(f, f )dv =
q(v v , )(v0 )(f0 f 0 f f )dv dvd
(16)
Adding up equations (13)-(16) we obtain the identity:
Z
Z Z Z
1
(v)Q(f, f )dv =
q(v v , )(f0 f 0 f f )[ + 0 0 ]dv dvd
4
This motivates the following definition.
16

(17)

Definition 17. A function : Rd R is said to be collision invariant if

+ = 0 + 0
This definition has a simple interpretation. Observe that, by (17), if is collision invariant then
Z
Q(f, f )(v)dv = 0 .
Since this integral represents the rate of change in the average value of due to collisions, the above
relation tells us that this average value stays unchanged.
Then (8) immediately implies that any function of the form
(v) = a + b v + c|v|2

(18)

is collision invariant1 , and hence gives 0 when integrated against the collision operator. To make use of
that, we go back to equation (B), multiply it by (18) and integrate with respect to v to get

Z
Z 
f
+ v x f dv
0 = Q(f, f )(v)dv =
t
Z

Z

=
f dv + divx
vf dv
t
Next, we integrate over x Rd and use the total time-derivative to see that:
Z
d
f dv = 0 .
dt
This gives the following conservation laws:
Conservation of mass
Z Z

Z Z
f (t, x, v)dxdv =

Conservation of momentum
Z Z

f0 (x, v)dxdv

(19)

f0 (x, v)v dxdv

(20)

Z Z
f (t, x, v)v dxdv =

Conservation of kinetic energy

Z Z
Z Z
f (t, x, v)|v|2 dxdv =
f0 (x, v)|v|2 dxdv

(21)

Another conservation law is obtained integrating the Boltzmann equation against (t, x, v) = |xtv|2
with respect to x and v. By (8) and (10), in fact, we deduce that
|x tv 0 |2 + |x tv0 |2 = 2|x|2 + t2 (|v 0 |2 + |v0 |2 ) 2t x (v 0 + v0 )
= 2|x|2 + t2 (|v|2 + |v |2 ) 2t x (v + v )
= |x tv|2 + |x tv |2
so that

Q(f, f )|x tv|2 dv = 0. Apply the same reasoning as before to get

Z Z
d
f (t, x, v)|x tv|2 dxdv = 0
dt

and hence the conservation law

Z Z
Z Z
2
f (t, x, v)|x tv| dxdv =
f0 (x, v)|x|2 dxdv .
1 In

fact, it can be shown that all collision invariants have this form, see [2] for details.

17

(22)

3.3.1

Entropy identity

Multiply the Boltzmann equation by (1 + log f ) and integrate over v:

Z
Z
(1 + log f )(t f + v x f )dv = Q(f, f )(1 + log f )dv
So

Z

Z
T f dv + t

Z

f log f dv + divx

v f log f =

Z
Q(f, f )dv +

Q(f, f ) log f dv

and using that f solves the Boltzmann equation we get

Z

Z
 Z
t
f log f dv + divx
v f log f = Q(f, f ) log f dv
Z Z Z
1
f 0 f0
=
q(v v , )(f0 f 0 f f ) log
dv dvd
4
f f
where in the last equality we have used relation (17) with = log f . Now integrate over x Rd , to
obtain the identity
Z Z
Z Z
d
f log f dxdv +
e(f )dxdv = 0
(23)
dt
with
e(f ) =

1
4

Z Z

q(v v , )(f0 f 0 f f ) log

Note that the non-negativity of e(f ) implies that

f 0 f0
dv d .
f f

(24)

RR

f log f is non-increasing in time.

Z Z
Lemma 8. Let g be any non negative function on Rd Rd such that
g| log g|dxdv < . Then
Z Z

Z Z
g| log g|dxdv

Z Z
g log g dxdv + 2

g(|x|2 + |v|2 )dxdv + Cd

where Cd is a constant depending only on the dimension d of the space.

Proof. Start by splitting the integral
Z Z
Z Z
Z Z
g| log g|dxdv =
g log g dxdv 2

g log g dxdv

(25)

g1

Then notice that

g log g 1(g(0,1)) = g log g 1(e(|x|2 +|v|2 ) g<1) + g log g 1(g<e(|x|2 +|v|2 ) )
= g log g 1((|x|2 +|v|2 )log g<0) + g log g 1(log 1 >|x|2 +|v|2 )
g

g(|x|2 + |v|2 ) 1((|x|2 +|v|2 )log g<0) g log

1
1
(|x|2 +|v|2 )
)
g (g<e

where 1() denotes the indicator function. By substituting this into (25) we get
Z Z
Z Z
Z Z
g| log g|dxdv =
g log g dxdv 2
g log g dxdv
g1
Z Z
Z Z
Z Z
1

g log g dxdv + 2
g(|x|2 + |v|2 )dxdv + 2
g log 1(g<e(|x|2 +|v|2 ) ) dxdv .
g

By the inequality t log 1t C t, which holds for t (0, 1), we can bound the last term from above so
that
Z Z
Z Z
2
2
1
1
g log 1(g<e(|x|2 +|v|2 ) ) dxdv C
e 2 (|x| +|v| ) dxdv =: Cd
g
which completes the proof of the lemma.
18

Apply the above result with g = f (t, x + vt, v) = f ] (t, x, v), to get
Z Z
Z Z
Z Z
]
]
]
]
f | log f |dxdv
f log f dxdv + 2
f ] (|x|2 + |v|2 )dxdv + Cd
and the change of variables (x, v) 7 (x vt, v) gives us
Z Z
Z Z
Z Z
f | log f |dxdv
f log f dxdv + 2
f (|x vt|2 + |v|2 )dxdv + Cd .

(26)

Lemma 9. (A priori estimates)

d
d
d1
Suppose that q is a non negative measurable function in L
), q = q(x, |vv |, |(vv )|)
loc (R R S
and q grows at most polynomially in x and v v . Assume further that f is a positive solution to the
Cauchy problem (B) with initial datum satisfying Assumptions 2, such that f C 1 ((0, ), S(Rd Rd )),
where S denotes the Schwartz space. If, moreover, | log f | grows at most polynomially in (x, v) uniformly
in compact time intervals, then the following hold:
Z Z
Z Z
f (t, x, v)dxdv =
f0 (x, v)dxdv
Z Z
Z Z
f (t, x, v)|v|2 dxdv =
f0 (x, v)|v|2 dxdv
Z Z
Z Z
2
f (t, x, v)|x tv| dxdv =
f0 (x, v)|x|2 dxdv
Z Z
Z Z
Z tZ Z
e(f )(s)dxdvds =
f0 log f0 dxdv .
f (t, x, v) log f (t, x, v)dxdv +
0

These identities in turn imply:

Z Z
Z Z
f (t, x, v)(1 + |x|2 + |v|2 )dxdv
f0 (x, v)(1 + 2|x|2 + (2t2 + 1)|v|2 )dxdv
Z Z

Z Z

Z Z
e(f )(s, x, v)dxdvds

f (t, x, v)| log f (t, x, v)|dxdv +

(27)

f0 (| log f0 | + 2|x|2 + 2|v|2 ) + Cd

(28)
Note that the assumptions of the lemma are enough to carry out all the a priori computations made in
the previous section, and these computations are not affected by allowing the collision kernel to depend
on the position.
The assumptions on the growth of the logarithm is needed to make sure that integrals
RR
such as
f log f are convergent, as f being in S(Rd Rd ) in (x, v) potentially allows for a decay that
is too rapid for the integral to converge.
Let us now turn to the proof.
Proof. The first three relations have already been shown. The fourth is obtained integrating (23) in the
time interval [0, t]. It only remains to prove (27) and (28).
For the first one we split the integral, and use the conservation of mass and kinetic energy to get
Z Z
Z Z
2
f (t, x, v)(1 + |v| )dxdv =
f0 (x, v)(1 + |v|2 )dxdv
To bound the remaining term we use the inequality |x|2 (|x tv| + t|v|)2 2(|x tv|2 + t2 |v|2 ) and
(21)-(22), to get
Z Z
Z Z
2
f (t, x, v)|x| dxdv 2
f0 (x, v)(|x|2 + t2 |v|2 )dxdv
which yields the required result. Finally, for the last inequality combine (26) with (23) and use the
conservation laws to get
Z Z
Z Z
Z Z
f (t, x, v)| log f (t, x, v)|dxdv
f log f dxdv + 2
f (|x vt|2 + |v|2 )dxdv + Cd
Z tZ Z
Z Z
Z Z
=
e(f )(s)dxdvds +
f0 log f0 dxdv + 2
f0 (|x|2 + |v|2 )dxdv + Cd
0
Z Z Z
Z Z
Z Z

e(f )(s)dxdvds +
f0 | log f0 |dxdv + 2
f0 (|x|2 + |v|2 )dxdv + Cd
0

19

R
where in the last inequality we have used the fact that intf log f is non increasing in time, and the
lemma is proved.

