Professional Documents
Culture Documents
CRM Barcelona
Centre de Recerca Matemtica
Managing Editor:
Carles Casacuberta
Luis A. Caffarelli
Franois Golse
Yan Guo
Carlos E. Kenig
Alexis Vasseur
Luis A. Caffarelli
Department of Mathematics
University of Texas at Austin
1 University Station, C1200
Austin, TX 78712-0257, USA
Franois Golse
cole Polytechnique
Centre de Mathmatiques L. Schwartz
F-91128 Palaiseau Cedex, France
Yan Guo
Division of Applied Mathematics
Brown University
Providence, RI 02912, USA
Carlos E. Kenig
Department of Mathematics
University of Chicago
Chicago, IL 60637, USA
Alexis Vasseur
Department of Mathematics
University of Texas at Austin
1 University Station, C1200
Austin, TX 78712-0257, USA
ISBN 978-3-0348-0190-4
e-ISBN 978-3-0348-0191-1
DOI 10.1007/978-3-0348-0191-1
Springer Basel Dordrecht Heidelberg London New York
Library of Congress Control Number: 2011940208
Mathematics Subject Classification (2010): Primary: 35Q35, 82C40, 35Q20, 35Q55, 35L70;
Secondary: 86A10, 76P05, 35P25
Springer Basel AG 2012
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in other ways, and storage in data banks. For any kind of use permission
of the copyright owner must be obtained.
Printed on acid-free paper
Foreword
This book contains expository lecture notes for some of the courses and talks given
at the school Topics in PDEs and Applications 2008. A CRM & FISYMAT Joint
Activity, which took place at the FisyMat-Universidad de Granada (April 7 to 11,
2008) and at the Centre de Recerca Matem`atica (CRM) in Bellaterra, Barcelona
(May 5 to 9, 2008).
The goal of the school was to present some of the main advances that were
taking place in the eld of nonlinear Partial Dierential Equations and their applications. Oriented to Master and PhD students, recent PhD doctorates, and
researchers in general, the courses encompassed a number of areas in order to
open new perspectives to researchers and students.
The program in the Granada event consisted of ve courses taught by Luigi
Ambrosio, Luis Caarelli, Francois Golse, Pierre-Louis Lions, and Horng-Tzer Yau,
as well as two talks given by Yan Guo and Pierre-Emmanuel Jabin. The event
at the Centre de Recerca Matem`
atica consisted of ve courses taught by Henri
Berestycki, Ham Brezis, Carlos Kenig, Robert V. Kohn, and Gang Tian.
The volume covers several topics of current interest in the eld of nonlinear
Partial Dierential Equations and its applications to the physics of continuous media and of particle interactions. The lecture notes describe several powerful methods introduced in recent top research articles, and carry out an elegant description
of the basis for, and most recent advances in, the quasigeostrophic equation, integral diusions, periodic Lorentz gas, Boltzmann equation, and critical dispersive
nonlinear Schrodinger and wave equations.
L. Caarelli and A. Vasseurs lectures describe the classical De Giorgi truncation method and its recent applications to integral diusions and the quasigeostrophic equation. The lectures by F. Golse concern the Lorentz model for the
motion of electrons in a solid and, more particularly, its BoltzmannGrad limit in
the case of a periodic conguration of obstacles like atoms in a crystal. Y. Guos
lectures concern the Boltzmann equation in bounded domains and a unied theory
in the near Maxwellian regime to establish exponential decay toward a normalized Maxwellian for all four basic types of boundary conditions. The lectures by
C. Kenig describe a recent concentration-compactness/rigidity method for critical
dispersive and wave equations, in both defocusing and focusing cases. The issues
studied center around global well-posedness and scattering.
vi
Foreword
We are very thankful to the CRM and FisyMat for hosting these advanced
courses. In particular, we thank the CRM director Joaquim Bruna for making
the CRM event possible. The CRM administrative sta and FisyMat coordinators
were very helpful at all times. We acknowledge nancial support from the Ministerio de Educaci
on y Ciencia, Junta de Andaluca, and Generalitat de Catalunya,
institutions that supported the two events in the school.
Finally and above all, we thank the authors for their talks, expertise, and
kind collaboration.
Xavier Cabre and Juan Soler
Contents
Foreword
1
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
2
3
4
11
11
12
12
13
15
22
23
30
34
34
37
39
. .
. .
39
41
. .
. .
. .
44
51
57
. .
66
vii
viii
Contents
2.6
2.7
2.8
2.9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
90
95
96
101
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
71
79
84
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
101
103
104
105
107
107
108
112
117
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
117
121
129
146
Chapter 1
Introduction
In 1957, E. De Giorgi [7] solved the 19th Hilbert problem by proving the regularity
and analyticity of variational (energy minimizing weak) solutions to nonlinear
elliptic variational problems. In so doing, he developed a very geometric, basic
method to deduce boundedness and regularity of solutions to a priori very discontinuous problems. The essence of his method has found applications in homogenization, phase transition, inverse problems, etc.
More recently, it has been successfully applied to several dierent problems
in uid dynamics: by one of the authors [12], to reproduce the partial regularity
results for NavierStokes equations originally proven in [1]; by several authors [5],
to study regularity of solutions to the NavierStokes equations with symmetries;
and by the authors, to prove the boundedness and regularity of solutions to the
quasi-geostrophic equation [2].
These notes are based on minicourses that we gave at the school Topics
in PDEs 2008 in Fisymat-Granada and CRM Barcelona, as well as at schools in
Ravello and Cordoba (Argentina). The structure of the notes consists of two parts.
The rst one is a review of De Giorgis proof, stressing the important aspects of
his approach, and the second one is a discussion on how to adapt his method to
the regularity theory for the quasi-geostrophic equation.
1.1
The 19th Hilbert problem consisted in showing that local minimizers of an energy
functional
E(w) =
F (w) dx
But the class of uniformly elliptic equations (I Aij I) with no regularity assumptions is a scaling invariant class in itself, and never gets close to the
Laplacian.
De Giorgi then studied solutions u of
Di aij (x)Dj u = 0
with no assumption on aij , except uniform ellipticity (I aij I), and showed
that u is C . Applying this theorem to (w)e , he solved the Hilbert problem.
Hence, we need to prove the following:
Theorem 1. Let u be a solution of Di aij Dj u = 0 in B1 of RN with 0 < I
aij (x) I (i.e., aij is uniformly elliptic). Then u C (B1/2 ) with
uC (B1/2 ) CuL2 (B1 ) ,
where = (, , n).
Proof. The proof is based on the interplay between the Sobolev inequality, which
says that uL2+ is controlled by uL2 , and the energy inequality, which says
that, in turn, since u is a solution of the equation, u L2 is controlled by u L2
for every truncation u = (u )+ .
1.1.1
B1 supp
u2 .
(Note that there is a loss going from one term to the other: (u) versus u.)
We denote by the term = () .
(Try it!)
1.1.2
The proof of the De Giorgi theorem is now split into two parts:
Step 1: From L2 to L .
Step 2: Oscillation decay.
We start with Step 1.
Lemma 2 (From an L2 to an L bound). If u+ L2 (B1 ) is small enough, namely
smaller than 0 (n, , ), then supB1/2 u+ 1.
Before going into the proof of this lemma, let us give a simple, geometric
analogy that avoids some of the technicalities of the proof.
Suppose that is a domain in Rn and is a minimal surface when restricted
to B1 , in the sense that the boundary of any perturbation inside B1 will have larger
area.
For this purpose, we will take dyadic balls Brk = B 12 +2k converging to B1/2 ,
and rings Rrk = Brk1 Brk ; and we will nd a nonlinear recurrence relation for
Vk = Vol( Brk ) for k even, that will imply that Vk goes to zero. In particular,
never reaches B1/2 .In this analogy, volume replaces the square of the L2 norm
of u, area the energy |u|2 , the isoperimetric inequality the Sobolev inequality,
and the minimality of the area the energy inequality. The argument is based on
the interplay between area and volume, as follows.
Ar controls Vk for r rk in a nonhomogeneous way
We have
Vk Vr (isoperimetric inequality) [Area(Vr )]N/N 1
= (two parts) = [Ar + Area( Br )]N/N 1 (2Ar )N/N 1 .
By minimality,
Area( Br ) Ar .
This is the energy inequality.
We have
Vol(Vk \ Vk+1 )
rk
rk+1
Ar 2k inf Ar
rk+1 rrk
N/(N 1) !!
If V0 < 0 , the build up in the exponent as we iterate beats the large geometric
coecient in the recurrence relation above, and Vk goes to zero. In particular,
B1/2 is clean.
We now pass to the proof of Lemma 2. The origin becomes now plus innity,
uL2 plays the role of volume, and uL2 that of area. We have the added
complication of having to truncate in space.
Proof of Lemma 2. We will consider a sequence of truncations k uk , where k is
a sequence of shrinking cut-o functions converging to B1/2 . More precisely:
B1/2
while uk is a sequence of monotone truncations converging to (u 1)+ :
uk = [u(1 2k )]+ .
Note that, where uk+1 > 0, uk > 2(k+1) . Therefore {(k+1 uk+1 ) > 0} is
contained in {(k uk ) > 2(k+1) }.
We will now show that, if uL2 (B1 ) = A0 is small enough, then
Ak = (k uk )2 0.
1/2
(k+1 uk+1 )p
2/p
C
(k+1 uk+1 )
so we get
Ak+1 C
2/p
(k+1 uk+1 )
(k uk )2 = C 22k Ak .
(k uk )2
So we get
Ak+1 C 24k (Ak )1+ .
Then, for A0 = small enough, Ak 0 (prove it). The build up of the exponent
in Ak forces Ak to go to zero. In fact, Ak has faster than geometric decay, i.e., for
any M > 0, Ak < M k if A0 (M ) is small enough.
Corollary 4. If u is a solution of Lu = 0 in B1 , then
uL (B1/2 ) CuL2 (B1 ) .
|B1 |,
|{u+ = 0}| 1
2
then
Then, if
|w(y)| dy
n1
D |x0 y|
|w|r n1 dr d
wr dr d
.
r n1
Integrating x0 on C,
|A| |C|
D
|w(y)|
dx0
|x0 y|n1
dy.
Among all C with the same measure |C|, the integral in x0 is maximized by the
ball of radius |C|1/n , centered at y:
|C|1/n .
C
Hence,
|A| |C| |C|
Since
1/n
D
|w|
1/2
|D|1/2 .
wk = 2k vk .
10
B1/2
v
v3
v2
v1
v
v
Ck
Dk
Ck-1
We will show that, by applying the isoperimetric inequality and the previous
lemma in a nite number of steps k0 = k0 (, , ),
|Ck0 | = 0.
Then () = 2
k0
. Note that:
4
11
1.2
1.2.1
Quasi-geostrophic ow equation
12
1.2.2
i1 ).
v = (i2 ,
and
In particular,
v =
i2 i1
,
.
||
||
The multipliers ij /|| are classical operators, called Riesz transforms, that correspond in physical space x to convolution with kernels
Rj (x) =
i.e.,
vij (x) =
xj
,
|x|n+1
Rj (x y)(y) dy.
1.2.3
BMO spaces
What is BMO? It is the space of functions with bounded mean oscillation. That is,
in any cube Q the average of u minus its average is bounded by a constant C,
1
1
u(y) dy dx C.
u(x)
|Q| Q
Q Q
The smallest C good for all cubes denes a seminorm (as it does not distinguish
a constant that we may factor out). The space of functions u in BMO of the unit
cube is smaller than any Lp (p < ) but not included in L (for this, (log |x|)
is a typical example).
13
1.2.4
()
Note that the kernel
K = C()|x|(n+2)
is singular near zero, so, in principle, some cancellation in u is expected for the
integral to converge. For instance, bounded and in C 2 near x0 suces. Also,
C() (1 ) guarantees that, as 1, converges to .
A particularly interesting case is the case = 1/2, since in this case ()1/2 u
coincides with the Dirichlet to Neumann map. More precisely, given dened for
x in Rn , we extend it to dened for (x, y) in (Rn+1 )+ by combining it with the
Poisson kernel:
C y
n
P1 (x/y).
Py (x) =
n+1 = y
2
2
2
(y + |x| )
Then (x, y) satises
x,y = 0 in Rn R+
and it can be checked that 1/2 (x0 ) = Dy (x0 , 0) in two ways:
(a) Represent (x0 , h) as
(x0 , h) = [Ph ](x0 )
and take the limit on the dierence quotient
(x0 , h) (x0 , 0)
,
h0
h
Dy (x0 , 0) = lim
or
(b) Fourier-transform in x:
14
y||
.
(, y) = ()e
In particular,
1/2 )().
0) = ()||
= (
Dy (,
Hence, we can make sense of the Greens and energy formula for the half
Laplacian: Let (x), (x) be two nice, decaying functions dened in Rn , and
y) decaying extensions into (Rn+1 )+ . Then we have
() =
x,y
(x,y) +
x,y .
Rn
(Rn+1 )+
(Rn+1 )+
, and 0, giving us
(1/2 ) =
.
1/2
Rn
(Rn+1 )+
Further, if we choose
= ( )+
and
= ( )+
(Rn+1 )+
| |2 ,
E( ) =
Rn+1
+
and that this minimum denes the H 1/2 norm of . In particular, we obtain
( )+ (1/2 ) =
[( )+ ]2 dx dy
Rn
[( )+ ]2 dx dy = ( )+ 2H 1/2
since the harmonic extension of the truncation has less energy than the truncation
of the harmonic extension.
15
To recapitulate
The operator 1/2 is interesting because:
(a) It can be understood as a surface diusion process.
(b) It is the EulerLagrange equation of the H 1/2 energy.
(c) Being of order 1, diusion competes with transport.
In fact, the derivatives of are
DXj = Rj (1/2 )
1/2 () =
and
Rj (DXj ),
where Rj , the Riesz transform, is the singular integral operator with symbol ij /||.
1.2.5
Regularity
The regularity theory for the quasi-geostrophic equation is based on two linear
transport regularity theorems: Theorems 10 and 11.
Theorem 10. Let be a (weak) solution of
t + v = (1/2 )
in
Rn [0, )
for some incompressible vector eld v (with no a priori bounds) and initial data
0 in L2 . Then
( , 1)L (Rn ) C( , 0)L2 (Rn ) .
Remarks. (i) All we ask from v is that the energy inequality makes sense for
any function h() with linear growth. Formally, if we multiply and integrate,
we may write
T2
h()v =
vH() =
div vH() = 0,
T1
Rn
where H () = h().
Therefore, the contribution of the transport term in the energy inequality vanishes. In the case of the Q-G equation, this can be attained by rigorously constructing in a particular way, for instance as a limit of solutions
in increasing balls BK .
(ii) From the scaling of the equation: For any ,
=
1
(x, t)
0 L2 .
16
dt
( ) (y, 0, t)Dz ( )(y, 0, t) dy
T1
Rn
T2
Rn+1
+
T1
=
T2
T1
( ) (y, z, t) (y, z, t) dy dz
dt
Rn+1
+
[ ] dy dz.
Note that ( ) is not the harmonic extension of , but the truncation of the
extension of , i.e., ( )+ .
Nevertheless, it is an admissible extension of (going to zero at innity)
and, as such,
H 1 (Rn+1 ) H 1/2 (Rn ) .
+
We will denote by A = (AT1 ,T2 ) the term on the left and by B = BT1 the one on
the right. Therefore, BT1 controls in particular (from the Sobolev inequality) all
of the future:
sup (t)2L2 +
tT1
T1
(t)2Lp BT1 .
17
for some q with 2 < q < p, in the following way. Every such q is a convex combination
1
1
q = 2 + (1 )p = 2 + p
r
s
for r, s appropriate conjugate exponents.
Therefore, xing such a q, we have, for each time t,
1/r
1/s
p
T1
2/q
(k )2
I
kq
2/q
|{k > 0} I|
1/
q
(by H
older with 2 and k >0 ) with q the conjugate exponent to q/2.
In turn, CTk1 (k1 ) and, by going from k to k 1, we can estimate
(this should sound familiar by now):
1/q
1/q
k
qk
q
= {k1 > 2 } I
2
(k1 )
(by Chebyshev),
I
k > 0 implies
k1 > 2k .
since k k1 and further
Ck
CTk1 (k1 ) and, putting together the estimates for
That is, 2
and ,
1+
(k )2 2Ck CTk1 (k1 )
.
I
But then,
inf
[Tk1 <t<Tk ]
1+
Bt (k ) 2k 2Ck CTk1 (k1 )
.
18
0 L2
for t 1.
0
We now pass to the issue of regularity, i.e., the oscillation lemma. Having
shown boundedness for the Q-G equation, our situation is now the following. We
have a solution that satises the energy bound:
sup
+
t
(#)
We decouple v from , and will prove a linear theorem, where for v satisfying ()
and satisfying (#) and the equation
t + v = 1/2 ,
we have:
Theorem 11. is locally C .
To simplify the notation, we will assume that exists for t 4 and will
focus on the point (X, t) = (0, 0). The H
older continuity will be proven through
an oscillation lemma, i.e., we will prove that on a geometric sequence of cylinders
k = B4k [4k , 0]
the oscillation of ,
k = sup inf ,
k
19
T2
2
T1
2 t dx dt
2 ( )2 (T2 ) dx
2 ( )2 (T1 ) dx.
(I)
Next we have the transport term, an extra term not usually present in the
energy inequality:
20
We split the term into the two factors ()v and , the logic being that v
is almost bounded and thus the rst term is almost like the standard right-hand
side in the energy inequality while the second would be absorbed by the energy.
For each xed t, we get
T2
|II|
L2n/(n1) [v ]L2n/(n+1)
T1
T2
T1
1
2L2n/(n1) + [v ]2L2n/(n+1) .
T1
T1
p
T1
T1
Finally, III is our energy term, i.e.,
III =
2 1/2 .
Using the harmonic extension , we get that
x,z ( 2 ) dx dz dt.
III =
2 dx dt =
By the standard energy inequality computation, we get that
()2 ( )2 .
III
[x,z ]2 dx dz dt +
We may choose to be a cut-o in x and z or to integrate to innity in z, if we
have control of in z.
Remark. If, for some reason, we know that ( ) 0 in B1 {z0 } for some z0 ,
then we may cut o only in x and still stop the integration at z0 .
21
T1 tT2
2L2 +
T1
(T1 )2L2
1
+ C
T2
T2
( )2L2
T2
T1
2L2n/(n1)
() 2L2 +
T1
T2
T1
() 2L2x,z .
Notice that is one extension of and therefore the term in the left
T2
T1
( )2L2
T2
controls
T1
2H 1/2 ,
T2
T1
2L2n/(n1)
2L2
T2
(T1 )2L2
T1
( )2L2
T2
+
T1
() 2L2x,z
+
T2
T1
(1.2.1)
() 2L2
where the constant depends only on the BMO seminorm and mean value of the
velocity v (namely the constant in ()).
In fact, if it werent because of the term on the right
T2
T1
() 2L2x,z
involving the extra variable z, everything would reduce to Rn , and we would have
the usual interplay between the Sobolev and energy inequalities, as in the global
case, and a straightforward adaptation of the second-order case would work.
In light of this obstruction, let us reassess the situation. As we mentioned
before, our rst lemma (Lemma 12) would be (following the iterative scheme) to
show that if, say, the L2 norm of + is very small (and + 2) in B4 [4, 0],
then + is strictly less than 2 in B1 [1, 0], i.e., + 0 < 2 for some 0 .
Let us see how this can work.
22
1.2.6
Therefore, the rst truncation 0 is controlled by its trace, with very small L2
norm, and the very narrow sides (of size ), whose inuence decays exponentially
moving inwards in X:
23
The idea is that, on one hand, the fast cut-o in Z will make the inuence
of the tiny sides decay so much in X that at the level of the next cut-o (in X)
it will be wiped out by the dyadic cut-o in , while the contribution of the trace
k will decay so fast (faster than M k if we choose c0 very small) that
k (X, k ) k Pk k L2 Pk L2
will also be wiped out by the consecutive truncation.
