14634.harmonic Analysis Method For Nonlinear Evolution Equations, I by Baoxiang Wang

p =0
u +∇
u ·∇
u+
-Δ
ut
iut + Δu = F(u)
8209.9789814360739- tp.indd 1 7/1/11 3:37 PM

This page intentionally left blank
0
∇p
=
u+
u·
∇
u+
-Δ
ut
iut + Δu = F(u)
Baoxiang Wang
Peking University, China
Zhaohui Huo
Chinese Academy of Sciences, China
Chengchun Hao
Chinese Academy of Sciences, China
Zihua Guo
Peking University, China
World Scientific
NEW JERSEY • LONDON • SINGAPORE • BEIJING • SHANGHAI • HONG KONG • TA I P E I • CHENNAI
8209.9789814360739- tp.indd 2 7/1/11 3:37 PM

Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
HARMONIC ANALYSIS METHOD FOR NONLINEAR EVOLUTION

EQUATIONS, I
Copyright © 2011 by World Scientific Publishing Co. Pte. Ltd.
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN-13 978-981-4360-73-9
ISBN-10 981-4360-73-2
Printed in Singapore.
ZhangJi - Harmonic analysis Method.pmd 1 6/15/2011, 1:23 PM

June 15, 2011 10:2 World Scientific Book - 9in x 6in booken-main
To our parents
v
Preface
The sole purpose of science is the glory of human spirit. —— C. G.

J. Jacobi
During the past thirty years, the nonlinear evolution equations (NLE)
have made a great progress by using the harmonic analysis techniques. This
book is devoted to introduce the harmonic analysis method for NLE, which
is based on lectures given by B. X. Wang at the Peking University during
the past seven years.
In order to solve an NLE, one needs to choose suitable function spaces as
working spaces. It is well known that most function spaces are established
by using the Lebesgue integrals as basic tools. However, the Lebesgue inte-
gration was not recognized to be of importance in its early stage. As long as
Lebesgue attempted to attend a conference, some mathematicians working
in analysis said to him: “This is not interested for you, we are discussing
the differentiable functions.”, and some experts engaged in geometry told
him: “We are talking the surface which has a tangent plane.” One can
imagine how Lebesgue was hurt inside. New ideas are usually not so easy
to grow up and mathematics brings mathematicians too much sadness and
blessedness, which is hard to express by languages.
However, the authors believe that, at least from the local history point
of view, many important progresses for nonlinear evolution equations have
brought us a series of perpetual surprises over the recent thirty years. The
harmonic analysis techniques of NLE can be gone back to the pioneering
work of R. S. Strichartz in 1977 who discovered the time-space decay of the
solutions of the linear wave equation, so called the Strichartz inequality,
which is now a fundamental tool in the study of nonlinear dispersive equa-
tions. Since 1980s, the NLE, especially the nonlinear dispersive equations
have gained a great development by using the harmonic analysis techniques.
vii
viii Preface
On the other hand, the theory of harmonic analysis is also promoted by

the study of NLE. J. Bourgain in 1993 systematically developed the X s,b
method which is powerful in handling the derivative nonlinearity in nonlin-
ear dispersive equations, where the X s,b space was previously discovered by
J. Rauch, M. Reed and M. Beals for the wave equation. J. Bourgain in 1999
invented the method, so called “separation of the localized energy”, to study
the energy scattering of nonlinear Schrödinger equations. The “I-method”
was introduced by T. Tao’s group to study the global wellposedness of non-
linear dispersive equations for a class of initial data with lower regularities.
C. E. Kenig and F. Merle developed the concentration compactness method
to study the sharp well posedness and scattering for the focusing nonlinear
dispersive equations. Recently, the frequency-uniform decomposition tech-
niques are also applied in the study of NLE. A poem describes the current
status of the harmonic analysis techniques of NLE:
Resonance to a Book1
—Xi Zhu
A square of pool likes an opening mirror
Blue sky and white cloud map in it, freely
wander
Why is the pond crystal and bottomed out
It is from the source of alive water
More precisely, we will study a class of nonlinear dispersive equations,
such as nonlinear Schrödinger equations, nonlinear Klein-Gordon eqautions,
KdV equations, as well as the Navier-Stokes equations and the Boltzmann
equation. As the first book of this series we will mainly study the local
and global wellposedness to the Cauchy problem for those equations. In
Chapter 1 we briefly introduce the theories of various function spaces, say
Besov and Triebel-Lizorkin spaces. In Chapter 2 we study the Navier-
Stokes equation by using the Littlewood-Paley decomposition to establish
some time-space estimates for the linear heat equation. Strichartz type
estimates for a class of linear dispersive equations will be systematically
set up in Chapter 3. Applying those Strichartz’ inequalities, in Chapter 4
we will consider the wellposedness of the solutions of nonlinear Schrödinger
and Klein-Gordon equations. In Chapter 5 we introduce the X s,b -method
and study the KdV and derivative Schrödinger equations. In Chapter 6 we
introduce the frequency-uniform decomposition method. It is known that
1A pool indicates a book.
Preface ix
Morawetz’ estimates are basic tools for the scattering theory of nonlinear
dispersive equations, which will be summarized in Chapter 7 for nonlinear
Schrödinger equations. Finally, in Chapter 8 we introduce some fundamen-
tal results for the Boltzmann equation. Chapters 1, 2, 3, 4 and 6 are written
by B. X. Wang, Chapter 5 is written by Z. H. Guo and Z. H. Huo, Chapter
7 is written by C. C. Hao and Chapter 8 is written by Z. H. Huo. Many
results in Chapters 2, 3, 4, 5 and 6 have been reproved or simplified by the
authors.
The authors would like to genuinely thank Professors Yulin Zhou, Hes-
heng Sun, Boling Guo, Ling Hsiao, Lizhong Peng and Carlos E. Kenig for
their constant supports and they are grateful to Professors Kong Ching
Chang, Weiyue Ding and Gang Tian for their helps. B. X. Wang thanks
Professors Zhouping Xin and Jiecheng Chen for their invitation to give a
series of lectures based on the book at Hong Kong and Hangzhou, respec-
tively. B. X. Wang deeply cherishes the memory of Professor Tingfu Wang
for his rudimental guidance in mathematics. Thanks are also due to Ms Ji
Zhang for her excellent editorial work.
Beijing, May 1, 2011

Contents
Preface vii
s
1. Fourier multiplier, function space Xp,q 1
1.1 Schwartz space, tempered distribution, Fourier transform 1
1.2 Fourier multiplier on Lp . . . . . . . . . . . . . . . . . . . 4
1.3 Dyadic decomposition, Besov and Triebel spaces . . . . . 8
s
1.4 Embeddings on Xp,q . . . . . . . . . . . . . . . . . . . . . 13
s
1.5 Differential-difference norm on Xp,q . . . . . . . . . . . . . 16
s
1.6 Homogeneous space Ẋp,q . . . . . . . . . . . . . . . . . . . 19
1.7 Bessel (Riesz) potential spaces Hps (Ḣps ) . . . . . . . . . . 22
1.8 Fractional Gagliardo-Nirenberg inequalities . . . . . . . . 25
s
1.8.1 GN inequality in Ḃp,q . . . . . . . . . . . . . . . . 25
s
1.8.2 GN inequality in Ḟp,q . . . . . . . . . . . . . . . . 29
2. Navier-Stokes equation 33
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.1.1 Model, energy structure . . . . . . . . . . . . . . . 34
2.1.2 Equivalent form of NS . . . . . . . . . . . . . . . 34
2.1.3 Critical spaces . . . . . . . . . . . . . . . . . . . . 35
2.2 Time-space estimates for the heat semi-group . . . . . . . 35
2.2.1 Lr → Lp estimate for the heat semi-group . . . . 35
2.2.2 Time-space estimates for the heat semi-group . . 36
2.3 Global well-posedness in L2 of NS in 2D . . . . . . . . . . 39
2.4 Well-posedness in Ln of NS in higher dimensions . . . . . 42
2.5 Regularity of solutions for NS . . . . . . . . . . . . . . . 44
s
2.5.1 Gevrey class and function space E2,1 . . . . . . . 44
xi
xii Contents
s
2.5.2 Estimates of heat semi-group in E2,1 . . . . . . . 46
s
2.5.3 Bilinear estimates in E2,1 . . . . . . . . . . . . . . 47
2.5.4 Gevrey regularity of NS equation . . . . . . . . . 48
3. Strichartz estimates for linear dispersive equations 51

0
3.1 Lp → Lp estimates for the dispersive semi-group . . . . . 52
3.2 Strichartz inequalities: dual estimate techniques . . . . . 63
3.3 Strichartz estimates at endpoints . . . . . . . . . . . . . . 68
4. Local and global wellposedness for nonlinear dispersive equations 75

4.1 Why is the Strichartz estimate useful . . . . . . . . . . . . 75
4.2 Nonlinear mapping estimates in Besov spaces . . . . . . . 78
4.3 Critical and subcritical NLS in H s . . . . . . . . . . . . . 83
4.3.1 Critical NLS in H s . . . . . . . . . . . . . . . . . 83
4.3.2 Wellposedness in H s . . . . . . . . . . . . . . . . 84
4.4 Global wellposedness of NLS in L2 and H 1 . . . . . . . . 87
4.5 Critical and subcritical NLKG in H s . . . . . . . . . . . . 88
5. The low regularity theory for the nonlinear dispersive

equations 91
5.1 Bourgain space . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Local smoothing effect and maximal function estimates . 99
5.3 Bilinear estimates for KdV and local well-posedness . . . 104
5.4 Local well-posedness for KdV in H −3/4 . . . . . . . . . . 113
5.5 I-method . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.6 Schrödinger equation with derivative . . . . . . . . . . . . 145
5.7 Some other dispersive equations . . . . . . . . . . . . . . . 152
6. Frequency-uniform decomposition techniques 157

6.1 Why does the frequency-uniform decomposition work . . . 157
6.2 Frequency-uniform decomposition, modulation spaces . . 159
6.2.1 Basic properties on modulation spaces . . . . . . 160
6.3 Inclusions between Besov and modulation spaces . . . . . 164
6.4 NLS and NLKG in modulation spaces . . . . . . . . . . . 176
6.4.1 Schrödinger and Klein-Gordon semigroup in
modulation spaces . . . . . . . . . . . . . . . . . . 176
6.4.2 Strichartz estimates in modulation spaces . . . . . 179
6.4.3 Wellposedness for NLS and NLKG . . . . . . . . . 183
Contents xiii
6.5 Derivative nonlinear Schrödinger equations . . . . . . . . 187

6.5.1 Global linear estimates . . . . . . . . . . . . . . . 191
6.5.2 Frequency-localized linear estimates . . . . . . . 192
6.5.3 Proof of global wellposedness for small rough data 197
7. Conservations, Morawetz’ estimates of nonlinear

Schrödinger equations 205
7.1 Nöther’s theorem . . . . . . . . . . . . . . . . . . . . . . . 205
7.2 Invariance and conservation law . . . . . . . . . . . . . . . 212
7.3 Virial identity and Morawetz inequality . . . . . . . . . . 214
7.4 Morawetz’ interaction inequality . . . . . . . . . . . . . . 220
7.5 Scattering results for NLS . . . . . . . . . . . . . . . . . . 222
8. Boltzmann equation without angular cutoff 227

8.1 Models for collisions in kinetic theory . . . . . . . . . . . 228
8.1.1 Transport model . . . . . . . . . . . . . . . . . . . 228
8.1.2 Boltzmann model . . . . . . . . . . . . . . . . . . 229
8.1.3 Cross section . . . . . . . . . . . . . . . . . . . . . 233
8.2 Basic surgery tools for the Boltzmann operator . . . . . . 235
8.3 Properties of Boltzmann collision operator without cutoff 240
8.4 Regularity of solutions for spatially homogeneous case . . 246
Appendix A Notations 259
Appendix B Definition of scattering operator 261
Appendix C Some fundamental results 263

C.1 Gagliardo-Nirenberg inequality in Sobolev
spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
C.2 Convexity Hölder inequality in sequence spaces `sp . . . . 263
C.3 Inclusion between homogeneous Triebel-Lizorkin spaces . 264
C.4 Riesz-Thorin interpolation theorem . . . . . . . . . . . . . 265
C.5 Hardy-Littlewood-Sobolev inequality . . . . . . . . . . . . 266
C.6 Van der Corput lemma . . . . . . . . . . . . . . . . . . . . 266
C.7 Littlewood-Paley square function theorem . . . . . . . . . 266
C.8 Complex interpolation in modulation spaces . . . . . . . . 267
C.9 Christ-Kiselev lemma . . . . . . . . . . . . . . . . . . . . 267
xiv Contents
Bibliography 269
Index 281
Chapter 1
s
Fourier multiplier, function space Xp,q
The intensive study to the nature is the most plentiful source for the
mathematical discovery. —— J. Fourier1
In this chapter we begin by considering the Fourier transform on the

Schwartz space S and its dual space S 0 , the material on Fourier transforms
is standard and our treatment here will be sketched, see Stein [202], Yosida
[256] for instance. On the basis of the Fourier transform, we introduce
s s
the Fourier multiplier space Mp , Besov spaces Bp,q and Triebel spaces Fp,q
and then discuss their elementary properties. The embeddings between
these function spaces are very useful in the study of partial differential
equations. For the convenience to the readers, we give a self-contained
treatment to the theory of function spaces. For some further results on
function spaces, one can refer to [13; 224]. Some notations are well-known
for readers and not stated in the text, which will be listed in the Appendix
A. For convenience to readers, we would like to remind some important
relations between Triebel-Lizorkin, Bessel potential spaces Hps (H s = H2s )
and Sobolev spaces Wpm with m ∈ Z+ :
s
Fp,2 = Hps , Hpm = Wpm , F2,2
s
= H s , ∀ s ∈ R, m ∈ Z+ , 1 < p < ∞.
1.1 Schwartz space, tempered distribution, Fourier trans-

form
Let α = (α1 , ..., αn ) be a multi-index, denote

Dα = ∂xα11 ...∂xαnn , ∂xαii = ∂ αi /∂xα
i , |α| = α1 + ... + αn ,
i
xα = xα αn
1 ...xn .
1
1 Joseph Fourier (1768–1830), a French mathematician and physicist, is best known for
initiating the investigation of Fourier series and their application to problems of the heat
transfer. He is also generally credited with the discovery of the greenhouse effect. He is
one of the authority in French analysis school.
1
2 s
We denote by C ∞ (Rn ) the set of all infinitely differentiable functions de-

fined on Rn . Denote
S = {φ ∈ C ∞ (Rn ) : pk (φ) < ∞},
X
pk (φ) = sup (1 + |x|2 )k/2 |Dα φ(x)|. (1.1)
x∈Rn
|α|6k
It is easy to see that pk (·) is a norm on S and so, it is a semi-norm on

S . From the theory of the linear topology, S generates a locally convex
linear topological space according to the semi-norm system {pk }∞ k=0 , which
is said to be the Schwartz space, any function in S is said to be a Schwartz
function. The base of zero neighborhoods of S is
Bk,ε = {φ ∈ S : pk (φ) < ε}, k ∈ Z+ = {0, 1, 2, ...}, ε > 0.
We denote by S 0 := S 0 (Rn ) the dual space of S and it is a locally convex
linear topological space, which is said to be a tempered distribution space.
The base of zero neighborhoods of S 0 is
( )
UB,ε = f ∈ S : sup |f (φ)| < ε ,
φ∈B
where B is a bounded set in S . Let φ, ψ ∈ S , the Fourier (inverse)

transform is defined as follows.
Z
b
φ(ξ) = (F φ)(ξ) = (2π) −n/2
e−ix·ξ φ(x)dx,
Rn
Z
ψ̌(x) = (F −1 ψ)(x) = (2π)−n/2 eix·ξ ψ(ξ)dξ, (1.2)
Rn
where x · ξ = x1 ξ1 + ... + xn ξn (if there is no confusion, we will write
x·ξ = xξ). S plays a crucial role in the theory of the Fourier analysis, which
is compatible with the Fourier transform. In fact, we have the following
Proposition 1.1. Let φ ∈ S . Then we have

d
D b
α φ(ξ) = i|α| ξ α φ(ξ), b = (−i)|α| xd
Dα φ(ξ) α φ(ξ),
F −1 (F φ) = φ.
Moreover, F (F −1 ) : S → S is a continuously linear bijection, i.e., an
isomorphism.
Proposition 1.2. Let φ, ψ ∈ S . Then

Z Z
b
φ(x)ψ(x)dx = b
φ(x)ψ(x)dx.
Rn Rn
1.1. Schwartz space, tempered distribution, Fourier transform 3
We define the convolution φ ∗ ψ on S in the following way

Z
φ ∗ ψ(x) = φ(x − y)ψ(y)dy.
Rn
Proposition 1.3. Let φ, ψ ∈ S . Then

φ[
∗ ψ = (2π)n/2 φb · ψ,
b φd
· ψ = (2π)−n/2 φb ∗ ψ.
b
We denote by σλ and τh the dilation and the translation operators:

σλ φ = φ(λ·), λ ∈ R \ {0},
τh φ = φ(· − h), h ∈ Rn .
Proposition 1.4. Let φ ∈ S , λ ∈ R \ {0}, h ∈ Rn . Then

τd
hφ = e
−ihξ b
φ,
e[ b
ihx φ = τ φ,
h
σd
λ φ = |λ|
−n b
(σ1/λ φ).
For the proofs of the above propositions, except that the proof of
F −1 φb = φ needs a special technique, say
e\
2
−|x|2 /2 = e−|ξ| /2 ,
the other proofs are easily obtained.

In the following we consider the Fourier transform of f ∈ S 0 . Inspired
by Proposition 1.2, we define fb as follows.
(F f )(φ) = fb(φ) = f (φ),
b φ ∈ S,
−1 ˇ
(F f )(φ) = f (φ) = f (φ̌), φ ∈ S .
Proposition 1.5. F : S 0 → S 0 is a continuously linear bijection map,

i.e., an isomorphism.
Now we can define the convolution of f ∈ S 0 and φ ∈ S . Let us recall

that for any f, φ, ψ ∈ S , we have
Z Z
(f ∗ φ)(x)ψ(x)dx = f (x)(φ̃ ∗ ψ)(x)dx,
Rn Rn
where φ̃ = φ(−·). Following this idea, we define

(f ∗ φ)(ψ) = f (φ̃ ∗ ψ), f ∈ S 0 , φ, ψ ∈ S .
Taking notice of the definition of the Fourier transform and the convo-
lution operator on S 0 , we can generalize the other operators to S 0 . We
4 s
first investigate the corresponding operation for Schwartz functions, which

can be converted to an integral operation. Then taking into account the
abstract form of this integration operation, we can generalize the operation
to S 0 . Let f ∈ S 0 , we define
Dα f (φ) := f ((−D)α φ),
τh f (φ) := f (τ−h φ), h ∈ Rn ,
σλ f (φ) := f (λ−n σ1/λ φ), λ > 0.
0
It is known that f ∈ S does not mean that f is a function, the following
result indicates that the convolution of f and a Schwartz function must be
an infinitely differentiable function.
Proposition 1.6. Let f ∈ S 0 and φ ∈ S . Then f ∗ φ is an infinitely

differentiable function and
(f ∗ φ)(x) = f (τx φ̃).
Moreover, f ∗ φ grows at most in a polynomial way, i.e., for any α ∈ Zn+
there exists a polynomial Pα , such that |Dα (f ∗ φ)(x)| 6 Pα (x), x ∈ Rn .
Proposition 1.7 (Paley-Wiener-Schwartz theorem).

(1) φ ∈ S and supp φb ⊂ {ξ : |ξ| 6 b} if and only if, φ(z) (z = x + iy)
is an entire analytic function of n complex variables and for any
ε > 0, λ > 0, there exists Cε,λ > 0, such that
|φ(z)| 6 Cε,λ (1 + |x|)−λ e(b+ε)|y| , x, y ∈ Rn .
(2) f ∈ S 0 and supp fb ⊂ {ξ : |ξ| 6 b} if and only if, f (z) (z = x + iy)
is an entire analytic function of n complex variables and for some
λ ∈ R and for any ε > 0, there exists Cε > 0 such that
|f (z)| 6 Cε (1 + |x|)λ e(b+ε)|y| , x, y ∈ Rn .
Proposition 1.7 is a fundamental result in the framework of the Fourier

analysis, which says that if the Fourier transform of a tempered distribution
f has a compact support, then it is an analytic function.
1.2 Fourier multiplier on Lp
We denote by Lp := Lp (Rn ) the Lebesgue space, 1 6 p 6 ∞, the norm on

Lp is written as k · kp . Let us consider the linear heat equation in Lp ,
ut − 4u = 0, u|t=0 = u0 . (1.3)
1.2. Fourier multiplier on Lp 5
2
It is easy to see that u = F −1 e−t|ξ| F u0 =: H(t)u0 is the solution of (1.3).
For any u0 ∈ Lp , we want to know if H(t)u0 also belongs to Lp . More
precisely, does
2
kF −1 e−t|ξ| F u0 kp . ku0 kp , 16p6∞ (1.4)
hold for all u0 ∈ Lp ? The answer is affirmative (see Chapter 2). One can
generalize (1.4) to the following notion.
Definition 1.1. Let ρ ∈ S 0 . If there exists C > 0 such that2
kF −1 ρF f kp 6 Ckf kp , ∀ f ∈ S, (1.5)
then ρ is said to be a multiplier on Lp . The set of all multipliers on Lp is

denoted by Mp , for which the norm is given by
kρkMp = sup{kF −1 ρF f kp : f ∈ S , kf kp = 1}. (1.6)
If ρ ∈ Mp , 1 6 p < ∞, since S is dense in Lp , one can extend F −1 ρF

into a bounded operator on Lp and the norm of the extension operator is
preserved. The extension operator is still written as F −1 ρF or (F −1 ρ)∗.
The multiplier theory is a fundamental tool in the harmonic analysis,
and very useful in the definition of various function spaces. In what follows
we discuss some basic properties of the multiplier space Mp .
Proposition 1.8. Let 1 6 p 6 q 6 2. Then we have
(1) Mp = Mp0 , k · kMp = k · kMp0 ;

(2) Mp ⊂ Mq , k · kMq 6 k · kMp ;
(3) M2 = L∞ , k · kM2 = k · k∞ ;
(4) M1 = {ρ ∈ S 0 : F −1 ρ is a bounded measure}, kρkM1 equals the
total variation of F −1 ρ.
Proof. First, we prove (1). Let ρ ∈ Mp . For any f, g ∈ S , kf kp =

kgkp0 = 1, we have3
|(F −1 ρF f, g)| 6 kF −1 ρF f kp kgkp0 6 kρkMp . (1.7)
Noticing that (F −1 ρF f, g) = (F −1 ρ ∗ f ∗ e
g )(0), we see that (1.7) implies
0
F −1 ρF g ∈ Lp and ρ ∈ Mp0 with kρkMp0 6 kρkMp . It follows that (1)
holds true.
2 Notice that F −1 ρF f = (F −1 ρ) ∗ f , in view of Proposition 1.6, we see that F −1 ρF f
is a smoothRfunction for any f ∈ S .
3 (f, g) = f (x)g(x)dx.
6 s
Secondly, we show the result of (2). Due to 1 6 p 6 q 6 2, we see that

p0 > 2. So, one can choose θ ∈ [0, 1] satisfying 1/q = (1 − θ)/p + θ/p0 . If
ρ ∈ Mp , in view of (1) we have ρ ∈ Mp0 . Namely,
0 0
F −1 ρF : Lp → Lp ; F −1 ρF : Lp → Lp (1.8)
are both bounded operators. It follows from Riesz-Thorin’s interpolation
theorem (see Appendix) that F −1 ρF : Lq → Lq is bounded and
kρkMq 6 kρk1−θ θ
Mp kρkMp0 = kρkMp . (1.9)
Thirdly, we prove (3). If ρ ∈ M2 , by Plancherel’s identity,
kF −1 ρF f k2 = kρF f k2 6 kρk∞ kf k2 . (1.10)
So, kρkM2 6 kρk∞ . Conversely, for any ε > 0, one can take a non-zero
measurable closed subset E of Rn , satisfying |ρ(ξ)| > kρk∞ − ε in E. Let
f ∈ L2 satisfy supp F f ⊂ E. It follows that kρkM2 > kρk∞ − ε.
Finally, we show that (4) holds. By the translation property of F −1 ρF ,
we easily see that ρ ∈ M∞ if and only if
|(F −1 ρ ∗ f )(0)| = (2π)n/2 |(F −1 ρF f )(0)| 6 Ckf k∞ , ∀ f ∈ S . (1.11)
Noticing that S is dense in C0 (Rn ), from (1.11) and Proposition 1.6 one
sees that F −1 ρ is a continuous functional on C0 (Rn )4 . From the construc-
tion of C0 (Rn )∗ , one has the result.
Proposition 1.9. Mp is a Banach algebra.
Proof. Obviously, k · kMp is a norm. By (3) in Proposition 1.8, one sees

that Mp ⊂ L∞ . If {ρk } is a Cauchy sequence in Mp , so does in L∞ . We
may assume, without loss of generality that it converges to ρ in L∞ . Since
L∞ ⊂ S 0 , we conclude that for any f ∈ S , F −1 ρk F f → F −1 ρF f
according to the strong topology on S 0 . Recalling that F −1 ρk F f is a
Cauchy sequence in Lp ⊂ S 0 , it has a limit ponit g. By the uniqueness of
the limit in S 0 one obtains that g = F −1 ρF f . So, kρk − ρkMp → 0 as
k → ∞. It follows that Mp is a Banach space. Let ρ1 , ρ2 ∈ Mp . For any
f ∈ S , we have
kF −1 ρ1 ρ2 F f kp 6 kρ1 kMp kF −1 ρ2 F f kp 6 kρ1 kMp kρ2 kMp kf kp , (1.12)
which implies that ρ1 ρ2 ∈ Mp and
kρ1 ρ2 kMp 6 kρ1 kMp kρ2 kMp . (1.13)
Hence, Mp is a Banach algebra.
4 We denote by C0 (Rn ) the space of all continuous functions f (x) that vanish as x → ∞.
1.2. Fourier multiplier on Lp 7
In order to emphasize that Mp depends on Rn , we write Mp = Mp (Rn ).