3.4

Truncated problems

We now want to make use of the a priori estimates. The idea is to define some nicer problems, for which
we can prove existence and uniqueness of solutions in the class of functions satisfying all the assumptions
of Lemma 9. For these solutions the conclusions of the above lemma will still hold.
To implement this strategy, we have to identify the problems in proving the existence of a solution to
(B), and find a way to avoid them. These are the following:
Large relative velocity collisions.
Note that Assumptions 1 on the collision kernel do not control its behaviour for large velocities.
Non-linearity of the collision operator.
In order to deal with the convergence of Q(f, f ) we will introduce a suitable additional renormalization, aimed to slow down the growth of the collision operator.
Regularity of the initial datum.
Lemma 9 takes the initial datum in the Schwartz space, but the theorem claims the existence of a
solution when f0 is merely L1+ (Rd Rd ). We will have to use some approximation argument.
We take care of these problems as follows:
Assume that the collision kernel q satisfies Assumptions 1, and consider a sequence of truncated kernels
qn such that:

qn C0,+
(Rd S d1 ) for all n 1,

the (qn ) satisfy the second and third assumptions in (1) uniformly in n,
each qn vanishes for |(v v ) | < n , where (n ) is a sequence of positive real numbers such that
n & 0,
qn q a.e.
For each n we then define the truncated (and renormalized) collision operator
RR
qn (f 0 f0 f f )ddv
n
R
.
Q (f, f ) =
1 + n |f |dv
Finally, if f0 satisfies Assumptions 2 we approximate it by a sequence (f0n ) such that:
f0n S(Rd Rd ) for all n 1,
2

for each fn , there exists a positive real number n such that f0n n e(|x|

Z Z

f0n (1 + |x|2 + |v|2 )dxdv

Z Z

Z Z

f0 (1 + |x|2 + |v|2 )dxdv

+|v|2 )

as n

(29)

Z Z
f0 | log f0 |dxdv

f0 | log f0 |dxdv

We have then obtained the sequence of truncated problems

n
f
+ v x f n = Qn (f n , f n )
t
f n (0, ) = f n
0
R
Recall that A(z) = S d1 q(z, )d and set
Z
An (z) =
qn (z, )d
S d1

20

as n

(30)

(Bn)

so that
Qn (f n , f n ) = Qn+ (f n , f n ) Qn (f n , f n )

where

Qn (f n , f n ) =

f n (An f n )
R
.
1 + n f n dv

Proposition 4. For each n 1, the Cauchy problem (Bn) has a unique global solution
f n C 1 ((0, ), S(Rd Rd )). Moreover, f n satisfies all the assumptions of Lemma 9, and hence the
conservation laws (19)-(22) and inequalities (27)-(28).
R
To prove this result, we will establish the following lemma which we will use to bound Qn (f, f )dv.
Lemma 10. Suppose a, b L1v (Rd ). Then
ZZZ
ZZZ
qn |a0 ||b0 |ddv dv1 =
qn |a(v1 [ (v1 v )])| |b (v + [ (v1 v )])| ddv dv1
cn ||a||L1v ||b||L1v
for some constant cn .
Proof. To prove this result, we simply make the change of variables u() = [ (v1 v )] for fixed v1
d
and v . The Jacobian associated with this transformation is given by ( (v1 v )) and so
ZZZ
ZZZ
1
qn |a||b|ddv dv1 =
qn |a(v1 u)| |b (v + u)|
dudv dv1
d
| (v1 v )|
Note that the integrand is defined everywhere, as qn vanishes if | (v1 v )| < n . Thus
ZZZ
ZZZ
1
qn |a||b|ddv dv1 d
qn |a(v1 u)| |b (v + u)| dudv dv1
n
Moreover, qn is supported only on a compact set of u, which we will denote by Kn . Additionally, qn is
bounded on this set by some constant n . Thus
ZZZ
ZZZ
n
|a(v1 u)| |b (v + u)| dudv dv1
qn |a||b|ddv dv1 d
n
Kn
Applying Tonellis theorem yields
ZZZ
ZZZ
n
qn |a||b|ddv dv1 d
|a(v1 u)| |b (v + u)| dv dv1 du
n
Kn
 Z

Z Z
n
= d
|a(v1 u)| dv1
|b (v + u)| dv du
n Kn
n
d ||a||L1v ||b||L1v (Kn )
n
and setting cn =

n
d (Kn )
n

yields the desired inequality.

We will use this lemma to establish L1 bounds on the truncated collision kernel.
Lemma 11. There are constants Cn such that for any f, g L1 L (Rd ), we have
Z
|Qn (f, f )(v) Qn (g, g)(v)|dv Cn ||f g||L1v
and

|Qn (f, f )(v)|dv Cn ||f g||L1v

21

Proof. To show this, we split the collision kernel into Qn+ and Qn where
RR
qn (f 0 f0 )ddv
R
.
Qn+ (f, f ) =
1 + n |f |dv
and

RR
Qn (f, f ) =

qn (f f )ddv
R
.
1 + n |f |dv

It is simple to see that



R
R
qn (1 + n |g|dv)(f 0 f0 ) (1 + n |f |dv)(g 0 g0 ) ddv




R
R
1 + n |f |dv 1 + n |g|dv

 0 0
RR
R
R
qn (f f g 0 g0 ) + n (f 0 f0 |g|dv g 0 g0 |f |dv) ddv




R
R
=
1 + n |f |dv 1 + n |g|dv
RR

Qn+ (f, f ) Qn+ (g, g) =

We will bound the term



Z
Z


n (f 0 f0 |g|dv g 0 g0
|f |dv)


To do this, we first perform the following simple calculation


  Z

Z
Z

 

n (f 0 f0 |g|dv g 0 g0
 = n f 0 f0 |g| g 0 g0 |f |dv 
|f
|dv)

 

 Z

Z
Z


= n f 0 |g| (f0 g0 ) dv + f 0 g0 (|g| |f |) dv + |f |g0 (f 0 g 0 ) dv 


n ||g||L1v |f 0 | (|f0 g0 |) + ||g f ||L1v |f 0 ||g0 | + ||f ||L1v |g0 | |f 0 g 0 |
Each of our terms involves exactly one of f 0 , g 0 or (f 0 g 0 ), exactly one of f0 , g0 or (f0 g0 ) and exactly
one of ||f ||L1v , ||g||L1v or ||f 0 g 0 ||L1v . This was done deliberately so that we can apply Lemma 10 to see
that


Z Z Z
Z
Z




0 0
0 0

qn n (f f |g|dv g g |f |dv) ddv dv1 3cn n ||g||L1v ||f ||L1v ||f g||L1v
Similar arguments yield
Z Z Z

and so
Z

|Qn+ (f, f )(v)



qn |(f 0 f0 g 0 g0 )| ddv dv1 n (||f ||L1v + ||g||L1v )||f g||L1v

Qn+ (g, g)(v)|dv




||f g||L1v n (||f ||L1v + ||g||L1v ) + 3cn n ||g||L1v ||f ||L1v

1 + n (||f ||L1v + ||g||L1v ) + n2 ||f ||L1v ||g||L1v

Cn ||f g||L1v

A slightly easier argument (there is no need to change coordinates) can be used to show that
Z
|Qn (f, f )(v) Qn (g, g)(v)|dv Dn ||f g||L1v
Combining these two results leaves us
Z
|Qn (f, f )(v) Qn (g, g)(v)|dv Cn ||f g||L1v
The second bound follows by taking g = 0.
We can now prove Proposition 4.
22

Proof. We construct the following scheme to approximate the required f n . We define the sequence (fm )
recursively by
T f0 = 0
f0 (0, x, v) = f0n (x, v)
T fm+1 = Qn (fm , fm )
fm+1 (0, x, v) = f0n (x, v)
We can explicitly compute these as
f0] (t, x, v) = f0n (x, v)
]
fm+1
(t, x, v) = f0n (x, v) +

t
]
]
Qn (fm
, fm
) ds

We will show that the operator that maps fm fm+1 is a contraction on C([0, ); L1x,v ) for sufficiently
small . Then we have a solution to the integral form of the equation. In other words a mild solution.