At the end of the process, at time t0 , we have only information on the trace ,
but we can go inwards by harmonicity to complete the proof.
1.2.7
First part
In this rst part, we prove that a solution between 0 and 2, with very small L2
norm, separates from = 2 in a smaller cylinder.
Lemma 12. We assume that
vL (4, 0; BMO(RN )) + sup
4t0
B4
v(t, x) dx C0 .
(1.2.2)
Then there exist 0 > 0 and > 0 such that, for every solution to (1.2.1), the
following property holds true: If we have 2 in [4, 0] B4 and
0
4
B4
( )2+
dx dz ds +
0
4
B4
()2+ dx ds 0 ,
24
Proof. The proof follows the strategy discussed above. First, we introduce some
previous tools. Since the method was based on the control of by two harmonic
functions, one for the local data and the other for the far away, before starting the
proof we build two useful barriers.
Step 1: Barriers
The rst barrier is the following:
Barrier b1 (x, z)
b1 = 2 in z = 1
z = 1/2
b1 = 2 in |x| = 1
b1 2 4 < 2
b1 = 0 in z = 0
.
and, for some 0 < , we have b1 2 4 < 2 in B1/2
The barrier b1 implies that
b1 = 2 in |x| = 1
25
x1 = 1
7
z =
7
x1 = 1
(i) b2 is harmonic in D;
(ii) b2 = 0 for z ;
(iii) b2 = 1 for x1 = 1 and b2 = 0 for x1 = 1.
Then b2 C cos(z/) e(1x1 )/ . In particular, if 1 x1 = h
, we have
h/ and C = (cos 1)1 .
b2 Ce
Remark. Exponential decay also holds for Di aij Dj by applying the Harnack inequality to the intervals Ik = {k x1 k + 1}.
Now we are ready to set the main inductive steps as discussed above. When
we do so, we will realize that we have to start the process for some advanced value
k of the step. So we will go back and do a rst large step to cover the starting of
the process.
Step 2: Setting of the constants
We recall that > 0 is dened by the fact that the barrier function satises that
b1 < 2 4 in B1/2
. Next, C = (cos 1)1 is the constant in the bound for the
barrier function b2 . The smallness constant C0 in the hypothesis of Lemma 12 will
be chosen later as C0 (, M ).
We need to x constants M for the rate of decay of the L2 norm of the
truncation k and for the rate of decay of the support in z of k .
We require:
k
26
2
n
.
n M
4P (1)L2
It is sucient to take
M sup
2 8P (1)L2
,
n
n
M
C0N
k/N
M 3(1+1/N ) .
For this it is enough to take M sup(1, C02N ). Indeed, this ensures M 2 /C02N M
and hence
k/N
M
M k/(2N ) M 6 .
C0N
The main inductive step will be the following.
Step 3: Induction
We set
k = ( Ck )+ ,
with Ck = 2 (1 + 2
1{B1+2k1 } k 1{B1+2k } ,
|k | C2k ,
and we let
Ak = 2
0
12k
k
0
RN
|(k k )|2 dx dz dt +
sup
[12k ,1]
(k k )2 dx dt.
RN
(1.2.3)
k k is supported in 0 z k
(1.2.4)
27
with cut-o function k (x)(z) where is a xed cut-o function in z only. Taking
the mean value of (1.2.1) in T1 between 4 and 2, we nd that (1.2.3) is veried
for 0 k 12N if 0 is taken such that
C224N (1 + )0 M 12N .
(1.2.5)
We have used that |k |2 C224N for 0 k 12N . Let us consider now the
support property (1.2.4). By the maximum principle, we have
(+ 1B4 ) P (z) + b1 (x, z)
in R+ B4 , where P (z) is the Poisson kernel. Indeed, the right-hand side function
is harmonic, positive, and the trace on the boundary is bigger than the one of .
From Step 1 we have b1 (x, z) 2 4. Moreover,
for 1 z 2, t 0, x B2 ,
0 = ( (2 2))+ 0
Hence,
0 0
for 1 z 2, t 0, x B2 .
vanishes for 1 = z 2.
[(k k ) P (z)]k+1
k+1 k+1
on B k ,
(1.2.6)
x+
i and xi it is bigger than 2. Finally, by the maximum principle,
k
N
i=1
b2 ((xi x+ )/ k , z/ k ) + b2 ((xi + x )/ k , z/ k ) + (k k ) P (z).
28
b2 ((xi x+ )/ k , z/ k ) + b2 ((xi + x )/ k , z/ k )
i=1
2N Ce
4(
k
2
2+1) k
2k2
k+1
(k 2k1 )+ .
So,
M k/2
(k+1)N/2
P (1)L2 2k2 .
k+1 k+1
0
for k+1 z k .
Note, in particular, that with Step 4 this gives that (1.2.4) is veried up to k =
12N + 1 and (1.2.6) up to k = 12N .
Step 6: Propagation of property (1.2.3)
We show in this step that if (1.2.3) is true for k 3 and (1.2.4) is true for k 3,
k 2 and k 1, then (1.2.3) is true for k.
First notice that, from Step 5, (1.2.4) is true at k 2, k 1, and k. We just
need to show that
Ak C0k (Ak3 )1+1/N
with
C0 = C
for k 12N + 1,
21+2/N
.
2/N
Indeed, if we do the third inequality of Step 2, this will give us the result.
(1.2.7)
(1.2.8)
29
|(k k ) |2
RN
0
=
|1/2 (k k )|2 .
RN
Note that we have used (1.2.4) in the rst equality. The Sobolev and H
older
inequalities give
Ak3 Ck3 k3 2L2(N +1)/N ([12k3 ,0]RN ) .
From (1.2.6),
k2 k2
2L2(N +1)/N P (1)2L1 k3 k3 2L2(N +1)/N .
Hence,
C k2 k1
2L2(N +1)/N + k2 k1 2L2(N +1)/N .
2
2k
2
2
2
k1 k +
k1 k .
Ak C2 ( + 2)
0
RN
RN
2
We have used here the fact that ||2 C22k k1
. If k > 0, then k1 2k .
So,
C2k
k1 ,
1{k >0}
and
C2k
1{k1 >0} 1{k >0}
k2 k1 .
Therefore,
2
2
2
k1 k + k1
k 2
C22k/N
2(N +1)/N
2(N +1)/N
+ (k2 k1 )
(k2 k1 )
,
2/N
30
and so
Ak
C2k(2+2/N ) 1+1/N
Ak3 .
2/N
1.2.8
Second part
In the rst part, we have established that, if 0 + 2 and its energy or norm is
very small in B4 , then + 2 in Bj , i.e., the oscillation of actually decays.
We now want to get rid of the very small hypothesis.
This second lemma (Lemma 13) proves that, if + 0 half of the time and
it only needs very little room (say, ) to go from {+ 0} to { 1}, it is because
( 1)+ has already very small norm to start with. This produces a dichotomy:
either the support of decreases substantially, or becomes small anyway, in the
same spirit as De Giorgis lemma.
Lemma 13. For every 1 > 0 there exists a constant 1 > 0 with the following
property: For every solution to (1.2.1) with v verifying (1.2.2), if 2 in Q4
and
|Q |
|{(x, z, t) Q4 : (x, z, t) 0}| 4 ,
2
then the following implication holds true: If
|{(x, z, t) Q4 : 0 < (x, z, t) < 1}| 1 ,
then
Q1
1)2+
dx dt +
Q
1
( 1)2+ dx dz dt C 1 .
This lemma is, of course, the adapted version of the De Giorgis isoperimetric
inequality.
The idea of theproof is the following. We rst throw away a small set of
times for which It = B |u |2 dx dz is very large:
1
It
This is a tiny set of times:
since
K2
.
21
|S| C 2 /k 2 ,
|u |2 dx dz dt C0 .
31
1
1
, we may choose a slice where |A| > 64
and
But for some t, say t < 64
|D| . Then |B| (64)2 K/1 K1 if 2 . In particular, ( 1)+ has
very small L2 norm for that t: ( 1)+ L2 k1 . But the energy inequality then
controls the L2 norm of ( 1)+ into the future. We now give a more detailed
proof.
B1
Let
K=
2
|+
| dx dz dt
.
1
1
2
|+ | (t) dx dz K .
t :
4
B1
Then
For all t {t :
B1
(1.2.9)
2
|+
| (t) dx dz K}, the isoperimetric inequality gives that
1/2
t [4, 0] : |C(t)|
31
and
B1
2
|+
| (t) dx dz
|C(t)|1/2 K 1/2
5/2
4C1 21 .
|A(t)|
(1.2.10)
32
In particular,
B1
But
B1
2
+
(t) dx
2
+
(t) dx dz 4(|B(t)| + |C(t)|) 821 .
=
B1
2
+
(t, x, z) dx
2
B1
+
(t)z dx d
z
2
+
(t) dx
B1
2
+
(t, x, z) dx dz
+2 K
B1
2 (t) dx dz C .
+
1
We want to show that |A(t)| 1/4 for every t I [1, 0]. First, since
|{(t, x, z) : 0}| |Q4 |/2,
2
(t0 ) dx C 1 .
there exists t0 1 such that |A(t0 )| 1/4. For this t0 , +
Using the energy inequality, for any r > 0 (where is of order 1/r) we have, for
every t t0 ,
C(t t0 )
2
2
+ (t) dx
+
(t0 ) dx +
+ Cr.
r
B1
B1
Let us choose r such that
Cr + C 1 1/128.
(z) = + +
+
0
z +
d
z + +
z
0
1/2
2
|z +
| d
z
.
(t, x, z) + (t, x) + 21
+
1/2
2
|z +
| d
z
.
The integral, in x only, of the square of the right-hand side term is less than
(t, x, z) 1}|
|{x B1 , +
1
.
4
33
(t) 1}|
|{z 21 , x B1 , +
21 (1 1/4) 61 21 /2.
In the same way as in (1.2.10), we nd that
|B(t)|
and
|C(t)|1/2 K 1/2
C 1 ,
|A(t)|
z dz.
(
1)2+
2 ( (z)
1)2+
+
0
2
| | dz
1)2+
Q1
1)2+
dz + 2 1
( 1)2+ dx ds C 1 .
| |2 dz.
34
1.2.9
Oscillation lemma
We are now ready to iterate the process for the oscillation lemma.
Lemma 14. There exists > 0 such that, for every solution to (1.2.3) with v
satisfying (1.2.2), if 2 in Q1 and
|{(t, x, z) Q1 : 0}|
1
,
2
then 2 in Q1/16 .
Note that depends only on N and C in (1.2.2).
Proof. For every k N, k K+ = E(1/1 + 1) (where 1 is dened in Lemma 13
with 0 = .
Hence, |{K+ 0}| 1 and K+ < 0 almost everywhere, which means that
2K+ ( 2) + 2 < 0, or < 2 2K+ , and in this case we are done.
Else, there exists 0 k0 K+ such that |{0 < k0 < 1}| 1 . From
Lemma 13 and Lemma 12 (applied on k0 +1 ) we get ( k0 +1 )+ 2 , which
means that
2 2(k0 +1) 2 2K+
in Q1/8 . Consider the function b3 dened by
(i) b3 = 0 in B1/8
;
1.2.10
Proof of Theorem 11
35
x (s) = 1
v(t + st0 /4, x + yt0 /4) dy,
0
|B4 | x0 (s)+B4
x0 (0) = 0.
Note that x0 (s) is uniquely dened from the CauchyLipschitz theorem. We set
supQ4 F0 + inf Q4 F0
4
0 (s, y) =
F
0
supQ4 F0 inf Q4 F0
2
v0 (s, y) = v(t + st0 /4, x + t0 /4(y x0 (s))) x 0 (s),
and then, for every k > 0,
Fk (s, y) = Fk1 (
s,
(y xk (s))),
supQ4 Fk + inf Q4 Fk
4
F
,
k (s, y) =
k
supQ4 Fk inf Q4 Fk
2
1
x k (s) =
vk1 (
s,
y) dy,
|B4 | xk (s)+B4
xk (0) = 0,
s,
(y xk (s))) x k (s),
vk (s, y) = vk1 (
where
will be chosen later. We divide the proof into several steps.
Step
1. For k = 0, 0 is a solution to (1.2.3) in [4, 0] RN , v0 BMO = vBMO ,
v0 (s) dy = 0 for every s, and |0 | 2. Assume that it is true at k 1. Then
s k1
x k (s) k1.
s Fk =
Cvk1 (
y)Lp
C
N/p vk1 Lp
Cp
N/p vk1 BMO .
So, for 0 s 1, y B4 and p > N ,
(1 + Cp
N/p ) C
1N/p .
|
(y xk (s))| 4
For
small enough, this is smaller than 1.
36
Step 2. For every k we can use the oscillation lemma. If |{k 0}| 12 |Q4 |,
then we have k 2 . Else we have |{k 0}| 12 |Q4 | and applying the
oscillation lemma on k gives k 2 + . In both cases, this gives
| sup k inf k | 2 ,
and so
k
sup Fk inf Fk (1 /2) sup F0 inf F0 .
Q1
Q1
Q1
Q1
Step 3. For s
2n ,
n
k=0
nk xk (s)
2n
n
nk
,
2
N/p
k=0
for
small enough. So
inf
sup
(1 /2)n .
[2n ,0]Bn /2
[
2n ,0]B
n /2
This gives that is C at (t, x, 0), and so is C at (t, x).
Bibliography
[1] L. Caarelli, R. Kohn, and L. Nirenberg, Partial regularity of suitable weak
solutions of the NavierStokes equations, Comm. Pure Appl. Math. 35 (1982),
no. 6, 771831.
[2] L. Caarelli and A. Vasseur, Drift diusion equations with fractional diusion
and the quasi-geostrophic equation, Ann. of Math. 171 (2010), no. 3, 1903
1930.
[3] D. Chae, On the regularity conditions for the dissipative quasi-geostrophic
equations, SIAM J. Math. Anal. 37 (2006), no. 5, 16491656 (electronic).
[4] D. Chae and J. Lee, Global well-posedness in the super-critical dissipative
quasi-geostrophic equations, Commun. Math. Phys. 233 (2003), no. 2, 297
311.
[5] C.-C. Chen, R. M. Strain, H.-T. Yau, and T.-P. Tsai, Lower bound on the
blow-up rate of the axisymmetric NavierStokes equations, Int. Math. Res.
Notices 9 (2008), 31 pp.
[6] P. Constantin, Euler equations, NavierStokes equations and turbulence, in:
Mathematical Foundation of Turbulent Viscous Flows, Lecture Notes in
Math., vol. 1871, pp. 143, Springer, Berlin, 2006.
[7] E. De Giorgi, Sulla dierenziabilit`
a e lanaliticit`
a delle estremaili degli integrali multipli regolari, Mem. Accad. Sci. Torino Cl. Sci. Fis. Mat. Natur. 3
(1957), 2543.
[8] G. Duvaut and J.-L. Lions, Les inequations en mecanique et en physique,
Travaux et Recherches Mathematiques, vol. 21, Dunod, Paris, 1972.
[9] A. Kiselev, F. Nazarov, and A. Volberg, Global well-posedness for the critical
2D dissipative quasi-geostrophic equation, Invent. Math. 167 (2007), 445453.
[10] P. Laurence and S. Salsa, Regularity of the free boundary of an American
option on several assets, Comm. Pure Appl. Math. 62 (2009), no. 7, 969994.
[11] G. Seregin and V. Sverak, On Type I singularities of the local axi-symmetric
solutions of the NavierStokes equations, Comm. Partial Dierential Equations 34 (2009), no. 2, 171201.
37
38
Bibliography
Chapter 2
39
40
time of the order of a fraction of the reciprocal collision frequency. (One should also
mention an earlier, formal derivation by C. Cercignani [12] of the Boltzmann
equation for a hard sphere gas, which considerably claried the mathematical
formulation of the problem.) Shortly after Lanfords derivation of the Boltzmann
equation, R. Illner and M. Pulvirenti managed to extend the validity of his result
for all positive times, for initial data corresponding with a very rareed cloud of
gas molecules [27].
An important assumption made in Boltzmanns attempt at justifying the
equation bearing his name is the Stosszahlansatz, to the eect that particle pairs
just about to collide are uncorrelated. Lanfords argument indirectly established
the validity of Boltzmanns assumption, at least on very short time intervals.
In applications of kinetic theory other than rareed gas dynamics, one may
face the situation where the analogue of the Boltzmann equation for monatomic
gases is linear, instead of quadratic. The linear Boltzmann equation is encountered
for instance in neutron transport, or in some models in radiative transfer. It usually
describes a situation where particles interact with some background medium
such as neutrons with the atoms of some ssile material, or photons subject to
scattering processes (Rayleigh or Thomson scattering) in a gas or a plasma.
In some situations leading to a linear Boltzmann equation, one has to think
of two families of particles: the moving particles whose phase space density satises
the linear Boltzmann equation, and the background medium that can be viewed
as a family of xed particles of a dierent type. For instance, one can think of
the moving particles as being light particles, whereas the xed particles can be
viewed as innitely heavier, and therefore unaected by elastic collisions with the
light particles. Before Lanfords fundamental paper, an important unfortunately
unpublished preprint by G. Gallavotti [19] provided a rigorous derivation of
the linear Boltzmann equation assuming that the background medium consists
of xed entities, like independent hard spheres whose centers are distributed in
the Euclidean space under Poissons law. Gallavottis argument already possessed
some of the most remarkable features in Lanfords proof, and therefore must be
regarded as an essential step in the understanding of kinetic theory.
However, Boltzmanns Stosszahlansatz becomes questionable in this kind of
situation involving light and heavy particles, as potential correlations among heavy
particles may inuence the light particle dynamics. Gallavottis assumption of a
background medium consisting of independent hard spheres excluded this possibility. Yet, strongly correlated background media are equally natural, and should
also be considered.
The periodic Lorentz gas discussed in these notes is one example of this type
of situation. Assuming that heavy particles are located at the vertices of some
lattice in the Euclidean space clearly introduces about the maximum amount of
correlation between these heavy particles. This periodicity assumption entails a
dramatic change in the structure of the equation that one obtains under the same
scaling limit that would otherwise lead to a linear Boltzmann equation.
41
Figure 2.1: Left: Paul Drude (18631906); right: Hendrik Antoon Lorentz (18531928)
Therefore, studying the periodic Lorentz gas can be viewed as one way of
testing the limits of the classical concepts of the kinetic theory of gases.
Acknowledgements. Most of the material presented in these lectures is the result of collaboration with several authors: J. Bourgain, E. Caglioti, H. S. Dumas,
L. Dumas and B. Wennberg, whom I wish to thank for sharing my interest for this
problem. I am also grateful to C. Boldighrini and G. Gallavotti for illuminating
discussions on this subject.
2.1
In the early 1900s, P. Drude [16] and H. Lorentz [30] independently proposed to
describe the motion of electrons in metals by the methods of kinetic theory. One
should keep in mind that the kinetic theory of gases was by then a relatively new
subject: the Boltzmann equation for monatomic gases appeared for the rst time
in the papers of J. Clerk Maxwell [35] and L. Boltzmann [5]. Likewise, the existence
of electrons had been established shortly before, in 1897 by J. J. Thomson.
The basic assumptions made by H. Lorentz in his paper [30] can be summarized as follows.
42
1
m F (t, x)
2
v )f (t, x, v) = Nat rat
|v| C(f )(t, x, v).