The following proposition indicates that Mp (Rn ) is isometrically invariant
under affine transforms5 of Rn .
Proposition 1.10. Let a : Rn → Rm (n > m) be a surjective affine

transform, and ρ ∈ Mp (Rm ). Then
kρ(a(·))kMp (Rn ) = kρkMp (Rm ) . (1.14)
In particular,
kρ(c ·)kMp (Rn ) = kρkMp (Rn ) , c 6= 0; (1.15)
kρ(hx, ·i)kMp (Rn ) = kρkMp (R) , x 6= 0, (1.16)
Pn
where hx, ξi = i=1 xi ξi .
Proof. It suffices to consider the case that a : Rn → Rm is a linear

transform. Make the coordinate transformation
ηi = ai (ξ), 1 6 i 6 m; ηj = ξj , m + 1 6 j 6 n, (1.17)
−1
which is written as η = A ξ or ξ = Aη. Let A∗ be the transposed matrix
of A. It is easy to see, for any f ∈ S , that

F −1 ρ(a(ξ))F f = F −1 ρ(η1 , ..., ηm )F f (A−1
∗ ·) (A∗ ·). (1.18)
It follows from ρ ∈ Mp (Rm ) that for any f ∈ S
kF −1 ρ(a(·))F f kLp (Rn ) = |A|−1 kF −1 ρF f (A−1
∗ ·)kLp (Rn )
6 kρkMp (Rm ) kf kLp(Rn ) . (1.19)
Thus, we have
kρ(a(·))kMp (Rn ) 6 kρkMp (Rm ) . (1.20)
Taking f (A−1
∗ ·) = f1 (x1 , ..., xm )f2 (xm+1 , ..., xn ), one can conclude that the
inverse inequality of (1.20) also holds.
In the following we give some criteria how to determine the multiplier
in Lp .
Proposition 1.11 (Bernstein multiplier theorem). Let L > n/2 be

an integer, ∂xαi ρ ∈ L2 , i = 1, ..., n and 0 6 α 6 L. Then we have ρ ∈ Mp ,
1 6 p 6 ∞ and
n
!n/2L
1−n/2L
X L
kρkMp . kρk ∂x ρ . (1.21)
2 i 2
i=1
5 An affine transform of Rn is a map F : Rn → Rn of the form F (p) = Ap + q for all
p ∈ Rn , where A is a linear transform of Rn .

8 s
Proof. By Proposition 1.8, it suffices to consider the case p = 1. Obvi-

ously,
Z
kρkM1 6 |F −1 ρ(x)|dx. (1.22)
Rn
For any t > 0,
Z
|F −1 ρ(x)|dx . tn/2 kρk2 . (1.23)
|x|<t
Pn
Denote J(x) = i=1|xi |L . One has that
Z Z

|F −1 ρ(x)|dx = J(x)−1 J(x) F −1 ρ(x) dx
|x|>t |x|>t
n
X L
. tn/2−L ∂x ρ . (1.24)
i 2
i=1
Pn
Taking t such that kρk2 = t−L i=1 k∂xLi ρk2 , from (1.22)–(1.24) we have
the result, as desired.
Proposition 1.12 (Mihlin multiplier theorem). Let L > n/2 be an

integer and ρ ∈ L∞ satisfy
|ξ||α| |Dα ρ(ξ)| 6 A, |α| 6 L, ξ ∈ Rn \ {0}. (1.25)
Then ρ ∈ Mp for any 1 < p < ∞ and
kρkMp 6 Cp A, (1.26)
where Cp may depend on p.
Proof. See [13].
1.3 Dyadic decomposition, Besov and Triebel spaces

Ingenious ideas are usually beautiful and resonating ideas, the dyadic
decomposition is one of them.
We consider the decomposition on Rn ,

R0 = {ξ : |ξ| < 1}, Rk = {ξ : 2k−1 6 |ξ| < 2k }, k ∈ N. (1.27)
It is easy to see that {Rk }∞
k=0 is a pairwise disjoint sequenceand Rn =
∪∞ R
k=0 k , namely, {R } ∞
k k=0 constitutes a decomposition of Rn .According
to this decomposition, we can roughly define the Littlewood-Paley decom-
position operators in the following way
4k ∼ F −1 χRk F , k ∈ N ∪ {0}, (1.28)
1.3. Dyadic decomposition, Besov and Triebel spaces 9
where χRk denotes the characteristic function on Rk . The dyadic decom-

position operator is a frequency-localized operator, which is one of the most
delicate and clever ideas in harmonic analysis. If 4k takes the form as in
(1.28), it has an advantage that 4k ’s (k > 0) orthogonalize each other, but
the characteristic function in (1.28) has no smoothness and it is hard to
use. In order to make multiplier calculations, we need to adopt the smooth
version of (1.28).
Let ψ : Rn → [0, 1] be a smooth radial cut-off function, say

 1, |ξ| 6 1,
ψ(ξ) = smooth, 1 < |ξ| < 2, (1.29)

0, |ξ| > 2.
Denote
ϕ(ξ) = ψ(ξ) − ψ(2ξ), (1.30)
and we introduce the function sequence {ϕk }∞
k=0 :

ϕk (ξ) = ϕ(2−k ξ), k ∈ N,
P∞ (1.31)
ϕ0 (ξ) = 1 − k=1 ϕk (ξ) = ψ(ξ).
Since supp ϕ ⊂ {ξ : 2−1 6 |ξ| 6 2}, we easily see that supp ϕk ⊂ {ξ :
2k−1 6 |ξ| 6 2k+1 }, k ∈ N, supp ϕ0 ⊂ {ξ : |ξ| 6 2}. Define
4k = F −1 ϕk F , k ∈ N ∪ {0}, (1.32)
{4k }∞
k=0 is said to be the Littlewood-Paley (or dyadic) decomposition op-
erator. Formally, we find that6
∞
X
4k = I. (1.33)
k=0
Combining the dyadic decomposition operator with the function spaces7

`q (Lp ) and Lp (`q ), we can introduce Besov spaces Bp,q
s
and Triebel-Lizorkin
s
spaces Fp,q , respectively. Let
−∞ < s < ∞, 1 6 p, q 6 ∞. (1.34)
6 Actually, this decomposition works just for any locally integrable function which has
some decay at the infinity, and one usually has all the convergence properties of the
summation that one needs. In many applications, one can make the a priori assumption
that f is Schwartz, in which case the convergence is uniform. However, if the function
does not decay, this formula fails. For instance, if f ≡ 1, then all 4k f ’s vanish because
4k 1 = ϕk (0) = ϕ(0) = 0.
P q 1/q P
7 k(a )k q p :=
k ` (L ) k kak (x)kp , k(ak )kLp (`q ) := ( k |ak (x)|q )1/q p .
10 s
We define

s
Bp,q = f ∈ S 0 (Rn ) : kf kBp,q
s <∞ , (1.35)
X
∞ 1/q
kf kBp,q
s := 2ksq k4k f kqp , (1.36)
k=0
s
Bp,q is said to be the Besov space8 . Assume that
−∞ < s < ∞, 1 6 p < ∞, 1 6 q 6 ∞, (1.37)
we define the following

s 0 n
Fp,q = f ∈ S (R ) : kf kFp,q
s <∞ , (1.38)
X 1/q
∞ ksq
kf kFp,q
s :=
2 |4 k f | q ,
(1.39)
k=0 p
s
Fp,q is said to be the Triebel-Lizorkin space9 . Note that we need replace
the `q -norm by the `∞ -norm in the above definition if q = ∞.
Besov and Triebel-Lizorkin spaces had been formulated during 1960s–
1980s, which have been widely applied in recent years. Roughly speaking,
s is the spatial regularity index; 4k f is the frequency localization of f at
the frequency |ξ| ∼ 2k ; the norms of the function sequence {2sk |4k f |} in
`q (Lp ) and Lp (`q ) generate Besov and Triebel norms of f , respectively.
The dyadic decomposition in the frequency space goes back to the
equivalent norm on Lp (Rn )
kf kp ∼ kf kFp,2
0 , 1 < p < ∞. (1.40)
(1.40) is the well known Littlewood-Paley square function theorem (see

Appendix).
If there is no explanation, we will assume that conditions (1.34) and
s s
(1.37) are satisfied for Besov and Triebel spaces Bp,q and Fp,q , respectively.
s s s 10
For simplicity, we will use Xp,q to denote Bp,q or Fp,q .
s s s s
Proposition 1.13. Let Xp,q = Bp,q (Xp,q = Fp,q ). There hold the follow-
ing inclusions.
8 In the definition of B s , one can also consider the case 0 < p ∧ q < 1, see Triebel
p,q
[224].
9F s
∞,q can be found in Triebel [224].
10 When X appears in the both sides of an inclusion, X should be the same, namely
X = B in both sides, or X = F in both sides.

1.3. Dyadic decomposition, Besov and Triebel spaces 11
(1) Let q1 6 q2 . Then

s s
Xp,q1
⊂ Xp,q2
. (1.41)
(2) Let ε > 0 and 1 6 q1 , q2 6 ∞. Then
s+ε s
Xp,q1
⊂ Xp,q2
. (1.42)
(3) Let p < ∞. Then
s s s
Bp,p∧q ⊂ Fp,q ⊂ Bp,p∨q , (1.43)
where p ∧ q = min(p, q) and p ∨ q = max(p, q).
Proof. Since `p ⊂ `p+a , a > 0, we can easily get the result of (1). Let
us observe that
∞
!1/q2
X
2skq2 |ak |q2 . sup 2(s+ε)k |ak |. (1.44)
k>0
k=0
Taking ak = k4k f kp or ak = |4k f |, we can show that (2) holds with the
help of (1). Finally, we prove (3). Denote bk = 2sk 4k f . We divide the
proof into the following two cases.
Case 1. q 6 p. In view of `q ⊂ `p and Minkowski’s inequality,11 we have
kbk k`p (Lp ) 6 kbk kLp (`q ) 6 kbk k`q (Lp ) , (1.45)
which implies the result.
Case 2. q > p. Analogous to Case 1, we can use Minkowski’s inequality
and `p ⊂ `q to get the result, as desired.
s s s
Proposition 1.14. Let Xp,q be Bp,q or Fp,q . Then
s
(1) Xp,q is a Banach space;
n s
(2) S (R ) ⊂ Xp,q ⊂ S 0 (Rn );
(3) If 1 6 p, q < ∞, then S (Rn ) is dense in Xp,q
s
.
Proof. Since `q (Lp ) and Lp (`q ) are normed spaces, one can conclude that
s
Xp,q is a normed space. In order to show the result of (1), it suffices to
s
prove that Xp,q is complete, whose proof will be left to the end. The proof
of (2) is separated into the following four steps.
s
Step 1. We show that S ⊂ Bp,∞ . In fact, for sufficiently large
L, M, N ∈ N,
kf kBp,∞
s = sup 2sk k4k f kp
k>0
11 Minkowski’s inequalities
P P∞ read
i) kP ∞
j=0 fj kp 6 Pj=0 kfj kp , for any p ∈ [1, ∞];
ii) ∞j=0 kfj kp 6 k
∞
j=0 fj kp , for any p ∈ (0, 1) and fj > 0.
12 s
. sup 2sk k(1 + |x|2 )L 4k f k∞

k>0
. sup 2sk k(I − 4)L ϕk F f k1

k>0
. k(1 + | · |2 )M (I − 4)L F f k∞
. pN (F f ). (1.46)
Since F : S → S is a continuous map, from (1.46) it follows the result,
as desired.
s s+ε
Step 2. We prove that S ⊂ Xp,q . Since S ⊂ Bp,∞ , by Proposition
s+ε s s s
1.13, we see that Bp,∞ ⊂ Bp,p∧q ⊂ Bp,q ∩ Fp,q . Thus, we have the desired
result.
s
Step 3. We show that Bp,∞ ⊂ S 0 . For convenience, we denote ϕ−1 ≡ 0.
From the construction of ϕk , one has that ϕk ϕk+` ≡ 0 for ` 6= −1, 0, 1. For
s
any f ∈ Bp,∞ , ψ ∈ S , we can take N ∈ N sufficiently large,
∞ X
X 1
|hf, ψi| 6 |h4k f, F ϕk+` F −1 ψi|
k=0 `=−1
∞ X
X 1
. k4k f kp k4k+` ψkp0
k=0 `=−1
∞
X
. kf kBp,∞
s 2−sk k4k ψkp0
k=0
. kf kBp,∞
s kψkB −s+ε
0 p ,∞
. kf kBp,∞
s pN (F ψ). (1.47)
For any bounded set B in S , we have pN (F ψ) . 1 for ψ ∈ B. The result
follows.
s
Step 4. Similarly as in Step 2, we can show that Xp,q ⊂ S 0 and the
details of the proof are omitted.
s
Finally, we prove the completeness of Bp,q . In an analogous way we can
get that Fp,q is also complete. Assume that {f` }∞
s
`=1 is a Cauchy sequence
in Bp,q . By (2) we see that it is also a Cauchy sequence in S 0 . Since
s
S 0 is a complete and locally convex topological linear space, there exists

an f ∈ S 0 such that f` → f in the sense of the strongly topology of S 0 .
On the other hand, that {f` }∞ s
`=1 is a Cauchy sequence in Bp,q implies that
{4k f` }∞ p p
`=1 is a Cauchy sequence in L . By the completeness of L , there
p
exists a gk ∈ L such that
k4k f` − gk kp → 0, as ` → ∞. (1.48)
s
1.4. Embeddings on Xp,q 13
Due to 4k f` → 4k f in S 0 as ` → ∞ and Lp ⊂ S 0 , we immediately have

gk = 4k f . Hence, (1.48) implies that
k4k (f` − f )kp → 0, as ` → ∞. (1.49)
(1.49) and Fatou’s lemma yield kf` − f kBp,q
s → 0 as ` → ∞.
s
1.4 Embeddings on Xp,q
It is known that there is no inclusion between Lp (Rn ) and Lq (Rn ) (p 6= q)

and we can not compare k · kp and k · kq . However, if we localize f in a
compact subset Ω of Rn in the frequency space, then we have
kF −1 χΩ F f kq . kF −1 χΩ F f kp , p 6 q. (1.50)
Combining this fact with the dyadic decomposition operator, we can ob-
tain the inclusions between Besov spaces in a simple way. This is a great
advantage of the frequency localized techniques.
Let Ω be a compact subset of Rn . Denote
SΩ = {f ∈ S : supp F f ⊂ Ω}. (1.51)
Proposition 1.15. Let 1 6 p 6 q 6 ∞. Then we have

kf kq . kf kp , ∀ f ∈ SΩ . (1.52)
Proof. Let ψ ∈ S satisfy F ψ(ξ) = 1, ∀ ξ ∈ Ω. In view of f ∈ SΩ we

see that F f = F f · F ψ. It follows that
Z
f (x) = C ψ(x − y)f (y)dy. (1.53)
Rn
Applied Hölder’s inequality,
kf k∞ . kψkp0 kf kp . kf kp , (1.54)
which implies that for q > p,
1−p/q
kf kq 6 kf k∞ kf kp/q
p . kf kp , (1.55)
the result follows.
Denote
LpΩ = {f ∈ Lp : supp F f ⊂ Ω}. (1.56)
Using (1.51) and standard approximate techniques, one obtains that
Proposition 1.16. Let 1 6 p 6 q 6 ∞. Then

kf kq . kf kp , ∀ f ∈ LpΩ . (1.57)
14 s
Corollary 1.1. Let 1 6 p 6 q 6 ∞, Bλ = {ξ : |ξ| 6 λ}. Then

kf kq . λn(1/p−1/q) kf kp , ∀ f ∈ LpBλ . (1.58)
Proof. Due to
kf kq = λ−n/q kf (·/λ)kq , (1.59)
n
F (f (·/λ)) = λ (F f )(λ·), (1.60)
we see, for any f ∈ LpBλ , that f (·/λ) ∈ LpB1 . In view of (1.59), (1.60) and
Proposition 1.16 we obtain the result, as desired.
Let ϕk be defined by (1.31). Since supp ϕk ⊂ B2k+1 , for any 1 6 p1 6
p2 6 ∞, we have, from (1.58), that Bernstein’s inequality holds
k4k f kp2 . 2n(1/p1 −1/p2 )k k4k f kp1 . (1.61)
Assume that s1 and s2 satisfy
n n
s1 − = s2 − . (1.62)
p1 p2
By (1.61) and (1.62), we immediately have
2s2 k k4k f kp2 . 2s1 k k4k f kp1 . (1.63)
Taking the `r -norm in both sides of (1.63), one has that
kf kBps2,r . kf kBps1 ,r . (1.64)
2 1
Thus, we have shown that
Theorem 1.1. Let 1 6 p1 6 p2 6 ∞, 1 6 r 6 ∞ satisfy s1 − n/p1 =

s2 − n/p2 . Then
Bps11 ,r ⊂ Bps22 ,r . (1.65)
s
Considering the inclusions on Fp,q , we have a similar result which is
slightly better than that of (1.65).
Theorem 1.2. Let 1 6 p1 < p2 < ∞, 1 6 q, r 6 ∞, −∞ < s2 < s1 < ∞

satisfy s1 − n/p1 = s2 − n/p2 . Then
Fps11,q ⊂ Fps22,r . (1.66)
Proof. By Proposition 1.13, we need to show

Fps11,∞ ⊂ Fps22,1 . (1.67)
s
1.4. Embeddings on Xp,q 15
We may assume that kf kFps1,∞ = 1. Let us recall an equivalent norm on

1
Lp :
Z ∞
kgkpp = p tp−1 |{x : |g(x)| > t}|dt, (1.68)
0
where |E| denotes the measure of the set E. So,

Z A ( )
X∞
p2 p2 −1 ks2
kf kF s2 ∼ t x: 2 |(4k f )(x)| > t dt
p2 ,1
0
Z ∞ ( k=0 )
X∞
p2 −1 ks2
+ t x: 2 |(4k f )(x)| > t dt
A
k=0
=:I + II, (1.69)

where A 1 is a constant which can be chosen as below. It is easy to see
that
∞
X
2ks2 |4k f | . 2K(s2 −s1 ) sup 2ks1 |4k f |. (1.70)
k>0
k=K+1
Applying (1.70) we can get the estimate of I (K = −1),

Z A

p2 −1

I. t x : sup 2 ks1
|(4 f )(x)| > ct dt
k
0 k>0
Z cA

p1 −1

ks1
. τ x : sup 2 |(4k f )(x)| > τ dτ
0 k>0
. 1. (1.71)
Now we estimate II. By Corollary 1.1,
k4k f k∞ . 2kn/p1 k4k f kp1 . 2k(n/p1 −s1 ) kf kFps1,∞ . (1.72)
1
Hence, for K ∈ N ∪ {0},

K
X K
X
2ks2 |4k f | . 2k(s2 −s1 +n/p1 ) . 2Kn/p2 . (1.73)
k=0 k=0
Taking K to be the largest natural number satisfying C2Kn/p2 6 t/2, we

have 2K ∼ tp2 /n . It is easy to see that such a K exists if t > A 1. For
P
t > A and ∞ k=0 2
ks2
|(4k f )(x)| > t, from (1.70) and (1.73) we get
∞
X
C2K(s2 −s1 ) sup 2ks1 |4k f | > 2ks2 |4k f | > t/2. (1.74)
k>0
k=K+1
16 s
Combining (1.68) and (1.74), we obtain that

Z ∞

p2 −1 ks1

p2 /p1
II . t x : sup 2 |(4k f )(x)| > ct dt
A k>0
Z ∞

. τ p1 −1 x : sup 2ks1 |(4k f )(x)| > τ dτ
A0 k>0
. 1. (1.75)
The estimates of I and II imply the conclusion.
s
Proposition 1.17. Let 1 6 p < ∞, s > n/p and 1 6 q 6 ∞. Let Xp,q be
s s
Bp,q or Fp,q . Then
s 0
Xp,q ⊂ B∞,1 ⊂ L∞ . (1.76)
Proof. Using the dyadic decomposition and (1.61), we have
∞
X X∞
kuk∞ 6 k4k uk∞ . 2kn/p k4k ukp
k=0 k=0
∞
!
X
. 2k(n/p−s) kukBp,∞
s 6 kukXp,q
s , (1.77)
k=0
which is the result, as desired.
s
1.5 Differential-difference norm on Xp,q
The earliest version of Besov spaces was defined by an integration with

a differential-difference form, not by dyadic decomposition operators. In
this section we consider the equivalent differential-difference norm, which
is quite convenient when we make nonlinear estimates in the study of non-
linear evolution equations.
Denote
m
X
Mm
h f (x) = Cmk
(−1)m−k f (x + kh), (1.78)
k=0
ωpm (t, f ) = sup k Mm

h f kp . (1.79)
|h|6t
We have
Proposition 1.18. Let s > 0, m, N ∈ N and m + N > s, 0 6 N < s;
1 6 p, q 6 ∞. Then
Xn Z ∞ q dt 1/q
N −s m N
kf kBp,q ∼ kf kp +
s t ωp (t, ∂xj f ) . (1.80)
j=1 0 t
s
1.5. Differential-difference norm on Xp,q 17
Proof. Since ωpm (t, f ) is an increasing function of t > 0, it suffices to

show that
!
n
X ∞
X q 1/q
kf kBp,q
s ∼ kf kp + 2i(s−N ) ωpm (2−i , ∂xNj f ) . (1.81)
j=1 i=−∞
s
Let f ∈ Bp,q . We write ρh (ξ) = eihξ − 1. One has that
∞
X
k Mm N
h ∂xj f kp = kF
−1 m
ρh F ∂xNj f kp 6 kF −1 ρm N
h ϕk F ∂xj f kp . (1.82)
k=0
In the following we show that12

kF −1 ρm N m km N k
h ϕk F ∂xj f kp . min(1, |h| 2 )2 k4k f kp . (1.83)
Noticing the support set of ϕk (ϕ−1 = 0),
1
X
kF −1 ρm N
h ϕk F ∂xj f kp . kF −1 ρm N
h ϕk+` F 4k ∂xj f kp . (1.84)
`=−1
By the definition of ρh we have ρh ∈ Mp . From the fact that Mp is a

Banach algebra, it follows that ρm
h ∈ Mp . In view of (1.15) in Proposition
1.10, we get ϕk ∈ Mp . Hence, ρmh ϕk ∈ Mp . By (1.84),
kF −1 ρm N N
h ϕk F ∂xj f kp . k4k ∂xj f kp . (1.85)
Observing the following identity
m
ρm h ξ
ρm
h ϕk =
h
, (|h||ξ|)m ϕk (1.86)
hh, ξim |h| |ξ|
and noticing that kρh /hh, ξikMp (Rn ) = k(eiξ − 1)/ξkMp (R) < ∞, it follows
that for all k > 1,
kρm
h ϕk kMp . 2
km
|h|m . (1.87)
Obviously, (1.87) also holds for k = 0. Hence, in view of (1.84), we have
kF −1 ρm N
h ϕk F ∂xj f kp . 2
km
|h|m k4k ∂xNj f kp . (1.88)
Applying the technique as in (1.84), we have from (1.15) that
k∂xNj 4k f kp . 2N k k4k f kp . (1.89)
Collecting (1.85), (1.88) and (1.89), we get (1.83). Inserting (1.83) into
(1.82), one has that
∞
X
2i(s−N ) ωpm (2−i , ∂xNj f ) . 2(i−k)(s−N ) (1 ∧ 2(k−i)m )2sk k4k f kp . (1.90)
k=0
12 Notice that Mh and 4k are different operators.
18 s
By (1.90) and Young’s inequality, we obtain

∞
!1/q
n
X X q
i(s−N )
2 ωpm (2−i , ∂xNj f ) . kf kBp,q
s . (1.91)
j=1 i=−∞
Clearly, kf kp . kf kBp,q
s . Thus, the right hand side of (1.80) can be con-
trolled by its left hand side.