]
]
fm
kfm+1 fm kC([0, );L1x,v ) = fm+1

C([0, );L1x,v )

Z t


]
n ]
]
n ]

Q (fm , fm )(s, x, v) Q (fm1 , fm1 )(s, x, v) ds
=

0

L1x,v

Z
=
0

C([0, );L1x,v )

n ] ]

]
]
Q (fm , fm ) Qn (fm1 , fm1 )

dt

kQn (fm , fm ) Qn (fm1 , fm1 )kL1x,v dt

Cn kfm fm1 kC([0, );L1v,x )

So we may take small enough for this to be a contraction. Then by the fixed point theorem we have a
solution in [0, ). As none of our estimates depended on the norm of the initial condition we may then
restart the process at /2 say, and continue to get existence for all time.
In fact we can prove that the limit is in C([0, ), S(Rd Rd )). To do this we bound the x and v
derivatives of fm+1 by moving them under the integral. When the derivatives hit qn they produce a
function that obeys all the same bounds as qn does, and when they hit fm they produce its derivatives. In
this way the same argument gives that any derivative of the solution is also in C([0, ), S(Rd Rd )). This
argument also works for multiplication by polynomials: we see that a polynomial in v can be absorbed
into the collision kernel which still obeys the same assumptions as before, and a polynomial in x is taken
care of by the fact that the initial data is Schwartz. This is enough
to give the Schwartz space result.
Rt
Finally we argue that as the limit obeys f ] (t, x, v) = f0n] + 0 Qn (f, f )] ds, and the integrand is smooth
and compactly supported we can deduce that f is time differentiable, i.e. f C 1 ([0, ), S(Rd Rd )).
To prove that | log f n | grows at most polynomially in (x, v), observe that fm C[0, T ] has initial
condition fm (0, x, v) = f0n (x, v) > 0. Thus fm is non-negative for some time interval. Hence (on this
time interval)
T fm+1 = Qn (fm , fm )
= Qn+ (fm , fm ) Qn (fm , fm )

Now because

(An f
R m)
1+n fm

fm (An fm )
R
1 + n fm

Cn for some Cn > 0, we have

T fm+1 Cn fm
23

Send m to get the inequality

T f n Cn f n
This implies t (f n] ) = (T f n )] Cn f n] and hence, by Gronwalls inequality, we obtain
f n] f0n (t, x, v)eCn t .
Going back along the characteristic, we finally get
2

f n f0n (t, x vt, v)eCn t n e(|xvt|

+|v|2 ) Cn t

as claimed.

3.5

Weak compactness results

For now, we have a sequence (f n ) of solutions to the problems (Bn) for which all the a priori estimates,
together with relations (29)-(30), hold. In this section we want to show that the two sequences
f n,

Qn (f n , f n )
1 + fn

n1

are contained is weakly compact sets: this will allow to apply the velocity averaging results in a later step.
The main tool we will use to prove weak compactness is the Dunford-Pettis (DP) criterion (Theorem
15). In particular, we will make use of it in the following form:
Corollary 3. Let h : R+ R+ be a continuous function, w : U R+ be in L
loc (U ), and assume
h(s)/s as s and w(y) as |y| . If, then, (f n ) is a sequence which is uniformly
bounded in L1 (U ) and such that
Z
sup
[h(|f n |) + |f n |(1 + w)] dy <
n

Rd

then (f n ) is uniformly integrable, and hence relatively weakly compact in L1 (U ).

3.5.1

Weak compactness of (f n )

The weak compactness of the sequence (f n ) of solutions to the truncated problems follows immediately
from the a priori estimates.
Lemma 12. The sequence (f n ) of solutions to (Bn) is weakly compact in L1 ((0, T ) Rd Rd ).
Proof. Observe that combining estimates (27)-(28) with equations (29)-(30) we have
Z Z
sup sup
f n (1 + |x|2 + |v|2 + | log f n |)dxdv CT
n t(0,T )

for all T > 0. The claim then follows by applying Corollary 3 with U = (0, T ) Rd Rd and
h(f n ) = f n | log f n |
w(x, v) = |x|2 + |v|2 .

Up to the extraction of a subsequence (that we still call (f n )), hence

fn * f

in L1 ((0, T ) Rd Rd ) .

24

3.5.2

Weak compactness of the collision term

Theorem 5. For any T > 0, R > 0, the sequences

Qn (f n , f n )
1 + fn

Qn+ (f n , f n )
1 + fn

n1

are relatively weakly compact in L1 ((0, T ) Rd BR ), where BR = {v Rd : kvk R}.

To prove this, we start from the negative part Qn (f n , f n ), and show that it is a uniformly integrable
sequence, so that weak compactness follows from DP criterion. We then obtain control on the positive
part via the following inequality.
Lemma 13 (Entropy inequality). For all K 1 we have
Qn (f n , f n ) KQn (f n , f n ) +

1
en (f n )
log K

K 1 ,

(31)

where en is defined as in (24) replacing q with qn .

Note that this remarkable inequality gives us control on the global part of the collision operator in
terms of the local one, and vice versa. This is, as will become clear from the proof, due to the particular
form of the collision kernel, and is therefore characteristic of the Boltzmann equation.
Proof. Since n is fixed, we drop it in the notation. Define the events
AK = {(v, v , ) : f 0 f0 Kf f }
BK = AcK .
Then
Z Z

q(v v , )f 0 f0 [ 1AK + 1BK ] dv d

Z Z
Z Z
Z Z
K
q(f f ) 1BK dv d +
q(f 0 f0 f f ) 1AK dv d +
q(f f ) 1AK dv d
Z Z
Z Z
Z Z
f 0 f0
1
q(f 0 f0 f f ) log
1AK dv d + K
q(f f ) 1AK dv d
K
q(f f ) 1BK dv d +
log K
f f
Z Z
f 0 f0
1
q(f 0 f0 f f ) log
1AK dv d
= KQ (f, f ) +
log K
f f

Q+ (f, f ) =

as claimed.
We are now ready to prove Theorem 5.
Qn (f n , f n )
satisfies the DP criterion in L1 ((0, T )
1 + fn
Rd BR ), and is therefore weakly compact. Note that, since

Proof. Let us start by proving that the sequence

Qn (f n , f n )
fn
An f n
R

An f n
n
n
1+f
1+f
1 + n f n dv
it suffices to prove that the sequence An f n is uniformly integrable, i.e. it satisfies properties (a)-(b)-(c)
of Definition 18.
To show (a) observe that
Z

Z Z
Z Z Z
Z Z
(An f n )dvdx =
An (v z)f n (z)dzdvdx =
f n (x, z)
An (z v)dv dzdx
Rd BR
Rd BR Rd
Rd Rd
BR
Z Z
Z Z
n
2
n
2
C
f (x, z)(1 + |z| )dzdx = C
f0 (x, z)(1 + |z| )dzdx
Rd

Rd

BR

25

BR

by (12), and the last term is uniformly bounded in n by (29). Let us now move to conditions (b) and
(c). We split the proof in several steps.
Step 1
Assume kAkL1 (Rd ) < . Then
kAkL1 (Rd ) = kqn kL1 (Rd S d1 ) =: an
and, using the properties of the qn , we can assume the an to be uniformly bounded in n. Let us
now introduce the function (s) = s log s 0, and observe that to show (b) it is enough to prove


s
n

1
d
+
log
a
=
that (f
)

L
((0,
T
);
L
(R

B
)).
To
this
end
note
that
(s)

s
log
s
=
s
log
R
a




s
s
a a + s log a a a + s| log a| for all a 0. Apply this inequality with a = an yields


Z Z
Z Z
Z Z
An f n
n
(An f )dvdx an
dvdx + | log an |
An f n dvdx

an
Rd BR
Rd BR
Rd BR
Z Z

An (f n )dvdx + | log an |an kf n kL1 (Rd Rd )