In this equation, C is the Lorentz collision integral, which acts on the only
variable v in the phase-space density f . In other words, for each continuous function (v), one has
C()(v) =
(v 2(v )) (v) cos(v, ) d,
||=1
v>0
43
44
2.2
Before discussing the BoltzmannGrad limit of the periodic Lorentz gas, we rst
give a brief description of Gallavottis result [18, 19] for the case of a Poisson
distribution of independent, and therefore possibly overlapping obstacles. As we
shall see, Gallavottis argument is in some sense fairly elementary, and yet brilliant.
First we dene the notion of a Poisson distribution of obstacles. Henceforth,
for the sake of simplicity, we assume a 2-dimensional setting.
The obstacles (metallic atoms) are disks of radius r in the Euclidean plane
R2 , centered at c1 , c2 , . . . , cj , . . . R2 . Henceforth, we denote by
{c} = {c1 , c2 , . . . , cj , . . .} = a conguration of obstacle centers.
We further assume that the congurations of obstacle centers {c} are distributed under Poissons law with parameter n, meaning that
Prob({{c} | #(A {c}) = p}) = en|A|
(n|A|)p
,
p!
where |A| denotes the surface, i.e., the 2-dimensional Lebesgue measure of a measurable subset A of the Euclidean plane R2 .
This prescription denes a probability on countable subsets of the Euclidean
plane R2 .
45
where Z > 0is a normalizing coecient. Since the term i>j0 1|ci cj |>2r is not
of the form k0 k (ck ), the obstacles are no longer independent under this new
probability measure.
Next we dene the billiard ow in a given obstacle conguration {c}. This
denition is self-evident, and we give it for the sake of completeness, as well as in
order to introduce the notation.
Given a countable subset {c} of the Euclidean plane R2 , the billiard ow in
the system of obstacles dened by {c} is the family of mappings
(X(t; , , {c}), V (t; , , {c})) : Zr S1 Zr S1
where
|| = 1.
in
in
f{c}
(x, v) on the singleThen, given an initial probability density f{c}
particle phase-space with support outside the system of obstacles dened by {c},
we dene its evolution under the billiard ow by the formula
in
(X(t; x, v, {c}), V (t; x, v, {c})),
f (t, x, v, {c}) = f{c}
t 0.
46
Let 1 (x, v, {c}), 2 (x, v, {c}), . . . , j (x, v, {c}), . . . be the sequence of collision
times for a particle starting from x in the direction v at t = 0 in the conguration
of obstacles {c}. In other words,
j (x, v, {c}) = sup{t | #{s [0, t] | dist(X(s, x, v, {c}); {c}) = r} = j 1}.
Letting 0 = 0 and k = k k1 , the evolved single-particle density f is
a.e. dened by the formula
f (t, x, v, {c}) = f in (x tv, v) 1t<1
j
+
f in x
k V (k ) (t j )V (j+ ), V (j+ ) 1j <t<j+1 .
j1
k=1
p0
pen||
(n||)p
= n||,
p!
so that
n = # obstacles per unit surface in R2 .
The average of the rst term in the summation dening f (t, x, v, {c}) is
f in (x tv, v)1t<1 = f in (x tv, v)en2rt
(where denotes the mathematical expectation) since the condition t < 1
means that the tube of width 2r and length t contains no obstacle center.
Henceforth, we seek a scaling limit corresponding to small obstacles, i.e.,
r 0 and a large number of obstacles per unit volume, i.e., n .
47
2r
v
t
x
Figure 2.3: The tube corresponding with the rst term in the series expansion
giving the particle density
There are obviously many possible scalings that satisfy this requirement.
Among all these scalings, the BoltzmannGrad scaling in space dimension 2 is
dened by the requirement that the average over obstacle congurations of the
rst term in the series expansion for the particle density f has a nontrivial limit.
BoltzmannGrad scaling in space dimension 2
In order for the average of the rst term above to have a nontrivial limit, one must
have
r 0+ and n + in such a way that 2nr > 0.
Under this assumption,
f in (x tv, v) 1t<1 f in (x tv, v)et .
Gallavottis idea is that this rst term corresponds with the solution at time
t of the equation
cos(v, ) d = 2nrf,
(t + v x )f = nrf
f
||=1
v>0
t=0
= f in
that involves only the loss part in the Lorentz collision integral, and that the
(average over obstacle conguration of the) subsequent terms in the sum dening
the particle density f should converge to the Duhamel formula for the Lorentz
kinetic equation.
After these necessary preliminaries, we can state Gallavottis theorem.
48
Theorem 2.2.1 (Gallavotti [19]). Let f in be a continuous, bounded probability density on R2 S1 , and let
fr (t, x, v, {c}) = f in ((X r , V r )(t, x, v, {c})),
where (t, x, v) (X r , V r )(t, x, v, {c}) is the billiard ow in the system of disks of
radius r centered at the elements of {c}. Assuming that the obstacle centers are
distributed under the Poisson law of parameter n = /2r with > 0, the expected
single particle density
fr (t, x, v, ) f (t, x, v) in L1 (R2 S1 )
uniformly on compact t-sets, where f is the solution of the Lorentz kinetic equation
2
f (t, x, R[]v) sin 2
(t + v x )f + f =
0
f t=0 = f in ,
d
4
in
k V
k=1
(k )
(t j )V
(j+ ), V r (j+ )
1j <t<j+1 ,
f
in
j
k=1
k V
(k )
(t j )V
(j+ ), V r (j+ )
nj dc1 . . . dcj
,
j!
where T (t; c1 , . . . , cj ) is the tube of width 2r around the particle trajectory colliding rst with the obstacle centered at c1 and whose j-th collision is with the
obstacle centered at cj .
As before, the surface of that tube is
|T (t; c1 , . . . , cj )| = 2rt + O(r2 ).
In the j-th term, change variables by expressing the positions of the j encountered obstacles in terms of free ight times and deection angles:
(c1 , . . . , cj ) (1 , . . . , j ; 1 , . . . , j ).
49
1
2
v
t 1 2
Figure 2.4: The tube T (t, c1 , c2 ) corresponding with the third term in the series
expansion giving the particle density
c1
r
1
2
c2
d1 . . . dj .
j!
2
2 2
2
The measure in the left-hand side is invariant by permutations of c1 , . . . , cj ; on
the right-hand side, we assume that
1 < 2 < < j ,
which explains why the 1/j! factor disappears in the right-hand side.
The substitution above is one-to-one only if the particle does not hit twice
the same obstacle. Dene therefore
50
and set
d
(2nr)j e2nrt
sin 21 sin 2j d4 1 4 j d1 . . . dj
0<1 <<j <t
in
k R
k=1
k1
[0,2]j
l v (t j )R
l=1
l v, R
j
l=1
It is dominated by
f in L O()j eO()t
l v .
l=1
tj
j!
+et
t
0
2
0
+
j et
sin 21 sin 2j
j2
in
k R
k=1
k1
l=1
l v (t j )R
[0,2]j
l v, R
j
l=1
d4 1
l v
l=1
dj
4
d1 . . . dj ,
which is the Duhamel series giving the solution of the Lorentz kinetic equation.
Hence, we have proved that
frM (t, x, v, ) f (t, x, v) uniformly on bounded sets as r 0+ ,
where f is the solution of the Lorentz kinetic equation. One can check by a straightforward computation that the Lorentz collision integral satises the property
C()(v) dv = 0 for each L (S1 ).
S1
2.3. Santal
os formula for the geometric mean free path
51
Integrating both sides of the Lorentz kinetic equation in the variables (t, x, v) over
[0, t] R2 S1 shows that the solution f of that equation satises
f (t, x, v) dx dv =
f in (x, v) dx dv
R2 S1
R2 S1
R2 S1
52
2r
1
2.3. Santal
os formula for the geometric mean free path
53
r (x,v)
dx dv
,
|Zr /ZD | |SD1 |
that is invariant under the billiard ow the notation |SD1 | designates the
(D 1)-dimensional uniform measure of the unit sphere SD1 . This measure is
obviously invariant under the billiard ow
(Xr , Vr )(t, , ) : Zr SD1 Zr SD1
dened by
X r = Vr
V r = 0
while
whenever X(t)
/ Zr
54
with R[n]v = v 2(v n)n denoting the reection with respect to the hyperplane
(Rn) .
The second such probability measure is the invariant measure of the billiard
map
(v nx )+ dS(x) dv
dr (x, v) =
(v nx )+ dx dv-meas(r+ /ZD )
where nx is the unit inward normal at x Zr , while dS(x) is the (D 1)dimensional surface element on Zr , and
r+ = {(x, v) Zr SD1 | v nx > 0}.
The billiard map Br is the map
r+ (x, v) Br (x, v) = (Xr , Vr )(r (x, v); x, v) r+ ,
which obviously passes to the quotient modulo ZD -translations:
Br : r+ /ZD r+ /ZD .
In other words, given the position x and the velocity v of a particle immediately
after its rst collision with an obstacle, the sequence (Brn (x, v))n0 is the sequence
of all collision points and post-collision velocities on that particles trajectory.
With the material above, we can dene a rst, very natural notion of mean
free path, by setting
Mean Free Path =
N 1
1
r (Brk (x, v)).
N + N
lim
k=0
D
Notice that, for r -a.e. (x, v) +
r /Z , the right-hand side of the equality above
is well-dened by the Birkho ergodic theorem. If the billiard map Br is ergodic
for the measure r , one has
N 1
1
r (Brk (x, v)) =
r dr ,
N + N
r+ /ZD
lim
k=0
2.3. Santal
os formula for the geometric mean free path
55
r =
r+ /ZD
r (x, v) dr (x, v) =
1 |BD |r D
|BD1 |r D1
where BD is the unit ball of RD and |BD | its D-dimensional Lebesgue measure.
In fact, one has the following slightly more general
Lemma 2.3.1 (H. S. Dumas, L. Dumas, F. Golse [17]). For f C 1 (R+ ) such that
f (0) = 0, one has
f (r (x, v))v nx dS(x) dv =
f (r (x, v)) dx dv.
r+ /ZD
Santal
os formula is obtained by setting f (z) = z in the identity above, and
expressing both integrals in terms of the normalized measures r and r .
Proof. For each (x, v) Zr SD1 one has
r (x + tv, v) = r (x, v) t,
56
so that
d
r (x + tv, v) = 1.
dt
Hence r (x, v) solves the transport equation
v x r (x, v) = 1,
x Zr ,
x Zr ,
r (x, v) = 0,
Since f C 1 (R+ ) and f (0) = 0, one has
v x f (r (x, v)) = f (r (x, v)),
f (r (x, v)) = 0,
v SD1 ,
v nx < 0.
x Zr ,
x Zr ,
v SD1 ,
v nx < 0.
Integrating both sides of the equality above, and applying Greens formula shows
that
f (r (x, v)) dx dv
(Zr /ZD )SD1
=
v x (f (r (x, v))) dx dv
(Zr /ZD )SD1
f (r (x, v))v nx dS(x) dv.
=
(Zr /ZD )SD1
Beware the unusual sign in the right-hand side of the second equality above, coming
from the orientation of the unit normal nx , which is pointing towards Zr .
With the help of Santalos formula, we dene the BoltzmannGrad limit for
the Lorentz gas with periodic as well as random distribution of obstacles as follows:
BoltzmannGrad scaling
The BoltzmannGrad scaling for the periodic Lorentz gas in space dimension D
corresponds with the following choice of parameters:
distance between neighboring lattice points = 1,
obstacle radius = r 1,
1
mean free path = r > 0.
Santal
os formula indicates that one should have
1
D1
D
r c D1 with c =
as 0+ .
|BD1 |
Therefore, given an initial particle density f in Cc (RD SD1 ), we dene
fr to be
x
x
t
t
in
D1
Xr D1 ; D1 , v , Vr D1 ; D1 , v
fr (t, x, v) = f
r
r
r
r
r
57
2.4
In the proof of Gallavottis theorem for the case of a Poisson distribution of obstacles in space dimension D = 2, the probability that a strip of width 2r and length
t does not meet any obstacle is e2nrt , where n is the parameter of the Poisson
distribution i.e., the average number of obstacles per unit surface.
This accounts for the loss term
f in (x tv, v)et
in the Duhamel series for the solution of the Lorentz kinetic equation, or of the
term f on the right-hand side of that equation written in the form
2
(t + v x )f = f +
f (t, x, R()v) sin 2 d
4 .
0
Things are fundamentally dierent in the periodic case. To begin with, there
are innite strips included in the billiard table Zr which never meet any obstacle.
The contribution of the 1-particle density leading to the loss term in the
Lorentz kinetic equation is, in the notation of the proof of Gallavottis theorem,
f in (x tv, v) 1t<1 (x,v,{c}) .
The analogous term in the periodic case is
f in (x tv, v) 1t<rD1 r (x/rD1 ,v)
where r (x, v) is the free-path length in the periodic billiard table Zr starting from
x Zr in the direction v S1 .
58
Figure 2.9: Open strips in the periodic billiard table that never meet any obstacle
Passing to the L weak- limit as r 0 reduces to nding
lim 1t<rD1 r (x/rD1 ,v) in w L (R2 S1 )
r0
59
The lower bound and the upper bound in this theorem are obtained by very
dierent means.
The upper bound follows from a Fourier series argument which is reminiscent
of Siegels proof of the classical Minkowski convex body theorem (see [39, 36]).
The lower bound, on the other hand, is obtained by working in physical space.
Specically, one uses a channel technique, introduced independently by P. Bleher
[2] for the diusive scaling.
This lower bound alone has an important consequence:
Corollary 2.4.2. For each r > 0, the average of the free path length (mean free
path) under the probability measure r is innite:
r (x, v) dr (x, v) = +.
(Zr /ZD )SD1
Recall that the average of the free path length under the other natural
probability measure r is precisely Santalos formula for the mean free path:
r =
D
+
r /Z
r (x, v) dr (x, v) =
1 |BD |r D
.
|BD1 |rD1
One might wonder why averaging the free path length r under the measures r
and r actually gives two so dierent results.
First observe that Santalos formula gives the mean free path under the probability measure r concentrated on the surface of the obstacles, and is therefore
irrelevant for particles that have not yet encountered an obstacle.
Besides, by using the lemma that implies Santal
os formula with f (z) = 12 z 2 ,
one has
1
2
1
r (x, v) dr (x, v) =
2 r (x, v) dr (x, v).
r
D
(Zr /ZD )SD1
+
r /Z
Whenever the components v1 , . . . , vD are independent over Q, the linear ow
in the direction v is topologically transitive and ergodic on the D-torus, so that
r (x, v) < + for each r > 0 and x RD . On the other hand, r (x, v) = + for
some x Zr (the periodic billiard table) whenever v belongs to some specic class
of unit vectors whose components are rationally dependent, a class that becomes
dense in SD1 as r 0+ . Thus, r is strongly oscillating (nite for irrational
directions, possibly innite for a class of rational directions that becomes dense as
r 0+ ), and this explains why r does not have a second moment under r .
60
Proof of the BourgainGolseWennberg lower bound. We shall restrict our attention to the case of space dimension D = 2.
As mentioned above, there are innite nonempty open strips included in Zr
i.e., never meeting any obstacle. Call a channel any such nonempty open strip
of maximum width, and let Cr be the set of all channels included in Zr .
If S Cr and x S, dene S (x, v) the exit time from the channel starting
from x in the direction v, dened as
(x, v) S S1 .
Obviously, any particle starting from x in the channel S in the direction v must
exit S before it hits an obstacle (since no obstacle intersects S). Therefore
r (x, v) sup{S (x, v) | S Cr such that x S},
so that
r (t) r
SCr
This observation suggests that one should carefully study the set of channels Cr .
Step 1: Description of Cr . Given S1 , we dene
Cr () = {channels of direction in Cr }.
We begin with a lemma which describes the structure of Cr ().
Lemma 2.4.3. Let r [0, 12 ) and S1 . Then:
1) if S Cr (), then
2) if Cr () = , then
Cr () = {S + k | k Z2 };
(p, q)
=
p2 + q 2
with
(p, q) Z2 \ {(0, 0)} such that gcd(p, q) = 1 and
p2 + q 2 <
1
.
2r
61
d2r
2r
L
d
Figure 2.10: A channel of direction = 15 (2, 1); minimal distance d between lines
L and L of direction through lattice points
Proof of the lemma. Statement 1) is obvious. As for statement 2), if L is an innite
line of direction S1 such that 2 /1 is irrational, then L/Z2 is an orbit of a
linear ow on T2 with irrational slope 2 /1 . Therefore L/Z2 is dense in T2 so
that L cannot be included in Zr .
Assume that
(p, q)
=
with (p, q) Z2 \ {(0, 0)} coprime,
p2 + q 2
and let L, L be two innite lines with direction , with equations
qx py = a and qx py = a respectively.
Obviously
|a a |
dist(L, L ) =
.
p2 + q 2
62
the equality above following from the assumption that p and q are coprime.
Since dist(L, L ) > 0 is minimal, then |a a | = 1, so that
1
dist(L, L ) =
.
p2 + q 2
Likewise, if L L = S with S Cr , then L and L are parallel innite lines
tangent to Zr , and the minimal distance between any such distinct lines is
dist(L, L ) =
1
p2 + q2
2r.
Cut S into three parallel strips of equal width and call S the middle one. For each
t > 1 dene
rw(, r)
(, r, t) = arcsin
.
3t
Lemma 2.4.4. If x S and v (R[], R[]), where R[] designates the rotation
of an angle , then
S (x, v) t/r.
Moreover
The proof of this lemma is perhaps best explained by considering Figure 2.11.
Step 3: Putting all channels together. Recall that we need to estimate
2
1
{(x, v) (S/Z ) S | S (x, v) > t/r} .
r
SCr
Pick
(p, q)
(p , q )
Ar =
=
= Ar .
p2 + q 2
p2 + q2
Observe that
|pq p q|
1
2
2
2
2
2
2
p +q p +q
p + q p2 + q 2
2r
2r
sin (, r, t) + sin ( , r, t)
max
,
p2 + q 2
p2 + q 2
| sin(,
)| =
sin((, r, t) + ( , r, t))
63
t
e
w
^
S
S
Figure 2.11: Exit time from the middle third S of an innite strip S of width w
whenever t > 1.
Then, whenever S Cr () and S Cr ( ),
(S (R[], R[]))) (S (R[ ] , R[ ] ))) =
with = (, r, t), = ( , r, t) and R[] = rotation of an angle .
Moreover, if = (p,q)
Ar then
2
2
p +q
2 | = 1 w(, r)
|S/Z
3
p2 + q 2 ,
while
#{S/Z2 | S Cr ()} = 1.
Conclusion: Therefore, whenever t > 1,
2 ) (R[], R[])
(S/Z
SCr
SCr
64
SCr
2 ) (R[], R[]))
r ((S/Z
Ar
1
w(, r)
3
p2 + q2 2(, r, t)
gcd(p,q)=1
p2 +q 2 <1/4r 2
2
3
p2 + q2 w(, r) arcsin
gcd(p,q)=1
p2 +q 2 <1/4r 2
gcd(p,q)=1
p2 +q 2 <1/4r 2
Now
2
3
rw(, r)2
.
p2 + q 2
3t
that, eventually,
rw(, r)
3t
1
p2 +q 2
2r > 12
2
p +q 2
, so
2
3
gcd(p,q)=1
p2 +q 2 <1/16r 2
r2
rw(, r)2
p2 + q 2
3t
18t
65
gcd(p,q)=1
p2 +q 2 <1/16r 2
.
p2 + q 2
1
1
.
=
2
2
2
16r
4r p + q
gcd(p,q)=1
p2 +q 2 <1/16r2
p2 +q 2 <1/16r 2
t 0.
1/4
r (t, v)
dr
(t) a.e. in v S1
r
1
as t +.
2 t
Shortly after [9] appeared, F. Boca and A. Zaharescu improved our method
and managed to compute (t) explicitly for each t 0. One should keep in mind
that their formula had been conjectured earlier by P. Dahlqvist [14], on the basis
of a formal computation.