−k
Next, we show the remaining part of (1.80). Let ρjk (ξ) = ei2 ξj − 1.
It suffices to prove that
n
X
k4k f kp . 2−N k kF −1 ρm N
jk F ∂xj f kp , k > 1. (1.92)
j=1
In fact, if (1.92) holds, in view of the definition of ωpm (t, f ), we have
kF −1 ρm N m −k
jk F ∂xj f kp . ωp (2 , ∂xNj f ), k > 1. (1.93)
It follows from (1.93) and the Besov norm that (ϕ0 ∈ Mp ),
∞
!1/q
n
X X q
kf kBp,q
s . kf kp + 2k(s−N ) ωpm (2−k , ∂xNj f ) , (1.94)
j=1 k=1
which implies that the left hand side of (1.80) can be controlled by its right
hand side.
Finally, we prove (1.92). We need the following lemma
Lemma 1.1. There exist smooth χj (j = 1, ..., n) satisfying

n
X
χj (ξ) = 1, ∀ ξ ∈ {ξ : 1/2 6 |ξ| 6 2}; (1.95)
j=1
√
supp χj ⊂ {ξ = (ξ1 , ..., ξn ) : |ξj | > 1/3 n}. (1.96)
Proof. We can choose that κ ∈ S (R), ζ ∈ S (Rn−1 ) with supp κ =

√
{ξ ∈ R : |ξ| > 1/3 n}, supp ζ = {ξ ∈ Rn−1 : |ξ| 6 3} and with positive
values in the interiors of supp κ and supp ζ, respectively. We write ξ j =
(ξ1 , ..., ξj−1 , ξj+1 , ..., ξn ) and
( j Pn
Pnκ(ξj )ζ(ξ ) j , j
κ(ξj )ζ(ξ ) j=1 κ(ξj )ζ(ξ ) 6= 0,
χj (ξ) = j=1
Pn (1.97)
j
0, j=1 κ(ξj )ζ(ξ ) = 0.
It is easy to see that χj satisfies (1.95) and (1.96).

s
1.6. Homogeneous space Ẋp,q 19
Now we can finish the proof of (1.92). By Lemma 1.1, for any k > 1,
n
X
k4k f kp 6 kF −1 ρ−m
jk ϕk χj (2
−k
·)ξj−N F (F −1 ρm N
jk F ∂xj f )kp . (1.98)
j=1
By Proposition 1.11,
(eiξj − 1)−m ϕχj ξj−N ∈ Mp , (1.99)
where ϕ is as in (1.30). It follows from Propositions 1.10 and (1.98) that
(1.92) holds.
As a consequence of Proposition 1.18, we have
Corollary 1.2. Let s > 0, s 6∈ N. we denote by [s] the integer part of s,

1 6 p, q 6 ∞. Then
Z ∞ !
Xn q dt 1/q
[s]−s [s]
kf kBp,q
s ∼ kf kp + t sup k Mh ∂xj f kp , (1.100)
j=1 0 |h|6t t
where Mh :=M1h .
It is very convenient to use (1.100) when we make nonlinear mapping

estimates.
s
1.6 Homogeneous space Ẋp,q
Let ϕ be as in (1.30). Analogous to (1.31), we write

ϕk (ξ) = ϕ(2−k ξ), k ∈ Z. (1.101)
It is easy to see that
X
ϕk (ξ) = 1, ξ ∈ Rn \ {0}. (1.102)
k∈Z
According to (1.32), we introduce the homogeneous dyadic decomposition

operators
4k = F −1 ϕk F , k ∈ Z. (1.103)
s
Similar to Xp,q , using {4k }k∈Z and function spaces `q (Lp ) and Lp (`q ), one
s
can define Ẋp,q . Notice that in (1.101)–(1.103), there is no restriction on
ξ = 0 and so, one needs to modify the Schwartz space S and it dual S 0 .
Denote
S˙ (Rn ) = {f ∈ S (Rn ) : (Dα fb)(0) = 0, ∀α}. (1.104)
20 s
As a subspace of S , S˙ := S˙ (Rn ) is equipped with the same topology as

S . We denote by S˙ 0 := S˙ 0 (Rn ) the dual space of S˙ . We now introduce
s
the homogeneous space Ẋp,q . Let
−∞ < s < ∞, 1 6 p, q 6 ∞. (1.105)
Denote

s
Ḃp,q ˙ 0 n
= f ∈ S (R ) : kf kḂ s < ∞ , (1.106)
p,q
X
∞ 1/q
kf kḂ s := 2ksq k4k f kqp , (1.107)
p,q
k=−∞
s
Ḃp,q is said to be a homogeneous Besov space. Let
−∞ < s < ∞, 1 6 p < ∞, 1 6 q 6 ∞. (1.108)
Define

s
Ḟp,q ˙ 0 n
= f ∈ S (R ) : kf kḞ s < ∞ , (1.109)
p,q
∞ 1/q
X ksq
kf kḞ s :=
2 |4 k f | q ,
(1.110)
p,q
k=−∞ p
s
Ḟp,q is said to be a homogeneous Triebel-Lizorkin space.
s s s
If there is no confusion, we will write Ẋp,q to stand for Ḃp,q or Ḟp,q .
Using the dilation, one has that
kf (2` ·)kẊp,q
s = 2`(n−s/p) kf kẊp,q
s , (1.111)
s
from which one can easily understand why we call Ẋp,q as a homogeneous
s s
space. Ẋp,q and Xp,q are quite similar. In the following we only state some
s
results on Ẋp,q , whose proofs follow an analogous way as the nonhomoge-
neous case. If there is no explanation, we will always assume that conditions
s s
(1.105) and (1.108) are satisfied for the spaces Ḃp,q and Ḟp,q , respectively.
s s s s
Proposition 1.19. Let Ẋp,q = Ḃp,q (or Ẋp,q = Ḟp,q ). Then
s
(1) Ẋp,q is a Banach space;
(2) ˙
S ⊂ Ẋp,qs
⊂ S˙ 0 ;
(3) If 1 6 p, q < ∞, then S˙ (Rn ) is dense in Ẋp,qs
;
s s
(4) If q1 6 q2 , then Ẋp,q1 ⊂ Ẋp,q2 ;
s s s
(5) Let 1 6 p < ∞. Then Ḃp,p∧q ⊂ Ḟp,q ⊂ Ḃp,p∨q .
s
1.6. Homogeneous space Ẋp,q 21
Theorem 1.3. Let −∞ < s2 < s1 < ∞, s1 − n/p1 = s2 − n/p2 and

1 6 r, q 6 ∞
Ḃps11 ,r ⊂ Ḃps22 ,r , Ḟps11,r ⊂ Ḟps22,q .
In the sequel, we will frequently use Theorem 1.3, where the first inclu-
sion can be shown in the same way as in Besov space, the second one is
slightly different from the Triebel-Lizorkin space, see Appendix.
Theorem 1.4. Let s > 0, m, N ∈ N, N < s, N + m > s and 1 6 p, q 6 ∞.

Then we have
Xn Z ∞ 1/q
dt
kf kḂ s ∼ tq(N −s) ωpm (t, ∂xNj f )q .
p,q
j=1 0 t
In particular, if s > 0 and s ∈/ N, then
n Z !1/q
X ∞
dt
kf kḂ s ∼ tq([s]−s) sup k Mh ∂x[s]j f kqp .
p,q
j=1 0 |h|6t t
The following result is a straightforward consequence of Proposition 1.18

and Theorem 1.4, which indicates the relation between homogeneous and
nonhomogeneous spaces.
Proposition 1.20. Let s > 0 and 1 6 p, q 6 ∞. Then we have

s
Bp,q = Lp ∩ Ḃp,q
s
.
The following is an interpolation inequality in Besov spaces, which are

very useful in nonlinear estimates, see [79; 89].
Proposition 1.21 (Convexity Hölder’s inequality). Let 1 6 pi , qi 6

PN PN
∞, 0 6 θi 6 1, σi , σ ∈ R (i = 1, . . . , N ), i=1 θi = 1, σ = i=1 θi σi ,
PN PN N σi σ
1/p = i=1 θi /pi , 1/q = i=1 θi /qi . Then ∩i=1 Ḃpi ,qi ⊂ Ḃp,q and for any
v ∈ ∩N σi
i=1 Ḃpi ,qi ,
N
Y
kvkḂ σ 6 kvkθḂiσi .
p,q pi ,qi
i=1
σ σ
This estimate also holds if one substitutes Ḃp,q by Ḟp,q (p, pi 6= ∞).
Proof. Applying Hölder’s inequality on Lp and `q respectively, one can

easily deduces the result:
!1/q
X
kvkḂp,q
σ = 2σkq k4k vkqp
k∈Z
22 s
N
!1/q
X Y
σkq
6 2 k4k vkθpiiq
k∈Z i=1
N
!1/q
XY
kσi θi
= (2 k4k vkθpii )q
k∈Z i=1
N
Y
6 kvkθḂiσi .
pi ,qi
i=1
σ
The case Ḟp,q is similar.
1.7 Bessel (Riesz) potential spaces Hps (Ḣps )
Recall that Js = (I − ∆)s/2 and Is = (−∆)s/2 are said to be the Bessel and
the Riesz potentials, respectively. Assume that
1 < p < ∞, −∞ < s < ∞. (1.112)
Denote
n o
Hps = f ∈ S 0 : kf kHps := kJs f kp < ∞ ; (1.113)
n o
Ḣps = f ∈ S˙ 0 : kf kḢps := kIs f kp < ∞ . (1.114)
Hps and Ḣps are said to be Bessel and Riesz potential spaces, respectively.
In the following we give the equivalent norms on Hps and Ḣps :
Theorem 1.5 (Littlewood-Paley square function theorem). Let s

and p satisfy (1.112). Then
Hps = Fp,2
s
, Ḣps = Ḟp,2
s
(1.115)
with equivalent norms.
Theorem 1.5 can be found in Stein [202] and Triebel [224].
Theorem 1.6. Let −∞ < s2 6 s1 < ∞ and 1 < p1 6 p2 < ∞ with

s1 − n/p1 = s2 − n/p2 . Then we have
Hps11 ⊂ Hps22 , Ḣps11 ⊂ Ḣps22 . (1.116)
Proof. It is a consequence of Theorems 1.5, 1.2 and 1.3.
Proposition 1.22. Let s and p satisfy (1.112). We have

1.7. Bessel (Riesz) potential spaces Hps (Ḣps ) 23
(1) Hps is a Banach space;

(2) S ⊂ Hps ⊂ S 0 and S (Rn ) is dense in Hps ;
s
(3) Bp,p ⊂ Hps ⊂ Bp,2
s s
(1 0);
(5) Hps ⊂ L∞ (s > n/p).
Remark 1.1. The results in (1)–(3) of Proposition 1.22 also hold for Riesz
potential spaces. The proof of Proposition 1.22 follows from Theorem 1.5
s
and the corresponding results of Fp,2 .
Proposition 1.23. Let s > 0 and 1 < p < ∞. We have Hps = Lp ∩ Ḣps .
Moreover, if s is an integer, then
X
kf kḢ s ∼ kDα f kp .
p
|α|=s
In the following we establish a modified Hölder inequality, as far as the

author have seen, Pecher [190] gave the first result on the modified Hölder
inequality. The next two results were obtained in Wang [234] by following
Pecher [190].
Proposition 1.24 (Modified Hölder inequality I). Let 1 6 p < ∞

and 1 < pi < ∞. Assume that αi is a multi-index, |αi | 6 si , ρi > 0,
and

1 si − |αi |
a i = ρi − , i = 0, 1, ..., N + 1. (1.117)
pi n
P +1
If ai > 0 and N i=0 ai = 1/p, then we have
NY
+1 N
Y +1

|Dαi ui |ρi 6 C kui kρḢisi . (1.118)
p pi
i=0 i=0
Substituted Ḣpsii by Hpsii in (1.118), the conclusion also holds.

PN +1
Proof. Let qi = 1/ai . Since 1/p = i=0 1/qi , we have, from Hölder’s
inequality, that
NY
+1 N
Y +1 N
Y +1

|Dαi ui |ρi 6 kDαi ui kρρii qi . kui kρi|αi | . (1.119)
p Ḣρi qi
i=0 i=0 i=0
Noticing that 1/ρi qi − |αi |/n = 1/pi − si /n, and using Theorem 1.6, we get
|α |
Ḣpsii ⊂ Ḣρi qii . By (1.119), we have the desired result.
24 s
Proposition 1.25 (Modified Hölder’s inequality II). Let 1 6 p <

∞, 1 < pi < ∞, |αi | 6 si , ρi > 0, and ai be as in (1.117). We have
P
(1) If each ai 6= 0, and ai >0 ai = 1/p, then
NY+1 N
Y +1

|Dαi ui |ρi 6 C kui kρHisi . (1.120)
p pi
i=0 i=0
P PN
(2) If ai >0 ai < 1/p 6 i=1 ρi /pi , then (1.120) also holds.
Proof. First, we prove (1). We may assume a0 , ..., aK > 0 and

aK+1 , ..., aN +1 < 0. By Hölder’s inequality,
NY +1 K
Y N
Y +1

|Dαi ui |ρi . kDαi ui kρρii qi kDαi ui kρ∞i . (1.121)
p
i=0 i=0 i=K+1
By Proposition 1.24,
K
Y K
Y
kDαi ui kρρii qi . kui kρḢisi . (1.122)
pi
i=0 i=0
Since ai < 0 (i = K + 1, ..., N + 1), in view of (5) of Proposition 1.22, we
have
N
Y +1 NY+1 NY+1
ρi
αi ρi
kD ui k∞ . αi
kD ui k si −|αi | . kui kρHisi . (1.123)
Hpi pi
i=K+1 i=K+1 i=K+1
Collecting (1.121)–(1.123), we obtain the result.
Next, we prove (2). We can assume that a0 , ..., aK > 0, aK+1 = ... =
aJ = 0 and aJ+1 , ..., aN +1 < 0.
P
Case 1. 1/p 6 K i=0 ρi /pi . Take qK+1 , ..., qN 1 (for any i > K and
qi > pi /ρi ) satisfying
XK N
X +1 K
1 1 X ρi
ai + < 6 . (1.124)
i=0
qi p p
i=0 i
i=K+1
Thus, we can choose appropriate qi (i = 0, ..., K) verifying pi /ρi 6 qi 6 1/ai
PN +1
and 1/p = i=0 1/qi . By Proposition 1.25, we obtain
NY
+1 N
Y +1
ρi
|D αi
u i | 6 C kui kρi|αi | . (1.125)
p Hρi qi
i=0 i=0
|α |
Since 1/qi ρi > 1/pi − (s − |αi |)/n and pi 6 qi ρi , we see that Hpsii ⊂ Hρi qii .
From (1.125) we have the result.
PK PN +1
Case 2. i=0 ρi /pi < 1/p 6 i=0 ρi /pi . Let qi = pi /ρi for i =
0, ..., K. For i = K + 1, ..., N + 1, we can find a qi ∈ [pi /ρi , ∞) satisfying
PN +1
1/p = i=0 1/qi . Analogous to the above discussions, we can obtain our
result.
1.8. Fractional Gagliardo-Nirenberg inequalities 25
1.8 Fractional Gagliardo-Nirenberg inequalities
We consider the Gagliardo-Nirenberg (GN) inequality in Besov and Triebel

spaces. Using the dyadic decomposition, we obtain a very simple proof for
the GN inequality. On the other hand, the GN inequality with fractional
order derivatives is of interest for its own sake and has an independent
significance. The results in this section are obtained in [99], earlier results
in this topic can be found in [23; 162; 183; 233; 242].
s
1.8.1 GN inequality in Ḃp,q
Lemma 1.2. Let 1 6 p, p0 , p1 , q, q0 , q1 6 ∞, s, s0 , s1 ∈ R, 0 6 θ 6 1.

Suppose that the following conditions hold:

n n n
− s = (1 − θ) − s0 + θ − s1 , (1.126)
p p0 p1
s 6 (1 − θ)s0 + θs1 , (1.127)
1 1−θ θ
6 + . (1.128)
q q0 q1
Then the fractional GN inequality of the following type
kukḂ s . kuk1−θ
s
Ḃ 0
kukθḂ s1 (1.129)
p,q p0 ,q0 p1 ,q1
holds for all u ∈ Ḃps00 ,q0 ∩ Ḃps11 ,q1 .
Proof. First, we consider the case 1/q 6 (1 − θ)/q0 + θ/q1 . By (1.127),

we have
1 1−θ θ s s0 s1
− − = − (1 − θ) − θ := −η 6 0. (1.130)
p p0 p1 n n n
Take p∗ and s∗ satisfying
1 1
= + η, s∗ = s + nη.
p∗ p
Applying the convexity Hölder inequality, we have
kf kḂ s∗∗ 6 kf k1−θ

s
Ḃ 0
kf kθḂ s1 . (1.131)
p ,q p0 ,q0 p1 ,q1
∗
Using the inclusion Ḃps∗ ,q ⊂ Ḃp,q
s
, we get the conclusion.
26 s
s
The result below clarify how the third indices q in Ḃp,q contribute the
validity of the GN inequalities.
Theorem 1.7. Let 1 6 q < ∞, 1 6 p, p0 , p1 6 ∞, 0 < θ < 1, s, s0 , s1 ∈

R. Then the fractional GN inequality of the following type
kukḂ s . kuk1−θ
s
Ḃ 0
kukθḂ s1 (1.132)
p,q p0 ,∞ p1 ,∞
holds if and only if

n n n
− s = (1 − θ) − s0 + θ − s1 , (1.133)
p p0 p1
n n
s0 − 6= s1 − , (1.134)
p0 p1
s 6 (1 − θ)s0 + θs1 , (1.135)
p0 = p1 if s = (1 − θ)s0 + θs1 . (1.136)
Proof. (Sufficiency) We can assume that s0 = 0 and the case s0 =6 0 can

be shown by a similar way.
Step 1. We consider the case p > p0 ∨ p1 . By definition13 ,
 1/q
X
kukḂ s =  N sq k4N ukqp  . (1.137)
p,q
N dyadic
From (1.135), it follows that

n n n n
θ − + s1 − s = (1 − θ) s + − . (1.138)
p p1 p0 p

Since 0 < θ < 1, (1.134) implies that np − pn1 + s1 − s s + n
p0 − n
p > 0.
Case 1. We consider the case
n n n n
s1 − s + − > 0, s + − > 0. (1.139)
p p1 p0 p
s s
Using the inclusion Ḃp,r 1
⊂ Ḃp,r 2
for any r1 6 r2 , it suffices to consider the
case q < 1/2, q ∈ N. For brevity, we write K := q −1 .
−1
X q2
kukḂp,q
s 6 (N1s ...NKs
k4N1 ukp ...k4NK ukp )
N1 >...>NK
q(1−q)
× (N1s . . . NK
s
k4N1 ukp . . . k4NK ukp ) . (1.140)
13 Here 4N is different from the notation as in Section 1.6, which is identical with 4logN
2
as in (1.103).
In view of Bernstein’s inequality,

n
−n n
−n
k4N ukp 6 N p0 p
k4N ukp0 , k4N ukp 6 N p1 p
k4N ukp1 . (1.141)
We can choose a ∈ (0, 1], k > 1 satisfying θK = k − 1 + a. Hence,
k4N1 ukp ...k4NK ukp
= (k4N1 ukp ...k4Nk−1 ukp k4Nk ukap )(k4Nk uk1−a
p k4Nk+1 ukp . . . k4NK ukp )
(1−a)( pn − n
p)
n
−n n
−n
. Nk 0 p0
Nk+1 p
. . . NKp0 p
k4Nk uk1−a
p0 k4Nk+1 ukp0 . . . k4NK ukp0
n
−n n
−n a( pn − n
p)
× N1p1 p p1
. . . Nk−1 p
Nk 1
k4N1 ukp1 . . . k4Nk−1 ukp1 k4Nk ukap1 .
(1.142)
Inserting (1.142) into (1.140), we have
X 2
kukḂ s . (N1s . . . NK
s
k4N1 ukp . . . k4NK ukp )q
p,q
N1 >...>NK
q(1−q)θK (1−θ)Kq(1−q)
× Λ(N1 , ..., NK )kukḂ s1 kukḂ 0 , (1.143)
p1 ,∞ p0 ,∞
where
− n + n −s1 +s −n n
a(− n + n −s1 +s)
p + p1 −s1 +s
Λ(N1 , . . . NK ) = N1 p p1 . . . Nk−1 Nk p p1
(1−a)(− n n
−n n
− n + n +s q(1−q)
p + p0 +s) p + p0 +s
× Nk Nk+1 . . . NK p p 0 .
(1.144)
By (1.143), we have
X K
X
kukḂp,q
s . Λ(N1 , . . . NK ) (Nis k∆Ni ukp )q (1.145)
N1 >...>NK i=1
(1−q)θ (1−θ)(1−q)
× kukḂ s1 kukḂ 0 . (1.146)
p1 ,∞ p0 ,∞
So, it suffices to prove