Rd BR
Z Z
an
(f n )dvdx + | log an |an kf n kL1 (Rd Rd )
Rd

Rd

The last term is then uniformly bounded in n, since

an , | log an | are uniformly bounded
RR
RR n

(f n )dvdx f n | log f n |dvdx

f0 (| log f0n | + 2|x|2 + 2|v|2 ) + Cd by (28), and the uniform
bound for the RHS follows from (29)-(30).
kf n kL1 (Rd Rd ) = kf0n kL1 (Rd Rd ) and the sequence kf n kL1 is uniformly bounded because of (29).
In conclusion, since this holds for each fixed t (0, T ), we have proved that (f n ) is uniformly bounded
in L ((0, T ); L1 (Rd BR )), and (b) follows2 .
We now move to (c). Observe that it is enough to show that
Z Z
(An f n )dvdx 0
as K
Rd

|v|>K

uniformly in n (and t). To this end, split the integral as follows:

Z Z
Z Z Z
n
(An f )dvdx
An (v v )f n (x, v ) 1(|v |K/2) dv dvdx
Rd |v|>K
Rd Rd Rd
Z Z Z
+
1(|v|K/2) 1(|v |K/2) An (v v )f n (x, v )dv dvdx
Rd

Rd

Rd

= I + II
For the first term, integrating over v and using the a priori bound (21), we get
Z Z
Z Z
an
n
|v |2 f n (x, v )dv dx
I an
f (x, v ) 1(|v |K/2) dv dx
2
(K/2)
d
d
d
d
R
R
R
Z RZ
4an
C
2 n
= 2
|v | f0 (x, v )dv dx 2 0
as K
K Rd Rd
K
where, by (29), the constant C is independent of n. For the second one:
!
!
Z
Z
n
II
An (z)dz kf kL1 (Rd Rd ) =
An (z)dz kf0n kL1 (Rd Rd )
|z|K/2

C0

|z|K/2

Z
An (z)dz
|z|K/2

2 Note

that, as pointed out in [2], we could have set R = in this argument. We will need, though, R to be finite in
the next step, where we remove the integrability assumption on A.

26

where C 0 is a uniform bound on kf0n kL1 (Rd Rd ) again given by (29). From the integrability assumption
on A, then, we deduce that the last term then becomes arbitrarily small as K , and (c) follows.
Step 2
We now remove the integrability condition on A. Note that if the An s are all supported in a compact
set then all the previous calculations continue to hold, and we are done. If not, take K > 0 and truncate
the An s by defining
An,K (z) := An (z) 1(|z|K) .
Then weak compactness will follow from
sup kAn,K f n An f n kL ((0,T );L1 (Rd BR )) 0

as K

(32)

To see that, we calculate explicitly that

kAn,K f n An f n kL1 (Rd BR ) =

Z
Rd

BR

Rd

An (v v ) 1|vv |K f n (x, v )dv dvdx

and observe that if K > R then

|v | = |v v + v| |v v| |v| K R
so that
1(|v v|K) 1(|v |KR) .
Using this inequality and integrating then An with respect to v we get

Z Z Z
An (v v )dv 1(|v |KR) f n (x, v )dv dx
kAn,K f n An f n kL1 (Rd BR )
BR
Rd Rd
Z Z
Z Z
C
|v |f n (x, v )dv dx

(1 + |v |2 )f n dv dx +
(K R)2 Rd Rd
Rd Rd
Z Z
Z Z
C
2 n
|v |f0n (x, v )dv dx

(1 + |v | )f0 dv dx +
(K R)2 Rd Rd
Rd Rd
and (32) follows by letting K and then 0, and observing that the convergence is uniform in n,
again by (29).
Step 3 We deduce the weak compactness of the (renormalized) positive part of the collision operator
from that of the negative part via entropy inequality, by noting that for each R > 0 and A Borel set of
Lebesgue measure (A), (31) implies
Z TZ Z

Qn+ (f n , f n ) 
sup
1A + 1(|x|+|v|R) dvdxdt 0
as (A) 0, R .
n
1+f
n
0
Rd Rd
and uniform integrability follows.

3.5.3

Temporal regularity of f

Combining the weak compactness of the above sequences with the a priori estimates, we can obtain some
information about the regularity in time of the limit f . Our goal is to show that
f C((0, T ), L1 (Rd Rd )) .
We start by introducing, for any fixed > 0, the new sequence
gn =

1
log(1 + f n )

and observing that each gn solves

T gn =

1
Qn (f n , f n ) ,
1 + f n

where T denotes, as usual, the transport operator.

27

(33)

Lemma 14. If the sequence (gn ) is defined as in (33), it holds:

Z Z
sup sup
|gn f n |dxdv 0
n t(0,T )

Rd

as & 0

(34)

Rd

d
1
Proof. Consider (s) = 1 (1 + s), so that gn = (f n ). Note that (s) =
, so that for each
dt
1 + s
s
. For any R > 0, therefore, we have
s > 0, (s)
1 + s


s
(s (s)) = (s (s)) 1(s<R) + (s (s)) 1(sR) s
1(s<R) + s 1(sR)
1 + R
sR () + s 1(sR)
where R () = 1
we get
n

kf

1
1+R

0 as & 0. Hence, replacing s with f n and splitting the integral as above,

gn kL1 (Rd Rd )

Z Z

Z Z

Z Z

|f (f )|dxdv
f R ()dxdv +
Z Z
Z Z
= R ()
f0n dxdv +
f n 1(f n R) dxdv
Z Z
CR () +
f n 1(f n R) dxdv
=

f n 1(f n R) dxdv

where C is a constant independent of n, R, , and it is finite because of the convergence (29). Now, since
f n is uniformly integrable in L1 ((0, T ) Rd Rd ), we conclude
Z Z
|gn f n |dxdv 0 as & 0
sup sup
n t(0,T )

Rd

Rd

as claimed.
In view of (34), we focus on the time regularity of the functions gn . Observe that, since
T gn =
we have

gn]
=
t

1
Qn (f n , f n )
1 + f n
]
1
n n
n
Q
(f
,
f
)
1 + f n

and hence
gn] (t + h) gn] (t) =

t+h

Qn (f n , f n )] (s)
ds .
1 + f n]

Therefore for all t, h such that 0 < t < t + h < T we have

n]

n]
g (t + h) g (t)

L1 (Rd BR )

Z

 t+h Qn (f n , f n )] (s) 


=

 dsdxdv
1 + f n] (s) 
Rd BR  t
Z t+h Z Z
|Qn (f n , f n )] (s)|

dsdxdv
1 + f n] (s)
t
Rd BR
Z t+h Z Z
|Qn (f n , f n )|
=
dsdxdv
1 + f n
t
Rd BR
Z

(f ,f )
Now, since the sequence Q 1+f
is uniformly integrable in L1 ((0, T ) Rd BR ), we can bound the
n
right hand side by (h), where is a function which is independent of n and such that (h) 0 as
& 0. Hence we deduce that




sup sup gn] (t + h) gn] (t) 1 d
0
(35)
L (R BR )

n t(0,T )

28

as h & 0. Moreover,


sup f n] (t + h) f n] (t) L1 (Rd Rd ) sup

Z Z

|f n] (t + h) f n] (t)|dxdv

t(0,T )

t(0,T )

Z Z

|f n] (t + h) gn] (t + h)| + |gn] (t + h) gn] (t)|( 1BR + 1BRc ) + |f n] (t) gn] (t)|dxdv

sup
t(0,T )

0 (h) + sup

Rd

BR

|gn] (t + h) gn] (t)|

and by taking the supremum over n and using (35) we conclude that


sup sup f n] (t + h) f n] (t) L1 (Rd Rd ) 0
n t(0,T )

as h & 0. Being the above convergence uniform in n, the weak limit f ] of f n] also satisfies


sup f ] (t + h) f ] (t) L1 (Rd Rd ) 0
t(0,T )

as h & 0, and therefore belongs to C((0, ), L1 (Rd Rd )).

From this information we deduce that f C((0, ), L1 (Rd Rd )) by the change of variables (x+vh, v) 7
(x, v):
Z Z
Z Z
0 sup
|f ] (t + h) f ] (t)|dxdv = sup
|f (t + h, x + vh, v) f (t, x + vh, v)|dxdv
t(0,T )

t(0,T )

Z Z
= sup
t(0,T )

|f (t + h, x, v) f (t, x, v)|dxdv = sup kf (t + h) f (t)kL1 (Rd Rd )

t(0,T )

and we conclude that

f C((0, ), L1 (Rd Rd ))
as claimed.