66
Figure 2.13: Black lines issued from the origin terminate at integer points with
coprime coordinates; red lines terminate at integer points whose coordinates are
not coprime
Theorem 2.4.6 (BocaZaharescu [3]). For each t > 0,
6
(s t)g(s) ds
r (t) (t) = 2
t
in the limit as r 0+ , where
1
g(s) =
1
1 2
ln(1 1s )
s +2 1 s
1
2 2
ln |1 2s |
2 1 s
if s [0, 1],
if s (1, ).
In the sequel, we shall return to the continued and Farey fractions techniques
used in the proofs of these two results, and generalize them.
2.5
The material at our disposal so far provides us with a rst answer albeit a
negative one to the problem of determining the BoltzmannGrad limit of the
periodic Lorentz gas.
67
Figure 2.14: Graph of (t) (blue curve) and (t) (green curve)
(x, v) Y SD1 ,
(x, v) Y SD1 ,
t=0
where nx is unit normal vector to Y at the point x, pointing towards the interior
of Y .
By the method of characteristics,
f (t, x, v) = f in D1 X
t
x
x
;
,
v
;
V
;
,
v
,
D1 D1
D1 D1
t
68
Theorem 2.5.1 (Golse [21, 24]). There exist initial data f in f in (x) C(TD )
such that no subsequence of f converges in L (R+ TD SD1 ) weak- to the
solution f of a linear Boltzmann equation of the form
(t + v x )f (t, x, v) =
p(v, v )(f (t, x, v ) f (t, x, v)) dv ,
SD1
f t=0 = f in ,
where > 0 and 0 p L2 (SD1 SD1 ) satises
p(v, v ) dv =
p(v , v) dv = 1 a.e. in v SD1 .
SD1
SD1
This theorem has the following important and perhaps surprising consequence: the Lorentz kinetic equation cannot govern the BoltzmannGrad limit of
the particle density in the case of a periodic distribution of obstacles.
Proof. The proof of the negative result above involves two dierent arguments:
a) the existence of a spectral gap for any linear Boltzmann equation, and
b) the lower bound for the distribution of free path lengths in the Bourgain
GolseWennberg theorem.
Step 1: Spectral gap for the linear Boltzmann equation
With > 0 and p as above, consider the unbounded operator A on L2 (TD SD1 )
dened by
(A)(x, v) = v x (x, v) (x, v) +
p(v, v )(x, v ) dv ,
SD1
with domain
D(A) = { L2 (TD SD1 ) | v x L2 (TD SD1 )}.
Then:
Theorem 2.5.2 (UkaiPointGhidouche [45]). There exist positive constants C and
such that
etA L2 (TD SD1 ) Cet L2 (TD SD1 ) ,
for each L2 (TD SD1 ), where
1
= D1
(x, v) dx dv.
|S
|
TD SD1
t 0,
69
Taking this theorem for granted, we proceed to the next step in the proof,
leading to an explicit lower bound for the particle density.
Step 2: Comparison with the case of absorbing obstacles
Assume that f in f in (x) 0 on TD . Then
f (t, x, v) g (t, x, v) = f in (x tv) 1Y (x) 1D1 (x/D1 ,v)>t .
Indeed, g is the density of particles with the same initial data as f , but assuming
that each particle disappears when colliding with an obstacle instead of being
reected.
Then
1Y (x) 1 a.e. on TD and |1Y (x)| 1
while, after extracting a subsequence if needed,
1D1 (x/D1 ,v)>t (t, v) in L (R+ TD SD1 ) weak-.
Therefore, if f is a weak- limit point of f in L (R+ TD SD1 ) as 0,
f (t, x, v) f in (x tv)(t, v) for a.e. (t, x, v).
Step 3: Using the lower bound on the distribution of r
Denoting by dv the uniform measure on SD1 ,
1
f (t, x, v)2 dx dv
|SD1 |
TD SD1
1
D1
f in (x tv)2 (t, v)2 dx dv
|S
|
TD SD1
1
f in (y)2 dy D1
(t, v)2 dv
=
|S
| SD1
TD
2
1
f in 2L2 (TD )
(t,
v)
dv
|SD1 | SD1
= f in 2L2 (TD ) (t)2 .
By the BourgainGolseWennberg lower bound on the distribution of free path
lengths,
CD in
f (t, , )L2 (TD SD1 )
f L2 (TD ) , t > 1.
t
On the other hand, by the spectral gap estimate, if f is a solution of the linear
Boltzmann equation, one has
f (t, , )L2 (TD SD1 )
f in (y) dy + Cet f in L2 (TD )
TD
70
so that
f in L1 (TD )
CD
in
+ Cet
t
f L2 (TD )
in
f (x) dx = 1, while
TD
in
f (y) dy =
D/2
(x) dx.
TD
t
Conclude by choosing so that
D/2
(x) dx < sup
t>1
CD
Cet
t
> 0.
Remarks.
1) The same result (with the same proof) holds for any smooth obstacle shape
included in a shell
{x RD | CD < dist(x, D1 ZD ) < C D }.
2) The same result (with the same proof) holds if the specular reection boundary condition is replaced by more general boundary conditions, such as absorption (partial or complete) of the particles at the boundary of the obstacles, diuse reection, or any convex combination of specular and diuse
reection in the classical kinetic theory of gases, such boundary conditons
are known as accommodation boundary conditions.
3) But introducing even the smallest amount of stochasticity in any periodic
conguration of obstacles can again lead to a BoltzmannGrad limit that is
described by the Lorentz kinetic model.
71
2.6
With the BourgainGolseWennberg lower bound for the distribution of free path
lengths in the periodic Lorentz gas, we have seen that the 1-particle phase space
density is bounded below by a quantity that is incompatible with the spectral
gap of any linear Boltzmann equation in particular with the Lorentz kinetic
equation.
In order to further analyze the BoltzmannGrad limit of the periodic Lorentz
gas, we cannot content ourselves with even more rened estimates on the distribution of free path lengths, but we need a convenient way to encode particle
trajectories.
More precisely, the two following problems must be answered somehow:
First problem: For a particle leaving the surface of an obstacle in a given direction,
nd the position of its next collision with an obstacle.
Second problem: Average in some sense to be dened in order to eliminate
the direction dependence.
From now on, our discussion is limited to the case of spatial dimension D = 2,
as we shall use continued fractions, a tool particularly well adapted to understanding the rational approximation of real numbers. Treating the case of a space
dimension D > 2 along the same lines would require a better understanding of
simultaneous rational approximation of D 1 real numbers (by D 1 rational
numbers with the same denominator), a notoriously more dicult problem.
We rst introduce some basic geometrical objects used in coding particle
trajectories.
The rst such object is the notion of impact parameter.
For a particle with velocity v S1 located at the position x on the surface
of an obstacle (disk of radius r), we dene its impact parameter hr (x, v) by the
formula
nx , v).
hr (x, v) = sin(
72
nx
x
Figure 2.15: The impact parameter h corresponding with the collision point x at
the surface of an obstacle, and a direction v
In other words, the absolute value of the impact parameter hr (x, v) is the distance
of the center of the obstacle to the innite line of direction v passing through x.
Obviously
hr (x, R[nx ]v) = hr (x, v),
where we recall the notation R[n]v = v 2(v n)n.
The next important object in computing particle trajectories in the Lorentz
gas is the transfer map.
For a particle leaving the surface of an obstacle in the direction v and with
impact parameter h , dene
Tr (h , v) = (s, h) with
n x
73
x
v
nx
74
Figure 2.17: Three types of orbits: the blue orbit is the shortest, the red one is
the longest, while the green one is of the intermediate length. The black segment
removed is orthogonal to the direction of the trajectories.
Theorem 2.6.2 (3-length theorem). Let (0, 1) \ Q and N 1. The sequence
{n | 0 n N }
denes N + 1 intervals on the circle of unit length R/Z. The lengths of these
intervals take at most three dierent values.
This striking result was conjectured by H. Steinhaus, and proved in 1957
independently by P. Erd
os, G. Hajos, J. Suranyi, N. Swieczkowski, P. Sz
usz
reported in [42], and by Vera S
os [41].
As we shall see, the 3-length theorem (in either form) is the key to encoding
particle paths in the 2-dimensional Lorentz gas. We shall need explicitly the formulas giving the lengths and widths of the three strips in the partition of the at
2-torus dened by the BlankKrikorian theorem. As this is based on the continued
fraction expansion of the slope of the linear ow considered in the BlankKrikorian
theorem, we rst recall some basic facts about continued fractions. An excellent
reference for more information on this subject is [28].
75
Figure 2.18: The 3-term partition. The shortest orbits are collected in the blue
strip, the longest orbits in the red strip, while the orbits of intermediate length
are collected in the green strip.
Continued fractions
Assume 0 < v2 < v1 and set = v2 /v1 , and consider the continued fraction
expansion of :
1
= [0; a0 , a1 , a2 , . . .] =
.
1
a0 +
a1 + . . .
Dene the sequences of convergents (pn , qn )n0 , meaning that
pn+2
= [0; a0 , . . . , an ],
qn+2
n 2,
p0 = 1, p1 = 0,
q0 = 0, q1 = 1.
76
n 0,
so that
dn+1 = an dn + dn1 ,
d0 = 1, d1 = .
51
1
=
.
= [0; 1, 1, . . .] =
1
2
1+
1 + ...
The sequence of errors associated with satises dn+1 = dn + dn1 for each
n 1 with d0 = 1 and d1 = , so that dn = n for each n 0.)
By induction, one veries that
qn dn+1 + qn+1 dn = 1,
n 0.
77
Collision patterns
The BlankKrikorian 3-length theorem has the following consequence, of fundamental importance in our analysis.
Any particle leaving the surface of one obstacle in some irrational direction v
will next collide with one of at most three exceptionally two other obstacles.
Any such collision pattern involving the three obstacles seen by the departing particle in the direction of its velocity is completely determined by exactly 4
parameters, computed in terms of the continued fraction expansion of v2 /v1 in
the case where 0 < v2 < v1 , to which the general case can be reduced by obvious
symmetry arguments.
2Ar
2r
Q/
2rB
Q/
Q/
Figure 2.20: Collision pattern seen from the surface of one obstacle; = 2r/v1
Assume
therefore 0 < v2 < v1 with = v2 /v1
/ Q. Henceforth, we set
= 2r 1 + 2 and dene
N (, ) = inf{n 0 | dn () },
dN (,)1 ()
.
k(, ) =
dN (,) ()
The parameters dening the collision pattern are A, B, Q as they appear on the
previous gure together with an extra parameter {1}. Here is how they
78
dN (,) ()
,
B(v, r) = 1
Once the structure of collision patterns elucidated with the help of the Blank
Krikorian variant of the 3-length theorem, we return to our original problem,
namely that of computing the transfer map.
In the next proposition, we shall see that the transfer map in a given, irrational direction v S1 can be expressed explicitly in terms of the parameters
A, B, Q, dening the collision pattern corresponding with this direction.
Write
K = (0, 1)3 {1}
and let (A, B, Q, ) K be the parameters dening the collision pattern seen by
a particle leaving the surface of one obstacle in the direction v. Set
if 1 2A < h 1,
(Q, h 2(1 A))
79
With this notation, the transfer map is essentially given by the explicit formula
TA,B,Q, , except for an error of the order O(r2 ) on the free-path length from
obstacle to obstacle.
Proposition 2.6.3 (CagliotiGolse [10, 11]). One has
Tr (h , v) = T(A,B,Q,)(v,r) (h ) + (O(r 2 ), 0)
in the limit as r 0+ .
In fact, the proof of this proposition can be read on the gure above that
represents a generic collision pattern. The rst component in the explicit formula
T(A,B,Q,)(v,r) (h )
represents exactly 2r times the distance between the vertical segments that are
the projections of the diameters of the 4 obstacles on the vertical ordinate axis.
Obviously, the free-path length from obstacle to obstacle is the distance between
the corresponding vertical segments, minus a quantity of the order O(r) that is the
distance from the surface of the obstacle to the corresponding vertical segment.
On the other hand, the second component in the same explicit formula is
exact, as it relates impact parameters, which are precisely the intersections of the
innite line that contains the particle path with the vertical segments corresponding with the two obstacles joined by this particle path.
If we summarize what we have done so far, we see that we have solved our
rst problem stated at the beginning of the present section, namely that of nding
a convenient way of coding the billiard ow in the periodic case and for space
dimension 2, for a.e. given direction v.
2.7
It remains to solve the second problem, namely, to nd a convenient way of averaging the computation above so as to get rid of the dependence on the direction v.
Before going further in this direction, we need to recall some known facts
about the ergodic theory of continued fractions.
80
Moreover, the Gauss map T is ergodic for the invariant measure dg(x). By
Birkhos theorem, for each f L1 (0, 1; dg),
1
N 1
1
f (T k x)
f (z) dg(z) a.e. in x (0, 1)
N
0
k=0
as N +.
How the Gauss map is related to continued fractions is explained as follows:
for
1
(0, 1) \ Q,
= [0; a0 , a1 , a2 , . . .] =
1
a0 +
a1 + . . .
the terms ak () of the continued fraction expansion of can be computed from
the iterates of the Gauss map acting on . Specically,
1
ak () =
, k 0.
T k
As a consequence, the Gauss map corresponds with the shift to the left on
innite sequences of positive integers arising in the continued fraction expansion
of irrationals in (0, 1). In other words,
T [0; a0 , a1 , a2 , . . .] = [0; a1 , a2 , a3 . . .],
equivalently recast as
an (T ) = an+1 (),
n 0.
Thus, the terms ak () of the continued fraction expansion of any (0, 1)\Q
are easily expressed in terms of the sequence of iterates (T k )k0 of the Gauss
map acting on . The error dn () is also expressed in terms of that same sequence
(T k )k0 , by equally simple formulas.
Starting from the induction relation on the error terms
dn+1 () = an ()dn () + dn1 (),
d0 () = 1, d1 () = ,
n 0.
n1
T k ,
k=0
n 0.
dn () 2[n/2] ,
81
n 0,
which establishes the exponential decay mentioned above. (As a matter of fact,
exponential convergence is the slowest possible for the continued fraction algorithm, as it corresponds with the rational approximation of algebraic numbers of
degree 2, which are the hardest to approximate by rational numbers.)
Unfortunately, the dependence of qn () in is more complicated. Yet one can
nd a way around this, with the following observation. Starting from the relation
qn+1 ()dn () + qn ()dn+1 () = 1,
we see that
qn ()dn1 () =
(1)nj
j=1
n
j=1
(1)nj
dn ()dn1 ()
dj ()dj1 ()
n1
T k1 T k .
k=j
Using once more the inequality T < 12 for [0, 1] \ Q, one can truncate the
summation above at the cost of some exponentially small error term. Specically,
one nds that
n
nj dn ()dn1 ()
qn ()dn1 ()
(1)
dj ()dj1 ()
j=nl
n1
n
(1)nj
T k1 T k 2l .
= qn ()dn1 ()
j=nl
k=j
More information on the ergodic theory of continued fractions can be found in the
classical monograph [28] on continued fractions, and in Sinais book on ergodic
theory [40].
An ergodic theorem
We have seen in the previous section that the transfer map satises
Tr (h , v) = T(A,B,Q,)(v,r) (h ) + (O(r 2 ), 0) as r 0+
for each v S1 such that v2 /v1 (0, 1) \ Q.
Obviously, the parameters (A, B, Q, ) are extremely sensitive to variations
in v and r as r 0+ , so that even the explicit formula for TA,B,Q, is not too
useful in itself.
Each time one must handle a strongly oscillating quantity such as the free
path length r (x, v) or the transfer map Tr (h , v), it is usually a good idea to
82
consider the distribution of that quantity under some natural probability measure
rather than the quantity itself. Following this principle, we are led to consider the
family of probability measures in (s, h) R+ [1, 1],
((s, h) Tr (h , v)),
or equivalently
n 0,
so that
N (, ) = inf{n 1 | dn () }
is transformed by the Gauss map as follows:
N (a, ) = N (T , /) + 1.
In other words, the transfer map for the 2-dimensional periodic Lorentz gas
in the billiard table Zr (meaning with circular obstacles of radius r centered at
the vertices of the lattice Z2 ) in the direction v corresponding with the slope is
essentially the same as for the billiard table Zr/ but in the direction corresponding
with the slope T . Since the problem is invariant under the transformation
T ,
r r/,
this suggests the idea of averaging with respect to the scale invariant measure in
the variable r, i.e., dr/r on R+ .
The key result in this direction is the following ergodic lemma for functions
that depend on nitely many dn s.
Lemma 2.7.1 (CagliotiGolse [9, 22, 11]). For (0, 1) \ Q, set
N (, ) = inf{n 0 | dn () }.
For each m 0 and each f C(Rm+1
), one has
+
1
| ln |
1/4
dN (,)m ()
dN (,) ()
,...,
d
Lm (f )
83
With this lemma, we can average over obstacle radius any function that
depends on collision patterns, i.e., any function of the parameters A, B, Q, .
Proposition 2.7.2 (CagliotiGolse [11]). Let K = [0, 1]3 {1}. For each F
C(K), there exists L(F ) R independent of v such that
1/2
dr
1
L(F )
F (A(v, r), B(v, r), Q(v, r), (v, r))
ln(1/)
r
for a.e. v S1 such that 0 < v2 < v1 in the limit as 0+ .
Sketch of the proof. First eliminate the dependence by decomposing
F (A, B, Q, ) = F+ (A, B, Q) + F (A, B, Q).
Hence it suces to consider the case where F F (A, B, Q).
Setting = v2 /v1 and = 2r/v1 , we recall that
dN (,) ()
,
dN (,)1 ()
dN (,) ()
and
.
B(v, r) is a function of
A(v, r) is a function of
dN (,)1
N (,)
(1)N (,)j
j=N (,)m
to which we apply the ergodic lemma above: its Ces`aro mean converges, in the
small radius limit, to some limit Lm (F ) independent of .
By uniform continuity of F , one nds that
84
and with the error estimate above for the integrand, one nds that
1/2
dr
1
F (A(v, r), B(v, r), Q(v, r), (v, r))
L(F )
ln(1/)
r
as 0+ .
With the ergodic theorem above, and the explicit approximation of the transfer map expressed in terms of the parameters (A, B, Q, ) that determine collision
patterns in any given direction v, we easily arrive at the following notion of a
probability of transition for a particle leaving the surface of an obstacle with an
impact parameter h to hit the next obstacle on its trajectory at time s/r with
an impact parameter h.
Theorem 2.7.3 (CagliotiGolse [10, 11]). For each h [1, 1], there exists a
probability density P (s, h|h ) on R+ [1, 1] such that, for each f belonging to
C(R+ [1, 1]),
1
| ln |
1/4
dr
f (Tr (h , v))
r
0
a.e. in v S1 as 0+ .
In other words, the transfer map converges in distribution and in the sense
of Ces`aro, in the small radius limit, to a transition probability P (s, h|h ) that is
independent of v.
We are therefore left with the following problems:
a) to compute the transition probability P (s, h|h ) explicitly and discuss its
properties, and
b) to explain the role of this transition probability in the BoltzmannGrad limit
of the periodic Lorentz gas dynamics.
2.8
Most unfortunately, our argument leading to the existence of the limit L(F ), the
core result of the previous section, cannot be used for computing explicitly the
value L(F ). Indeed, the convergence proof is based on the ergodic lemma in the last
section, coupled to a sequence of approximations of the parameter Q in collision
patterns that involve only nitely many error terms dn () in the continued fraction
expansion of . The existence of the limit is obtained through Cauchys criterion,
precisely because of the diculty in nding an explicit expression for the limit.