X K
X
Λ(N1 , . . . NK ) (Nis k∆Ni ukp )q . kukqḂ s . (1.147)
p,q
N1 >...>NK i=1
In fact, (1.144)–(1.147) imply the result. Finally, we prove (1.147). Apply-

ing the condition (1.139), we have
X
Λ(N1 , . . . NK )(Nks k∆Nk ukp )q
N1 >...>NK
X (k−1)(s−s1 + pn − np ) (K−k+1−a)(s+ pn − np )+a(s−s1 + pn − np ) q(1−q)
. Nk−1 1
Nk 0 1
Nk−1 >Nk
28 s
× Nksq k∆Nk ukqp

X Nk−1 (k−1)(s−s1 + p1 − p )q(1−q) sq
n n
. Nk k∆Nk ukqp
Nk
Nk−1 >Nk
. kukqḂ s . (1.148)
p,q
Case 2. We consider the case

n n n n
s1 − s + − < 0, s + − < 0. (1.149)
p p1 p0 p
P P
Substituting the summation N1 >...>NK by N1 6...6NK in (1.140) and
repeating the procedure as in Case 1, we can get the result, as desired.
Up to now, we have shown the results for the following two cases: (i)
s = (1 − θ)s0 + θs1 and p0 = p1 ; (ii) s < (1 − θ)s0 + θs1 and p > p0 ∨ p1 .
Step 2. We consider the case p < p0 ∨ p1 and s < (1 − θ)s0 + θs1 .
Due to θ ∈ (0, 1) and 1/p 6 (1 − θ)/p0 + θ/p1 , we see that p0 6= p1 and
p0 ∧ p1 < p < p0 ∨ p1 . Let 0 < ε 1. In view of the result as in Step 1, we
see that
1/2 1/2
kf kḂ s . kf kḂ s−ε kf kḂ s+ε . (1.150)
p,q p,∞ p,∞
Since s0 − n/p0 6= s1 − n/p1 , we can assume that s0 − n/p0 < s1 − n/p1 . It

follows that 1/p − s/n ∈ (1/p0 − s0 /n, 1/p1 − s1 /n). Hence, for sufficiently
small ε > 0,

1 s±ε 1 s0 1 s1
− ∈ − , − .
p n p0 n p1 n
It follows that there exist θ± ∈ (0, 1) satisfying

1 s±ε 1 s0 1 s1
− = (1 − θ± ) − + θ± − .
p n p0 n p1 n
Due to limε→0 θ± = θ, we see that for sufficiently small ε > 0,
s ± ε 6 (1 − θ± )s0 + θ± s1 .
Therefore, by Lemma 1.2, we have
1−θ
− θ
s−ε . kf k s0
kf kḂp,∞ Ḃ
kf kḂ−s1 , (1.151)
p0 ,∞ p1 ,∞
1−θ
+ θ
kf kḂp,∞
s+ε . kf k s0
Ḃ
kf kḂ+s1 . (1.152)
p0 ,∞ p1 ,∞
We easily see that θ = (θ+ + θ− )/2. Inserting (1.151) and (1.152) into
(1.150), we have the result, as desired.
We omit the proof of the necessity, one can refer to [99] for details.
For the most general case, we have the following

Theorem 1.8. Let 0 < p, p0 , p1 , q, q0 , q1 6 ∞, s, s0 , s1 ∈ R, 0 6 θ 6 1.
Then the fractional GN inequality of the following type
kukḂp,q
s . kuk1−θ
s
Ḃ 0
kukθḂ s1 (1.153)
p0 ,q0 p1 ,q1
holds for all u ∈ Ḃps00 ,q0 ∩ Ḃps11 ,q1 if and only if

n n n
− s = (1 − θ) − s0 + θ − s1 , (1.154)
p p0 p1
s 6 (1 − θ)s0 + θs1 , (1.155)
1 1−θ θ
6 + , if p0 6= p1 and s = (1 − θ)s0 + θs1 , (1.156)
q q0 q1
1 1−θ θ
s0 6= s1 or 6 + , if p0 = p1 and s = (1 − θ)s0 + θs1 ,
q q0 q1
(1.157)
n n 1 1−θ θ
s0 − 6= s − or 6 + , if s < (1 − θ)s0 + θs1 . (1.158)
p0 p q q0 q1
Proof. See [99].
s
1.8.2 GN inequality in Ḟp,q
s
In homogeneous Triebel-Lizorkin spaces Ḟp,q , we have the following (cf.
[99]):
Theorem 1.9. Let 1 6 p, pi , q < ∞, s, s0 , s1 ∈ R, 0 < θ < 1. Then the

fractional GN inequality of the following type
kukḞp,q
s . kuk1−θ
s
Ḟ 0
kukθḞps1,∞ (1.159)
p0 ,∞ 1

n n n
− s = (1 − θ) − s0 + θ − s1 , (1.160)
p p0 p1
s 6 (1 − θ)s0 + θs1 , (1.161)
s0 6= s1 if s = (1 − θ)s0 + θs1 . (1.162)
Proof. We only prove the sufficiency, which is a reformulation of Oru’s
[183] (see also [23]) and the proof of the necessity can be found in [99].
First, we consider the case s < (1 − θ)s0 + θs1 . We can take sufficiently
small ε > 0 satisfying
s 6 (1 − θ)s∗0 + θs∗1 , s∗0 := s0 − ε, s∗1 := s1 − ε.
30 s
Since ε 1, we can assume that

1 1 ε 1 1 ε
∗ := − > 0, ∗ := − > 0.
p0 p0 n p1 p1 n
Hence,

n n ∗ n ∗
− s = (1 − θ) − s 0 + θ − s 1 , (1.163)
p p∗0 p∗1
which implies that
1 1−θ θ s s∗ s∗
− ∗ − ∗ = − (1 − θ) 0 − θ 1 := −η 6 0. (1.164)
p p0 p1 n n n
Putting
1 1
= + η, s∗ = s + nη, (1.165)
p∗ p
we see that
1 1−θ θ
∗
= ∗ + ∗, s∗ = (1 − θ)s∗0 + θs∗1 . (1.166)
p p0 p1
Using Hölder’s inequality, in an analogous way as in Besov spaces, we have
kf kḞ s∗∗ . kf k1−θ
0
θ
s∗ kf k s∗
1
.
p ,q Ḟp∗ ,q Ḟp∗ ,q
0 1
Recalling the inclusions

s∗ s∗
Ḟps00,∞ ⊂ Fp∗0,q , Ḟps11,∞ ⊂ Fp∗1,q
0 1
we immediately get the conclusion.

Next, we consider the case s = (1 − θ)s0 + θs1 and s0 6= s1 . In this case
we easily see that 1/p = (1 − θ)/p0 + θ/p1 . The result follows from Lemma
C.1.
The following is the GN inequality with fractional derivatives.
Corollary 1.3. Let 1 < p, p0 , p1 < ∞, s, s1 ∈ R, 0 6 θ 6 1. Then the

fractional GN inequality of the following type
kukḢ s . kuk1−θ θ
Lp0 kukḢ s1 (1.167)
p p1

n n n
− s = (1 − θ) + θ − s1 , s 6 θs1 . (1.168)
p p0 p1
As the end of this chapter, we state some recent results on the general-
izations of Besov’s and Triebel-Lizorkin’s spaces.
Remark 1.2.
(1) In [5], the Qα space was introduced for which the norm is given by
Z Z 1/2
|f (x) − f (y)|2
kf kQα (Rn ) = sup n+2α
dxdy < ∞,
I I I |x − y|
where I ranges over all cubes in Rn . Xiao [253] considered the well
posedness of the Navier-Stokes equation in Q type spaces.
(2) Yang and Yuan [254] gave a unified way to handle the Qα and
Triebel-Lizorkin spaces. Indeed, for any p ∈ (1, ∞), q ∈ (1, ∞],
s,τ
τ, s ∈ R, they introduced the space Ḟp,q in the following way
  p/q 1/p

 Z 

1   X 
∞
sj q
kf kḞp,q = sup
s,τ
τ  (2 |4j f (x)|)  dx ,
P dyadic |P |  

 P j=− logl(P
2
) 
where l(P ) denotes the side length of the cube P . They showed
α,1/2−α/n s,τ
that Qα = Ḟ2,2 . In a similar way as Ḟp,q , Yang and Yuan
s,τ
[255] introduced Besov type space Ḃp,q . Some recent progress on
s,τ s,τ
Bp,q and Fp,q spaces can be found in [257].
s
(3) Function space like Ḟ∞,q are not considered in this chapter. It is
of importance in the case s = 0 and q = 2, which is equivalent
to BMO space introduced by F. John and L. Nirenberg in 1961.
An interesting generalization of BMO type on homogeneous-type
spaces or on measurable subsets of such spaces was given by X.
T. Duong and L. X. Yan [69]. Their spaces are defined by certain
maximal functions associated with generalized approximation of
the identity, which coincide with the classical BMO spaces of F.
John and L. Nirenberg in some special cases.
Chapter 2
Navier-Stokes equation
To be without some of the things you want is an indispensable part of

happiness. —— B. Russell
We consider the initial value problem for a nonlinear evolution equation,

which is said to be locally well-posed in Lp if for any initial value u|t=0 ∈ Lp ,
there exists a unique solution in C(0, T ; Lp) for some T > 0, moreover,
the solution map u|t=0 → u is at least a continuous mapping from Lp to
C(0, T ; Lp ). Furthermore, if T = ∞, then we say that it is globally well-
posed.
The Navier-Stokes (NS) equation is a fundamental equation in the the-
ory of fluid mechanics. However, the existence for its global smooth solu-
tions in three spatial dimensions has been open for many years.
In two spatial dimensions, the global well-posedness for the NS equation
was established by Ladyzhenskaya [154], and Kato [119] gave an alternate
proof based on the semi-group method and he also obtained the local well
posedness of solutions in Ln for n > 3.
We will use harmonic analysis techniques to study the NS equation.
First, we consider the Lr → Lp estimate for the heat semi-group H(t) = et∆
and establish the time-space estimates in mixed Lebesgue spaces Lqt Lpx for
the solutions of the linear heat equation, those estimates are similar to the
Strichartz estimates for the dispersive equations. Next, applying the time-
space estimates we show the local well-posedness in Ln for the NS equation
in a very simple way. In 2D case, in view of the a priori estimate in L2 ,
we immediately obtain the global well-posedness of the NS equation in L2 .
The method here is also useful for the other nonlinear parabolic equations,
such as the semi-linear heat equation, the Ginzburg-Landau equation and
so on.
33
34 Navier-Stokes equation
Finally, the regularity behavior for the NS equation will be investigated

by applying the frequency-uniform decomposition techniques.
2.1 Introduction
2.1.1 Model, energy structure

We study the Cauchy problem for the (incompressible) Navier-Stokes (NS)
equation
ut − ∆u + (u · ∇)u + ∇p = 0, div u = 0, u(0, x) = u0 (x), (2.1)

Pn 2
where ∆ = i=1 ∂xi , ∇ = (∂x1 , ..., ∂xn ), div u = ∂x1 u1 + ... + ∂xn un ,
u = (u1 , ..., un ) and p are real-valued unknown functions of (t, x) ∈ R+ ×Rn ,
u0 = (u10 , ..., un0 ) denotes the initial value of u at t = 0.
In view of its own structure, one can easily deduce that the smooth
solution of (2.1) satisfies the following conservation law:
Z t
1 2 1
ku(t)k2 + k∇u(s)k22 ds = ku0 k22 , (2.2)
P 2 0 P 2
n n
where kuk22 := i=1 kui k22 , k∇uk22 := i,j=1 k∂xj ui k22 for u = (u1 , ..., un ).
Indeed, multiplying u in (2.1) and then integrating by part, we have (2.2).
According to (2.2), it is natural to ask what happens if u0 ∈ L2 (Rn ).
Using the compactness method, we can obtain the existence of the weak
solutions if u0 ∈ L2 (Rn ), however, the uniqueness, the persistence of the
regularity of the solution and the continuity of the solution map are hard
to obtain, see [159].
2.1.2 Equivalent form of NS

Let (u, p) be a smooth solution of the NS equation. Taking the divergence
of the first equation in (2.1) and noticing that div u = 0, we immediately
obtain that
∆p + div[(u · ∇)u] = 0. (2.3)
−1
It follows that ∇p = (−∆) ∇div[(u · ∇)u]. For convenience, we write
P = I + (−∆)−1 ∇div. (2.4)
Solving p from (2.3) and inserting it into (2.1), we get
ut − ∆u + P [(u · ∇)u] = 0, u(0, x) = u0 (x). (2.5)
From (2.5), we see that the NS equation belongs to a nonlinear parabolic
equation.
2.2. Time-space estimates for the heat semi-group 35
2.1.3 Critical spaces

If u solves Eq. (2.5), so does uλ = λu(λ2 t, λx) (∀ λ > 0) with initial data
uλ (0, x) = λu0 (λx). We say that X := X(Rn ) is a critical space for the
NS equation, if the norm of uλ (0, x) in X is invariant for all λ > 0. Taking
notice of
kuλ (0, ·)kLr (Rn ) = λku0 (λ ·)kLr (Rn ) = λ1−n/r ku0 kLr (Rn ) , (2.6)
we see that r = n is the unique index so that the norm of uλ (0, x) in L (Rn )
r
is invariant for all λ > 0. So, Ln is a critical space of the NS equation.

Using the same way as in the above, we can verify that Ḣ n/2−1 (Rn ) is
also the unique space so that the norm of uλ (0, x) in Ḣ s (Rn ) is invariant.
Ḣ n/2−1 is another critical space in all Ḣ s .
On the other hand, noticing that Ḣ n/2−1 ⊂ Ln is a sharp inclusion, one
can easily understand Ḣ n/2−1 to be a critical space.
We can calculate that kuλ (0, ·)kḂ∞,∞
−1 ∼ ku0 kḂ∞,∞
−1 , from this point of
−1
view, we see that Ḃ∞,∞ is the largest critical space.
2.2 Time-space estimates for the heat semi-group
We consider the Cauchy problem for the heat equation:

ut − ∆u = f, u(0, x) = u0 (x).
Taking the Fourier transform, we get
u b = fb,
bt + |ξ|2 u b(0) = u
u b0 .
Solving the ordinary differential equation, we obtain that
Z t
u(t) = H(t)u0 + H(t − τ )f (τ, ·)dτ, (2.7)
0
2
where H(t) = et∆ = F −1 e−t|ξ| F .
2.2.1 Lr → Lp estimate for the heat semi-group

2
In H(t), e−t|ξ| as an exponential decay function, which corresponds to the
dissipation for H(t), determines that the semi-group H(t) has very good
properties. Let us start with an Lr → Lp estimate of H(t), which can be
found in [193].
Proposition 2.1. Let 1 6 r 6 p 6 ∞. Then
k n 1 1
k∇k H(t)f kp . t− 2 − 2 ( r − p ) kf kr , k = 0, 1, t > 0. (2.8)
Proof. First, we show that H(t) : Lr → Lr . Using Young’s inequality,

one has that
2
kH(t)f kr 6 kF −1 e−t|ξ| k1 kf kr . kf kr . (2.9)
Next, we consider kH(t)kLr →L∞ . Again, by Young’s inequality,
2 n
kH(t)f k∞ 6 kF −1 e−t|ξ| kr0 kf kr . t− 2r kf kr . (2.10)
Now, for any p > r, using Hölder’s inequality, (2.9) and (2.10), we have
n 1 1
1−r/p
kH(t)f kp 6 kH(t)f k∞ kH(t)f kr/p
r . t− 2 ( r − p ) kf kr . (2.11)
We estimate k∂x1 H(t)f kr . By Young’s inequality,
2 1
k∂x1 H(t)f kr 6 kF −1 (ξ1 e−t|ξ| )k1 kf kr . t− 2 kf kr . (2.12)
So,
1
k∇H(t)f kr . t− 2 kf kr . (2.13)
Combining (2.11) with (2.13), we immediately have
n 1 1
k∇H(t)f kp = kH(t/2)∇H(t/2)f kp . t− 2 ( r − p ) k∇H(t/2)f kr
1 n 1 1
. t− 2 − 2 ( r − p ) kf kr . (2.14)
This is the result, as desired.
As a generalization, we can show that for any s > 0 and 1 < r 6 p 6 ∞,
s n 1 1
k(−∆)s/2 H(t)f kp . t− 2 − 2 ( r − p ) kf kr . (2.15)
2.2.2 Time-space estimates for the heat semi-group

We study the heat semi-group in the mixed space Lγt Lpx and the technique is
to extensively use the dyadic decomposition together with the exponential
2
decay of e−t|ξ| . On the heat semi-group, the frequency localization idea
goes back to Chemin [29] (see also [31]). The techniques used in this section
θ
are also adapted to more general semi-groups, say e−t(1+ia)|ξ| with 0 < θ <
∞, see [240]. The following result is due to [240].
Proposition 2.2. Let a > 0, 1 6 r 6 p 6 ∞, 0 < λ 6 ∞ and 2/γ =

a + n(1/r − 1/p). Then we have
kH(t)f kLγ (R+ ;Ḃ 0 6 Ckf kḂ −a . (2.16)
p,λ ) r, λ∧γ
2.2. Time-space estimates for the heat semi-group 37
2
Proof. Noticed that e−t|ξ| is exponentially decaying, in order to exten-
sively use this fact, the frequency localization techniques will be applied.
We have
2
k4j H(t)f kr 6 kϕj e−t|ξ| kMr kf kr . (2.17)
−j
Using the multiplier criteria and noticing that ϕj = ϕ(2 ξ), suppϕ ⊂ {ξ :
1/2 6 |ξ| 6 2}, one has that
2 2j
|ξ|2 2j
|ξ|2
kϕj e−t|ξ| kMr = kϕe−t2 kMr 6 kϕe−t2 kH L , (2.18)
where L > n/2. Due to supx>0 xk /ex 6 C, it is easy to see that
2j
|ξ|2 2j
kϕe−t2 kH L . e−ct2 . (2.19)
So, by (2.17)–(2.19) we get
2j
k4j H(t)f kr . e−ct2 kf kr . (2.20)
P
Since 4j = 4j ( `=0,±1 4j+` ), (2.20) implies that
2j
k4j H(t)f kr . e−ct2 k4j f kr . (2.21)
Taking the `λ -norm in (2.21), we obtain that
 1/λ
X 2j
kH(t)f kḂ 0 .  e−ct2 k4j f kλr  . (2.22)
r,λ
j
In what follows the proof is separated into two cases. The first case is
that γ > λ. Taking the norm in Lγt (R+ ) on (2.22) and using Minkowski’s
inequality, we have
∞ 1/λ
X −ctλ22k
kH(t)f kLγ (R+ ,Ḃ 0 ) .
e k4 k f k λ
r γ/λ
r,λ
k=−∞ Lt
X
∞ 1/λ
−ctλ22k
. e γ/λ
k4k f kλr . (2.23)
Lt
k=−∞
Noticing that

−ctλ22k
e γ/λ
. 2−λ2k/γ , (2.24)
Lt
from (2.23) and (2.24) we get the consequence in the case p = r and γ > λ.
−2/γ+n(1/r−1/p) −2/γ
As r < p, applying the inclusion Ḃr,λ ⊂ Ḃp,λ , we can obtain
the result.
Secondly, we consider the case γ < λ. By (2.22),
Z X
kH(t)f kγḂ 0 dt . 2−2j k4j f kγp . (2.25)
R+ p,λ
j
We get the result in the case p = r and γ < λ. If r < p, in an analogous
way to the first case one can prove the conclusion.
In fact, the above estimates (2.15) and (2.16) can be developed to the
case 0 < r 6 p 6 ∞ and we have no restriction on λ, γ > 0. We have
a little bit of surprise about this fact, see [240] for details. Some earlier
results related to (2.16) were due to Weissler [250] and Giga [77], where
their proof is based on Marcinkiewicz’ interpolation theorem.
Corollary 2.1. Let 2/γ(p) = n(1/2 − 1/p). For any 2 6 p < ∞, we have
k∇H(t)f kL2 (R+ ;L2 ) . kf kL2 , (2.26)
kH(t)f kLγ(p) (R+ ;Lp ) . kf kL2 . (2.27)
0
Proof. Taking λ = r = 2 in Proposition 2.2 and noticing that Ḃp,2 ⊂ Lp ,
one immediately has the consequence.
For simplicity, we write
Z t
(A f )(t, x) := H(t − τ )f (τ, x)dτ. (2.28)
0
In what follows we consider the estimate of A f . By Proposition 2.1,

Z t
k n 1 1
k∇k A f kp . (t − τ )− 2 − 2 ( r − p ) kf (τ )kr dτ, k = 0, 1. (2.29)
0
Applying the Hardy-Littlewood-Sobolev inequality, we obtain that
Proposition 2.3. Let 1 6 r 6 p 6 ∞ and 1 < γ, γ1 < ∞ satisfy

1 1 k n 1 1 k n 1 1
= + + − − 1, + − < 1, k = 0, 1. (2.30)
γ γ1 2 2 r p 2 2 r p
Then we have
k∇k A f kLγ (R+ ;Lp ) . kf kLγ1 (R+ ;Lr ) . (2.31)
Proposition 2.3 can not handle the case γ = ∞, but we have (see [240])
Proposition 2.4. Let 1 6 r 6 ∞, 1 6 q 0 6 λ 6 ∞1 . Then

kA f kL∞ (R+ ,Ḃ 0 . kf kLq0 (R+ ,Ḃ −2/q ) . (2.32)
r,λ ) r,λ
Proof. Using (2.21) and Young’s inequality,

Z t
2k
k4k A f kr . e−c(t−τ )2 k4k f (τ )kr dτ
0
−2k/q
.2 k4k f kLq0 (R+ ,Lr ) . (2.33)
1 p0 stands for the conjugate number of p, i.e. 1/p + 1/p0 = 1.
2.3. Global well-posedness in L2 of NS in 2D 39
Taking the `λ -norm in (2.33) and using Minkowski’s inequality, we get

( ∞ )
X λ 1/λ
kA f kḂ 0 . 2−2k/q k4k f kLq0 (R+ ,Lr )
r,λ
k=−∞
. kf kLq0 (R+ ,Ḃ −2/q ) , (2.34)

r,λ
which implies the result, as desired.
Corollary 2.2. Let 2/γ(p) = n(1/2 − 1/p). For any 2 6 p < ∞, 2/γ(p) <
1, we have
k∇A f kL2 (R+ ;L2 ) . kf kLγ(p)0 (R+ ,Lp0 ) , (2.35)
kA f kL∞ (R+ ,L2 ) ∩ Lγ(p) (R+ ;Lp ) . kf kLγ(p)0 (R+ ,Lp0 ) . (2.36)
Clearly, all of the conclusions hold if we substitute the time interval R+

by [0, T ].
2.3 Global well-posedness in L2 of NS in 2D
For any Banach function space X := X(Rn ) and u = (u1 , ..., un ) ∈ X n ,