3.6

Velocity averaging

To make use of the compactness results proved in the preceding section we go back to the sequences
gn =

1
log(1 + f n )

indexed by > 0, and recall that each gn solves

T gn =

1
Qn (f n , f n ) .
1 + f n

(36)

The main advantage in using this sequences is that now we know that the RHS of (36) is weakly relatively
compact in L1 ((0, T ) Rd BR ) and such is, therefore, T gn . Moreover, since
0 T gn f n
and f n is weakly compact in L1 ((0, T ) Rd Rd ), we deduce from the DP criterion that T gn is also
weakly compact in the same space. Therefore the velocity averaging results apply to (gn ), and we deduce
the following.
Proposition 6. For each T > 0 we have:
Z
Z
(i)
f n dv f dv a.e. and in L1 ((0, T ) Rd ).
(ii) An f n A f a.e. and in L1 ((0, T ) Rd BR ), for all R > 0.
29

(iii) For each function compactly supported in L ((0, T ) Rd Rd ),

R n n n
R
Q (f , f )dv
Q (f, f )dv
R
R

n
1 + f dv
1 + f dv
as n in L1 ((0, T ) Rd ).
Proof. Note that the family (indexed by > 0) of functions (s) = 1 log(1 + s) satisfies all the
assumptions of Lemma 5, namely ( ) is a uniformly Lipschitz family in C(R; R) with (0) = 0 and
(z) z as 0 uniformly on compact sets. Moreover, as already noted, the sequence T gn = T (f n )
is weakly relatively compact in L1 ((0, T ) Rd BR ). Hence we can apply Corollary 1 with n 1
to obtain the convergence of the velocity averages
Z
Z
n
f dv f dv
in L1 ((0, T ) Rd ). For (ii) we use the same truncation argument of step 2 in the proof of Theorem 5.
By (12), which holds uniformly in n, for any K > 0 we have
!
Z
Z


An (v v )dv =
An (v v )dv
1(|v|K) + 1(|v|>K)
|v |R

|v |R

An (v v )dv

sup
|v|K

(37)

+ C(1 + |v| ) 1(|v|>K)

|v |R

d
1
where the last integral is finite since An L
loc (R ; L (BR )). We write An,K (v) := An (v v ) 1(|v|>K) .
We have

kAn f n A f kL1 ((0,T )Rd Rd ;L1 (BR )) kAn f n An,K f n kL1 ((0,T )Rd Rd ;L1 (BR ))
+ kAn,K f n AK f kL1 ((0,T )Rd Rd ;L1 (BR ))
+ kAK f n A f kL1 ((0,T )Rd Rd ;L1 (BR )) .
But by (12)
kAn f n An,K f n kL1 ((0,T )Rd Rd ;L1 (BR )) =
Z

Z Z

Z Z

An f n dvdxdt
|v|>K

f n (1 + |v|2 )dv 0

C
0

|v|>K

as K (recall that supn

f n (1 + |v|2 )dv < ). So to prove the claim it is enough to show that

kAn,K f n A f kL1 ((0,T )Rd Rd ;L1 (BR )) 0

as n . To this end, define the sequence
K
n (t, x, v, v ) = An,K (v v)
and note that by (37) these functions are uniformly bounded in L ((0, T ) Rd Rd ; L1 (BR )). We can
therefore apply Corollary 2 to get the strong convergence in L ((0, T ) Rd BR )
Z
Z
Z
n
K
K
(t,
x,
v,
v
)f
(t,
x,
v)dv

f
dv
=
AK (v v)f (t, x, v)dv = AK (t, x, v ) .

n
By finally letting K we obtain the strong convergence of An f n to A f in L ((0, T ) Rd BR )
as claimed.
To show the convergence (iii), fix any compactly supported function L ((0, T ) Rd Rd ), and
define the sequence
An f n
R
.
n =
1 + f n dv
30

R
R
By (i) and (ii) An f n and (1 + f n dv) converge almost everywhere to A f and 1 + f dv respectively,
and hence
Af
An f n
R
R

a.e.
1 + f n dv
1 + f dv
We can therefore apply again Corollary 1 to get
R n n n
R
Z
Z
Q (f , f )dv
Q (f, f )dv
n
n
R
R
= f (An f )n dv f (A f )dv =
1 + f n dv
1 + f dv
in L1 ((0, T ) Rd ), and this concludes the proof.

3.7

Exponential formulation

Another kind of solutions to the Boltzmann equation can be obtained by viewing (A f )f as a linear
feedback term with variable coefficient c(t, x, v) = A f which we think of as a given function. The
equation is then
T f + cf = Q+ (f, f ) = Q+
We then have the following form of solutions:
 
Z t


Z t
d
]
]
]
c ds Q+]
c ds f = exp
exp
dt
0
0
Duhamels principle then gives
t

Z
exp


Z
Z t
c] ds f ] = f0 +
exp

]
c] (s, x, v) ds Q+] (, x, v) d

This is unpleasant looking, but with some notation we can make it nicer. R
t
Define T 1 g as the solution to T f = g, f |t=0 = 0. Then (T 1 g)] = 0 g ] (s, x, v) ds. Also define
R
t
F = T 1 c so F ] = 0 c] (s, x, v) ds, then TF1 = eF T 1 eF (an operator composition viewing eF as
multiplication operators). Then we see that
Z t

]
]
exp
c (s, t, v) ds = eF (t,x,v)
0

Z
0

(eF )] Q+] (, x, v) d = (T 1 eF Q+ )] = (eF TF1 Q+ )]

So the equation is (eF f )] = f0 + (eF TF1 Q+ )] , and undoing the characteristic map ] and dividing by eF
we obtain
f = f0 eF + TF1 Q+ (f, f )

where F = T 1 (A f )

= eF (f0 + T 1 [eF Q+ (f, f )])

This form has some advantages, in particular due to the explicit formula for T 1 we can see that it
is a continuous and weakly continuous map from L1 ((0, T ) Rd Rd ) to C([0, T ]; L1 (Rd Rd )) which
preserves non-negativity. Furthermore for F C((0, T ); L1 (Rd Rdloc )), we see that TF1 is strongly and
weakly continuous from L1 ((0, T ) Rd Rdloc ) into C((0, T ); L1 (Rd Rdloc )).
Theorem 7. We have
f = f0 eF + TF1 Q+ (f, f )
where F = T 1 (A f ).
Proof. To prove this result, we first show that f f0 eF + TF1 Q+ (f, f ). To this end, we once again
consider the sequence gn = 1 log(1 + f n ). As shown earlier, this sequence is weakly (relatively) compact
as n , so we assume that
gn g
31

for some g weakly as n . It is easy to see, as in the proof of Lemma 5, that lim0 g = f as 0
strongly in L1 .
Similar arguments to the above tell us that the sequence
ln :=

fn
1 + fn

is weakly compact. Thus we may assume that ln l weakly as n . Moreover, l f as 0

strongly in L1 .
We will also approximate F by a sequence Fn = T 1 (An fn ) and we see that T (eFn ) = eFn (An fn )
Finally, we notice that
Qn+ (fn , fn )
1 + fn
is weakly compact, so we can assume
Qn+ (fn , fn )
Q+,
1 + fn
weakly for some Q+, .
Now that we have built up some convergent sequences, we perform the following computations
n
Fn
n
n
Fn
T (eF
n g ) = e T (g ) + g T (e )

eFn
Qn (f n , f n ) + gn eFn (An fn )
1 + f n


fn
eFn
n
n
n
Fn
n
Q
(f
,
f
)
+
e
(A

f
)
g

=
n
n

1 + f n +
1 + f n

By integrating, we see that

(eFn gn )(t, x, v)

(eFn gn )(0, x, v)

eFn
Qn (f n , f n )(s, x (t s)v, v)ds
n +
0 1 + f


Z t
fn
Fn
n
e (An fn ) g
+
(s, x (t s)v, v)ds
1 + f n
0
 Fn n n n 
e Q+ (f , f )
(t, x, v)
= T 1
1 + f n



fn
+ T 1 eFn (An fn ) gn
(t, x, v)
1 + f n
Z

If we define TF1
by TF1
:= eFn T 1 eFn then the above equation reads
n
n
gn




 n n n 
Q+ (f , f )
1
fn
1
1
n Fn
n
= log(1 + f0 )e
+ TFn (An fn ) g
+ TFn

1 + f n
1 + f n

(38)