Nevertheless, we have arrived at the following expression for the transition
probability P (s, h|h ):
85
Theorem 2.8.1 (CagliotiGolse [10, 11]). The transition density P (s, h|h ) is expressed in terms of a = 12 |h h | and b = 12 |h + h | by the explicit formula
3
(s 12 sa) (1 + 12 sa) 1 ( 12 s + 12 sb) +
P (s, h|h ) = 2
sa
+ (s 12 sa) 1 ( 12 s + 12 sb) 1 12 sa +
+sa |1 s| 1s<1 + (sa |1 s|)+ ,
with the notations x y = min(x, y) and x y = max(x, y).
Moreover, the function
(s, h, h ) (1 + s)P (s, h|h ) belongs to L2 (R+ [1, 1]2 ).
In fact, the key result in the proof of this theorem is the asymptotic distribution of 3-obstacle collision patterns i.e., the computation of the limit L(f ),
whose existence has been proved in the last sections proposition.
Theorem 2.8.2 (CagliotiGolse [11]). Dene K = [0, 1]3 {1}. Then, for each
F C(K),
1/4
dr
1
F ((A, B, Q, )(v, r))
L(F )
| ln |
r
F (A, B, Q, ) dm(A, B, Q, ) a.e. in v S1
=
K
as 0 , where
dm(A, B, Q, ) = dm0 (A, B, Q) 12 (=1 + =1 ),
12
dA dB dQ
1
.
dm0 (A, B, Q) = 2 10<A<1 10<B<1A 10<Q< 2AB
1A
Before giving an idea of the proof of the theorem above on the distribution of
3-obstacle collision patterns, it is perhaps worthwhile explaining why the measure
m above is somehow natural in the present context.
To begin with, the constraints 0 < A < 1 and 0 < B < 1 A have an
obvious geometric meaning (see Figure 2.20 on collision patterns.) More precisely,
the widths of the three strips in the 3-term partition of the 2-torus minus the slit
constructed in the penultimate section (as a consequence of the BlankKrikorian
3-length theorem) add up to 1. Since A is the width of the strip consisting of the
shortest orbits in the BlankKrikorian theorem, and B that of the strip consisting
of the next to shortest orbits, one has
0<A+B 1
with equality only in the exceptional case where the orbits have at most two
dierent lengths, which occurs for a set of measure 0 in v or r. Therefore, one has
0 < B(v, r) < 1 A(v, r),
86
Likewise, the total area of the 2-torus is the sum of the areas of the strips
consisting of all orbits with the three possible lengths:
1 = QA + Q B + (Q + Q )(1 A B) = Q(1 B) + Q (1 A)
Q(2 A B)
as Q Q (see again the gure above on collision patterns).
Therefore, the volume element
dA dB dQ
1A
B
in the expression of dm0 implies that the parameters A, 1A
or equivalently B
mod (1 A) and Q are uniformly distributed in the largest subdomain of [0, 1]3
that is compatible with the geometric constraints.
The rst theorem is a consequence of the second: indeed, P (s, h|h ) ds dh is
the image measure of dm(A, B, Q, ) under the map
F (A, B, Q, ) dm(A, B, Q, )
K
|v|=1
0<v2 <v1
87
Farey fractions
Put a ltration on the set of rationals in [0, 1] as follows:
FQ = pq | 0 p q Q, gcd(p, q) = 1 ,
indexed in increasing order:
0=
0
pj
1
< 1 < < j =
< < (Q) = = 1,
1
qj
1
p prime
p|n
1
1
p
.
p
p
and =
q
q
p
q
< =
p
q
a
q a
q = 1 and q + q > Q.
Conversely, q, q are denominators of adjacent fractions in FQ if and only if
0 q, q Q,
q + q > Q,
gcd(q, q ) = 1.
p(, Q)
p(, Q)
< < (, Q) =
.
q(, Q)
q(, Q)
88
Figure 2.21: The SternBrocot tree. Each fraction on the n-th line is the mediant
of the two fractions closest to on the (n 1)-st line. The rst line consists of 0
and 1 written as 0 = 01 and 1 = 11 . Each rational in [0, 1] is obtained in this way.
At this point, we recall the relation between Farey and continued fractions.
Pick 0 < < 1; we recall that, for each (0, 1) \ Q,
N (, ) = min{n N | dn () },
p(, Q))
p(, Q))
< (, Q) =
q(, Q)
q(, Q)
= [0; a1 , a2 , . . .] =
a0 +
1
a1 + . . .
89
D(v, r)
B(v, r) = 1
D(v, r)
= 1 dN (,)1 ()/ mod D(v, r)
r), Z)/ mod D(v, r).
= 1 dist( 1 Q(v,
To summarize, we have
r), D(v, r))
F (A(v, r), B(v, r), Q(v, r)) = G(Q(v, r), Q(v,
and we are left with the task of computing
r), D(v, r)) dv
lim
G(Q(v, r), Q(v,
r0+
S1+
where S1+ is the rst octant in the unit circle. The other octants in the unit circle
give the same contribution by obvious symmetry arguments.
More specically:
Lemma 2.8.3. Let (0, 1) \ Q, and let pq < < pq be the two adjacent Farey
fractions in FQ surrounding , with Q = [1/]. Then:
a) If
p
q
<
,
q
then
Q(v, r) = q,
b) If
p+
q
p+
q
D(v, r) = 1 (q p).
r) = q,
Q(v,
p
q ).
D(v, r) = 1 (
c) If
r) =
Q(v,
q,
<
q ,
then
Q(v, r) = (q q),
r) = (q q),
Q(v,
90
one has
S1+
q
q q
g
=
,
H (Q(q p)) d
Q Q
p/q
0<q,qQ<q+
gcd(q,q)=1
g
0<q,qQ<q+
gcd(q,q)=1
q q
,
Q Q
1
qQ
H
0<q,qQ<q+
gcd(q,q)=1
q q
,
Q Q
.
(x, y) dx dy
Q2 0<q,qQ<q+
Q Q
0<x,y<1<x+y
gcd(q,q)=1
in the limit as Q .
This is precisely the path followed by F. Boca and A. Zaharescu to compute
the limiting distribution of free path lengths in [3] (see Theorem 2.4.6); as explained above, their analysis can be greatly generalized in order to compute the
transition probability that is the limit of the transfer map as the obstacle radius
r 0+ .
2.9
We are now ready to propose an equation for the BoltzmannGrad limit of the
periodic Lorentz gas in space dimension 2. For each r (0, 12 ), denote the billiard
map by
+
Br : +
r (x, v) Br (x, v) = (x + r (x, v)v, R[x + r (x, v)v]v) r .
91
n N .
where (x0 , v0 ) = (x r (x, v)v, v) and hr (x1 /r, v1 ) = sin(nx1 , v1 ), and m is the
probability measure on K obtained in Theorem 2.8.2.
If this holds, the iterates of the transfer map Tr are described by the Markov
chain with transition probability 2P (2s, h|h ). This leads to a kinetic equation on
an extended phase space for the BoltzmannGrad limit of the periodic Lorentz
gas in space dimension 2:
F (t, x, v, s, h) =
density of particles with velocity v and position x at time t
that will hit an obstacle after time s, with impact parameter h.
Theorem 2.9.1 (CagliotiGolse [10, 11]). Assume (H), and let f in 0 belong to
Cc (R2 S1 ). Then one has
1
fr
F ( , , , s, h) ds dh in L (R+ R2 S1 ) weak-
0
F (0, x, v, s, h) = f in (x, v)
2s
P (, h|h ) dh d,
with (x, v, s, h) running through R2 S1 R+ (1, 1). The notation R[] designates the rotation of an angle .
92
Let us briey sketch the computation leading to the kinetic equation above
in the extended phase space Z = R2 S1 R+ [1, 1].
In the limit as r 0+ , the sequence (bnr (x, v))n1 converges to a sequence
of i.i.d. random variables with values in K = [0, 1] {1}, according to assumption (H).
Then, for each s0 > 0 and h0 [1, 1], we construct a Markov chain
(sn , hn )n1 with values in R+ [1, 1] in the following manner:
(sn , hn ) = Tbn (hn1 ),
n 1.
n 1,
and
vn = R[2 arcsin(hn1 ) ]vn1 ,
n 1.
With the sequence (vn , sn , hn )n1 so dened, we next introduce the formulas
for (Xt , Vt , St , Ht ):
While 0 t < , we set
Xt (x, v, s, h) = x + t,
Vt (x, v, s, h) = v,
St (x, v, s, h) = s t,
Ht (x, v, s, h) = h.
93
Then
E[1s<t E[(T(ts)0 (x + sv, R[(h)]v, s1 , h1 ))|s1 , h1 ]]
= 1s<t g(t s, x + sv, R[(h)]v, Tb1 (h))] dm(b1 )
= 1s<t g(t s, x + sv, R[(h)]v, s1 , h1 )]2P (2s1 , h1 |h) ds1 dh1 .
94
Finally,
g(t, x, v, s, h) = (x + tv, v, s t, h) 1t<s
+ 1s<t g(t s, x + sv, R[(h)]v, s1 , h1 )]P (s1 , h1 |h) ds1 dh1 .
This formula represents the solution of the problem
(t v x + s )g = 0, t, s > 0, x R2 , s S1 , |h| < 1,
P (s1 , h1 |h)g(t, x, v, s1 , h1 ) ds1 dh1 ,
g(t, x, s, 0, h) =
g
R
+ (1,1)
t=0
= .
The boundary condition for s = 0 can be replaced with a source term that is
proportional to the Dirac measure s=0 :
P (s1 , h1 |h)g(t, x, v, s1 , h1 ) ds1 dh1 ,
(t v x + s )g = s=0
g
R
+ (1,1)
t=0
= .
One concludes by observing that this problem is precisely the adjoint of the Cauchy
problem in the theorem.
Let us conclude this section with a few bibliographical remarks. Although
the BoltzmannGrad limit of the periodic Lorentz gas is a fairly natural problem,
it remained open for quite a long time after the pioneering work of G. Gallavotti
on the case of a Poisson distribution of obstacles [18, 19].
Perhaps the main conceptual diculty was to realize that this limit must
involve a phase-space other than the usual phase-space of kinetic theory, i.e., the
set R2 S1 of particle positions and velocities, and to nd the appropriate extended
phase-space where the BoltzmannGrad limit of the periodic Lorentz gas can be
described by an autonomous equation.
Already Theorem 5.1 in [9] suggested that, even in the simplest problem
where the obstacles are absorbing i.e., holes where particles disappear forever
the limit of the particle number density in the BoltzmannGrad scaling cannot be
described by an autonomous equation in the usual phase space R2 S1 .
The extended phase space R2 S1 R+ [1, 1] and the structure of the limit
equation were proposed for the rst time by E. Caglioti and the author in 2006,
and presented in several conferences see for instance [23]; the rst computation
of the transition probability P (s, h|h ) (Theorem 2.8.1), together with the limit
equation (Theorem 2.9.1) appeared in [10] for the rst time. However, the theorem
concerning the limit equation in [10] remained incomplete, as it was based on the
independence assumption (H).
Shortly after that, J. Marklof and A. Strombergsson proposed a complete
derivation of the limit equation of Theorem 2.9.1 in a recent preprint [32]. Their
95
analysis establishes the validity of this equation in any space dimension, using
in particular the existence of a transition probability as in Theorem 2.8.1 in any
space dimension, a result that they had proved in an earlier paper [31]. The method
of proof in this article [31] avoided using continued or Farey fractions, and was
based on group actions on lattices in the Euclidean space, and on an important
theorem by M. Ratner implying some equidistribution results in homogeneous
spaces. However, explicit computations (as in Theorem 2.8.1) of the transition
probability in space dimension higher than 2 seem beyond reach at the time of
this writing see however [33] for computations of the 2-dimensional transition
probability for more general interactions than hard sphere collisions.
Finally, the limit equation obtained in Theorem 2.9.1 is interesting in itself;
some qualitative properties of this equation are discussed in [11].
Conclusion
Classical kinetic theory (Boltzmann theory for elastic, hard sphere collisions) is
based on two fundamental principles:
a) Deections in velocity at each collision are mutually independent and identically distributed.
b) Time intervals between collisions are mutually independent, independent of
velocities, and exponentially distributed.
The BoltzmannGrad limit of the periodic Lorentz gas provides an example
of a non-classical kinetic theory where velocity deections at each collision jointly
form a Markov chain, and the time intervals between collisions are not independent
of the velocity deections.
In both cases, collisions are purely local and instantaneous events: indeed the
BoltzmannGrad scaling is such that the particle radius is negligible in the limit.
The dierence between these two cases is caused by the degree of correlation
between obstacles, which is maximal in the second case since the obstacles are
centered at the vertices of a lattice in the Euclidean space, whereas obstacles are
assumed to be independent in the rst case. It could be interesting to explore
situations that are somehow intermediate between these two extreme cases for
instance, situations where long range correlations become negligible.
Otherwise, there remain several outstanding open problems related to the
periodic Lorentz gas, such as
i) obtaining explicit expressions of the transition probability whose existence is
proved by J. Marklof and A. Strombergsson in [31], in all space dimensions,
or
ii) treating the case where particles are accelerated by an external force for
instance the case of a constant magnetic eld, so that the kinetic energy of
particles remains constant.
Bibliography
[1] S. Blank, N. Krikorian, Thoms problem on irrational ows. Internat. J. Math.
4 (1993), 721726.
[2] P. Bleher, Statistical properties of two-dimensional periodic Lorentz gas with
innite horizon. J. Statist. Phys. 66 (1992), 315373.
[3] F. Boca, A. Zaharescu, The distribution of the free path lengths in the periodic two-dimensional Lorentz gas in the small-scatterer limit. Commun.
Math. Phys. 269 (2007), 425471.
[4] C. Boldrighini, L. A. Bunimovich, Ya. G. Sinai, On the Boltzmann equation
for the Lorentz gas. J. Statist. Phys. 32 (1983), 477501.
[5] L. Boltzmann, Weitere Studien u
ber das W
armegleichgewicht unter Gasmolek
ulen. Wiener Berichte 66 (1872), 275370.
[6] J. Bourgain, F. Golse, B. Wennberg, On the distribution of free path lengths
for the periodic Lorentz gas. Commun. Math. Phys. 190 (1998), 491-508.
[7] L. Bunimovich, Ya. G. Sinai, Statistical properties of Lorentz gas with periodic conguration of scatterers. Commun. Math. Phys. 78 (1980/81), 479497.
[8] L. Bunimovich, N. Chernov, Ya. G. Sinai, Statistical properties of twodimensional hyperbolic billiards. Russian Math. Surveys 46 (1991), 47106.
[9] E. Caglioti, F. Golse, On the distribution of free path lengths for the periodic
Lorentz gas III. Commun. Math. Phys. 236 (2003), 199221.
[10] E. Caglioti, F. Golse, The BoltzmannGrad limit of the periodic Lorentz gas
in two space dimensions. C. R. Math. Acad. Sci. Paris 346 (2008), 477482.
[11] E. Caglioti, F. Golse, On the Boltzmann-Grad Limit for the Two Dimensional
Periodic Lorentz Gas. J. Stat. Phys. 141 (2010), 264317.
[12] C. Cercignani, On the Boltzmann equation for rigid spheres. Transport Theory Statist. Phys. 2 (1972), 211225.
[13] C. Cercignani, R. Illner, M. Pulvirenti, The mathematical theory of dilute
gases. Applied Mathematical Sciences, 106. Springer-Verlag, New York, 1994.
96
Bibliography
97
[14] P. Dahlqvist, The Lyapunov exponent in the Sinai billiard in the small scatterer limit. Nonlinearity 10 (1997), 159173.
[15] L. Desvillettes, V. Ricci, Nonmarkovianity of the BoltzmannGrad limit of a
system of random obstacles in a given force eld. Bull. Sci. Math. 128 (2004),
3946.
[16] P. Drude, Zur Elektronentheorie der Metalle. Annalen der Physik 306 (3)
(1900), 566613.
[17] H. S. Dumas, L. Dumas, F. Golse, Remarks on the notion of mean free path
for a periodic array of spherical obstacles. J. Statist. Phys. 87 (1997), 943950.
[18] G. Gallavotti, Divergences and approach to equilibrium in the Lorentz and
the windtreemodels. Phys. Rev. (2) 185 (1969), 308322.
[19] G. Gallavotti, Rigorous theory of the Boltzmann equation in the Lorentz gas.
Nota interna no. 358, Istituto di Fisica, Univ. di Roma (1972). Available as
preprint mp-arc-93-304.
[20] G. Gallavotti, Statistical Mechanics: a Short Treatise, Springer, BerlinHeidelberg (1999).
[21] F. Golse, On the statistics of free-path lengths for the periodic Lorentz gas.
Proceedings of the XIVth International Congress on Mathematical Physics
(Lisbon 2003), 439446, World Scientic, Hackensack NJ, 2005.
[22] F. Golse, The periodic Lorentz gas in the BoltzmannGrad limit. Proceedings
of the International Congress of Mathematicians, Madrid 2006, vol. 3, 183
201, European Math. Soc., Z
urich, 2006.
[23] F. Golse, The periodic Lorentz gas in the BoltzmannGrad limit (joint
work with J. Bourgain, E. Caglioti and B. Wennberg). Oberwolfach Report
54/2006, vol. 3 (2006), no. 4, 3214, European Math. Soc., Z
urich, 2006.
[24] F. Golse, On the periodic Lorentz gas in the BoltzmannGrad scaling. Ann.
Faculte des Sci. Toulouse 17 (2008), 735749.
[25] F. Golse, B. Wennberg, On the distribution of free path lengths for the periodic Lorentz gas II. M2AN Model. Math. et Anal. Numer. 34 (2000), 1151
1163.
[26] H. Grad, Principles of the kinetic theory of gases, in: Handbuch der Physik,
S. Fl
ugge ed. Band XII, 205294, Springer-Verlag, Berlin 1958.
[27] R. Illner, M. Pulvirenti, Global validity of the Boltzmann equation for twoand three-dimensional rare gas in vacuum. Erratum and improved result:
Global validity of the Boltzmann equation for a two-dimensional rare gas in
vacuum [Commun. Math. Phys. 105 (1986), 189203] and Global validity
of the Boltzmann equation for a three-dimensional rare gas in vacuum [ibid.
113 (1987), 7985] by M. Pulvirenti. Commun. Math. Phys. 121 (1989), 143
146.
98
Bibliography
Bibliography
99
[44] D. Szasz, Hard Ball Systems and the Lorentz Gas. Encyclopaedia of Mathematical Sciences, 101. Springer, 2000.
[45] S. Ukai, N. Point, H. Ghidouche, Sur la solution globale du probl`eme mixte
de lequation de Boltzmann non lineaire. J. Math. Pures Appl. (9) 57 (1978),
203229.
Chapter 3
Introduction
Boundary eects play a crucial role in the dynamics of gases governed by the
Boltzmann equation:
(3.1.1)
t F + v x F = Q(F, F ),
where F (t, x, v) is the distribution function for the gas particles at time t 0,
position x , and v R3 . Throughout this chapter, the collision operator takes
the form
|v u| F1 (u )F2 (v )q0 () d du
Q(F1 , F2 ) =
R3 S2
|v u| F1 (u)F2 (v)q0 () d du
R3
S2
(3.1.2)
101
102
(2) Bounce-back reection, in which the particles bounce back with reversed
velocity.
(3) Specular reection, in which the particles bounce back specularly.
(4) Diuse reection (stochastic), in which the incoming particles are a probability average of the outgoing particles.
Due to its importance, there have been many contributions in the mathematical
study of dierent aspects of the Boltzmann boundary value problems [1], [2], [3],
[4], [6], [9], [10], [11], [15], [30], [31], [34], [37], [39], [46], [49], among others. See
also the references in the books [8], [12] and [44].