P P
we write kuk2X = i kui k2X and k∇uk2X = i,j k∂xi uj k2X . For any vector
function u(t) = (u1 (t), ..., un (t)) defined in I and valued in X n , we write
Z 1/q
kukLq (I,X n ) = ku(t)kqX dt . (2.37)
I
If there is no confusion, we will write Lq (I, X) := Lq (I, X n ). We denote

by [X]n0 the completion of the set
{u ∈ S n : div u = 0} in X n
and by Lq (I, [X]n0 ) the space of all of the vector functions u(t) =
(u1 (t), ..., un (t)) in Lq (I, X n ) such that u(t) ∈ [X]n0 for a.e. t ∈ I. It
is easy to verify that
[Lp ]n0 = {u ∈ [Lp ]n : div u = 0},
Lq (I, [X]n0 ) = {u ∈ Lq (I, [X]n ) : div u(t) = 0, a.e. t ∈ I},
where “div” is in the sense of tempered distributions.
Taking n = 2 and p = 4 in Corollaries 2.1 and 2.2, we have

k∇H(t)u0 kL2 (0,T ; L2 ) . ku0 kL2 , (2.38)
kH(t)u0 kL4 (0,T ; L4 ) . ku0 kL2 , (2.39)

k∇A f kL2 (0,T ; L2 ) . kf kL4/3(0,T ; L4/3 ) , (2.40)
kA f kL∞ (0,T ; L2 )∩L4 (0,T ; L4 ) . kf kL4/3(0,T ; L4/3 ) . (2.41)
We write Lpx,t∈[0,T ] p p
= L (0, T ; L ). Define the metric space:
n o
D = u : kukL4x,t∈[0,T ] + k∇ukL2x,t∈[0,T ] 6 M , (2.42)
d(u, v) = ku − vkL4x,t∈[0,T ] + k∇(u − v)kL2x,t∈[0,T ] . (2.43)
Assume that u0 ∈ [L2 ]20 . We consider the mapping:

M : u(t) → H(t)u0 + A P[(u · ∇)u]. (2.44)
Now we fix an M > 0 such that CM = 1/2, where C is the largest constant
appeared in the following inequalities. We show that M : (D, d) → (D, d)
is a contraction mapping. In fact, for any u ∈ D,
kMukL4x,t∈[0,T ] 6 kH(t)u0 kL4x,t∈[0,T ] + CkP[(u · ∇)u]kL4/3 , (2.45)
x,t∈[0,T ]
k∇MukL2x,t∈[0,T ] 6 k∇H(t)u0 kL2x,t∈[0,T ] + CkP[(u · ∇)u]kL4/3 . (2.46)

x,t∈[0,T ]
In view of Mihlin’s multiplier theorem, P : L4/3 → L4/3 . So, it follows from

Hölder’s inequality that,
kP[(u · ∇)u]kL4/3 . kukL4x,t∈[0,T ] k∇ukL2x,t∈[0,T ] 6 M 2 . (2.47)
x,t∈[0,T ]
By (2.38) and (2.39), there exists a T > 0 such that

k∇H(t)u0 kL2 (0,T ; L2 ) + kH(t)u0 kL4 (0,T ; L4 ) 6 M/2. (2.48)
So,
kMukL4x,t∈[0,T ] 6 M/2 + CM 2 6 M, (2.49)
k∇MukL2x,t∈[0,T ] 6 M/2 + CM 2 6 M. (2.50)
Similarly,
1
d(Mu, Mv) 6
d(u, v). (2.51)
2
In view of Banach’s contraction mapping principle, there exists a u ∈ D
satisfying
u(t) = H(t)u0 + A P[(u · ∇)u]. (2.52)
In view of the regularity theory, we see that the solution is smooth in the
domain (0, T ) × R2 . Since u0 ∈ [L2 ]20 and div P = 0, it follows from the
2.3. Global well-posedness in L2 of NS in 2D 41
integral equation that div u = 0 holds in (0, T ) × Rn . Again, by (2.41),

one sees that u ∈ C(0, T ; [L2 ]20 ). The conservation (2.2) can be deduced
by a straightforward calculation to the integral equation. It follows that
ku(T )k2 6 ku0 k2 . According to the standard semi-group theory, one can
extend the solution above step by step and finally finds a maximal time
Tm . It suffices to show that Tm = ∞. If not, then Tm < ∞. Tm is maximal
implies that
kukC([0,Tm); [L2 ]20 )∩L4x,t∈(0,T + k∇ukL2 (0,Tm ; L2 ) = ∞.

m)
By the energy estimate,

Z t
1 1
ku(t)k22 + k∇u(s)k22 ds = ku0 k22 , ∀ t < Tm . (2.53)
2 0 2
In view of the Gagliardo-Nirenberg inequality,
1/2 1/2
kukL4x,t∈[0,T ] . kukL∞(0,T ; L2 ) k∇ukL2 (0,T ; L2 ) . ku0 k2 , ∀ T < Tm . (2.54)
A contradiction. We have shown
Theorem 2.1. Let u0 ∈ [L2 ]20 . Then the NS equation (2.5) has a unique
solution u satisfying
u ∈ C(0, ∞; [L2 ]20 ) ∩ L2 (0, ∞; [Ḣ 1 ]20 ), (2.55)
and
Z t
1 1
ku(t)k22 + k∇u(s)k22 ds = ku0 k22 , 0 < t < ∞. (2.56)
2 0 2
The contraction mapping argument implies that the solution map is
analytic and of course, is continuous and the persistence of the regularity
of solutions is also easy to obtain by using the integral equation. So the NS
equation is globally well-posed in L2 .
Finally, we point out that the method used here can be applied to some
other kinds of nonlinear parabolic equations. For examples,
ut − ∆u + |∇u|u = 0, u(0, x) = u0 (x)
is globally well posed in L2 (R2 ); the Hamilton-Jacobi equation
ut − ∆u + |∇u|3/2 = 0, u(0, x) = u0 (x)
is locally well-posed in L2 (R2 ).

2.4 Well-posedness in Ln of NS in higher dimensions
For convenience, we still write Lpx,t∈[0,T ] = Lp (0, T ; Lp (Rn )) and use the
same notations as in the previous section. By Corollaries 2.1 and 2.2, we
have for n > 3,
k∇H(t)u0 kL2 (0,T ; L2 ) . ku0 kL2 , (2.57)
kH(t)u0 kL2+4/n . ku0 kL2 , (2.58)
x,t∈[0,T ]
k∇A f kL2 (0,T ; L2 ) . kf kL(2+4/n)0 , (2.59)

x,t∈[0,T ]
kA f kL∞ (0,T ; 2+4/n

L2 )∩Lx,t∈[0,T ]
. kf kL(2+4/n)0 . (2.60)
x,t∈[0,T ]
One may expect that the techniques in the previous section also work very
well. Unfortunately,
k(u · ∇)ukL(2+4/n)0 6 k∇ukL2x,t∈[0,T ] kukLn+2 . (2.61)
x,t∈[0,T ] x,t∈[0,T ]
Ln+2
x,t∈[0,T ] is out of the control of the above linear estimates. So, we need
the following
Corollary 2.3. We have

kH(t)u0 kLn+2 . ku0 kn , (2.62)
x,t∈[0,T ]
kH(t)u0 kL∞ (0,T ; Ln ) . ku0 kn , (2.63)

k∇A f k Ln+2 . kf kL(n+2)/2 , (2.64)
x,t∈[0,T ] x,t∈[0,T ]
k∇A f kL∞ (0,T ; Ln ) . kf kL(2+n)/2 . (2.65)

x,t∈[0,T ]
Proof. Taking p = r = λ = n and γ = 2 + n in Proposition 2.2, we get

kH(t)u0 kL2+n (R+ ;Ḃ 2/(2+n) ) . ku0 kḂ 0 . ku0 kn . (2.66)
n,n n, n
2/(2+n) 0 2/(2+n)
Using the inclusions Ḃn,n = Ḟn,n ⊂ Ḟn+2,2 = Ln+2 , we see that
(2.66) implies that (2.62) holds. Obviously, we have (2.63).
(2.64) is a straightforward consequence of Proposition 2.3. Taking r =
λ = (n + 2)/2 and q 0 = (n + 2)/2 in Proposition 2.4, we get
kA f kL∞ (R+ ; Ḃ 2n/(2+n) )
. kf kL(n+2)/2 (R+ ; Ḃ 0 ). (2.67)
(n+2)/2, (n+2)/2 (n+2)/2, (n+2)/2
This implies
k∇A f kL∞ (R+ ; Ḃ (n−2)/(2+n) )
. kf kL(n+2)/2 . (2.68)
(n+2)/2, (n+2)/2 x,t∈R+
(n−2)/(2+n) 0(n−2)/(2+n)
Using the embedding Ḃ(n+2)/2, (n+2)/2 = Ḟ(n+2)/2, (n+2)/2 ⊂ Ḟn,2 = Ln , we
see that (2.68) implies that (2.65) holds.
2.4. Well-posedness in Ln of NS in higher dimensions 43
The above linear estimates can derive the local well-posedness and the
global well-posedness with small data in Ln for n > 3. We will use the
same notations as in the case n = 2. We have
Theorem 2.2. Let u0 ∈ [Ln ]n0 . Then there exists a Tm > 0 such that the
NS equation (2.5) has a unique solution u satisfying
u ∈ C([0, Tm ); [Ln ]n0 ) ∩ L2+n
loc (0, Tm ; [L
2+n n
]0 ). (2.69)
If Tm < ∞, then we have kukL2+n(0,Tm ; L2+n ) = ∞. If ku0 kn is sufficiently
small, then Tm = ∞. Moreover, if u0 ∈ [L2 ]n0 , then u ∈ C(0, Tm ; [L2 ]n0 ),
∂xi u ∈ L2 (0, Tm ; [L2 ]n0 ), and
Z t
1 1
ku(t)k22 + k∇u(s)k22 ds = ku0 k22 , 0 < t < Tm . (2.70)
2 0 2
Proof. Put
n o
D = u : kukL2+n 6 δ, kukL∞ ([0,T ];Ln ) 6 2Cku0 kn , (2.71)
x,t∈[0,T ]
d(u, v) = ku − vkL2+n . (2.72)

x,t∈[0,T ]
We consider the mapping:

M : u(t) → H(t)u0 + A P div (u ⊗ u), (2.73)
by Corollary 2.3, we have
kMukL2+n . kH(t)u0 kL2+n + ku ⊗ ukL(2+n)/2
x,t∈[0,T ] x,t∈[0,T ] x,t∈[0,T ]
. kH(t)u0 kL2+n + kuk2L2+n

x,t∈[0,T ] x,t∈[0,T ]
. kH(t)u0 kL2+n + δ2, (2.74)

x,t∈[0,T ]
kMukL∞ (0,T ; Ln ) . ku0 kn + ku ⊗ ukL(2+n)/2

x,t∈[0,T ]
. ku0 kn + δ 2 . (2.75)
If Cδ 6 1/4, we can show that M is a contraction mapping from D into
itself. So, there exists a u satisfying
u(t) = H(t)u0 + A P∇ · (u ⊗ u). (2.76)
By a standard argument, we see that u is unique in L2+n (0, T ; [L2+n ]n0 ).
Moreover, one can extend the solution step by step and find a maximal Tm
such that u ∈ C([0, Tm ); [Ln ]n0 ) ∩ L2+n
loc (0, Tm ; [L
2+n n
]0 ).
If u0 ∈ [L ]0 , then we can use (2.57)–(2.61) to obtain u ∈ C(0, T ; [L2 ]n0 )
2 n
and ∂xi u ∈ L2 (0, T ; [L2 ]n0 ) in a similar way as in the 2D case. According
to the regularity theory (see [67] for instance), we see that u is infinitely
smooth in the domain (0, T ) × Rn . So, a straightforward computation will
lead to (2.2). Moreover, if ku0 kn is small enough, we can take T = ∞ in
(2.71) and (2.72).
The result of Theorem 2.2 has a long history. Weissler [251] gives a de-
tailed Lp -theory in half-space for local solutions. Kato [119] used a different
way to obtain the well posedness result in Ln by constructing the resolution
norm like supt tθ ku(t)kp . The idea only taking Ln+2 x,t as a resolution space
was hidden in some references, see for instance [67]. Here the proof is to
take Ln+2 n
x,t∈[0,T ] ∩ C([0, T ]; L ) as a resolution space which enables us to get
the local well posedness of solutions in Ln , where (2.65) plays a crucial role.
Similar bilinear estimate to (2.65) in 3D was shown in [70] in a different
way:
kA P div (u ⊗ v)kL∞ ([0,T ];L3 ) . kukL5x,t kvkL5x,t . (2.77)
In 3D case, Kenig and Koch [123] obtained the result of Theorem 2.2 by
applying (2.77).
2.5 Regularity of solutions for NS
We will show that, for any t > 0, the solution of the NS equation is infinitely
differentiable and in fact, is really analytic.
s
2.5.1 Gevrey class and function space E2,1
First, we consider the Gevrey classes. Let s > 0, denote
s
n ∞ n ∃ m!
Gs (R ) = f ∈ C (R ) : ρ, M > 0 s.t. kf kḢ m 6 M ∀m ∈ Z+ .
ρm
(2.78)
Gs (Rn ) is said to be a Gevrey s-class.
Our aim is to show that the solution of the NS equation belongs to the
Gevrey 1-class. In order to show this fact, we will use Wiener’s decomposi-
tion of Rn . In Chapter 6, we will continuously use Wiener’s decomposition
to study the nonlinear dispersive equations. Let Qα be the unit cube with
the center at α ∈ Zn , i.e., Qα = α + Q0 , Q0 = {x = (x1 , ..., xn ) : −1/2 6
xi < 1/2}. It is easy to see that
Rn = ∪α∈Zn Qα , Qα ∩ Qβ = ∅, α 6= β.
n
So, {Qα } constitutes a decomposition of R , which is said to be the uniform
decomposition (or Wiener’s decomposition) of Rn . Comparing it with the
dyadic decomposition, we see that it is more delicate than the dyadic de-
composition of Rn . Putting the uniform decomposition into the frequency
2.5. Regularity of solutions for NS 45
space and combining it with the `1 (L2 ) space, we have the following
( )
X
s 0 n s|k| −1 b
E2,1 = f ∈ S (R ) : kf kE2,1
s = 2 F χQk f < ∞ . (2.79)
2
k∈Zn
s
Such a kind of space and its generalized version Ep,q were studied in [247],
which is a kind of generalized modulation spaces [86]. It is easy to see
s
that E2,1 is a Banach space, which has an exponential regularity weight.
s
From PDE’s point of view, E2,1 has infinite smoothness, which is different
from any Besov spaces and any classical modulation spaces. The classical
s
modulation space M2,1 takes the following norm:
( )
X

s 0 n
M = f ∈ S (R ) : kf kM s =
2,1 hki F χQ fb < ∞ , (2.80)
2,1
s −1
k
2
k∈Zn
s s
where hki = 1 + |k|. M2,1 and its generalizations Mp,q have finite smooth-
s
ness, which are similar to Besov spaces (see Chapter 6). E2,1 is of impor-
tance in our regularity argument for the parabolic equation, whose infinite
regularity can be seen in the following (cf. [247]):
Proposition 2.5. We have

[
s
G1 = E2,1 . (2.81)
s>0
s
Proof. First, we show that for any s > 0, E2,1 is a subset of G1 . Indeed,
s
for any f ∈ E2,1 and 0 < c 1,
kf kḢ m . k∇m f kE2,1
0
X
. kχQk |ξ|m fb k2
k∈Zn
X p
. (|k| + n/2)m kχQk fb k2
k∈Zn
m! X (cs)m (|k| + C)m s|k|
. 2 kχQk fb k2
(cs)m n
m! · 2s|k|
k∈Z
m!
. kf kE2,1
s .
(2s)m
On the other hand, let f ∈ G1 and L > n/2. Using Taylor’s expansion and
Hölder’s inequality,
X
kf kE2,1
s = 2s|k| kχQk fb k2
k∈Zn
X∞ X
(s ln 2)m
6 kf k2 + |k|m kχQk fb k2
m=0
m!
k∈Zn , |k|1
∞
X m X
(s ln 2)
. kf k2 + C m |k|−L kχQk ∇\
m+L f k
2
m=0
m!
k∈Zn , |k|1
∞
X m
(Cs)
. kf k2 + kf kḢ L+m
m=0
m!
X ∞
(Cs)m (m + L)!
. kf k2 + · .
m=0
m! ρm
P∞
Choosing 0 < s 1, we see that m=0 (Cs/ρ)m (m + L)L is a convergent
s
series. Thus, f ∈ E2,1 .
s
2.5.2 Estimates of heat semi-group in E2,1
Using Proposition 2.5, we want to get the solution of the NS equation
s
u ∈ G1 . It suffices to show that u belongs to E2,1 (s > 0). Hence, one
needs to estimate kH(t)u0 kE2,1
s and kA f kE2,1
s .
Proposition 2.6. Let H(t) = et∆ . There exists a c > 0 such that for
j = 0, 1 and T 6 1,
k∇j H(t)u0 kE2,1ct . t
−j/2
ku0 kE2,1
0 , t 6 T, (2.82)
Z t
j
sup
∇ H(t − τ )f (τ )dτ
. (T + T 1−j/2 ) sup kf (τ )kE2,1
cτ .
t∈[0,T ] 0 ct
E2,1 τ ∈[0,T ]
(2.83)
Proof. For convenience, we will use the notation
k = F −1 χQk F . (2.84)
By Plancherel’s identity, for |k| > 1,
2 2
kk H(t)u0 k2 = kχQk (ξ)e−t|ξ| F u0 k2 6 e−ct|k| kk u0 k2 . (2.85)
If k = 0, we see that (2.85) also holds for t 6 1. Taking the summation
over all k ∈ Zn , we can get (2.82) hold for j = 0.
We now prove (2.82) for j = 1. If k = 0 and t 6 1,
2
k∇k H(t)u0 k2 . kχQk (ξ)e−t|ξ| F u0 k2 6 kk u0 k2 . (2.86)
For |k| > 1,
2
k∇k H(t)u0 k2 . k|ξ|χQk (ξ)e−t|ξ| F u0 k2 6 t−1/2 e−ct|k| kk u0 k2 . (2.87)
Combining (2.86) with (2.87), we have

k∇k H(t)u0 k2 . t−1/2 e−ct|k| kk u0 k2 . (2.88)
This proves (2.82) for j = 1. For t ∈ [0, T ] with T 6 1,
Z t Z t

k ∇ H(t − τ )f (τ )dτ . (t − τ )−1/2 e−c(t−τ )|k| kk f (τ )k2 dτ,

0 2 0
(2.89)
which implies that (2.83) holds in the case j = 1. Another case j = 0 is
similar to the case j = 1.
s
2.5.3 Bilinear estimates in E2,1
By Proposition 2.6, our goal is to show that the solution of the NS equation
ct
belongs to E2,1 , it is necessary to estimate ku ⊗ ukE2,1
ct . We have
λ
Proposition 2.7. E2,1 is a Banach algebra. More precisely, we hace
kuvkE2,1
λ 6 C2Cλ kukE2,1
λ kvkE λ ,
2,1
(2.90)
λ
where C is independent of λ > 0 and u, v ∈ E2,1 .
Proof. Noticing that for k = (k1 , ..., kn ) ∈ Zn with |k| = |k1 | + ... + |kn |,
we have
X
kuvkE2,1
λ = 2λ|k| k k (uv)k2 . (2.91)
k∈Zn

X
uv = (i u)(j v), (2.92)
i,j∈Zn
which implies that

X
k (uv) = k (i u j v). (2.93)
i,j∈Zn
b) ∗ (χQj vb) and

We have F (i u j v) = (χQi u
√
b) ∗ (χQj vb) ⊂ Ω := {ξ : |ξ − i − j| 6 2 n}.
supp(χQi u (2.94)
So,
√
k (i u j v) = 0, |k − i − j| > 3 n. (2.95)
It follows that
kk (i u j v)k2 = kk (i u j v)k2 χ(|k−i−j|63√n) . (2.96)
In view of Plancherel’s identity and Hölder’s inequality,

kk (i u j v)k2 6 k i u j vk2 χ(|k−i−j|63√n)
6 k i uk∞ kj vk2 χ(|k−i−j|63√n)
. k i uk2 kj vk2 χ(|k−i−j|63√n) . (2.97)
Hence,
X X
kuvkE2,1
λ 6 2λ|k| kk (i u j v)k2
k∈Zn i,j∈Zn
X X
. 2λ|k| k i uk2 kj vk2 χ(|k−i−j|63√n)
k∈Zn i,j∈Zn
X
. 2Cλ 2λ(|i|+|j|) k i uk2 kj vk2 . (2.98)
i,j∈Zn
The result follows.
2.5.4 Gevrey regularity of NS equation

ct
Once we get that the solution of the NS equation is in E2,1 , then it belongs
to the Gevrey 1-class. Let
D = {u : sup ku(t)kE2,1
ct 6 M }, (2.99)
06t6t0
d(u, v) = sup ku(t) − v(t)kE2,1

ct . (2.100)
06t6t0
Considering the mapping

Z t
T : u(t) → H(t)u0 − H(t − τ )[P ∇ · (u ⊗ u)](τ )dτ, (2.101)
0
0
we show that for u0 ∈ E2,1 , there exists a t0 > 0 such that T : (D, d) →
(D, d) is a contraction mapping. For convenience, we denote
|||u||| = sup ku(t)kE2,1
ct . (2.102)
06t6t0
Let u ∈ D. By Proposition 2.6 and 2.7,

1/2
|||T u||| . ku0 kE2,1
0 + t0 |||u||| + t0 |||(u ⊗ u)|||
1/2
. ku0 kE2,1
0 + t0 |||u|||2 . (2.103)
1/2
Take M = 2Cku0 kE2,1 0 and assume t0 < 1. If Ct0 M 6 1/4, then T u ∈ D.
Similarly, for any u, v ∈ D,
1
|||T u − T v||| 6 |||u − v|||. (2.104)
2
So, there exists a u ∈ D satisfying T u = u.