We will need to bound the final term on the right hand side. To do this, we see that for any non-negative
L with compact support


Z  n n n 
Z 
Q+ (f , f )

n
n
n
R
R
dv

Q
(f
,
f
)
dv
+
1 + f n
1 + fn dv1
1 + fn dv1
From the prior results on velocity averaging Proposition 6 (ii), sending n yields


Z  n n n 
Z 
Q+ (f , f )

R
R
lim sup
dv

Q
(f,
f
)
dv
+
1 + f n
1 + fn dv1
1 + f dv1
n
But the left hand side converges to
Z
Q+,

1+
32

R
dv
f dv1

and so

R
dv
Q+,
1 + f dv1

Z
Q+ (f, f )

1+

R
dv
f dv1

Because this holds for any 0, we conclude that

Q+, Q+ (f, f )
almost surely in v, and so because T 1 is a positive functional we have
T 1 (Q+, ) T 1 (Q+ (f, f ))
It is easy to see that if xn x weakly in L1 then TF1
(xn )(t) TF1 (x)(t) in L1 . We will now return
n
to equation (38). Sending n gives
1
log(1 + f0 )eF + TF1 ((A f ) [g l ]) + TF1 (Q+, )

1
log(1 + f0 )eF + TF1 ((A f ) [g l ]) + TF1 (Q+ (f, f ))

g =

Notice that
h
i
1
1
log(1 + f0 ) = log (1 + f0 ) log(ef0 ) = f0

as 0.
Also by the continuity of TF1 ,
TF1 ((A f ) [g l ]) TF1 ((A f ) [f f ]) = 0
and so
f f0 eF + TF1 Q+ (f, f )
It remains to show that
f f0 eF + TF1 Q+ (f, f )
Our argument to show this will be similar to the one we have just performed. This time, we consider
the sequence of cut-off functions (hn ) such that hn = min(f n , 1 ). It is clear that hn h weakly for
some h satisfying
h f
We also get Qn+ (hn , hn ) Q+ (h , h ) weakly. A similar computation to the one earlier (except this
time considering T (eF fn )) gives
f n = f0n eFn + TF1
(Qn+ (f n , f n ))
n
But f n hn , and again because TF1
is positive
n
f n f0n eFn + TF1
(Qn+ (hn , hn ))
n
We send n in this and conclude that
f f0 eF + TF1 (Q+ (h , h ))
Next, by the monotone convergence theorem,
Q+ (h , h ) Q+ (f, f )
and so
f f0 eF + TF1 (Q+ (f, f ))
as was claimed.
Theorem 8. The f constructed above is indeed a renormalized solution.
33

Proof. Firstly, we have

0

Q (f, f )
(A f )f
=
Af
1+f
1+f

But we know from Lemma 6 that limn An f n A f in L1 ((0, T ) Rd BR ) sense for all T, R > 0.
Thus
Q (f, f )
L1 ((0, T ) Rd Rd loc )
1+f
for any T > 0.
We will show the (harder) corresponding result for
inequality (13),

Q+ (f,f )
1+f .

To do this, recall that, from the entropy

en (f n )
log(2)
R
Next, for any positive , we divide by the positive function (1 + f n dv) to get
Qn (f n , f n ) 2Qn (f n , f n ) + 4

Qn (f n , f n )
Qn (f n , f n )
en (f n )
R
R
2 R n +4
(39)
n
1 + f dv
1 + f dv
log(2)(1 + f n dv)
R
R
We have shown in Proposition 6 that f n dv f dv. We also showed in that same proposition that
for any function L ((0, T ) Rd Rd ) with compact support
R
R
Qn (f n , f n )dv
d Q (f, f )dv
Rd
R
R
R
n
1 + f dv
1 + f dv
Moreover, since en (f n ) are L1 (R+ Rd Rd ) uniformly bounded, they converge vaguely to some finite
measure . The absolutely continuous part of (which we denote by e) is in L1 ((0, T ) Rd Rd ) and
is non-negative and so, by taking limits in (39)
R
R
Q (f, f )dv
e()
d Q (f, f )dv
Rd
R
R
R
R
+4
1 + f dv
1 + f dv
log(2)(1 + f dv)
By sending 0, we see that
Z

Z
Q (f, f )dv

Rd

Q (f, f )dv +
Rd

4e()
log(2)

and because e is absolutely continuous we get

Q (f, f ) Q (f, f ) + E

(40)

where E L1 ((0, T ) Rd Rd ) is non-negative. Thus

Q+ (f, f )
Q (f, f )
E
Q (f, f )

+E
1+f
1+f
1+f
1+f
and so

Q+ (f,f )
1+f

L1 ((0, T ) Rd Rdloc ).

Next, we show that Q+ (f, f )# (t, x, v) is in L1 (0, T ) for almost all (x, v). To do this, recall that by

#
Lemma 7 we have TF1 (Q+ (f, f )(t, x, v)) L1 (Rd Rdloc ) for every t > 0 and so TF1 (Q+ (f, f )(t, x, v))
L1 (Rd Rdloc ). Writing this out fully, we see that for compact sets K Rd
Z Z

h
i#
(eF (t) )# (x, v) T 1 eF (t) (Q+ (f, f )(t, x, v)) dxdv
K Rd
Z Z
Z t


#
>
(eF (t) )# (x, v)
eF (s) Q#
(f,
f
)(s,
x,
v)
dsdxdv
+
K Rd
0
Z Z Z t


#
#
>
e(F (t)F (s)) (x, v) Q#
(f,
f
)(s,
x,
v)
dsdxdv
+

>

Rd

34

Rt
But F # (t, x, v) = 0 (A f )(s, x, v). We know that A f is non-negative, so F # increases in t and is also
non-negative. Thus
Z Z Z t


#
>
e(F (t)) (x, v) Q#
(41)
+ (f, f )(s, x, v) dsdxdv
K

Rd

It is clear that for each t > 0 we have F # (t) L1 (Rd Rdloc ). Define 0 to be the set of (x, v) Rd Rd
such that F # (t) < . Then c0 is null.
We also define 1 to be the set of (x, v) such that
Z t


#
e(F (t)) (x, v) Q#
(f,
f
)(s,
x,
v)
ds <
+
0

By (41), c1 is also null. On 0 1 , which has null complement, we have

Z t


#
e(F (t)) (x, v) Q#
+ (f, f )(s, x, v) ds <
0

and
eF

(t)

>0

Thus, on this set

Z t


Q#
+ (f, f )(s, x, v) ds <

0
#

which proves that Q+ (f, f ) (t, x, v) is in L1 (0, T ) for almost all (x, v). By (40), Q (f, f )# (t, x, v) is in
L1 (0, T ) for almost all (x, v). Then by Theorem 1, f is a renormalized solution.

Rough coefficients

4.1

Another use for renormalised solutions

In the previous section we saw how the notion of renormalised solutions allowed a quadratic non-linearity
to be controlled. Another use of renormalised solutions is to make sense of very irregular solutions. In fact,
we can have solutions that are only measurable and finite a.e. in the space L. In this space distributional
solutions do not make sense. We will give a different definition of a solution in such a space. To do this,
instead of using the derivative of to bring down a non-linearity, we will use bounded to control the
spikes of the solution. We will show that even in this space we have the stability of solutions subject
to perturbations of the coefficients, which allow us to gain knowledge about the characteristic curves
even when standard ODE theory would not tell us about their existence. In particular, we will state a
generalised version of Cauchy-Lipschitz that applies even when the force-field term is not Lipschitz. This
result is due to [3], and an adapted version of the time independent divergence free case is presented
here.
4.1.1

Cauchy-Lipschitz with rough coefficients

The well known Cauchy-Lipschitz3 theorem for ODEs gives existence and uniqueness for locally Lipschitz
forcefields. We will obtain a slightly weaker result under the assumption that the force-fields are merely
1,1
Wloc
, (i.e. almost Lipschitz). The solutions will exist for only almost all initial data.
1,1
Theorem 9 (Main result). Let b : RN RN , b Wloc
, div b = 0 a.e., and4
the equation

X = b(X)
X(0) = x
has a unique solution X(t, x) in the sense that for almost all x RN ,
3 Also
4 i.e.

called Picard-Lindelof
= b1 + b2 , b1 L1 , b2 L . See section 2 for more details.

b
1+|b|

35

b
1+|x|

L1 + L . Then
(42)

 The equation is satisfied for t R.