According to Grad [28, p. 243], one of the basic problems in the Boltzmann study is to prove existence and uniqueness of its solutions, as well as their
time-decay toward an absolute Maxwellian, in the presence of compatible physical boundary conditions in a general domain. In spite of those contributions to
the study of Boltzmann boundary problems, there are fewer mathematical results of uniqueness, regularity, and time decay-rate for Boltzmann solutions toward a Maxwellian. In [41], it was announced that Boltzmann solutions near a
Maxwellian would decay exponentially to it in a smooth bounded convex domain
with specular reection boundary conditions. Unfortunately, we are not aware of
any complete proof for such a result [45]. In [30], global stability of the Maxwellian
was established in a convex domain for diusive boundary conditions. Recently,
important progress has been made in [16] and [47] to establish an almost exponential decay rate for Boltzmann solutions with large amplitude for general
collision kernels and general boundary conditions, provided that certain a-priori
strong Sobolev estimates can be veried. Even though these estimates had been
established for spatially periodic domains [22], [23] near Maxwellians, their validity is completely open for the Boltzmann solutions, even local in time, in a
bounded domain. As a matter of fact, this kind of strong Sobolev estimates may
not be expected for a general non-convex domain [23]. This is because even for
simplest kinetic equations with the dierential operator v x , the phase boundary
R3 is always characteristic but not uniformly characteristic at the grazing set
0 = {(x, v) : x and v n(x) = 0} where n(x) is the outward normal at x.
Hence it is very challenging and delicate to obtain regularity from the general
theory of hyperbolic PDE. Moreover, in comparison with the half-space problems
studied, for instance in [34], [49], the geometrical complication makes it dicult
to employ spatial Fourier transforms in x. There are many cycles (bouncing characteristics) interacting with the boundary repeatedly, and analysis of such cycles
is one of the key mathematical diculties.
We aim to develop a unied L2 L theory in the near Maxwellian regime,
2
1
to establish exponential decay toward a normalized Maxwellian = e 2 |v| , for all
four basic types of boundary conditions in rather general domains. Consequently,
uniqueness among these solutions can be obtained. For convex domains, these
solutions are shown to be continuous away from the singular grazing set 0 .
3.2
103
We let = {x : (x) < 0} be connected and bounded with (x) a smooth function.
We assume that (x) = 0 at the boundary (x) = 0. The outward normal vector
at is given by
(x)
n(x) =
,
(3.2.1)
|(x)|
and it can be extended smoothly near = {x : (x) = 0}. We say that is real
analytic if is real analytic in x. We dene to be strictly convex if there exists
c > 0 such that
ij (x) i j c ||2
(3.2.2)
for all x such that (x) 0, and all R3 . We say that has a rotational
symmetry if there are vectors x0 and such that, for all x ,
{(x x0 ) } n(x) 0.
(3.2.3)
dV (s)
= 0,
ds
(3.2.4)
with the initial condition [X(t; t, x, v), V (t; t, x, v)] = [x, v].
such that there exists
Denition 1 (Backward exit time). For (x, v) with x
some > 0 for which xsv for 0 s , we dene tb (x, v) > 0 to be the last
moment at which the back-time straight line [X(s; 0, x, v), V (s; 0, x, v)] remains in
the interior of :
tb (x, v) = sup{ > 0 : x sv for 0 s }.
(3.2.5)
(3.2.6)
104
3.3
f , the Boltzmann
Lf f Kf = {Q(, f ) + Q( f, )} = f k(v, v )f (v ) dv
(3.3.1)
with the collision frequency (v) |v u| (u)q0 () du d {1 + |v|} for
0 1; and
1
(3.3.2)
(3.3.3)
(3.3.4)
(3.3.5)
and
f (t, x, v)| = f (x, v, v 2(n(x) v)n(x)) = f (x, v, R(x)v).
(3.3.6)
(4) Diuse reection: assume the natural normalization with some constant c >
0,
c
vn(x)>0
(v)|n(x) v| dv = 1.
(3.3.7)
(v)
v n(x)>0
f (t, x, v ) (v ){n(x) v } dv .
(3.3.8)
105
For both the bounce-back and specular reection conditions (3.3.4) and
(3.3.6), it is well-known that both mass and energy are conserved for (3.1.1). Without loss of generality, we may always assume that the mass-energy conservation
laws hold for t 0, in terms of the perturbation f :
f (t, x, v) dx dv = 0,
(3.3.9)
R3
R3
|v|2 f (t, x, v) dx dv = 0.
(3.3.10)
Moreover, if the domain has any axis of rotation symmetry (3.2.3), then we
further assume that the corresponding conservation of angular momentum is valid
for all t 0:
{(x x0 ) } vf (t, x, v) dx dv = 0.
(3.3.11)
R3
For the diuse reection (3.3.8), the mass conservation (3.3.9) is assumed to
be valid.
3.4
Main results
(3.4.1)
that, if F0 = + f0 0 and
||wf0 || + sup e0 t ||wg(t)||
0t
0t
106
Theorem 3. Assume that w2 {1 + |v|}3 L1 in (3.4.1). Assume that the conservation of mass (3.3.9) and energy (3.3.10) are valid for f0 . Then there exists
0t
for some > 0. Moreover, if is strictly convex (3.2.2), and initially f0 (x, v) is
continuous except on 0 , and
f0 (x, v) = f0 (x, v) on R3 \ 0 ,
R3 \ 0 }.
then f (t, x, v) is continuous in [0, ) {
Theorem 4. Assume that w2 {1 + |v|}3 L1 in (3.4.1). Assume that is both
strictly convex (3.2.2) and analytic, and the mass (3.3.9) and energy (3.3.10) are
conserved for f0 . If has any rotational symmetry (3.2.3), we require that the
corresponding angular momentum (3.3.11) is conserved for f0 . Then there exists
0t
0t
f0 (x, v)| = c
f0 (x, v ) (v ){n(x) v } dv ,
nx v >0
R3 \ 0 }.
then f (t, x, v) is continuous in [0, ) {
Our result extends the result in [30] to non-convex domains with a dierent,
more direct approach.
3.5
107
The Velocity Lemma plays the most important role in the study of continuity
for cycles (bouncing generalized trajectories) in the specular case. It states that,
in a strictly convex domain (3.2.2), the singular set 0 cannot be reached via
the trajectory dx/dt = v, dv/dt = 0 from interior points inside , and hence 0
does not really participate or interfere with the interior dynamics. No singularity
would be created from 0 and it is possible to perform calculus for the back-time
exit time tb (x, v). This is the foundation for future regularity study. Moreover,
the Velocity Lemma also provides the lower bound away from the singular set 0 ,
which leads to estimates for repeating bounces in the specular reection case. Such
a Velocity Lemma was rst discovered in [22], [23], in the study of regularity of
the VlasovPoisson (Maxwell) system with at geometry. It was then generalized
in [32] for the VlasovPoisson system in a ball, and it is the starting point for the
construction of regular solutions to the VlasovPoisson system in a general convex
domain [33] with specular boundary condition.
L2 decay theory
3.6
Since no spatial Fourier transform is available, we rst establish linear L2 exponential decay estimates in Section 3.4 via a functional analytical approach. It turns
out that it suces to establish the following nite-time estimate:
0
||Pf (s)||2
ds M
||{I
0
P}f (s)||2
ds + boundary contributions
(3.6.1)
f (0, x, v) = f0 (x, v)
(3.6.2)
with all four boundary conditions (3.3.3), (3.3.4), (3.3.6) and (3.3.8). Here, for any
xed (t, x), the standard projection P onto the hydrodynamic part is given by
(3.6.3)
Pf = {a(t, x) + b(t, x) v + c(t, x)|v|2 } (v),
Pa f = a(t, x) (v), Pb f = b(t, x)v (v), Pc f = c(t, x)|v|2 (v),
and || || is the weighted L2 norm with the collision frequency (v).
Similar types of estimates like (3.6.1), but with strong Sobolev norms, have
been established in recent years [20] via the so-called macroscopic equations for
the coecients a, b and c. The key of the analysis was based on the ellipticity for b
which satises Poissons equation
b = 2 {I P}f,
(3.6.4)
108
3.7
L decay theory
We study linear L decay for all four dierent types of boundary conditions:
inow, bounce-back, specular and diuse (stochastic) reection. In order to control
the nonlinear term (f, f ), we need to estimate the weighted L of wf . We recall
that L = K, and study the L (pointwise) decay of the linear Boltzmann
equation (3.6.2). We choose a weight function
h(t, x, v) = w(v)f (t, x, v),
(3.7.1)
109
where
Kw h = wK
h
,
w
(3.7.2)
(3.7.3)
(3.7.4)
Notice that neither G(t) nor U (t) is a strongly continuous semigroup in L [44].
We rst obtain an explicit representation of G(t) in the presence of various
boundary conditions. Then we can obtain the explicit exponential decay estimate
for G(t). Moreover, we also establish the continuity for G(t) with a forcing term
q if is strictly convex (3.2.2) based on the Velocity Lemma. To study the L
decay for U (t), we make use of the Duhamel Principle:
U (t) = G(t) +
(3.7.5)
Following the pioneering work of Vidav [48] almost 40 years ago, we iterate (3.7.5)
back to get:
t s1
t
G(ts1 )Kw G(s1 s)Kw U (s) ds ds1 ,
U (t) = G(t)+ G(ts1 )Kw G(s1 ) ds1 +
0
0 0
(3.7.6)
where certain compactness property was discovered for the last double integration. Such a compactness is a feature of the so-called A-smoothing operators
introduced in [48], which have many applications in the Boltzmann theory [26].
Moreover, such an A-smoothing property provides an eective way to estimate
the sharp growth rate of a wide class of semigroups, which is important in the
study of nonlinear instability problems [27]. Recently, a similar iteration was employed and new compactness was observed in the so-called Mixture Lemma for
L decay of the Boltzmann equation, either for a whole or a half space [34], [35],
[36]. Our idea here is to estimate the last double integral in terms of the L2 norm
of f = h/w, which decays exponentially by L2 decay theory. The presence of different boundary conditions now leads to complicated bouncing trajectories. Each
of the boundary conditions presents a dierent diculty, as illustrated below.
For the inow boundary condition (3.3.3), the back-time trajectory comes
either from the initial plane or from the boundary. Even though, when g = 0, the
solution operators for (3.7.4) and (3.7.2) are not semigroups, for any (t, x, v) a
110
similar representation as G(t s1 )Kw G(s1 s)Kw U (s) is still possible. With the
compactness property of Kw , the main contribution in (3.7.6) is roughly of the
form
t s1
|h(s, X(s; s1 , X(s1 ; t, x, v), v ), v )| dv dv ds ds1 . (3.7.7)
0
v ,v bounded
(3.7.8)
|h(s, y, v )| dy dv C
,v bounded
|f (s, y, v )|2 dy dv
1/2
.
For bounce-back, specular or diuse reections, the characteristic trajectories repeatedly interact with the boundary. Instead of X(s; t, x, v), we should use
the generalized characteristics, dened as cycles, Xcl (s; t, x, v) in (3.7.7). The key
question is, for any xed (t, x, v), whether or not the change of variable
y Xcl (s; s1 , Xcl (s1 ; t, x, v), v )
is valid, i.e., to determine if it is almost always true that
dXcl (s; s1 , Xcl (s1 ; t, x, v), v )
= 0.
det
dv
(3.7.9)
(3.7.10)
The bounce-back cycles Xcl (s; t, x, v) from a given point (t, x, v) are relatively simple, just going back and forth between two xed points xb (x, v) and
xb (xb (x, v), v). Now the change of variable (3.7.9) and (3.7.10) can be established by the study of the set Sx (v).
The specular cycles Xcl (s; t, x, v) reect repeatedly with the boundary, and
dXcl (s; s1 , Xcl (s1 ; t, x, v), v )/dv is very complicated to compute and (3.7.10) is
extremely dicult to verify, even in a convex domain. This is in part due to
the fact that there is no apparent way to analyze dXcl (s; s1 Xcl (s1 ; t, x, v), v )/dv
inductively with nite bounces. To overcome such a diculty, det(dvk /dv1 ) can be
computed asymptotically in a delicate iterative fashion for special cycles almost
tangential to the boundary, which undergo many small bounces near the boundary.
It then follows that det{dXcl (s; s1 , Xcl (s1 ; t, x, v), v )/dv } = 0 for these special
cycles. This crucial observation is then combined with analyticity of to conclude
that the set of det{dXcl (s; s1 , Xcl (s1 , x, v), v )/dv } = 0 is arbitrarily small, and
the change of variable (3.7.9) is almost always valid. Analyticity plays an important
role in our proof, and it certainly is an interesting open question to remove such
a restriction in the future.
111
The diuse cycles Xcl (s; t, x, v) contain more and more independent variables
and (3.7.7) involves their integrations. A change of variable similar to (3.7.8) is
expected with respect to one of those independent variables. However, the main
diculty in this case is the L control of G(t) which satises (3.7.4). The most
1
natural L estimate for G(t) is for the weight w = 2 , in which the diuse
boundary condition takes the form
h(t, x, v )(v ){v n(x)} dv
h(t, x, v) = c
v n(x)>0
with c v n(x)>0 (v ){v n(x)} dv = 1. However, such a weight makes the linear
Boltzmann theory break down. For any (t, x, v), since there are always particles
moving almost tangential to the boundary in the bounce-back reection, it is
impossible to reach down the initial plane no matter how many cycles the particles
take. In other words, there is no explicit expression for G(t) in terms of initial data
completely. To establish the L estimate for the dierent weight w = {1+2 |v|2 }
in (3.4.1), we make the crucial observation that the measure of those particles that
cannot reach the initial plane after k bounces is small when k is large. We make
use of the freedom parameter in our weight function to control the L norm.
Therefore we can obtain an approximate representation formula for G(t) by the
initial datum, with only a nite number of bounces.
Bibliography
[1] Arkeryd, L., On the strong L1 trend to equilibrium for the Boltzmann equation, Studies in Appl. Math. 87 (1992), 283288.
[2] Arkeryd, L., Cercignani, C., A global existence theorem for the initial boundary value problem for the Boltzmann equation when the boundaries are not
isothermal, Arch. Rational Mech. Anal. 125 (1993), 271288.
[3] Arkeryd, L., Esposito, R., Marra, R., and Nouri, A., Stability of the laminar
solution of the Boltzmann equation for the Benard problem, Bull. Inst. Math.
Academia Sinica 3 (2008), 5197.
[4] Arkeryd, L., Esposito, R., Pulvirenti, M., The Boltzmann equation for weakly
inhomogeneous data, Commun. Math. Phys. 111 (1987), no. 3, 393407.
[5] Alexandre, R., Desvillettes, L., Villani, C., and Wennberg, B., Entropy dissipation and long-range interactions, Arch. Rational. Mech. Anal. 152 (2000),
no. 4, 327355.
[6] Arkeryd, L., Heintz, A., On the solvability and asymptotics of the Boltzmann equation in irregular domains, Comm. Partial Dierential Equations
22 (1997), no. 11-12, 21292152.
[7] Beals, R., Protopopescu, V., Abstract time-depedendent transport equations,
J. Math. Anal. Appl. 212 (1987), 370405.
[8] Cercignani, C., The Boltzmann Equation and Its Applications, SpringerVerlag, 1988.
[9] Cercignani, C., Equilibrium states and the trend to equlibrium in a gas according to the Boltzmann equation, Rend. Mat. Appl. 10 (1990), 7795.
[10] Cercignani, C., On the initial-boundary value problem for the Boltzmann
equation, Arch. Rational Mech. Anal. 116 (1992), 307315.
[11] Cannoe, R., Cercignani, C., A trace theorem in kinetic theory, Appl. Math.
Letters 4 (1991), 6367.
[12] Cercignani, C., Illner, R., and Pulvirenti, M., The Mathematical Theory of
Dilute Gases, Springer-Verlag, 1994.
112
Bibliography
113
114
Bibliography
[31] Hamdache, K., Initial boundary value problems for the Boltzmann equation.
Global existence of weak solutions, Arch. Rational Mech. Anal. 119 (1992),
309353.
[32] Hwang, H.-J., Regularity for the VlasovPoisson system in a convex domain,
SIAM J. Math. Anal. 36 (2004), no. 1, 121171.
[33] Hwang, H.-J., Vel
azquez, J., Global existence for the VlasovPoisson system
in bounded domains, Arch. Rational Mech. Anal. 195 (2010), no. 3, 763796.
[34] Liu, T.-P.; Yu, S.-H., Initial-boundary value problem for one-dimensional
wave solutions of the Boltzmann equation, Comm. Pure Appl. Math. 60
(2007), no. 3, 295356.
[35] Liu, T.-P., Yu, S.-H., Greens function of the Boltzmann equation, 3-D waves,
Bull. Inst. Math. Acad. Sin. (N.S.) 1 (2006), no. 1, 178.
[36] Liu, T.-P., Yu, S.-H., The Greens function and large-time behavior of solutions for the one-dimensional Boltzmann equation, Comm. Pure Appl. Math.
57 (2004), no. 12, 15431608.
[37] Maslova, N. B., Nonlinear Evolution Equations, Kinetic Approach, World
Scientic Publishing, Singapore, 1993.
[38] Mischler, S., On the initial boundary value problem for the VlasovPoisson
Boltzmann system, Commun. Math. Phys. 210 (2000) 447466.
[39] Masmoudi, N., Saint-Raymond, L., From the Boltzmann equation to the
StokesFourier system in a bounded domain, Comm. Pure Appl. Math. 56
(2003), no. 9, 12631293.
[40] Shizuta, Y., On the classical solutions of the Boltzmann equation, Comm.
Pure Appl. Math. 36 (1983), 705754.
[41] Shizuta, Y., Asano, K., Global solutions of the Boltzmann equation in a
bounded convex domain, Proc. Japan Acad. 53A (1977), 35.
[42] Strain, R., Guo, Y., Exponential decay for soft potentials near Maxwellians,
Arch. Rational Mech. Anal. 187 (2008), no. 2, 287339.
[43] Tabachnikov, S., Billiards, SMF, 1995.
[44] Ukai, S., Solutions of the Boltzmann equation. Pattern and Waves Qualitative Analysis of Nonlinear Dierential Equations, Studies of Mathematics
and its Applications vol. 18, Kinokuniya-North-Holland, Tokyo, 1986, 3796.
[45] Ukai, S., private communications.
[46] Ukai, S., Asano, K., On the initial boundary value problem of the linearized
Boltzmann equation in an exterior domain, Proc. Japan Acad. 56 (1980),
1217.
Bibliography
115
[47] Villani, C., Hypocoercivity, Memoirs of the AMS, vol. 202, no. 950, 2009.
[48] Vidav, I., Spectra of perturbed semigroups with applications to transport
theory, J. Math. Anal. Appl. 30 (1970), 264279.
[49] Yang, T., Zhao, H.-J., A half-space problem for the Boltzmann equation with
specular reection boundary condition, Commun. Math. Phys. 255 (2005),
no. 3, 683726.
Chapter 4
The Concentration-Compactness
Rigidity Method for Critical
Dispersive and Wave Equations
Carlos E. Kenig1
4.1
Introduction
In these lectures I will describe a program (which I will call the concentrationcompactness/rigidity method) that Frank Merle and I have been developing to
study critical evolution problems. The issues studied center around global wellposedness and scattering. The method applies to nonlinear dispersive and wave
equations in both defocusing and focusing cases. The method can be divided into
two parts. The rst part (the concentration-compactness part) is in some sense
universal and works in similar ways for all critical problems. The second part
(the rigidity part) has a universal formulation, but needs to be established
individually for each problem. The method is inspired by the elliptic work on the
Yamabe problem and by works of Merle, MartelMerle and MerleRaphael in the
nonlinear Schrodinger equation and generalized KdV equations.