0 n
Theorem 2.3. Let u0 ∈ [E2,1 ]0 . Then there exists a Tm > 0, such that the
0 n
NS equation (2.5) has a unique solution u ∈ C [0, Tm ); [E2,1 ]0 . Moreover,
there exists a t0 ∈ (0, Tm ) satisfying
c(t∧t0 ) n
u(t) ∈ [E2,1 ]0 , t ∈ [0, Tm ). (2.105)
If Tm < ∞, then sup06t<Tm ku(t)kE2,1 0 = ∞. Moreover, u ∈

C 0, Tm ; [L2 ]n0 , ∂xi u ∈ L2 0, Tm ; [L2 ]n0 , and
Z
1 1 t 1
ku(t)k22 + k∇u(s)k22 ds = ku0 k22 , 0 < t < Tm . (2.106)
2 2 0 2
If ku0 kn is sufficiently small, then Tm = ∞.
Remark 2.1.
(1) By Theorem 2.3, we can get that the solution of the NS equation
belongs to C ∞ ((0, Tm ) × Rn ).
(2) Theorem 2.3 implies that the solution of the NS equation belongs
to the Gevrey 1-class and so, is really analytic.
(3) Theorem 2.3 also describes the disappearing process of the regu-
larity of the solutions to the NS equation.
Remark 2.2. The Gevrey regularity for the evolution equations is of im-
portance for its own sake. The Gevrey regularity of the weak solutions for
a class of linear and semi-linear Fokker-Planck equations
(∂t + v · ∇x − ∆v )u = F (t, x, v, u, ∇v u)
was recently studied by Chen, Li and Xu [35], see also [37] and references
therein for a class of the linear model of spatially inhomogeneous Boltzmann
equations without an angular cutoff.
Remark 2.3. As the end of this chapter, we state some recent progress on
NS equation without proofs.
(1) Let us mention the result by Koch and Tartaru [151] where the
global solutions for NS in 3D are obtained with the small data in
the space BM O−1 with the norm:
Z !1/2
−3/2 t∆
kukḂ∞,∞
−1 + sup R |e u(y)|dydt ,
x∈R3 , R>0 [0,R2 ]×{y:|x−y|6R}
see also Chemin and Gallagher’s generalizations in [30; 31; 32] for
a class of large data.
(2) Recently, Escauriaza, Seregin and Šverák [70] showed that any
“Leray-Hopf” weak solution in 3D which remains bounded in
L3 (R3 ) cannot develop a singularity in finite time. Their proof
used a blow-up procedure and reduction to a backwards unique-
ness question for the heat equation, and was then completed using
Carleman-type inequalities and the theory of unique continuation.
Kenig and Koch [123] gave an alternative proof by substituting L3
with Ḣ 1/2 . Dong and Du [68] generalized their results in higher
spatial dimensions n > 3.
(3) Noticing that L3 ⊂ B∞,∞ −1
in 3D is a sharp embedding, for any
solution u of the NS equation in C([0, T ∗ ); L3 ), we see that u ∈
C([0, T ∗ ); B∞,∞
−1
). May [163] prove that if T ∗ < ∞, then there
exists a constant c > 0 independent of the solution of NS equation
such that
lim sup ku(t) − ωkB∞,∞
−1 >c
t→T ∗
for all ω ∈ S . For the Leray-Hopf weak solution, Cheskidov and

Shvydkoy [39] obtained similar result.
(4) Recently, the Cauchy problem for the 3D anisotropic Navier-Stokes
equation
ut − (∂x21 + ∂x22 )u + u · ∇u + ∇p = 0, ∇ · u = 0, u|t=0 = u0
is considered in [187; 33] (see also references therein). These equa-
tions come from meteorology models. More recent results based on
the harmonic analysis method can be found in [155].
Chapter 3
Strichartz estimates for linear

dispersive equations
First, let us imagine the large-time-decaying phenomena in a disper-

sive system... .
We begin to study the dispersive equations. Roughly speaking, the

dispersive equation takes the following form
∂t u − iP (D)u = F, (3.1)
√
where u(t, x) is an unknown function, i = −1, P\ (D)u = P (ξ)b u, and
P (ξ) is a real-valued function, which is said to be the dispersion relation of
(3.1)1 . If F is a nonlinear function of u, then (3.1) is said to be a nonlinear
dispersive equation.
In this chapter we study the time-space estimates for the solutions of
the linear dispersive equation in mixed Lebesgue spaces Lqt Lpx (or more
general spaces Lq (0, T ; Bp,r
s
)), so called the Strichartz inequalities, which is
a starting point to the study of nonlinear dispersive equations. For instant,
we consider the Schrödinger equation2
iut + ∆u = f, u(0, x) = u0 (x), (3.2)

Pn
where ∆ = i=1 ∂x2i , u(t, x) is a complex valued function of (t, x) ∈ R×Rn ,
u0 denotes the initial value at t = 0. f is a known complex function of
(t, x) ∈ R × Rn .
We will use its integral form. Taking the Fourier transform to (3.2), we
get
ib b = fb,
ut − |ξ|2 u b(0) = u
u b0 .
1 We emphasize that P (·) must be a real function.
2 Erwin Schrödinger (1887-1961) was an Austrian theoretical physicist who achieved
fame for his fundamental contributions to quantum mechanics, especially the Schrödinger
equation, for which he received the Nobel Prize in 1933.
51
52 Strichartz estimates for linear dispersive equations
Solving the ordinary differential equation, we obtain that

Z t
u(t) = S(t)u0 − i S(t − τ )f (τ, ·)dτ, (3.3)
0
2
where S(t) = eit∆ := F −1 e−it|ξ| F . By Plancherel’s identity, we see that
kS(t)u0 k2 = ku0 k2 , which means that S(t)u0 is invariant in L2 for any
t ∈ R. However, we can show that S(t) satisfies the following Strichartz
estimates:
kS(t)u0 kLγ(r) (R,Lr ) . ku0 k2 ,
Z t

S(t − τ )f (τ, ·)dτ . kf kLγ(ρ)0 (R,Lρ0 ) ,

0 Lγ(r) (R,Lr )
where 2 6 r, ρ 6 ∞, 2/γ(·) = n(1/2 − 1/·) ∈ [0, 1), and p0 is the conjugate

number of p. Such kinds of estimates are time-decaying versions, which are
of importance in the study of nonlinear dispersive equations. In the next
chapter, we will further indicate why Strichartz inequalities are useful.
0
3.1 Lp → Lp estimates for the dispersive semi-group
2
Let us consider the Schrödinger semi-group eit∆ := F −1 eit|ξ| F . Since
Z
|x−y|2
eit∆ u0 = ct−n/2 e 4it u0 (y)dy, (3.4)
Rn
1 ∞
we immediately have the L → L decay
keit∆ u0 k∞ . t−n/2 ku0 k1 .
Taken noticing of keit∆u0 k2 = ku0 k2 , an interpolation yields
keit∆u0 kp . t−n/2 ku0 kp0 , p > 2, 1/p + 1/p0 = 1.
0
This is the fundamental Lp → Lp estimate, or the time decay estimate for
the Schrödinger semi-group. For the general semi-group, it is not expected
to have an analytic expression as in (3.4), we need to look for other ways
0
to get the Lp → Lp estimates.
0
In this section we study the Lp → Lp estimates for a class of disper-
sive semi-groups U (t) := F −1 exp(itP (ξ))F and use two different ways to
0
consider its Lp → Lp decay. The first method is very effective for homo-
geneous functions P (·), the second method can deal with nonhomogeneous
radial functions P (·).
First, we consider the decay estimates for the semi-group Um (t) :=
F −1 exp(it|ξ|m )F . Some earlier decay estimates on Um (t) were obtained
0
3.1. Lp → Lp estimates for the dispersive semi-group 53
0
by Pecher [191]. Following Pecher’s proof, we can get an Lp → Lp decay
estimate with smooth effects in the case m > 2. Some nontrivial generalized
dispersive Lq − Lp estimates can be founded in Cui [52; 53], Sugimoto [207;
208] (see also [195]).
We need the following lemma, which is due to Littman [161].
Lemma 3.1. Let v ∈ C0∞ (Rn ), supp v = Ω. P (ξ) : Rn → R is an infinitely

smooth function on Ω. For any ξ ∈ Ω, the rank of (∂ 2 P (ξ)/∂ξi ∂ξj )ni,j=1 is
at least ρ > 0, Then there exists a K ∈ N such that for any λ ∈ R,
X
kF −1 eiλP (ξ) vk∞ . (1 + |λ|)−ρ/2 kDα vk∞ . (3.5)
|α|6K
Now we derive the decay estimates of Um (t) in homogeneous Besov

spaces. Let {4k }k∈Z be as in Sec. 1.6. In view of Young’s inequality (see
Appendix), one has that
m m
kF −1 eit|ξ| F 4k f k∞ 6 kF −1 eit|ξ| ϕ(2−k ξ)k∞ kf k1 . (3.6)
−1 it|ξ|m −k
Now we estimate kF e ϕ(2 ξ)k∞ . Applying Proposition 1.4 and
Lemma 3.1, we have
m
kF −1 eit|ξ| ϕ(2−k ξ)k∞
km
|ξ|m
= 2kn kF −1 eit2 ϕ(ξ)k∞ . t−ρ/2 2k(n−mρ/2) , (3.7)
where ρ denotes the rank of (∂ 2 |ξ|m /∂ξi ∂ξj )ni,j=1 on the support of ϕ
(suppϕ ⊂ {ξ : |ξ| ∈ (2−1 , 2)}). It is easy to see that

n − 1, m = 1,
ρ= (3.8)
n, m > 2.
So,
m
kF −1 eit|ξ| F 4k f k∞ . t−ρ/2 2k(n−mρ/2) kf k1 . (3.9)
Obviously, by Plancherel’s identity,
m
kF −1 eit|ξ| F 4k f k2 6 kf k2 . (3.10)
In view of Riesz-Thorin’s interpolation theorem, (3.9) and (3.10) imply that
for any 2 6 p 6 ∞, 1/p + 1/p0 = 1,
2−k(2n−mρ)(1/2−1/p) kUm (t)4k f kp . t−ρ(1/2−1/p) kf kp0 . (3.11)
For convenience, we write
2σ(m, p) := (2n − mρ)(1/2 − 1/p). (3.12)
P1
Substituting f by `=−1 4k+` f in (3.11), we get
1
X
2−2σ(m,p)k kUm (t)4k f kp . t−ρ(1/2−1/p) k4k+` f kp0 . (3.13)
`=−1
Taking the `q -norm in (3.13), we have

kUm (t)f kḂ −2σ(m,p) . t−ρ(1/2−1/p) kf kḂ 00 , 2 6 p 6 ∞. (3.14)
p,q p ,q
s s
Using the inclusion Ḃp,2 ⊂ Ḟp,2 = Ḣps and Ḃps0 ,2 ⊃ Ḟps0 ,2 = Ḣps0 , we have
kUm (t)f kḢ −2σ(m,p) . t−ρ(1/2−1/p) kf kp0 , 2 6 p < ∞. (3.15)
p
Taking m = 1, 2, we immediately have
Proposition 3.1. Let n > 2, W (t) = F −1 eit|ξ| F , 2 6 p < ∞, 1 6 q 6 ∞,

and 1/p + 1/p0 = 1. Then
kW (t)f kḂ −(n+1)(1/2−1/p) . t−(n−1)(1/2−1/p) kf kḂ 00 , (3.16)
p,q p ,q
kW (t)f kḢ −(n+1)(1/2−1/p) . t−(n−1)(1/2−1/p) kf kp0 . (3.17)

p
2
Proposition 3.2. Let n > 1, S(t) = F −1 eit|ξ| F , 2 6 p < ∞, 1 6 q 6 ∞,
and 1/p + 1/p0 = 1. Then we have
kS(t)f kḂ 0 . t−n(1/2−1/p) kf kḂ 00 , (3.18)
p,q p ,q
kS(t)f kp . t−n(1/2−1/p) kf kp0 . (3.19)
In the higher order case m > 2, we have
Proposition 3.3. Let n > 1, m > 2, 2 6 p < ∞, 1 6 q 6 ∞, and

1/p + 1/p0 = 1. Then we have
kUm (t)f kḂ −2σ(m,p) . t−n(1/2−1/p) kf kḂ 00 , (3.20)
p,q p ,q
kUm (t)f kḢ −2σ(m,p) . t−n(1/2−1/p) kf kp0 . (3.21)

p
Noticed that for m > 2, 2σ(m, p) = n(2 − m)(1/2 − 1/p) < 0, Um (t) gains
some regularities in Proposition 3.3.
Below, we consider another decay estimate of Um (t) without the smooth-
ness in the case m > 2. We have
Proposition 3.4. Let n > 1, m > 2, 2 6 p < ∞, and 1/p + 1/p0 = 1.

Then
0
kUm (t)f kp . t−n(1/p −1/p)/m kf kp0 . (3.22)
0
Proof. In view of the dilation property, it suffices to consider the case

t = 1. Let {ϕk }∞
k=0 be as in Sec. 1.3. (3.13) implies that for any k > 1,
1
X
kUm (1)4k f kp . 22σ(m,p)k k4k+` f kp0 . (3.23)
`=−1
By supp ϕ0 ⊂ {ξ : |ξ| 6 2}, we have

1
X
kUm (1)40 f kp 6 kUm (1)40 4` f kp
`=0
X1 1
X
. kϕ0 F 4` f kp0 . k4` f kp0 . (3.24)
`=0 `=0
We write 4−1 = 0. It follows that (3.23) holds for all k > 0. Taking the `2
in (3.23), one has that
kUm (1)f kBp,2
0 . kf kB 00 . (3.25)
p ,2
Similar to (3.15), (3.25) implies that

kUm (1)f kp . kf kp0 . (3.26)
In view of the dilation property, we obtain the result, as desired.
Using (3.21), (3.22) and the convexity Hölder inequality (Proposition
1.21), we immediately have
Proposition 3.5. Let n > 1, m > 2, 2 6 p < ∞, and 1/p + 1/p0 = 1.

Then for any θ ∈ [0, 1],
kUm (t)f kḢ −2σ(m,p)θ . t−(nθ+2n(1−θ)/m)(1/2−1/p) kf kp0 . (3.27)
p
The above method is also valid for some other homogeneous non-radial
functions, say P (ξ) = ξ14 ± ... ± ξk4 , k 6 n. Noticing that, if P (ξ) is not
a homogeneous function, (3.7) can not be obtained by scaling, we need to
look for another way to handle the nonhomogeneous case.
Our idea is to simplify P (·) as a radial function P (ξ) := P (|ξ|), which
is essentially reduced to one dimensional case. P (|ξ|) can be separated into
two parts, lower and higher frequency parts, which correspond to |ξ| . 1
and |ξ| 1, respectively. We further assume that P (ξ) has a different
growth as |ξ| . 1 and |ξ| 1, which is sufficient for many semi-groups.
In what follows, we always assume that P : (0, ∞) → R is a smooth
radial function satisfying
(H1) there exists an m1 > 0, such that for any α > 2, α ∈ N,

|P 0 (r)| ∼ rm1 −1 , |P (α) (r)| . rm1 −α , r > 1;
(H2) there exists an m2 > 0, such that for any α > 3, α ∈ N,
|P 0 (r)| ∼ rm2 −1 , |P (α) (r)| . rm2 −α , 0 < r < 1;
(H3) there exists an α1 such that
|P 00 (r)| ∼ rα1 −2 , r > 1;
(H4) there exists an α2 such that
|P 00 (r)| ∼ rα2 −2 , 0 < r < 1.
Remark 3.1. The following are somep examples for which conditions (H1)–
(H4) are satisfied: (1) P (ξ) = 1 + |ξ|2 (relevant to the Klein-Gordon
4 2
semi-group); (2) P (ξ) = |ξ|
p + |ξ| (relevant to the fourth order Schrödinger
semi-group); (3) P (ξ) = 1 + |ξ|4 (relevant
p to the beam semi-group).
Comparing P (|ξ|) = |ξ| with P (|ξ|) = 1 + |ξ|2 , althoughp
they have the
2
same growth as t → ∞, we easily see that the rank of (∂ij 1 + |ξ|2 )n×n
p
2
is n and 1 + |ξ| is better than |ξ| at ξ = 0. So, we can expect that the
0
Klein-Gordon semi-group has a better Lp → Lp decay.
We first consider 1D case, which is easier to find the ideas. The following
result is due to [94].
Proposition 3.6. Let n = 1 and U (t) = F −1 eitP (|ξ|) F . We have the

following results.
(a) Let {4k }k∈Z be defined in Sec. 1.6. Then for any k > 0,
kU (t)4k u0 k∞ . 2k ku0 k1 .
Moreover, if P satisfies (H3), then
kU (t)4k u0 k∞ . |t|−θ/2 2k(1−α1 θ/2) ku0 k1 , 0 6 θ 6 1.
(b) Let {4k }k∈Z be as in Sec. 1.6. Then for any k < 0, one has that
kU (t)4k u0 k∞ . 2k ku0 k1 .
kU (t)4k u0 k∞ . |t|−θ/2 2k(1−α2 θ/2) ku0 k1 , 0 6 θ 6 1.
(c) Let 40 be as in Sec. 1.3. Assume that P satisfies (H2) and (H4)
with m2 = α2 , then

−θ 1 1
kU (t)40 u0 k∞ . (1 + |t|) ku0 k1 , 0 6 θ 6 min , .
m2 2
0
We need the following Van der Corput lemma, see Appendix.
Lemma 3.2. Let ϕ ∈ C0∞ (R), and P ∈ C 2 (R) satisfy |P 00 (ξ)| > λ > 0 for
all ξ ∈ supp ϕ. Then
Z

eiP (ξ) ϕ(ξ)dξ . λ−1/2 (kϕk∞ + kϕ0 k1 ).

Proof. [Proof of Proposition 3.6] First, we prove (a). Using Young’s

inequality, one has that
kU (t)4k u0 k∞ . kJk k∞ ku0 k1 ,
where
Jk (x) = F −1 (eitP (|ξ|) ϕ(2−k |ξ|))(2−k x). (3.28)
By (3.28), we have
kJk k∞ . 2k . (3.29)
k
Denote P1 (ξ) = xξ + tP (2 |ξ|). We have |P100 (ξ)| & |t|2 kα1
for ξ ∈ supp ϕ.
Using Van der Corput lemma, we get
kJk k∞ . |t|−1/2 2k(1−α1 /2) . (3.30)
Making an interpolation between (3.29) and (3.30), we get that for any
0 6 θ 6 1,
kJk k∞ . |t|−θ/2 2k(1−θα1 /2) .
It follows that (a) holds.
The proof of (b) is analogous to (a) and the details will be omitted. We
now prove (c). In view of the first conclusion in (b), we have
X
kU (t)40 u0 k∞ . 2k ku0 k1 . ku0 k1 . (3.31)
k<0
We first consider the case m2 < 2. Since min(1/m2 , 1/2) = 1/2, from (b)
it follows that
X
kU (t)40 u0 k∞ . |t|−1/2 2k(1−m2 /2) ku0 k1 . |t|−1/2 ku0 k1 . (3.32)
k<0
From (3.31) and (3.32), we can get the conclusions.

Next, we discuss the case m2 > 2. It suffices to consider the case m2 > 2
and θ = min(1/m2 , 1/2) = 1/m2 . A straightforward calculation yields

d 1
. 2−k(m2 −1) , ξ ∈ suppϕ. (3.33)
dξ P 0 (2k |ξ|)
So, if |x| 6 1, then we have |∂ξm (eixξ ϕ(ξ))| . 1. Integrating by part to

Jk (·), we have
|Jk (x)| . |t|−1 2k(1−m2 ) .
If |x| > 1, let k0 be the minimal integer so that |x| 6 |t|2k0 m2 , then |x| ≈
|t|2k0 m2 . For |k − k0 | > C 1, we have |P10 (ξ)| > c|t|2km2 . Integrating by
part to Jk (·), we have
|Jk (x)| . |t|−1 2k(1−m2 ) .
For |k − k0 | 6 C, noticing that |x| > 1 and m2 > 2, we have
(1−m2 /2)/m2
|x|
|Jk (x)| . |t|−1/2 2k(1−m2 /2) . |t|−1/2 . |t|−1/m2 .
|t|
So,
X X X
| Jk (x)| . |Jk (x)| + |Jk (x)|
k60 |k−k0 |6C |k−k0 |>C
X X X
. |t|−1/m2 + 2k + |t|−1 2k(1−m2 )
|k−k0 |6C 2k <|t|−1/m2 2k >|t|−1/m2
−1/m2
. |t| ,
which finishes the proof of (c).
In higher spatial dimensions, we have
Proposition 3.7. Let n > 2. Denote U (t) = F −1 eitP (|ξ|) F . We have the
following decay estimates.
(a) Let {4k }k∈Z be as in Sec. 1.6, k > 0, and P satisfy (H1). Then
n−1
kU (t)4k u0 k∞ . |t|−θ 2k(n−m1 θ) ku0 k1 , 0 6 θ 6 . (3.34)
2
m1 n α1 −m1
kU (t)4k u0 k∞ . |t|−n/2 2k(n− 2 − 2 )
ku0 k1 . (3.35)
(b) Let {4k }k∈Z be as in Sec. 1.6, k < 0, and P satisfy (H2). Then
we have
n−1
kU (t)4k u0 k∞ . |t|−θ 2k(n−m2 θ) ku0 k1 , 0 6 θ 6 . (3.36)
2
m2 n α2 −m2
kU (t)4k u0 k∞ . |t|−n/2 2k(n− 2 − 2 )
ku0 k1 . (3.37)
0
(c) Let 40 be as in Sec. 1.3, and P satisfies (H2). Then we have

−θ n n−1
kU (t)40 u0 k∞ . (1 + |t|) ku0 k1 , 0 6 θ 6 min , .
m2 2
(3.38)
Moreover, if P satisfies (H4) with α2 = m2 , then

n n
kU (t)40 u0 k∞ . (1 + |t|)−θ ku0 k1 , 0 6 θ 6 min , .
m2 2
(3.39)
Proof. The idea is to use the Bessel function. After making the polar
coordinate transform, the estimate is reduced to an oscillating integration
in one spatial dimension. Next, using the decay and cycle properties of the
Bessel function, analogous to the 1D case, we can get the conclusion. Let
Jm (r) be the Bessel function
Z 1
(r/2)m
Jm (r) = eirt (1 − t2 )m−1/2 dt, m > −1/2.
Γ(m + 1/2)π 1/2 −1
Let us state some properties on the Bessel functions, see [82] and [203].
Lemma 3.3. For any 0 < r < ∞, we have

(i) Jm (r) 6 Crm ,
d
(ii) dr (r−m Jm (r)) = −r−m Jm+1 (r).
(i) is obvious. Integrating by part to Jm , we can get (ii). It is known that,

the Fourier transform for a radial f is also radial (see [202]):
Z ∞
ˆ
f (ξ) = 2π f (r)rn−1 (r|ξ|)−(n−2)/2 J n−2 (r|ξ|)dr. (3.40)
2
0
By Lemma 3.3, for any 0 6 s 6 2 and any k > 0,
k
∂
(ϕ(r)r n−1
(rs)−(n−2)/2
J n−2 (rs)) 6 Ck . (3.41)
∂rk 2
If m = − n−2 [ ]
2 , Jm (r) has the following property (see 117, Chapter 1, (1.5) )
n−2
r− 2 J n−2 (r) = cn R(eir h(r)), (3.42)
2
where h satisfies
n−1
|∂rk h(r)| 6 ck (1 + r)− 2 −k . (3.43)
So, for any s > 2 and k > 0,
n−1
|∂rk ϕ(r)rn−1 h(rs) | 6 ck s− 2 . (3.44)
Now we prove (a). By Young’s inequality,

kU (t)4k u0 k∞ . kF −1 eitP (|ξ|) ϕ(2−k |ξ|)k∞ ku0 k1 .
Using (3.40), we have
F −1 (eitP (|ξ|) ϕ(2−k |ξ|))(x)
k
= 2kn F −1 (eitP (|2 ξ|) ϕ(|ξ|))(2k |x|)
Z ∞
k
= 2kn eitP (2 r) ϕ(r)rn−1 (r2k s)−(n−2)/2 J n−2 (r2k s)dr
2
0
k
=: Ik (2 s),
where s = |x| and J n−2 (r) denotes the Bessel function. It suffices to show
2
kIk (s)k∞ 6 |t|−θ 2k(n−m1 θ) .