X(, x) C 1 (R), b(X) C(R).
4.1.2

Strategy

The Cauchy-Lipschitz theorem is usually proven using a fixed point argument involving the Lipschitz
condition. Here we will use a completely different method; we will view Cauchy-Lipschitz as a statement
about PDEs.
Theorem 10 (Cauchy-Lipshitz (smooth case)). Let b C (RN ; RN ). Then the ODE stated in (42)
has a unique smooth solution (in both variables) X(t, x) and
Xi (t, x)
= bi (X)
t
In particular, the solutions of the ODE are the characteristics of the transport equation
u
u
bi
=0
t
xi
u(0, x) = x

(43)

Instead of looking at the original ODE we will attack the transport equation. We will derive stability
results using renormalised solutions, and then return to the original ODE via the characteristics.
4.1.3

Techniques and ideas used

Other than the notion of renormalised solutions, the proofs that follow use standard regularising and
approximating methods.

4.2

A proof using the stability result

We now have the machinery to state the stability result that allows us to prove the main theorem.
Theorem 11 (Stability). Let bn b, cn c, div bn div b respectively in L1loc and assume that b, div b
and c satisfy
1,1
b Wloc
, div b L , c L

|b|
L1 + L
1 + |x|

(44)
(45)

Furthermore, let un be the renormalized solution to (1) (with b replaced by bn etc.) with initial data
un0 L such that un0 converges in L to some u0 L. We also assume that un is a bounded sequence in
L (L). Then un u in C(L) where u is the renormalized solution of the original transport equation
(1).
Theorem 12 (Existence and measure preservation). Let u0 L. Then there is a unique renormalized
solution u to (1) such that u L (L) C(L).
We shall now restate and prove the main theorem using these results. Later, we will show that these
results are indeed valid.
Theorem 13 (Main result). Assume that (44) and (45) are satisfied and that c, div b = 0 and b depend
only on x. Then there is a unique X(t, x) C(R; L) such that a.e. in x we have
(i) X(, x) C 1 (R), b(X(, x)) C(R)
(ii)

Xi
t

= bi (X) i.e. the original ODE.

Also (with no a.e. restriction) we have

36

(a) For a fixed t, X(t, x) preserves Lebesgue measure

R (i.e. X(t,R) = where is the Lebesgue
measure)so that for all test functions we have (X(t, x))dx = (x)dx.
(b) (Group property) For all t, s and for almost all x we have X(t + s, x) = X(t, X(s, x)).
(c) The equation
i
[i (X)]
=
bj (X)
t
Xj

(46)

(X(0, x)) = (x)

holds in the sense of distributions for all C 1 (RN ; RN ) with , |(D)(z)|(1 + |z|) L . This is
the notion of renormalized solutions restated in the RN RN case.5
(d) (X) L1loc (RN ; C(R)) for all such .
Proof. We start with existence. The plan is as follows:
1. Mollify b and use the smooth version of Cauchy-Lipschitz and characteristics to formulate a sequence of solutions to the transport equation.
2. Fix a , show compactness and apply the stability result to gain existence and uniqueness of the
solution to (46).
3. Deduce measure preservation and group property from the transport equation.
4. Use properties of a particular to deduce continuity.
5. Use similar compactness ideas to deduce that b(X) is continuous.
6. Use (46) to deduce the statements that hold for almost all x.
We will now execute this plan:
= b(X)
where X
and b are obtained
1. Apply the smooth version of Cauchy-Lipschitz to solve X
R
1
N
by a convolution with the function N (/) where D+ (R ) and = 1. Doing this yields
x) which obeys X = b(X)
and X(0,
x) = x. Furthermore, provided u0 is
a unique smooth X(t,
t
to see
smooth, we can use the chain rule, the group property and the measure preservation of X
0
that u
= u (X) solves
u
u

= bi
, u
(0, x) = u0 (x)
t
xi
satisfies
and so by choosing u0 (x) = xi (i.e. u0 is a projection) we deduce that X
i
i
X
X
= bj
,
t
xj

x) = x
X(0,

2. As everything is smooth, we know that a function is a renormalised solution to the perturbed

ODE if and only if it is a solution to the perturbed ODE. We wish to extract a subsequence
that converges for any that obeys the conditions on (46). By combining the equations
from X
i
X

X i
t = bi (X) = bj xj and the chain rule we see that
h
i

i (X)
t

= bj

h
i

i (X)
xj

i
b (X)
j j
X

which is the perturbed version of (46). Our assumptions on and (45) tell us that for every
bj (z) L1 + L . By the measure preservation property this is true for z = X,
i.e.
i, zi (z)
j
i
L1 + L uniformly in t, and we obtain
bj (X)
j
X

5 Here

and throughout,

i
Xj

means

i
xj

evaluated at X.

37

]
[i (X)
t
]
[i (X)
t

is bounded in L (R; L1 + L )
is relatively compact in the sets L (T, T ; L1 (BR )) for all R, T < .

bounded in L (R; L1 ). This is enough to apply the stability result, hence

This tells us that (X)
loc
converges in C([T, T ]; L)N for all 0 < T < to an X that solves the original (46).
X
3. We can use these to approximate the test functions and obtain (a) using the measure preservation
result on the transport equation. The group property follows from the time invariance of the
u
transport equation. We observe that X(t, X(s, x)) solves u
s bi xi = 0 for u(0, x) = X(t, x) in
the sense of a renormalised solution.
4. We repeat the reasoning given in step 2 on equation (46) to obtain that [(X)]
L (R; L1 +L ) ,
t
1,1
1
N
1
1
N
1
N
Lloc (R ; Lloc (R)). Hence (X) Lloc (R ; Wloc (R)) , Lloc (R ; C(R)). Thus for a.e. x we have
that (X) is continuous. Fixing any compact interval and choosing a such that is strictly
increasing on this interval tells us that X is continuous on that interval, and hence X is continuous
on all of R.
5. Doing the same for b(X) is slightly more tricky. We introduce a positive even C 1 (R) function
and look at
i

X
h

+ (X)
i bj Xk
(X)i (b(X))
=
i (b(X))

t
Xj t
bj Xk t

+ (X)
i bj bk (X)

=
bj (X)i (b(X))
j
k
X
bj X
N

N 1
By a Sobolev embeddings argument and the fact that Db L1loc we know that b Lloc
. Thus
|b| log(1 + |b|) L1loc , because log grows sufficiently slowly. We can now pick so that:

(|z|)|Db(z)| f (z) L1+

0 (|z|)|b| log(1 + |b|) g(z) L1
+

for some f ,g and uniformly for  [0, 1]. We can use this to bound the time derivative as follows:
 h
i




where z = b(X)
 t (X)i (b(X))  g(X) kk + f (X) |zD(z)|
+ f (X)
kD(z)(1 + |z|)k
Cg(X)

+ C 0 f (X)
L1+
Cg(X)
h
i
preserves measure we have that (X)
i (b(X))

Because X
is bounded in L (R; L1 ) and weakly
t
relatively compact in L1 (T, T ; L1 (BR )) for all finite R, T . We then send  0 and obtain

1
N
t [(X)i (b(X))] L (R; L ), and so by using the same method as before, for almost all x R ,
1
(X)(b(X)) is continuous on R. Additionally, for almost all x, X C(R), C (R) and so we
can choose to see that b(X) is continuous on R.
6. Now we just let tend to the identity, and use our almost everywhere continuity of X and b(X)
to deduce the integral form of X = b(X). This gives the almost everywhere existence of solutions
to the originally ordinary differential equation.
Now we will show uniqueness. It is sufficient to show that for arbitrary u0 D(RN ), u(t, x) = u0 (X(t, x))
solves the transport equation with c = 0 (which we already know has a unique solution). Essentially we
are showing that X is the characteristic map for the transport equation with a test function as initial
data.
We have already proved the existence of an X such that for almost all x, X(, x) C 1 (R) and X
satisfies the ODE. As u0 is smooth, we can apply the chain rule to obtain that for almost all x,
du
u
=
bi (X)
dt
Xi
38

u
1
N
But this is just the statement that u is a mild solution to u
t bi xi = 0, and certainly u Lloc (R R )
as it has compact support. The forcing term is 0 L1loc (R RN ). Also X is invertible by the group
property and preserves measure, so we can apply theorem 1 to show that u is a distributional solution
and hence unique.