To focus on the issues, let us rst concentrate on the energy critical nonlinear Schr
odinger equation (NLS) and the energy critical nonlinear wave equation
(NLW). We thus have:
(x, t) RN R,
i t u + u |u|4/N 2 u = 0,
(4.1.1)
n
1
u t=0 = u0 H (R ),
N 3,
1 Supported
117
118
and
ut=0 = u0 H 1 (Rn ),
t ut=0 = u1 L2 (Rn ),
(x, t) RN R,
(4.1.2)
N 3.
In both cases, the sign corresponds to the defocusing case, while the +
sign
1
corresponds to the focusing case. For (4.1.1), if u is a solution, so is N 2/2
u x , t2 .
1
For (4.1.2), if u is a solution, so is N 2/2
u x , t . Both scalings leave invariant
the energy spaces H 1 , H 1 L2 respectively, and that is why they are called energy
critical. The energy which is conserved in this problem is
1
1
E (u0 ) =
|u0 |2 ,
|u0 |2
(NLS)
2
2
1
1
1
|u0 |2 ,
|u0 |2 +
|u1 |2
(NLW)
E ((u0 , u1 )) =
2
2
2
2
. The + corresponds to the defocusing case while the
where 21 = 12 N1 = N2N
corresponds to the focusing case.
In both problems, the theory of the local Cauchy problem has been understood for a while (in the case of (4.1.1), through the work of CazenaveWeissler
[7], while in the case of (4.1.2) through the works of Pecher [37], GinibreVelo [14],
GinibreVeloSoer [13], and many others, for instance [3], [20], [34], [41], etc.).
These works show that, say for (4.1.1), for any u0 with u0 H 1 , there exists a
unique solution of (4.1.1) dened for all time and the solution scatters, i.e., there
1
exist u+
0 , u0 in H such that
lim
u(t) eit u
0 H 1 = 0.
t
A corresponding result holds for (4.1.2). Moreover, for any initial data u0 ((u0 , u1 ))
in the energy space, there exist T+ (u0 ), T (u0 ) such that there exists a unique solution in (T (u0 ), T+ (u0 )) and the interval is maximal (for (4.1.2), (T (u0 , u1 ),
T+ (u0 , u1 ))). In both problems, there exists a crucial space-time norm (or Strichartz norm). For (4.1.1), on a time interval I, we dene
||u||S(I) = ||u||L2(N +2)/N 2 Lx2(N +2)/N 2 ,
I
This norm is crucial, say for (4.1.1), because, if T+ (u0 ) < +, we must have
||u||S((0,T+ (u0 ))) = +;
4.1. Introduction
119
120
i t u + u |u|4/N u = 0,
ut=0 = u0 ,
(x, t) RN R,
(4.1.3)
N 3.
Here, ||u0 ||L2 is the critical norm. The analog of Theorem 1 was obtained, for u0
radial, by TaoVisanZhang [50], KillipTaoVisan [28], KillipVisan-Zhang [30],
using our proof scheme for N 2. (In the focusing case one needs to assume
||u0 ||L2 < ||Q||L2 , where Q is the ground state, i.e., the non-negative solution of
the elliptic equation Q + Q1+4/N = Q.) The case N = 1 is open.
Corotational wave maps into S 2 , 4D YangMills in the radial case:
wave map system
2u = A(u)(Du, Du) Tu M
Consider the
121
i t u + u |u|2 u = 0,
(x, t) R3 R,
ut=0 = u0 H 1/2 (R3 ).
|u0 | ,
1
E(u0 ) =
2
1
|u0 |
4
2
|u0 |4 ,
and Q is the ground state, i.e., the positive solution to the elliptic equation
Q + Q + |Q|2 Q = 0,
then if ||u0 ||L2 ||u0 ||L2 > ||Q||L2 ||Q||L2 , we have blow-up in nite time, while
if ||u0 ||L2 ||u0 ||L2 < ||Q||L2 ||Q||L2 , then u exists for all time and scatters. In
joint work with Merle [25] we have considered the defocusing case. We have shown,
using this circle of ideas, that if sup0<t<T+ (u0 ) ||u(t)||H 1/2 < , then T+ (u0 ) = +
and u scatters. We would like to point out that the fact that T+ (u0 ) = + is
analogous to the L3, result of EscauriazaSereginSverak for NavierStokes [11].
4.2
The Schr
odinger equation
i t u + u + |u|4/N 2 u = 0,
(x, t) RN R,
(4.2.1)
ut=0 = u0 H 1 .
Let us start with a quick review of the local Cauchy problem theory. Besides the
norm f S(I) = f L2(N +2)/N 2 Lx2(N +2)/N 2 introduced earlier, we need the norm
I
f W (I) = f L2(N +2)/N 2 L2(N +2)/N 2 +4 .
I
1
< +,
a unique solution to (4.2.1) in R I, with u C(I; H ) and u
W (I)
122
The proof is by xed point. The key ingredients are the following Strichartz
estimates [43], [21]:
eit u0
C u0 H 1 ,
W (,+)
t
C gL2 L2N/N +2 ,
(4.2.2)
t x
W (,+)
t i(tt )
sup
g(, t )dt
C gL2 L2N/N +2
t
0 e
L2
4/N 2
(|u|4/N 2 u)
2 2N/N +2 uS(I) uW (I) .
(4.2.3)
LI Lx
1
1
|u|2 = E(u0 ).
|u(t)|2
E(u(t)) =
2
2
We also have the standard nite-time blow-up criterion: if T+ (u0 ) < , then
||u||S([0,T+ (u0 )) = +.
We next turn to another fundamental result in the local Cauchy theory,
the so-called Perturbation Theorem.
Perturbation Theorem 15 (see [49], [23], [22]). Let I = [0, L), L +, and u
uS(I) M,
uW (I) < +,
on RN I,
where > 0.
and
sup u(t) u
(t)H 1 C(A, A , M )(A + ) ,
tI
123
For the details of the proof, see [22]. This result has several important consequences:
,T
Corollary 1. Let K H 1 be such that K is compact. Then there exist T
+,K
,K
Corollary 2. Let u
0 H 1 ,
u0 H 1 A, and let u
be the solution of (4.2.1),
with maximal interval (T (
u0 ), T+ (
u0 )). Assume that u0,n u
0 in H 1 , with
u0 ) lim T+ (u0,n ), T (
u0 ) lim T (u0,n )
corresponding solution un . Then T+ (
and for t (T (
u0 ), T+ (
u0 )), un (t) u
(t) in H 1 .
Before we start with our sketch of the proof of Theorem 1, we will review the
classic
of Glassey [15] for blow-up in nite time. Thus, assume u0 H 1 ,
2 argument
2
|x| |u0 (x)| dx < and E(u0 ) < 0. Let
I be the maximal interval of existence.
One easily shows that, for t I, y(t) = |x|2 |u(x, t)|2 dx < +. In fact,
2
2
|u(x, t)| |u(x, t)|
.
y (t) = 4 Im u u x, and y (t) = 8
Hence, if E(u0 ) < 0, E(u(t)) = E(u0 ) < 0, so that
1
1
1
|u(t)|2 E(u0 ) < 0,
|u(t)|2 |u(t)|2 = E(u0 ) +
2
2
2
and y (t) < 0. But then, if I is innite, since y(t) > 0 we obtain a contradiction.
We now start with our sketch of the proof of Theorem 1.
Step 1: Variational estimates. (These are not needed in defocusing problems.)
Recall that W (x) = (1+|x|2 /N (N 2))(N 2)/2 is a stationary solution of (4.2.1).
It solves the elliptic equation W + |W |4/N 2 W = 0, W 0, W is radially
decreasing, W H 1 . By the invariances of the equation,
N 2/2
W (0 (x x0 ))
is still a solution. Aubin and Talenti [1], [46] gave the following variational characterization of W : let CN be the best constant in the Sobolev embedding ||u||L2
CN ||u||L2 . Then ||u||L2 = CN ||u||L2 , u 0, if and only if u = W0 ,x0 ,0 for
Hence,
N
, and
|W |2 = 1/CN
E(W ) =
|W |
1
1
2 2
NN2
|W |2 =
1
.
N
N CN
124
Lemma 1. Assume that ||v|| < ||W || and that E(v) (1 0 )E(W ), 0 > 0.
Then there exists = (0 ) so that:
i) ||v||2 (1 )||W ||2 ;
ii)
|v|2 |v|2 ||v||2 ;
iii) E(v) 0.
Proof. Let
f (y) =
1
C 2 2 /2
y
,
y N
2
2
y = ||v||2 .
Note that f (0) = 0, f (y) > 0 for y near 0, y > 0, and that
f (y) =
1 CN
y 2 /21 ,
2
2
2 /2
2
2
|v|2 CN
|v|
2/N 2
2
2
2
= |v| 1 CN
|v|
|v| |v|2
2
(1 )2/N 2
|v|2 1 CN
|W |2
2/N 2
|v|2 1 (1 )2/N 2 ,
ii)
|u(t)|2 |u(t)|2 |u(t)|2 (coercivity);
125
iii) E(u(t)) ||u(t)||2 ||u0 ||2 , with comparability constants which depend
on 0 (uniform bound).
Proof. The statements follow from continuity of the ow, conservation of energy
and Lemma 1.
Note that iii) gives uniform bounds on ||u(t)||. However, this is a long way
from giving Theorem 1.
Remark 3. Let u0 H 1 , E(u0 ) < E(W ), but ||u0 ||2 > ||W ||2 . If we choose 0
0 ) (1 0 )E(W ), we can conclude, as in the proof of Lemma 1, that
so that E(u
|u(t)|2 (1 + ) |W |2 , t I. But then,
2
|u|2
|u(t)|2 |u(t)|2 = 2 E(u0 )
N 2
2 E(W )
=
1
1
2
2
N
N
N 2 CN
N 2 CN
2
< 0.
N
(N 2)CN
Hence, if |x|2 |u0 (x)|2 dx < , Glasseys proof shows that I cannot be innite. If
u0 is radial, u0 L2 , using a local virial identity (which we will see momentarily)
one can see that the same result holds.
Step 2: Concentration-compactness procedure. We now turn to the proof of i) in
Theorem 1. By our variational estimates, if E(u0 ) < E(W ), ||u0 ||2 < ||W ||2 ,
if 0 is chosen so that E(u0 ) (1 0 )E(W ), recall that
E(u(t)) ||u(t)||2 ||u0 ||2 ,
t I, with constants depending only on 0 . Recall also that if ||u0 ||2 < ||W ||2 ,
E(u0 ) 0. It now follows from the local Cauchy theory that if ||u0 ||2 <
||W ||2 and E(u0 ) 0 , 0 small, then I = (, +) and ||u||S(,+) < ,
so that u scatters. Consider now
G = {E : 0 < E < E(W ) :
if ||u0 ||2 < ||W ||2 and E(u0 ) < E, then uS(I) < }
and Ec = sup G. Then 0 < 0 Ec E(W ) and if ||u0 ||2 < ||W ||2 , E(u0 ) <
Ec , I = (, +), u scatters and Ec is optimal with this property. Theorem 1 i)
is the statement Ec = E(W ). We now assume Ec < E(W ) and will reach a
contradiction. We now develop the concentration-compactness argument:
Proposition 1. There exists u0,c H 1 , ||u0,c ||2 < ||W ||2 , with E(u0,c ) = Ec ,
such that, for the corresponding solution uc , we have uc S(I) = +.
126
it
e v0,n
j,n
j ,n
|tj,n tj ,n | |xj,n xj ,n |
+
+
+
,
n
j ,n
j,n
2j,n
j,n
for j = j (we say that {(j,n , xj,n , tj,n )} is orthogonal), such that
i) V0,1 H 1 0 (A) > 0.
ii) If Vjl (x, t) = eit V0,j , then we have, for each J,
v0,n =
J
Vl
N 2/2 j
j=1 j,n
tj,n
x xj,n
, 2
j,n
j,n
+ wnJ ,
where lim
eit wnJ
S(,+) 0, and for each J 1 we have
J
J
j=1
E(v0,n ) =
J
j=1
Vjl
tj,n
2
j,n
+ E(wnJ ) + o(1) as n .
127
yR (t) =
Then yr (t) = 2 Im
uuR , so that
(t)| C
|yR
|u|2
1/2
|u|2
|x|2
1/2
C||W ||2 ,
|x|<R
128
= A + B.
A (t)2 (R(t))2 ||v||2L2 C2 R2 ||W ||2 ,
which tends to 0 as tends to 0.
B (t)2 (R(t))2 ||v||2L2 (|y|R(t)) 0,
tT+ (u0 )
|u|2
|u|2 + |u|2 + 2 .
|x|
|x|>R()
Pick 0 > 0 so that E(u
). Recall that, by our variational esti 0 ) (1 0 )E(W
mates, we have that |u(t)|2 |u(t)|2 C0 ||u0 ||2L2 . If ||u0 ||L2 = 0, using
the smallness of tails, we see that, for R > R0 ,
Choose now C0 radial with (x) = |x|2 for |x| 1, supp {|x| 2}.
Dene
zR (t) = |u(x, t)|2 R2 (x/R) dx.
Similar computations to Glasseys blow-up proof give
(t) = 2R Im
uu(x/R)
zR
129
and
(t) = 4
zR
l,j
1
R2
Re
xl xj (x/R) xl u xj u
2 (x/R)|u|2
4
N
(x/R)|u|2 .
(t)| C0 R2 ||u0 ||2 , by CauchySchwartz, Hardys inequality and
Note that |zR
our variational estimates. On the other hand,
zR
(t)
2
|x|R
|u(t)| |u(t)|
|u|2
|u(t)| + 2 + |u(t)|2
|x|
R|x|2R
C||u0 ||2 ,
for R large. Integrating in t, we obtain zR
(t) zR
(0) Ct||u0 ||2 , but
(t) zR
(0)| 2CR2 ||u0 ||2 ,
|zR
4.3
t u u = |u|
ut=0 = u0 H 1 (Rn ),
N 3.
t ut=0 = u1 L2 (Rn ),
(4.3.1)
(N 2)/2
Recall that W (x) = 1 + |x|2 /N (N 2)
is a static solution that does
not scatter. The general scheme of the proof is similar to the one for Theorem 1.
We start out with a brief review of the local Cauchy problem. We rst consider
130
wt=0 = w0 H 1 (Rn ),
t wt=0 = w1 L2 (RN ).
(4.3.2)
The following are the relevant Strichartz estimates: for an interval I R, let
f S(I) = f L2(N +1)/N 2 L2(N +1)/N 2 ,
x
W (,+)
t
+
t D1/2 w
+ wS(,+)
W (,+)
(4.3.3)
Because of the appearance of D 1/2 in these estimates, we also need to use the
following version of the chain rule for fractional derivatives (see [26]).
Lemma 2. Assume F C 2 , F (0) = F (0) = 0, and that for all a, b we have
|F (a + b)| C {|F (a)| + |F (b)|} and |F (a + b)| C {|F (a)| + |F (b)|}. Then,
for 0 < < 1, 1p = p11 + p12 , p1 = r11 + r12 + r13 , we have
i) D F (u)Lp C F (u)Lp1 D uLp2 ,
ii) D (F (u) F (v))Lp C [F (u)Lp1 + F (v)Lp1 ] D (u v)Lp2
+C [F (u)Lr1 + F (v)Lr1 ] [D uLr2 + D vLr2 ] u vLr3 .
Using (4.3.3) and this lemma, one can now use the same argument as for
(4.2.1) to obtain:
131
(u0 , u1 )H 1 L2 A.
Then, for 3 N 6, there exists = (A) > 0 such that if S(t)(u0 , u1 )S(I)
, 0
I, there exists
a unique
solution
to
(4.3.1) in RN I, with (u, t u)
1/2
1
2
1/2
+
+
+
2
1
0) if
(u+
1 L2
0 , u1 ) H L such that (u(t), t u(t)) S(t)(u0 , u1 ) H
t+
1
1
1
2
2
,
u
))
=
|
+
It turns out that for (4.3.1) there is another very important conserved quantity
in the energy space, namely momentum. This
is crucial for usto be able to treat
non-radial data. This says that, for t I, u(t) t u(t) = u0 u1 . Finally,
the analog of the Perturbation Theorem also holds in this context (see [22]). All
the corollaries of the Perturbation Theorem also hold.
Remark 12 (Finite speed of propagation). Recall that if R(t) is the forward fundamental solution for the linear wave equation, the solution for (4.3.2) is given by
(see [42])
t
R(t)
w
+
R(t)
where stands for convolution in the x variable. The nite speed of propagation
is the statement that supp R( , t), supp t R(
, t) B(0, t). Thus, if supp w0
C
C
C
B(x
,
a),
supp
w
B(x
,
a),
supp
h
[
0
1
0
0ta B(x0 , a t) {t}], then w 0
on 0ta B(x0 , a t) {t}. This has important consequences for solutions of
(4.3.1).
0 , a), then the corresponding solutions agree
If (u0 , u1 ) (u0 , u1 ) on B(x
on 0ta B(x0 , a t) {t} RN (I I ).
We now proceed with the proof of Theorem 2. As in the case of (4.2.1), the
proof is broken up in three steps.
Step1: Variational estimates. Here these are immediate from
the corresponding
ones in (4.2.1). The summary is (we use the notation E(v) = 12 |v|2 21 |v|2 ):
132
ii)
|u(t)|2 |u(t)|2 |u(t)|2 .
iii) E(u(t)) 0 (and here E((u, t u)) 0).
iv) E((u, t u)) (u(t), t u(t))2H 1 L2 (u0 , u1 )2H 1 L2 , with comparability
constants depending only on 0 .
Remark 13. If E((u0 , u1 )) (10 )E((W, 0)), ||u0 ||2 > ||W ||2 , then, for t I,
2
2
u(t) (1 + ) W . This follows from the corresponding result for (4.2.1).
We now turn to the proof of ii) in Theorem 2. We will do it for the case when
u0 L2 < . For the general case, see [24]. We know that, in the situation of ii),
we have
2
|u(t)| (1 + ) |W |2 , t I,
E((W, 0)) E((u(t), t u)) + 0 .
Thus,
1
2
|u(t)|2
so that
|u(t)|2
(t u(t))2 +
1
2
N
|u(t)|2 2 E((W, 0)) + 2 0 .
N 2
Let y(t) = |u(t)|2 , so that y (t) = 2 u(t) t u(t). A simple calculation gives
(t u)2 |u(t)|2 + |u(t)|2 .
y (t) = 2
Thus,
N
N 2
1
2
(t u(t))2 +
2N
(t u)2 2 2 E((W, 0))
N 2
2N
+ 0 +
|u(t)|2 2 |u(t)|2
N 2
4(N 1)
4
2
(t u) +
|u(t)|2
=
N 2
N 2
4
|W |2 + 0
N 2
4(N 1)
(t u)2 + 0 .
N 2
y (t) 2
(t u)2 +
133
If I [0, +) = [0, +), there exists t0 > 0 so that y (t0 ) > 0, y (t) > 0, t > t0 .
For t > t0 we have
N 1
4(N 1)
(t u)2 u2
y (t)2 ,
y(t)y (t)
N 2
N 2
so that
or
y (t)
y (t)
N 1
N 2
y (t)
,
y(t)
u0,c < W ,
E((u0,c , u1,c )) = Ec
and such that for the corresponding solution uc on (4.3.1) we have uc S(I) = +.
Proposition 4. For any uc as in Proposition 3, with (say) uc S(I+ ) = +, I+ =
I [0, +), there exist x(t) RN , (t) R+ , t I+ , such that
xx(t)
1
K = v(x, t) = (t)N12/2 uc xx(t)
,
t
,
u
,
t
:
t
I
t
c
+
(t)
(t)
(t)N/2
has compact closure in H 1 L2 .