By (i) of Lemma 3.3,
kIk (s)k∞ . 2kn . (3.45)
We consider the following two separated cases.
k
Case 1. s 6 2. Denote Dr = ( itP 0 (21k r)2k ) dr
d
. We have Dr (eitP (2 r) ) =
k
eitP (2 r)
. By the condition (H1), for any m > 0 and r ∼ 1,

dm 1
6 Cm 2−k(m1 −1) . (3.46)
drm P 0 (2k r)
We write ϕ̃(r) = ϕ(r)rn−1 . Integrating by part to Ik , for any q ∈ Z+ , we
have Z ∞
k n−2
Ik (s) = 2kn eitP (2 r) ϕ̃(r)(rs)− 2 J n−2 (rs)dr
2
Z0 ∞
k n−2
= 2kn Dr (eitP (2 r) )ϕ̃(r)(rs)− 2 J n−2 (rs)dr
2
0
kn Z ∞

2 itP (2k r) d 1 − n−2
=− k e ϕ̃(r)(rs) 2 J n−2 (rs) dr
it2 0 dr P 0 (2k r) 2
q
2kn X X
= k q
Cq,m
(it2 ) m=0 q
l1 ,...lq ∈Λm
Z ∞ q
Y
itP (2k r) 1 n−2
× e ∂rlj ∂rq−m ϕ̃(r)(rs)− 2 J n−2 (rs) dr,
0 j=1
P 0 (2k r) 2
(3.47)
where Λqm +
= {l1 , . . . , lq ∈ Z : 0 6 l1 < . . . < lq 6 q, l1 + . . . lq = m}. By
(3.41), (3.46) and (3.47), we get that for any q ∈ Z+ ,
|Ik (s)| . |t|−q 2k(n−m1 q) . (3.48)
0
Interpolating (3.48) with (3.45), we get for any θ > 0, |Ik (s)| .
|t|−θ 2k(n−m1 θ) .
Case 2. s > 2. By (3.42),
Z ∞
k
Ik (s) = cn 2kn eitP (2 r) ϕ̃(r)(eirs h(rs) + e−irs h̄(rs))dr
Z0 ∞
k
kn
= cn 2 ei(tP (2 r)+rs) ϕ̃(r)h(rs)dr
0
Z ∞
k
+ cn 2kn ei(tP (2 r)−rs) ϕ̃(r)h̄(rs)dr
0
=: B1 + B2 .
We can assume, without loss of generality that t > 0 and P 0 (r) > 0. We
consider the estimate of B1 . Put P1 (r) = tP (2k r) + rs. Noticing that
P10 (r) = t2k P 0 (2k r) + s > ct2km1 , we see that (3.46) also holds if one
replaces P by P1 . In view of (3.44), analogous to Case 1, we can get that
for any θ > 0,
|B1 | . |t|−θ 2k(n−m1 θ) .
We now consider the estimate of B2 . Put P2 (r) = tP (2k r) − rs. Notice
that if s = t2k P 0 (2k r), then P20 (r) = 0. We divide Case 2 into two subcases.
Case 2a. s > 2 supr∈[1/2,2] t2k P 0 (2k r), or s < 21 inf r∈[1/2,2] t2k P 0 (2k r).
It is easy to see that |P20 (r)| > ct2km1 if r ∼ 1 and (3.46) still holds if one
substitutes P by P2 . Using (3.44), we have for any θ > 0,
|B2 | . |t|−θ 2k(n−m1 θ) .
1
Case 2b. 2 inf r∈[1/2,2] t2k P 0 (2k r) 6 s 6 2 supr∈[1/2,2] t2k P 0 (2k r). Us-
ing (3.44),
n−1 n−1 (n−1)m1
|B2 | . 2kn s− 2 . t− 2 2k(n− 2 )
. (3.49)
Making an interpolation between (3.49) and (3.45), we get that for any
0 6 θ 6 n−1
2 ,
|B2 | 6 t−θ 2k(n−m1 θ) . (3.50)

If (H3) holds, then |P200 (r)| > t2kα1 . Using Van der Corput lemma,
Z ∞
kα1 −1/2 d n α1 −m1
|B2 | . (t2 ) | (ϕ̃(r)h(rs))|dr . t−n/2 2k(n− 2 (m1 + n )) .
0 dr
(3.51)
Thus, we finish the proof of (a).
The proof of (b) is similar to that of (a) and we omit the details. Finally,
we prove (c). Let 0 6 θ 6 min( mn2 , n−1 n
2 ). If θ < m2 , then n − m2 θ > 0. In
view of (b), we immediately have
X2
kU (t)40 u0 k∞ . |t|−θ 2k(n−m2 θ) k40 u0 k1
k=−∞
. |t|−θ k40 u0 k1 .
Assume that n−1 n n
2 > m2 and θ = m2 . From the proof of (b), we see that
k0 m2
for k0 < 0 and s ∼ t2 > 2, there holds
n−1 (n−1)m2 n
|Ik0 (s)| . t− 2 2k0 (n− 2 ) . t− m2 .
If |k − k0 | > C 1, then
|Ik (s)| . t−α 2k(n−m2 α) , ∀ α > 0.
So, choosing α large enough, we have
X X
|I60 (s)| . |Ik (s)| + |Ik (s)|
|k−k0 |6C |k−k0 |>C
− mn
X X
.t 2 + 2kn + t−α 2k(n−m2 α)
− 1 − 1
2k <t m2 2k >t m2
n
−
. t m2 ,
which implies the result, as desired. If (H4) holds and m2 = α2 , the proof
is analogous and the details are omitted.
Using the above estimates to the Klein-Gordon equation
utt + u − ∆u = f (t, x), u(0, x) = u0 (x), ut (0, x) = u1 (x),
1/2
which corresponds to the semi-group G(t) = eit(I−∆) , we have the fol-
lowing
1/2
Proposition 3.8. Let G(t) = eit(I−∆) , θ ∈ [0, 1], 2σ(θ, ·) = (n + 1 +
θ)(1/2 − 1/·), and 2/β(θ, ·) = (n − 1 + θ)(1/2 − 1/·). Then we have
kG(t)f kB s−2σ(θ,p) . |t|−2/β(θ,p) kf kBps0 ,q .
p,q
Remark 3.2. If the wave equation contains a damping term, say

utt − ∆u + αut = 0, u(0) = u0 , ut (0) = u1 ,
where α > 0, the generating semi-group is quite different from the case
2
α = 0 and it is similar to F −1 e−αt/2−it|ξ|−iα t/|ξ| F . In fact, there are
many works have been devoted to study the decaying estimates for its
solutions; cf. W.K. Wang and Yang [248], W.K. Wang and W.J. Wang [249],
Ikehata, Nishihara and Zhao [112], Hayashi, Kaikina and Naumkin [103;
104], for details.
3.2. Strichartz inequalities: dual estimate techniques 63
3.2 Strichartz inequalities: dual estimate techniques
Slip of bamboo for writing is eventually replaced by paper.
In this section we study the Strichartz estimates for a class of dispersive

semi-groups. Such kinds of estimates go back to the pioneer work of R.
S. Strichartz [206] in 1977 for the wave semi-group, afterwards there are
a series of generalizations, see Kato [120], Cazenave and Weissler [26] for
the Schrödinger equation; Pecher [191], Ginibre and Velo [80] for the wave
equation; Brenner [22; 21] for the Klein-Gordon equation; Kenig-Ponce-
Vega [126] for KdV and more general dispersive equations. Below, we give
a unified approach to the Strichartz inequalities based on the dual estimate
method, see [235; 234], which simplifies various Strichartz estimates for the
above mentioned semi-groups. As corollaries, the Strichartz estimates for
a class of higher order semi-groups are also obtained. Denote
Z t
−1 itP (ξ)
U (t) = F e F, A = U (t − τ )f (τ, ·)dτ, (3.52)
0
n
where P (·) : R → R is a smooth function. In what follows we always
assume that
X = Lp , 0
or X = Bp,2 , 2 6 p < ∞. (3.53)
Assume that U (t) satisfies the following estimate
kU (t)f kX α . t−θ kf kX ∗ , (3.54)
where α ∈ R, θ ∈ (0, 1), X α := (I − 4)−α/2 X, and X ∗ is the dual space of

X. Supposing that (3.54) holds, we can get some interesting estimates for
U (·) and A. Using (3.52) and (3.54), we have
Z t
kAf kX α . |t − τ |−θ kf (τ )kX ∗ dτ. (3.55)
0
Using the Hardy-Littlewood-Sobolev inequality, we immediately obtain

that
Lemma 3.4. Assume that (3.53) and (3.54) are satisfied. For any T > 0
and s ∈ R, we have
kAf kL2/θ (−T,T ;X s+α ) . kf kL(2/θ)0 (−T,T ;(X ∗ )s ) , (3.56)
where (2/θ)0 denotes the conjugate number of 2/θ, i.e., θ/2 + 1/(2/θ)0 = 1.
For convenience, we denote by DT the set of all of the functions which

are defined in (−T, T ), valued in S and taken finitely many values, I =
(−T, T ). It is easy to see that for 2 6 p < ∞, DT is dense in Lq (I, (X ∗ )s )
(s ∈ R, 1 6 q < ∞).
Lemma 3.5. Assume that (3.53) and (3.54) are satisfied. Then
kU (t)f kL2/θ (R,X s+α/2 ) . kf kH s . (3.57)
Proof. First, we show that for any T > 0, I = (−T, T ), ϕ ∈ S , ψ ∈ DT ,

Z
T

(U (t)ϕ, ψ(t))dt . kϕk2 kψkL(2/θ)0 (I,(X ∗ )−α/2 ) . (3.58)
−T
In fact,
Z Z
T T

(U (t)ϕ, ψ(t))dt . kϕk2 U (−t)ψ(t)dt . (3.59)
−T −T
2
By Lemma 3.4,
Z 2
T

U (−t)ψ(t)dt
−T
2
Z Z T !
T

= ψ(t), U (t − τ )ψ(τ )dτ dt
−T −T
Z
T

. kψkL(2/θ)0 (I,(X ∗ )−α/2 ) U (t − τ )ψ(τ )dτ
−T
L2/θ (I,X α/2 )
. kψk2L(2/θ)0 (I,(X ∗ )−α/2 ) . (3.60)

Combining (3.59) with (3.60), we see that (3.58) holds. Since S is dense
0
in L2 , DT is dense in L(2/θ) (I, (X ∗ )−α/2 ), by (3.58) we immediately have
kU (t)ϕkL2/θ (I,X α/2 ) . kϕk2 . (3.61)
Noticing that the above estimates are independent of T , and letting T → ∞,
we get that those estimates also hold if (−T, T ) is substituted by R. Taking
ϕ = (I − ∆)s/2 f , we can get the result, as desired.
Lemma 3.6. Assume that (3.53) and (3.54) are satisfied. Then for any
T > 0, I = (−T, T ),
kAf kL∞ (I,H s+α/2 ) . kf kL(2/θ)0 (I,(X ∗ )s ) . (3.62)
Proof. Analogous to Lemma 3.5, it suffices to show that for any f ∈ DT ,

kAf kL∞ (I,H α/2 ) . kf kL(2/θ)0 (I,X ∗ ) . (3.63)
s/2
Let us write Js = (I − ∆) . We have
kAf k2H α/2 = (AJα/2 f, AJα/2 f )
Z t

. kf kL(2/θ)0 (I,X ∗ )
U (· − τ )f (τ )dτ
. (3.64)
0 L2/θ (I,X α )
Applying the same techniques as in Lemma 3.4, one has that
Z t

U (· − τ )f (τ )dτ . kf kL(2/θ)0 (I,X ∗ ) . (3.65)
2/θ
0 αL (I,X )
Combining (3.64) and (3.65), we have (3.63).
Lemma 3.7. Assume that (3.53) and (3.54)are satisfied. Then for any
T > 0, I = (−T, T ),
kAf kL2/θ (I,X s+α/2 ) . kf kL1(I,H s ) . (3.66)
Proof. Analogous to Lemma 3.5, it suffices to prove that for any f ∈ DT ,
kAf kL2/θ (I,X α/2 ) . kf kL1 (I,L2 ) . (3.67)
Let ψ, f ∈ DT , we have
Z Z
T T

(Af (t), ψ(t))dt . kf kL1 (0,T ;L2 ) U (· − t)ψ(t)dt . (3.68)
0 ·
L∞ (I,L2 )
Similar to Lemma 3.6, we have
Z
T

U (· − t)ψ(t)dt . kψkL(2/θ)0 (I,(X ∗ )−α/2 ) . (3.69)
· ∞
L (I,L2 )
R0
One can similarly estimate −T (Af (t), ψ(t))dt. So, (3.68) and (3.69) imply
that
Z T

(Af (t), ψ(t))dt . kf kL1(0,T ;L2 ) kψkL(2/θ)0 (I,(X ∗ )−α/2 ) . (3.70)
0
Using (3.70) and the duality, we can directly obtain (3.67).
Remark 3.3. If (3.54) is replaced by
kU (t)f kẊ α . t−θ kf kẊ ∗ , (3.71)
p 0
where Ẋ = L or Ẋ = Ḃp,2 ,
then the results in Lemmas 3.4–3.7 also hold
if one replaces the nonhomogeneous spaces by corresponding homogeneous
spaces, for instance, the substitution of (3.56) in Lemma 3.5 is that
kU (t)f kL2/θ (I,Ẋ s+α/2 ) . kf kḢ s . (3.72)
We omit the details of those results.
Now we apply the above results to a class of linear dispersive equations.

We can use (3.14) and (3.15) to get some general results, however, every
specific semi-group seems to be important in applications and so, we sepa-
rately discuss every concrete semi-group, say S(t), W (t) and Um (t). First,
we consider the Schrödinger semi-group S(t) = eit∆ and the higher order
m/2
semi-group Um (t) = eit(−∆) :
Theorem 3.1. Let m > 2,

∗ ∞, n 6 m,
m = (3.73)
2n/(n − m), n > m,

1 n 1 1
= − . (3.74)
γ(·) m 2 ·
m/2
Assume that 2 6 r, p < m∗ , Um (t) = eit(−∆) . Then we have
kUm (t)φkLγ(p) (I,Ḃ s . kφkḢ s , (3.75)
p,2 )
kAUm f kLγ(p) (I,Ḃ s . kf kLγ(r)0 (I,Ḃ s0 ), (3.76)

p,2 ) r ,2
Rt
where I ⊂ R is an interval, AUm := 0 Um (t − τ ) · dτ . In (3.75) and (3.76),
replacing homogeneous Besov spaces by corresponding Besov spaces, Bessel
potential spaces and Riesz potential spaces, respectively, the conclusions still
hold.
It is known that the solution of the wave equation

utt − ∆u = f (t, x), u(0, x) = u0 (x), ut (0, x) = u1 (x) (3.77)
is relevant to the semi-group W (t), we have
1/2
Theorem 3.2. Let W (t) = eit(−∆) , n > 2,

∗∗ ∞, n = 2, 3,
2 = (3.78)
2(n − 1)/(n − 3), n > 3,
2σ(·) 2 1 1
= = − . (3.79)
n+1 (n − 1)β(·) 2 ·
∗∗
Let 2 6 r, p < 2 . Then we have
kW (t)φkLβ(p) (I,Ḃ s−σ(p) ) . kφkḢ s , (3.80)
p,2
kAW f kLβ(p) (I,Ḃ s−σ(p) ) . kf kLβ(r)0 (I,Ḃ s+σ(r) ) , (3.81)

p,2 r0 ,2
Rt
where I ⊂ R is an arbitrary interval, AW := 0 W (t − τ ) · dτ . In (3.80)
and (3.81), substituting homogeneous Besov spaces by corresponding Riesz
potential spaces, the results also hold.
The solution of the Klein-Gordon equation
utt + u − ∆u = f (t, x), u(0, x) = u0 (x), ut (0, x) = u1 (x) (3.82)

1/2
is relevant to the semi-group G(t) = eit(I−∆) , we have the following
1/2
Theorem 3.3. Let G(t) = eit(I−∆) and θ ∈ [0, 1]. For n > 3 − θ,
we write 2∗∗ = 2(n − 1 + θ)/(n − 3 + θ) and for n 6 3 − θ, we denote
2∗∗ = ∞. Let 2 6 r, p < 2∗∗ , 2σ(θ, ·) = (n + 1 + θ)(1/2 − 1/·) and
2/β(θ, ·) = (n − 1 + θ)(1/2 − 1/·). Then we have the following estimates:
kG(t)φkLβ(θ,p) (I,B s−σ(θ,p) ) . kφkH s , (3.83)

p,2
kAG f kLβ(θ,p) (I,B s−σ(θ,p) ) . kf kLβ(θ,r)0 (I,B s+σ(θ,r) ) , (3.84)

p,2 r0 ,2
Rt
where I = (−T, T ) ⊂ R is arbitrary, AG := 0 G(t − τ ) · dτ . In (3.83) and
(3.84), replacing Besov spaces by corresponding Bessel potential spaces, the
results also hold.
For the higher order Schrödinger semi-group, if we consider the smooth

effect, we have
m/2
Theorem 3.4. Let Um (t) = eit(−∆) , m > 2 and 2∗ be as in (3.73). We
write

2 1 1
=n − , (3.85)
γ(·) 2 ·

1 1
2σ(m, ·) = n(2 − m) − . (3.86)
2 ·
Let 2 6 r, p < 2∗ . Then we have
kUm (t)φkLγ(p) (I,Ḃ s−σ(m,p) ) . kφkḢ s , (3.87)
p,2
kAUm f kLγ(p) (I,Ḃ s−σ(m,p) ) . kf kLγ(r)0 (I,Ḃ s+σ(m,r) ) , (3.88)

p,2 r0 ,2
Rt
where I ⊂ R is arbitrary, AUm := 0 Um (t−τ )·dτ . In (3.87) and (3.88), re-
placing homogeneous Besov spaces by corresponding Riesz potential spaces,
the conclusions also hold.
Proof. The proofs of Theorems 3.1–3.4 are analogous and we only prove
Theorem 3.4. By Proposition 3.3,
kUm (t)f kḂ −2σ(m,p) . t−2/γ(p) kf kḂ 00 . (3.89)

p,2 p ,2
0
Putting Ẋ = Ḃp,2 and α = −2σ(m, p), in view of Lemma 3.5 we can get
(3.87). Now we show that (3.88) holds. By Lemmas 3.4, 3.6, 3.7 and
Remark 3.3,
kAUm f kLγ(p) (I,Ḃ s−σ(m,p) ) . kf kLγ(p)0 (I,Ḃ s+σ(m,p) ) , (3.90)
p,2 p0 ,2
kAUm f kL∞ (I,Ḣ s ) . kf kLγ(p)0 (I,Ḃ s+σ(m,p) ) , (3.91)

p0 ,2
kAUm f kLγ(p) (I,Ḃ s−σ(m,p) ) . kf kL1 (I,Ḣ s ) . (3.92)

p,2
Case I. p ∈ [2, r]. Taking θ ∈ [0, 1] such that 1/p = (1 − θ)/2 + θ/r, we
deduce that
1 θ 1−θ
= + , σ(m, p) = θσ(m, r) + (1 − θ)σ(m, 2). (3.93)
γ(p) γ(r) ∞
In view of the convexity Hölder inequality, (3.90) and (3.91) imply that
kAUm f kLγ(p) (I,Ḃ s−σ(m,p) ) 6 kAUm f k1−θ
L∞ (I,Ḃ )
θ
s−σ(m,2) kAUm f k γ(r)
L (I,Ḃ
s−σ(m,r)
)
p,2 2,2 r,2
. kf kLγ(r)0 (I,Ḃ s+σ(m,r) ) . (3.94)

r0 ,2
Case II. p > r > 2. Take θ ∈ (0, 1) such that 1/r0 = (1 − θ)/2 + θ/p0 . It
follows that
1 θ 1−θ
0
= 0
+ , σ(m, r) = θσ(m, p) + (1 − θ)σ(m, 2). (3.95)
γ(r) γ(p) 1
Noticing the complex interpolation,
0 s+σ(m,p) 0 s+σ(m,r)
(Lγ(p) (I, Ḃp0 ,2 ), L1 (I, Ḃ2,2
s
))[θ] = Lγ(r) (I, Ḃr0 ,2 ), (3.96)
(3.90) and (3.92) imply that
0 s+σ(m,r) s−σ(m,p)
AUm : Lγ(r) (I, Ḃr0 ,2 ) → Lγ(p) (I, Ḃp,2 ) (3.97)
is a bounded operator. Theorem 3.4 follows.
3.3 Strichartz estimates at endpoints
Let us recall that the starting point in previous section is the estimate
(3.54), where we always assume θ ∈ (0, 1). We consider in this section the
case θ = 1. Assume that U (t) is defined in (3.52), and for any 2 0 such that
kU (t)f kH α(p) . t−θ(p) kf kp0 , ∀ 2 6 p < ∞. (3.98)
p
3.3. Strichartz estimates at endpoints 69
Suppose that there exists a p1 > 2 satisfying θ(p1 ) > 1, then there exists a
r ∈ (2, p1 ) such that θ(r) = 1 (see (3.118)), i.e.,
kU (t)f kH α(r) . t−1 kf kr0 . (3.99)
r
(3.99) is said to be an endpoint estimate. Recall that the condition θ < 1 in

previous section aries from the Hardy-Littlewood-Sobolev (HLS) inequality,
which is essential for (3.55) and (3.56). This is why we say that θ(r) = 1
is the endpoint case. A natural question is what happens in the endpoint
case θ(r) = 1. If fact, if (3.99) occurs, then the Strichartz estimates still
hold, i.e.,
kU (t)φkL2 (I,Br,2
s ) . kφk s−α(r)/2 .
H (3.100)
This estimate was essentially obtained by Keel and Tao [121],3 where they
used the techniques of the interpolation on bilinear operators.
According to the proof of (3.57), it suffices to show
Z
T

(U (t)ϕ, ψ(t))dt . kϕk2 kψkL2 (I,B −α(r)/2 ) . (3.101)
−T r0 ,2
In order to prove (3.101), analogous to (3.59) and (3.60), one needs to show
that
Z 2
T

U (−t)ψ(t)dt . kψk2L2 (I,B −α(r)/2 ) . (3.102)
−T r0 ,2
2
The left hand side of (3.102) can be rewritten as

Z T Z T

U (−s)ψ(s), U (−t)ψ(t) dsdt. (3.103)
−T −T
By (3.103), it is natural to introduce the following bilinear operator

ZZ

L(F, G) := U (−s)F (s), U (−t)G(t) dsdt, (3.104)
D
where
D := {(s, t) : s, t ∈ [−T, T ], s 6 t}. (3.105)
Our goal is to prove that
|L(F, G)| . kF kL2 (I,B −α(r)/2 ) kGkL2 (I,B −α(r)/2 ) . (3.106)
r0 ,2 r0 ,2
3 Keel and Tao did not consider the case α(r) 6= 0, here our proof is a modification of
Keel and Tao’s proof.