4.3

Proof of stability result

We have already seen that the transport equation has some stability in lemma 2. However, the most
general result requires many separate approximation arguments, so to simplify things we will only consider the case where c, cn = 0 and b depends only on x (which was all that was needed above). We will
look at the equation
u
u
bi
=0
t
xi
with bn b and div bn div b in L1loc . To remove some of the sub/super-scripts we will no longer
assume that u
is a convolution with as in the previous section: instead, we will write u
= un and adopt
a similar notation for other functions. The sequence of initial conditions will then be given by u
0 .
4.3.1

Stability

We have to prove that for all admissible , (

u) (u) in C(0, T ; L1loc ). To do this we will first prove
a.e. convergence, then show uniformity over compact sets. We will fix an admissible and set v = (
u).
4.3.2

Pointwise

The key idea here is that if (u) solves the equation then (u) itself is a renormalized solution, so (u)2
is a solution.
By the definition of u
we have that in D0 ([0, T ) RN ),

v
v
bi
=0
t
xi
and v|(t=0) = v0 = (
u0 ). Without loss of generality we can assume that v n converge in L ((0, T ) RN )
to some v in the weak sense. By the convergence of b and div b, v will solve the equation, and because
of the convergence of u
0 it has initial condition v 0 = (u0 ).
In the same way we can take w
= (
u)2 , to obtain a w with w
w such that w solves the equation
0
0 2
with initial condition w = (u ) . But then because v solves the equation it is a renormalised solution
by theorem 1. Hence v 2 (the square of v) solves the equation with initial condition (u0 )2 = w0 . By
uniqueness we then have that v 2 = w. However, v2 v 2 weak* in L ((0, T ) RN ) and hence in
u) where was arbitrary (under the
L2 (L2loc ). This gives us convergence locally in measure. As v = (
conditions imposed), we can vary over strictly increasing functions to deduce that u
u locally in
measure to some u. Therefore v (u) and so u is a renormalised solution.
4.3.3

On compact time sets

We now need to show that the convergence is uniform on our time interval [0, T ]. To begin with we will
show that for all such we have (
u) (u) in C(0, T ; L1loc ). It suffices to prove that if t t in [0, T ]
1
then (
u(t)) (u(t)) in L (BR ) for all balls BR . Because BR is of finite measure, it is sufficient to
show (
u(t)) (u(t)) in L2 (BR ). Applying our equation with a test cut off function R we have
d
dt

Z
w
R dx =

bi
R
w
R+w
bi
dx
xi
xi

The convergence of b and w

as n in L1loc then give
Z
Z
d
bi
R
w
R dx
wR + wbi
dx
dt
xi
xi
39

in L1 (0, T )

R
R
R
R
Hence v2 R dx v 2 R dx uniformly in [0, T ], and v(t)2 R dx v(t)2 R dx in C[0, T ]. Now we
consider s sufficiently large so that (
v ) is relatively compact in H s (BR ). Then v(t) v(t) in H s (BR )
2
and hence weakly in L (BR ). Combining this with the uniform convergence of integrals above we obtain
that v(t) v(t) in L2 (BR ), and because v C(0, T ; L2 (BR )) we have that v v in C(0, T ; L2 (BR )).
Now we have to show that this is enough to deduce that u
u in C(0, T ; L). For uniformly bounded
u
and u we can use (z) = |z| M to obtain local convergence in measure. It is therefore enough
to show that we can reduce the unbounded case to this. To do this we note that for any measurable
that is finite a.e., we have that for any fixed ball BR RN ,
f : [0, T ] RN R
(
)
!
sup ({|f (t, )| M } BR )
t[0,T ]

sup |f (t, )| M

BR

t[0,T ]

As g(x) := supt[0,T ] |f (t, x)| is measurable and finite a.e. we have that 1 1 {|g| M } 0. By the
dominated convergence theorem on the bounded domain BR we see that
sup ({|f (t, )| M } BR ) 0
t[0,T ]

We choose an M sufficiently large to make the this less than /2 uniformly over f = u
and use the
bounded case to obtain a total error that is bounded above by .
4.3.4

Existence and uniqueness

Existence and uniqueness are now easy. We just use mollify the coefficients and initial conditions with
parameters  and respectively, and then solve the equation to find (
u) which is unique by standard
results. We choose to be as in the definition of renormalized solutions, with k strictly increasing.
Taking 0 we get existence and uniqueness of u
as a renormalised solution to the equation with
mollified coefficients. The stability result then allows us to send  0.

4.4

Further remarks

In [3] the bounded divergence and non-autonomous cases are considered. The result is similar to what
is stated here, but the group property fails, and measure preservation is no longer exact, becoming
exponential upper and lower bounds based on the bounds on the divergence. They also prove versions
of the stability result when c 6= 0 the coefficients are time-dependent, and for initial conditions in Lp
spaces. The methods of proof are similar, and so we omitted them here for the sake of brevity. The
results for ODEs are also shown to be sharp, in the sense that there is non-uniqueness of solutions to the
s,1
1,p
ODE for some divergence free b Wloc
for any 0 s < 1, and also for some time-independent b Wloc
for any p without any divergence assumption.

Conclusions

In this project, we have seen how the idea of renormalised solutions has allowed us to provide a solution
to the Boltzmann equation. We have derived various stability results to help us achieve this. Next, we
explored applications of these techniques to transport equations with rough coefficients. These surprising
results were of great interest to us and we hope that the reader found this project to be informative and
enjoyable. We briefly mentioned some of the many interesting remaining open problems in this field.
The various difficulties encountered in this project should suggest why these interesting problems remain
unsolved despite plenty of attention.

A
A.1

Appendix
Weak* compactness

Theorem 14 (Banach-Alaoglu). Let X be a normed vector space. Then the closed unit ball of X is
weak* compact.
Corollary 4. A set is relatively weak* compact if and only if it is bounded.
40

A.2

Weak compactness

Definition 18 (Uniformly integrable). A set F L1 (U ) is uniformly integrable if and only if

(a) F is bounded in L1 (U ).
(b) For any  > 0 there is a > 0 such that
Z
Z
|f | <  for every f F and for every A U with
1<
A

(c) F is tight when we view

R f F as the measure |f |. In other words, for every  > 0 there is a
compact K U with U \K |f | .
Theorem 15 (Dunford-Pettis). A collection of functions is relatively weakly compact in L1 if and only
if it is uniformly integrable.
Corollary 5. Suppose that |fn | |gn | where (fn ), (gn ) L1 and (gn ) is relatively weakly compact. Then
(fn ) is relatively weakly compact.
Proof. By Dunford Pettis, (gn ) is uniformly integrable. Fix  > 0. Then we see that as |fn | < |gn |, the
same that satisfies b in the definition of uniform continuity for (gn ) will also apply for (fn ). Hence
(fn ) is uniformly integrable, and relatively weakly compact.
Lemma 15 (Weak compactness in Lp ). Let p (1, ) and A Lp be bounded. Then A is relatively
weakly compact.

A.3

Strong compactness

Theorem 16 (Rellich-Kondrachov (special case)). Let U RN be open bounded and Lipschitz, and
s < t. Then H s (U ) H t (U ).

A.4

Continuity in Banach spaces

Lemma 16. Let X be a real Banach space. Suppose the function u : [0, T ) X R satisfies
and u X for a.a. fixed t. Then u(t) C((0, T ); X).

u
t

References
[1] DiPerna, R. J.; Lions; On the Cauchy problem for Boltzmann equations: global existence and weak
stability. Ann. of Math. (2) 130 (1989), no. 2, 321366.
[2] C. Cercignani, R. Illner, M. Pulvirenti; The Mathematical Theory of Dilute Gases, Springer-Verlag
(1994).
[3] R.J. DiPerna, P.L. Lions; Ordinary differential equations, transport theory and Sobolev spaces. Inventiones mathematicae 98 (1989) 511-548

41