Remark 14. As in the case of (4.2.1), in Proposition 4 we can construct (t),
x(t) continuous in [0, T+ ((u0,c , u1,c ))). Moreover, by scaling and compactness of
K, if T+ ((u0,c , u1,c )) < , we have (t) C0 (K)/(T+ ((u0,c , u1,c )) t). Also,
if T+ ((u0,c , u1,c )) = +, we can always nd another (possibly dierent) critical
so that (t)
A > 0, for t [0, T+ ((v0,c , v1,c ))),
element vc , with a corresponding
using the compactness of K. We can also nd g : (0, Ec ] [0, +) decreasing so
2
2
that if u0 < W and E((u0,c , u1,c )) Ec , then uS(,+) g().
Up to here, we have used, in Step 2, only Step 1 and general arguments.
To proceed further we need to use specic features of (4.3.1) to establish further
properties of critical elements.
The rst one is a consequence of the nite speed of propagation and the
compactness of K.
134
|u(t)|2
|u(t)|2 + |t u(t)|2 + |u(t)|2 +
C.
|x|2
|x| 32 M +t
Indeed, choose M C , M 1 for |x| 32 M , with M 0 for
|x| M . Let u0,M = u0 M , u1,M = u1 M . From our assumptions, we have
where is as in the local Cauchy the(u0,M , u1,M )H 1 L2 C. If C0 < ,
ory, the corresponding solution uM of (4.3.1) has maximal interval (, +)
and supt(,+) (uM (t), t uM (t))H 1 L2 2C. But, by nite speed of propagation, uM u for |x| 32 M + t, t [0, T+ (u0 , u1 )), which proves the remark.
We turn to the proof of the lemma. Recall that (t) C0 (K)/(1 t). We
claim that, for any R0 > 0,
|uc (x, t)|2
|uc (x, t)|2 + |t uc (x, t)|2 +
= 0.
lim
t1 |x+x(t)/(t)|R
|x|2
0
, t , t uc xx(t)
,t ,
Indeed, if v (x, t) = (t)1N/2 uc xx(t)
(t)
(t)
|uc (x, t)|2 + |t uc (x, t)|2 =
|v (x, t)|2 dy 0,
|x+x(t)/(t)|R0
|y|(t)R0
t1
|x+x(t)/(t)| 32 R0 +(ts)
135
a contradiction.
Finally, pick tn 1 so that x(tn )/(tn ) x. Observe that, for every 0 > 0,
for n large enough, for all s [0, 1), {|x x| 1 + 0 s} {|x + x(tn )/(tn )|
3
2 R0 + (tn s)}, for some R0 = R0 (0 ) > 0. From this we conclude that
|u(x, s)|2 + |s u(x, s)|2 dx = 0,
|xx0 |1+0 s
wt=0 = w0 H 1 (Rn ),
t wt=0 = w1 L2 (RN ).
Then, for |a| 1/4, we have
x1 at
x1 at
t ax1
t ax1
, t w
,x,
,x,
sup
w
2
2
2
2
2
t
1a
1a
1a
1a
L (dx1 dx )
C w0 H 1 + w1 L2 + hL1 L2 .
t
The simple proof is omitted; see [24] for the details. Note that if u is a solution
(N +2)/N 2 2(N +2)/N 2
of (4.3.1), with maximal interval I and I I, u LI
Lx
,
4
N +2
4/N 2
1 2
u LI Lx . Thus, the conclusion of the lemma
and since N 2 + 1 = N 2 , |u|
x1 at
tax1
applies, provided the integration is restricted to
RN I .
,
x
,
1a2
1a2
Sketch of proof of Theorem 7. Assume
rst that T+ (u0,c , u1,c ) = 1. Assume, to
argue by contradiction, that (say) x1 (u0,c )u1,c = > 0. Recall that, in this
situation, supp uc , t uc B(0, 1 t), 0 < t < 1. For convenience, set u(x, t) =
uc (x, 1 + t), 1 < t < 0, which is supported in B(0, |t|). For 0 < a < 1/4, we
consider the Lorentz transformation
x1 at
t ax1
za (x1 , x , t) = u
, x,
,
1 a2
1 a2
136
and we x our attention on 1/2 t < 0. In that region, the previous lemma
and the following comment show, in conjunction with the support property of u,
that za is a solution in the energy space of (4.3.1). An easy calculation shows that
supp za ( , t) B(0, |t|), so that 0 is the nal time of existence for za . A lengthy
calculation shows that
E((za ( , 1/2), t za ( , 1/2))) E((u0,c , u1,c ))
=
a
and that, for some t0 [1/2, 1/4], |za (t0 )|2 < |W |2 , for a small (by
integration in t0 and a change of variables, together with the variational estimates
for uc ). But, since E((u0,c , u1,c )) = Ec , for a small this contradicts the denition
of Ec , since the nal time of existence of za is nite.
In the case when T+ (u0,c , u1,c ) = +, (t) A0 > 0, the niteness of the
energy of za is unclear, because of the lack of the support property. We instead
do a renormalization. We rst rescale uc and consider, for R large, uR (x, t) =
RN 2/2 uc (Rx, Rt), and for a small,
t ax1
x1 at
, x,
za,R (x1 , x , t) = uR
.
1 a2
1 a2
We assume, as before, that x1 (u0,c )u1,c = > 0 and hope to obtain a contradiction. We prove, by integration in t0 (1, 2), that if h(t0 ) = (x)za,R (x1 , x , t0 ),
with a xed cut-o function, for some a1 small and R large, we have, for some
t0 (1, 2), that
1
E((h(t0 ), t h(t0 ))) < Ec a1
2
and
|h(t0 )|2 < |W |2 .
lim
a0
{|x|<1}
But, by nite speed of propagation, we have that v = za,R on a large set and, after
a change of variables to undo the Lorentz transformation, we reach a contradiction
from these two facts.
From all this we see that, to prove Theorem 2, it suces to show:
Step 3: Rigidity Theorem.
Theorem 8 (Rigidity). Assume that E((u0 , u1 )) < E((W, 0)), |u0 |2 < |W |2 .
Let u be the corresponding solution of (4.3.1), and let I+ = [0, T+ ((u0 , u1 ))). Suppose that:
137
u0 u1 = 0.
a)
r(R) = r(t, R) =
|x|R
|u|2
|x|2
dx.
N
N 2
R u t u =
(t u)2 +
|u|2 |u|2 + O(r(R)).
i) t
2
2
ii) t
R u t u = (t u)2 |u|2 + |u|2 + O(r(R)).
iii) t
R
1
1
1
|u|2 + (t u)2 |u|2
2
2
2
u t u + O(r(R)).
We also have
|u(t)| |u(t)|
and
1
2
(t u(t))2 +
1
2
|u(t)|2 ,
|u(t)|2 |u(t)|2
t>0
CE,
t > 0.
138
The compactness of K and the fact that (t) A0 > 0 show that, given > 0,
we can nd R0 () > 0 so that, for all t > 0, we have
|u|2
|t u|2 + |u|2 + 2 + |u|2 E.
x(t)
|x|
|x+ (t) |R()
The proof of this case is accomplished through two lemmas.
Lemma 7. There exist 1 > 0, C > 0 such that, if 0 < < 1 , if R > 2R0 (),
there
with
exists t0 = t0 (R, )
0 < t0 CR, such that for 0 < t < t0 , we have
x(t)
x(t)
()
and
<
R
R
(t) = R R0 ().
(t)
0
Note that in the radial case, since we can take x(t) 0, a contradiction
follows directly from Lemma 7. This will be the analog of the local virial identity
proof for the corresponding case of (4.2.1). For the non-radial case we also need:
Lemma 8. There exist 2 > 0, R1 () > 0, C0 > 0, so that if R > R1 (), for
0 < < 2 , we have t0 (R, ) C0 R/, where t0 is as in Lemma 7.
From Lemma 7 and Lemma 8 we have, for 0 < < 1 , R > 2R0 (),
t0 (R, ) CR, while for 0 < < 2 , R > R1 (), t0 (R, ) C0 R/. This is
clearly a contradiction for small.
Proof of Lemma 7. Since
x(0) = 0, (0) = 1; if not, we have for all 0 < t < CR,
x(t)
with C large, that (t) < R R0 (). Let
zR (t) =
Then
R u t u +
N
1
2
2
R u t u.
(t) CE/2.
By our variational estimates, we also have
Thus, for small, zR
1
2
(t u)2
1
2
1
1
1
(t u)2 + |u|2 |u|2 .
yR (t) = R
2
2
2
x(t)
For |x| > R, x + (t)
R0 (), so that, since u0 u1 = 0 = u(t) t u(t),
|yR (0)| CR
0 () + R].
139
1
(t u)2 +
|yR (t0 )|
R
x+ x(t0 ) R0 ()
2
(t0 )
1
(t u)2 +
R
x(t
)
x+ 0 >R0 ()
2
(t )
0
1
1
|u|2 |u|2
2
2
1
1
|u|2 |u|2
2
2
.
In the rst integral, |x| R, so that R (x) = x. The second integral is bounded
by M RE. Thus,
1
1
1 2
2
2
|yR (t0 )|
x
(t u) + |u| |u|
M RE.
x+ x(t0 ) R0 ()
2
2
2
(t0 )
1
x(t0 )
1
1
(t u)2 + |u|2 |u|2
x(t0 )
(t0 ) x+ (t
2
2
2
R0 ()
0)
x(t0 )
1
1
1
x+
(t u)2 + |u|2 |u|2
+
x(t0 )
(t
)
2
2
2
0
x+ (t ) R0 ()
u t u,
z(t) = xu t u +
2
2
where we recall that z is well dened since supp u, t u B(0, 1 t). Then, for
0 < t < 1, we have
1
1
(t u)2
|u|2 |u|2 .
z (t) =
2
2
140
0<t<1
and z (t) CE, for 0 < t < 1. From the support properties of u, it is easy to
see that limt1 z(t) = 0, so that, integrating in t, we obtain
z(t) CE(1 t),
0 t < 1.
=
+
.
1 tn
1 tn
2
2
1 tn
Note that, for > 0 given, we have
x + x(tn ) |u(tn )||t u(tn )| + |u(tn )||t u(tn )| CE(1 tn ).
x(tn )
(t
)
n
|x+ (tn ) |(1tn )
Next we will show that |x(tn )/(tn )| 2(1tn ). If not, B(x(tn )/(tn ), (1
tn )) B(0, (1 tn )) = , so that
|u(tn )|2 + |t u(tn )|2 = 0,
B(x(tn )/(tn ),(1tn ))
while
x(tn )
|x+ (t
|(1tn )
n)
|y|(tn )(1tn )
2
u y x(tn ) , tn
(tn )
2
dy
y x(tn )
, tn
+ t u
0,
(tn )
(tn )N n
because of the
of K and the fact that (tn )(1 tn ) . Arguing
compactness
t u
,
using
Hardys
inequality (centered at x(tn )/(tn )), the proof
similarly for u1t
n
is concluded.
141
N (N 2)
1
div (w (y w)y)
w
4
+ |w|4/N 2 w 2y s w (N 1)s w,
142
which is elliptic with smooth coecients for |y| < 1, but degenerates at |y| = 1.
Here are some straightforward bounds on w( ; ) ( > 0): w H01 (B1 ) with
is increasing.
ds dy, so that E
i) E(w(s
2 )) E(w(s1 )) =
3/2
s1
B1 (1 |y|2 )
s2
1
dy
1+N 2
ii)
(s w) w
w
1/2
2 B1
2
(1 |y|2 )
s1
s2
s2
2
1
|w|
=
ds dy
E(w(s))ds
+
N s1 B1 (1 |y|2 )1/2
s1
s2
dy
s ww|y|2
+
(s w)2 + s wy w +
.
2
1/2
1
|y|
s1
B1
(1 |y|2 )
iii)
lim
slog 1/
E for
E(w(s))
= E((u0 , u1 )) = E, so that, by part i), E(w(s))
The proof is computational; see [24]. Our rst improvement over this is:
1
(s w)2
dy ds C log 1/.
Lemma 11.
2
0
B1 1 |y|
143
1
1
(s w)2 +
|w|2 (y w)2
2
2
$
(N 2)N 2 N 2 2 #
w
|w|
log(1 |y|2 ) dy
+
8
2N
$
#
+
log(1 |y|2 ) + 2 y w s w log(1 |y|2 )(s w)2
2 (s w)2 .
d
(s w)2
=
1 |y|2
ds
We next integrate in s, between 0 and 1, and drop the next to last term by
sign. The proof is nished by using CauchySchwartz and the support property of
w( ; ).
|w|2
a)
Corollary 4.
B1
1/2
(1 |y|2 )
dy ds C(log 1/)1/2 .
b) E(w(1))
C(log 1/)1/2 .
Proof. Part a) follows from ii), iii) above, CauchySchwartz and Lemma 11. Note
that we obtain the power 1/2 on the right-hand side by CauchySchwartz. Part b)
follows from i) and the fact that
1
0
E(w(s))
ds C(log 1/)1/2 ,
and a).
which is a consequence of the denition of E
log 1/
(s w)2
Lemma 12.
1
B1
(1 |y|2 )
3/2
C(log 1/)1/2 .
s +(log 1/)1/8
s
B1
(s w)2
(1
3/2
|y|2 )
C
.
(log 1/)1/8
Proof. Split 1, (log 1/)3/4 into disjoint intervals of length (log 1/)1/8 . Their
number is (log 1/)5/8 and 58 18 = 12 .
144
4
(y) = (1 |y|2 )1/2 .
We next point out that w satises the additional (crucial) estimates:
$
#
|w |2 (y w )2
|w |2
+
< .
1/2
1/2
B1 (1 |y|2 )
B1
(1 |y|2 )
Indeed, for the rst estimate it suces to show that, uniformly in j large, we have
s +
j
|w(y, s; j )|2
dy ds C,
1/2
s j
B1 (1 |y|2 )
which follows from ii) above, together with the choice of sj , by Corollary 5,
CauchySchwartz and iii). The proof of the second estimate follows from the rst
2 dr
|w (r, z)| 1/2 dz +
|z w (r, z)|2 1/2 dz < .
r
r
|z|<1
|z|<1
0
0
145
To conclude, we take advantage of the degeneracy of the equation. We desingularize the problem by letting r = a2 , setting v(a, z) = w (a2 , z), so that
a v(a, z) = 2r 1/2 r w (r, z). Our equation becomes:
a2 v + z v + cv + |v|4/N 2 v = 0,
and our bounds give:
1
2
|z v(a, z)| da dz =
|z|<1
da
|a v(a, z)|
dz =
a
|z|<1
2
0 < a < 1,
|z|<1
|z| < 1,
|z w (r, z)|2
|z|<1
v|a=0 = 0,
dr
dz < ,
r 1/2
Bibliography
146
Bibliography
147
[13] J. Ginibre, A. Soer, and G. Velo. The global Cauchy problem for the critical
nonlinear wave equation. J. Funct. Anal., 110(1):96130, 1992.
[14] J. Ginibre and G. Velo. Generalized Strichartz inequalities for the wave equation. J. Funct. Anal., 133(1):5068, 1995.
[15] R. T. Glassey. On the blowing up of solutions to the Cauchy problem for
nonlinear Schrodinger equations. J. Math. Phys., 18(9):17941797, 1977.
[16] M. G. Grillakis. Regularity and asymptotic behaviour of the wave equation
with a critical nonlinearity. Ann. of Math. (2), 132(3):485509, 1990.
[17] M. G. Grillakis. Regularity for the wave equation with a critical nonlinearity.
Comm. Pure Appl. Math., 45(6):749774, 1992.
[18] M. G. Grillakis. On nonlinear Schrodinger equations. Comm. Partial Dierential Equations, 25(9-10):18271844, 2000.
[19] L. H
ormander. The analysis of linear partial dierential operators. III, volume 274 of Grundlehren der Mathematischen Wissenschaften [Fundamental
Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1985. Pseudodierential operators.
[20] L. Kapitanski. Global and unique weak solutions of nonlinear wave equations.
Math. Res. Lett., 1(2):211223, 1994.
[21] M. Keel and T. Tao.
120(5):955980, 1998.
Amer. J. Math.,
[22] C. Kenig. Global well-posedness and scattering for the energy critical focusing non-linear Schrodinger and wave equations. Lecture Notes for a minicourse given at Analyse des equations aux derivees partialles, Evian-lesbains, June 2007.
[23] C. Kenig and F. Merle. Global well-posedness, scattering and blow-up for the
energy-critical, focusing, non-linear Schr
odinger equation in the radial case.
Invent. Math., 166(3):645675, 2006.
[24] C. Kenig and F. Merle. Global well-posedness, scattering and blow-up for the
energy critical focusing non-linear wave equation. Acta Math., 201(2):147
212, 2008.
[25] C. Kenig and F. Merle. Scattering for H 1/2 bounded solutions to the cubic
defocusing NLS in 3 dimensions. Trans. Amer. Math. Soc., 362(4):19371962,
2010.
[26] C. Kenig, G. Ponce, and L. Vega. Well-posedness and scattering results for the
generalized Korteweg-de Vries equation via the contraction principle. Comm.
Pure Appl. Math., 46(4):527620, 1993.
[27] S. Keraani. On the defect of compactness for the Strichartz estimates of the
Schr
odinger equations. J. Dierential Equations, 175(2):353392, 2001.
148
Bibliography
[28] R. Killip, T. Tao, and M. Visan. The cubic nonlinear Schrodinger equation
in two dimensions with radial data. J. Eur. Math. Soc., 11:12031258, 2009.
[29] R. Killip and M. Visan. The focusing energy-critical nonlinear Schrodinger
equation in dimensions ve and higher. Amer. J. Math., 132:361424, 2010.
[30] R. Killip, M. Visan, and X. Zhang. The mass-critical nonlinear Schrodinger
equation with radial data in dimensions three and higher. Analysis and PDE,
1:229266, 2008.
[31] J. Krieger, W. Schlag, and D. Tataru. Renormalization and blow up for
charge and equivariant critical wave maps. Invent. Math., 171(3):543615,
2008.
[32] J. Krieger, W. Schlag, and D. T
ataru. Slow blow-up solutions for the H 1 (R3 )
critical focusing semi-linear wave equation in R3 . Duke Math. J., 147(1):153,
2009.
[33] H. Levine. Instability and nonexistence of global solutions to nonlinear wave
equations of the form P utt = Au + F(u). Trans. Amer. Math. Soc., 192:1
21, 1974.
[34] H. Lindblad and C. Sogge. On existence and scattering with minimal regularity for semilinear wave equations. J. Funct. Anal., 130(2):357426, 1995.
[35] F. Merle and L. Vega. Compactness at blow-up time for L2 solutions of the
critical nonlinear Schrodinger equation in 2D. Internat. Math. Res. Notices,
(8):399425, 1998.
[36] F. Merle and H. Zaag. Determination of the blow-up rate for the semilinear
wave equation. Amer. J. Math., 125(5):11471164, 2003.
[37] H. Pecher. Nonlinear small data scattering for the wave and KleinGordon
equation. Math. Z., 185(2):261270, 1984.
[38] P. Raphael. Existence and stability of a solution blowing up on a sphere for an
L2 -supercritical nonlinear Schrodinger equation. Duke Math. J., 134(2):199
258, 2006.
[39] I. Rodnianski and J. Sterbenz. On the formation of singularities in the critical
O(3) -model. Ann. of Math. (2), 172(1):187242, 2010.
[40] E. Ryckman and M. Visan. Global well-posedness and scattering for the
defocusing energy-critical nonlinear Schr
odinger equation in R1+4 . Amer. J.
Math., 129(1):160, 2007.
[41] J. Shatah and M. Struwe. Well-posedness in the energy space for semilinear
wave equations with critical growth. Internat. Math. Res. Notices, (7):303.,
approx. 7 pp. (electronic), 1994.
Bibliography
149