Noticing that s and t in (3.102) have equal positions, and taking F = G =

ψ, we have from (3.106) that (3.100) holds.
Now the question reduces to show the bilinear estimate (3.106). Keel
and Tao’s idea is to use the dyadic decomposition on t − s to cancel the
singularity at t = s. Put
Dj := {(s, t) ∈ D : T 2j < t − s 6 T 2j+1 }. (3.107)
It is easy to see that D = ∪j60 Dj . Denote
ZZ

Lj (F, G) := U (t − s)F (s), G(t) dsdt. (3.108)
Dj
So, it suffices to show that

X
|Lj (F, G)| . kF kL2 (I,B −α(r)/2 ) kGkL2 (I,B −α(r)/2 ) . (3.109)
r0 ,2 r0 ,2
j
In the following we estimate Lj (F, G).
Lemma 3.8. Assume that (3.98) and (3.99) are satisfied, and P =
(1/r, 1/r). Let (1/a, 1/b) ∈ B(P, ε) and ε > 0 is sufficiently small. Then
we have
|Lj (F, G)| . (2j T )−β(a,b) kF kL2 (I,H −α(a)/2 ) kGkL2 (I,H −α(b)/2 ) , (3.110)
a0 b0
where
1
β(a, b) = (θ(a) + θ(b)) − 1. (3.111)
2
Proof. We can assume that F and G are Schwartz functions which have
compact support contained in [−T, T ] on the time variable. We will consider
the following three cases:
(1) a = b = p ∈ (2, ∞);
(2) 2 6 a < r, b = 2;
(3) a = 2, 2 6 b < r.
First, we consider case (1). Using Young’s and Hölder’s inequalities,
Z Z
|Lj (F, G)| . (t − s)−θ(p) kF (s)kH −α(p)/2 kG(t)kH −α(p)/2 dsdt
I t−s∼2j T p0 p0
j 1−θ(p)
. (2 T ) kF kL2(I, H −α(p)/2 ) kGkL2 (I, H −α(p)/2 ) . (3.112)
p0 p0
It follows that (3.110) holds.

Next, we consider the case (2). By Hölder’s inequality,
Z
·−2j T

|Lj (F, G)| . U (· − s)F (s)ds kGkL2 (I, L2 ) . (3.113)
·−2j+1 T 2 2 L (I,L )
Applying (3.98) and Hölder’s inequality, we have

Z 2
·−2j T

U (· − s)F (s)ds
·−2j+1 T 2 2 L (I,L )
Z Z t−2j T Z t−2j T
. |s − σ|−θ(a) kF (σ)kH −α(a)/2 kF (s)kH −α(a)/2 dσdsdt
I t−2j+1 T t−2j+1 T a0 a0
Z Z 2j T Z 2j T
. |s − σ|−θ(a) ×
I 0 0
kF (σ + t − 2j+1 T )kH −α(a)/2 kF (s + t − 2j+1 T )kH −α(a)/2 dσdsdt
a0 a0
j 2−θ(a)
. (2 T ) kF k2L2 (I, H −α(a)/2 ) . (3.114)
a0
By (3.113) and (3.114),

|Lj (F, G)| . (2j T )1−θ(a)/2 kF kL2 (I, H −α(a)/2 ) kGkL2 (I, L2 ) . (3.115)
a0
Case (3) is similar to case (2).

We denote by Σ the open quadrilateral domain with the vertices
(1/p, 1/p) (p 1), (1/r, 1/2), (1/2, 1/2) and (1/2, 1/r), which contains
B(P, ε). Now we show that (3.110) holds for any (1/a, 1/b) ∈ Σ. In fact,
it suffices to consider the case that 1/b > 1/a. We can choose η ∈ (0, 1),
(1/p0 , 1/p0 ) and (1/a0 , 1/2) satisfying (1) and (2), such that

1 1 1 1 1 1
, =η , + (1 − η) , . (3.116)
a b p0 p0 a0 2
Noticing that U (t) : L2 → L2 and
kU (t)ϕk2 = kϕk2 , (3.117)

i.e., θ(0) = α(0) = 0. From (3.117) and (3.98) we see that if α(p0 ) 6= 0,
then α(a) 6= 0 for any 2 < a < p0 ; if α(p0 ) = 0, then α(a) = 0 for any
2 < a < p0 . We consider the case α(p0 ) 6= 0. It is easy to see that
θ(p) α(p) 1/2 − 1/p
= = . (3.118)
θ(q) α(q) 1/2 − 1/q
(3.116) and (3.118) imply that
(θ(a), θ(b)) = η(θ(p0 ), θ(p0 )) + (1 − η)(θ(a0 ), θ(0)), (3.119)

(α(a), α(b)) = η(α(p0 ), α(p0 )) + (1 − η)(α(a0 ), α(0)). (3.120)
So, by the complex interpolation, we have the result, as desired.

We need the following (see [13]):
Lemma 3.9. Let


 A0 × B0 → C0 ,
T : A0 × B1 → C1 , (3.121)

A1 × B0 → C1
be bounded bilinear operators, where (A0 , A1 ), (B0 , B1 ), (C0 , C1 ) are com-
patible Banach pairs4 . Let 0 < η, ηi < 1 and 1 6 p, q 6 ∞ satisfy η = η1 +η2
and 1/p + 1/q > 1. Then
T : (A0 , A1 )η0 ,p × (B0 , B1 )η1 ,q → (C0 , C1 )η,1 (3.122)
is a bounded operator.
Theorem 3.5. Let U (t) and A be as in (3.52) and satisfy (3.98) and
(3.99). Assume that 2 6 p, q 6 r. Then
kU (t)φkL2/θ(p) (I,Bp,2
s ) . kφkH s−α(p)/2 , (3.123)
kAf kL2/θ(p) (I,B s−α(p)/2 ) . kf kL(2/θ(q))0 (I,Ḃ s+α(q) ) . (3.124)
p,2 q0 ,2
Proof. Now we can use Lemmas 3.8 and 3.9 to show (3.100). Take p =
q = 2, η = 2/3 and η0 = η1 = 1/3 in Lemma 3.9. Choose a0 = b0 and
a1 = b1 satisfying
θ(a0 ) = 1 + , θ(a1 ) = 1 − 2, (3.125)
where > 0 is sufficiently small. Put
−α(ai )/2
Ai = Bi = L2 (I, Ha0 ), Ci = `β(a
∞
0 ,bi )
, i = 0, 1. (3.126)
i
By θ(r) = 1, it follows from (3.118) that



 1/r = (1 − η0 )/a0 + η0 /a1 ,

α(r) = (1 − η0 )α(a0 ) + η0 α(a1 ),
(3.127)

 β(a0 , b1 ) = β(a1 , b0 ),

(1 − η)β(a0 , b0 ) + ηβ(a0 , b1 ) = 0.
In view of α(p0 ) 6= 0, we have α(a0 ) 6= α(a1 ). (3.127) implies that
−α(r)/2
(A0 , A1 )η0 ,2 = (B0 , B1 )η1 ,2 = L2 (I, Br0 ,2 ), (C0 , C1 )η,1 = `01 . (3.128)
So, by Lemma 3.9,
X
|Lj (F, G)| . kF kL2 (I,B −α(r)/2 ) kGkL2 (I,B −α(r)/2 ) . (3.129)
r0 ,2 r0 ,2
j
4 (A , A ) is said to be a compatible Banach pair, if there exists a linear topological
0 1
space A such that A0 , A1 ⊂ A.
We emphasize that in the right hand side of (3.129), the omitted constant
is independent of T .
As α(p0 ) = 0, the proof is slight different from the above discussions
and we omit the details. So, (3.100) is proved and the other cases of (3.123)
have been discussed in the previous section.
Below we prove (3.124). In fact, (3.129) implies that
kAf kL2 (I,B α(r)/2 ) . kf kL2 (I,B −α(r)/2 ) . (3.130)
r,2 r0 ,2
Using the same way as in the previous section, we have (see Lemmas 3.6,
3.7)
kAf kL∞ (I,L2 ) . kGkL2 (I,B −α(r)/2 ) , (3.131)
r0 ,2
kAf kL2 (I,B α(r)/2 ) . kf kL1(I,L2 ) . (3.132)

r,2
Analogous to the proof of Theorem 3.4, we can get (3.124).

Remark 3.4. We emphasize that Theorem 3.5 is independent of T > 0.
Thus, by a standard limit argument, we can get that Theorem 3.5 also
holds for I = R.
Remark 3.5. One needs the condition r < ∞ in Theorem 3.5 and we
may further ask what happens if r = ∞, which corresponds to the cases
n = 2 and n = 3 for the Schrödinger equation and the wave equation,
respectively. Generally speaking, Theorem 3.5 does not hold if r = ∞,
however, for the radial functions, the conclusion of Theorem 3.5 is still true
in the case r = ∞, see Tao [215].
Corollary 3.1. Let n > 3, S(t) = eit∆ , 2∗ = 2n/(n − 2), 2 6 r, p 6 2∗ ,

and 2/γ(·) = n (1/2 − 1/·). Then we have
kS(t)φkLγ(p) (I,Ḃ s ) . kφkḢ s , (3.133)
p,2
Z t

S(t − τ )f (τ )dτ . kf kLγ(r)0 (I,Ḃ s0
r ,2
), (3.134)
0 s )
Lγ(p) (I,Ḃp,2
where I = (−T, T ) ⊂ R is an arbitrary interval. In (3.133) and (3.92),
replacing homogeneous Besov spaces by corresponding Riesz potential spaces
Ḣρs (ρ = r, p), the conclusion still holds.
1/2
Corollary 3.2. Let W (t) = eit(−∆) , n > 3, 2∗∗ = 2(n − 1)/(n − 3),
2 6 r, p 6 2∗∗ , 2σ(·) = (n + 1)(1/2 − 1/·), and 2/β(·) = (n − 1)(1/2 − 1/·).
Then we have
kW (t)φkLβ(p) (I,Ḃ s−σ(p) ) . kφkḢ s , (3.135)
p,2
kAW f kLβ(p) (I,Ḃ s−σ(p) ) . kf kLβ(r)0 (I,Ḃ s+σ(r) ) , (3.136)

p,2 r0 ,2
Rt
where I = (−T, T ) ⊂ R is arbitrary, AW := 0 W (t−τ )·dτ . In (3.135) and
(3.136), substituting homogeneous Besov spaces by relevant Riesz potential
spaces Ḣρs , the result also holds.
1/2
Corollary 3.3. Let G(t) = eit(I−∆) , θ ∈ [0, 1], n > 3 − θ, 2∗∗ = 2(n −
1 + θ)/(n − 3 + θ), 2 6 r, p 6 2∗∗ , 2σ(θ, ·) = (n + 1 + θ)(1/2 − 1/·), and
2/β(θ, ·) = (n − 1 + θ)(1/2 − 1/·). Then
kG(t)φkLβ(θ,p) (I,B s−σ(θ,p) ) . kφkH s , (3.137)
p,2
kAG f kLβ(θ,p) (I,B s−σ(θ,p) ) . kf kLβ(θ,r)0 (I,B s+σ(θ,r) ) , (3.138)

p,2 r0 ,2
Rt
where I = (−T, T ) ⊂ R is arbitrary, AG := 0 G(t − τ ) · dτ . In (3.137) and
(3.138), replacing Besov spaces by corresponding Bessel potential spaces,
the conclusion also hold.
Chapter 4
Local and global wellposedness for

nonlinear dispersive equations
Stone from another mountain can be harder than jade. —— From

the Book of Songs
In this chapter we will study the local and global wellposedness in H s

for a class of nonlinear dispersive equations. In order to solve a nonlin-
ear dispersive equation, one needs to make a delicate balance between the
solution and its survival space. The weak solution has too large existing
space, which is hard to be unique. The smooth solution exists in a relatively
small space and its existence is somehow a problem. So, one of our main
task is to look for the most appropriate space to carry out the solution,
which strongly associates with the energy and dispersive structures of the
equation.
To some extent, Strichartz’ inequalities realize the balance between the
solution and its survival space, which enable us to obtain the global well-
posed results in the energy space for a class of nonlinear dispersive equa-
tions, such as nonlinear Schrödinger equations, nonlinear Klein-Gordon
equations and nonlinear wave equations and so on.
4.1 Why is the Strichartz estimate useful
Let us consider the Cauchy problem for the nonlinear Schrödinger equation
(NLS)
iut + ∆u = f (u), u(0, x) = u0 (x), (4.1)

√ Pn
where i = −1, f (u) = |u|α u, α > 0, ∆ = i=1 ∂x2i , u(t, x) is a complex
valued function of (t, x) ∈ R × Rn , u0 is the initial value of u at t = 0. We
easily deduce that the solutions of (4.1) formally satisfy the conservations
75
76 Local and global wellposedness for nonlinear dispersive equations
of mass and energy1 :

ku(t)k22 = ku0 k22 , E(u(t)) = E(u0 ), (4.2)
where
1 2
E(u(t)) = k∇u(t)k22 + ku(t)k2+α
2+α . (4.3)
2 2+α
According to the conservation laws, it is natural to ask the wellposedness
of solutions in L2 (Rn ) and H 1 (Rn ), respectively.
For any α > 0, using the compactness or the parabolic regularity
method, we can easily prove the global existence of weak solutions of (4.1),
however, the uniqueness, persistence and the continuity of the solution map
u0 → u are hard to obtain, see [159].
Noticed that S(t) = eit∆ : L2 → L2 , it is natural to solve NLS in
H (k = 0, 1, 2, ...) by using the boundedness of S(t) in L2 . According
k
to the standard semi-group theory, if one can solve NLS in H k , then the
uniqueness, persistence and the continuity of the solution map u0 → u can
also be obtained by a standard way. However, this method has much more
assumptions on nonlinearity and initial data. Let us consider the equivalent
integral form of (4.1):
Z t
u(t) = S(t)u0 − i S(t − τ )f (u(τ ))dτ, (4.4)
0
here we only assume that f (u) is a nonlinear function. Applying kS(t)f k2 =

kf k2 , we have
Z t
ku(t)k2 6 kS(t)u0 k2 + kS(t − τ )f (u(τ ))k2 dτ
0
Z t
= ku0 k2 + kf (u(τ ))k2 dτ. (4.5)
0
If we want to get the solution in C(0, T ; L2 ), then one needs

|f (u)| 6 C|u|,
from which we have
kukC(0,T ;L2) . ku0 k + T kukC(0,T ;L2) . (4.6)
We can easily construct a contraction mapping by choosing T > 0 suffi-
ciently small and show that NLS is well-posed in C(0, T ; L2).
1 Taking the inner product of (4.1) with u and u , and considering its imaginary and
t
real part, respectively, we can get (4.2).
4.1. Why is the Strichartz estimate useful 77
|f (u)| 6 C|u| is a rather strong condition on the nonlinearity, which

2
contains f (u) = sin u and f (u) = (e−|u| − 1)u as examples. The power
nonlinearity is not included in the above discussions.
Similarly, say for n = 1, one can solve NLS in C(0, T ; H 1 ). Assume that
f (0) = 0, |f 0 (u)| 6 C|u|α , 0 < α < ∞.
Using the embedding H 1 ⊂ L∞ , we have for k = 0, 1,
Z t
k∇k u(t)k2 6 kS(t)∇u0 k2 + kS(t − τ )∇k f (u(τ ))k2 dτ
0
Z t
k
. k∇ u0 k2 + k|u|α ∇k uk2 dτ
0
Z t
. k∇k u0 k2 + kukα k
∞ k∇ uk2 dτ. (4.7)
0
So,
k∇k ukC(0,T ;L2 ) . k∇k u0 k2 + T k∇k ukα+1
C(0,T ;L2 ) . (4.8)
Hence we obtain the local wellposedness in C(0, T ; H 1 ) for NLS. By the

conservation of energy, one can extend it to a global one.
The above idea can be generalized to any spatial dimensions and we can
solve NLS in C(0, T ; H s ) (s > n/2) by assuming that
|f (k) (u)| 6 C|u|α+1−k , [n/2] < α < ∞, k = 0, 1, ..., [n/2] + 1,
where s > n/2 is essential to guarantee the inclusion H s ⊂ L∞ .
According to the above discussions, for n > 2, we can not solve NLS in
H 1 if we only use the L2 estimates.
Now, our question is how to show the wellposedness of NLS in the energy
spaces in higher spatial dimensions. Using the Strichartz estimates as tools,
we can obtain the global wellposedness results of NLS in H 1 for the power
nonlinearity |u|α u with α < 2∗ − 2 (2∗ is as in (3.73)), see Kato [120]. If
α = 2∗ − 2, the local wellposedness for NLS can be established by resorting
to the Strichartz inequalities. On the basis of the local weposedness result,
Bourgain [19] in 1999 developed the localized energy separation method
and he obtained the global wellposedness of the radial solutions in 3D and
4D. Colliander, Keel, Staffilani, Takaoka and Tao [44] further developed
Bourgain’s technique and they removed the radial assumption. In higher
spatial dimensions n > 4 and α = 2∗ − 2, the global wellposedness of NLS
is solved by Ryckman and Visan [196] and Visan [232].
If α > 2∗ − 2, the global wellposedness of NLS is still open.
In order to indicate the general idea, now we briefly give an application

of the Strichartz inequality to NLS in 2D; cf. Tsutsumi [226], Cazenave
and Weissler [25]. We consider the integral equation:
Z t
u(t) = S(t)u0 − i S(t − τ )|u(τ )|2 u(τ )dτ. (4.9)
0
Recall the Strichartz inequalities:
kS(t)u0 kL4x,t (R2+1 ) . ku0 k2 , (4.10)
Z t

S(t − τ )f (τ )dτ . kf kL4/3(R2+1 ) . (4.11)
x,t
0 L4x,t (R2+1 )
So, we have
kukL4x,t (R2+1 ) . ku0 k2 + k|u|2 ukL4/3(R2+1 )
x,t
. ku0 k2 + kuk3L4 (R2+1 ) . (4.12)

x,t
So, we can solve (4.9) in the space L4x,t (R2+1 ) at least for small Cauchy
data u0 ∈ L2 . To realize (4.12), one needs to make a delicate balance
between the nonlinearity |u|2 u and the Strichartz space L4x,t . The above
performance contains a very general idea to show the well posedness of
NLS. In the next few sections we will mainly use this technique to study
NLS and nonlinear Klein-Gordon equations.
4.2 Nonlinear mapping estimates in Besov spaces
In order to solve a nonlinear dispersive equation, after establishing the

Strichartz inequalities for the relevant linear dispersive equation, we need
0
to estimate the nonlinear terms in the spaces Lγ(r) (0, T ; Brs0,2 ). The energy
structure is not necessary for the local wellposedness and for the global
wellposedness with small data. In this section, we consider a nonlinear
mapping estimate in Besov spaces. The relevant estimates go back to the
works of Pecher [190], Brenner [21], Ginibre and Velo [79]. For NLS, a
general nonlinear estimate can be found in Cazenave and Weissler [26].
Wang [236; 234] obtained the nonlinear estimates for the nonlinear Klein-
Gordon equations (NLKG). Now we prove a general result, which covers
NLS, NLKG and their higher order versions as special cases, which was
obtained in [241].
The nonlinear mapping estimates in Besov spaces rely upon two mod-
ified versions of the Hölder inequality, one is the convexity Hölder’s in-
equality as in Proposition 1.21, another is the modified Hölder inequality
4.2. Nonlinear mapping estimates in Besov spaces 79
in Sec. 1.7. We know that

|α| q
X XY
Dα f (u) = f (q) (u) Dαi u, (4.13)
q=1 Λqα i=1

where Λqα = α1 + · · · + αq = α, |αq | > · · · > |α1 | > 1 . Assume that f (u)
satisfies
|f (k) (u)| 6 C|u|p+1−k , k = 0, 1, ..., |α|. (4.14)
Noticing that2
N
Y N
Y N i−1
X Y N
Y
ai − bi = aj bj (ai − bj ),
i=1 i=1 i=1 j=1 j=i+1
we have for any 1 6 r0 < ∞ and |α| > 1,

k Mh Dα f (u)kLr0
|α|
X X n q
Y

. (|uh |p−q + |u|p−q )|uh − u| Dαi u
Lr 0
q=1 Λqα i=1
q
X i−1
Y q
Y o

+ |uh |p+1−q Dαj uh Dαj uDαi (uh − u)
Lr 0
i=1 j=1 j=i+1
|α|
X X q
X
=: kIq kLr0 + kIIqi kLr0 . (4.15)
q=1 Λqα i=1
Lemma 4.1. Let 2 6 r < ∞ and 0 6 δ 6 s0 ∧ s < ∞. Assume that

f ∈ C {s−δ} satisfies the following condition:
|f (k) (u)| . |u|p+1−k , k = 0, 1, ..., {s − δ}, {s − δ} 6 p + 1, (4.16)
where we assume that {a} = 1 + [a] if a is not an integer; {a} = a if a is
an integer. Suppose that
1 s 1 δ 1 1 s0
0
p − + − = 0, − > 0. (4.17)
r n r n r r n
/ N, then we have
If s − δ ∈
kf (u)kḂ s−δ
0
. kukpḂ s0 kukḂ s . (4.18)
r ,2 r,2 r,2
If s − δ ∈ N, substituting homogeneous Besov spaces by relevant Riesz po-

tential spaces, (4.18) also holds.
2 We
Q0 QN
assume j=1 aj = j=N+1 bj = 0.
Proof. If s − δ < 1, the proof is very easy and we omit the details of the
proof. It suffices to consider the case s − δ > 1. We assume that s − δ ∈ / N.
By the equivalent norm in homogeneous Besov spaces,
Z ∞ X 1/2
−2v α 2 dt
kf (u)kḂ s−δ = t sup k M h D f (u)k L r 0 , (4.19)
r0 ,2
0 |h|6t t
|α|=[s−δ]
where v = s − δ − [s − δ]. In view of (4.16) we see that k Mh Dα f (u)kLr0

satisfies (4.15).
Step 1. we estimate
Z ∞ !1/2
−2v 2 dt
Aq := t sup kIq kLr0 . (4.20)
0 |h|6t t
In (4.16), we consider the estimate of kIq kLr0 . Put

1 s 1 s 0 + β0 − v
0
a0 = (p − q) − , a00 = − ,
r n r n
1 s0 + βi − |αi |
ai = − , i = 1, · · · , q,
r n
where βi will be chosen below.
If s 6 s0 +1, then we can take β0 = β1 = · · · = βq−1 = 0 and βq = s−s0 .
Since s− δ > 1, we see that v = s− δ − [s− δ] 6 s− δ − 1 6 s0 , which implies
a00 > 0. In view of s − |αq | 6 s0 , we obtain that aq > 1/r − s0 /n > 0. If
q > 2, then |αi | 6 s0 for all i = 1, · · · , q − 1. Indeed, in the opposite case
one has that |αq | + |αq−1 | > s0 + 1 > s, which is impossible. So, we have
P
ai > 0, i = 0, 1, ..., q. Notice that a00 + qi=0 ai = 1/r0 . It follows from the
modified Hölder inequality that
kIq kLr0 6 Ckukp−1
s kuh − uk 1/a0 kukḂ s .
Ḃ 0 L 0
(4.21)
r,2 r,2
Hence,
Aq 6 Ckukp−1
s kukḂ v
Ḃ 0
kukḂ s 6 CkukpḂ s0 kukḂ s . (4.22)
r,2 1/a0 ,2 r,2 r,2 r,2
0
Below, we consider the case s > s0 + 1. We will choose suitable βi

(i = 0, 1, ..., q) satisfying the following four conditions:
(a) 0 6 β0 6 v, 0 6 βi 6 |αi |, i = 1, · · · , q,
(b) s0 + β0 > v, s0 + βi > |αi |, i = 1, · · · , q,
Pq
(c) i=0 βi = s − s0 ,
(d) s0 + βi 6 s, i = 0, · · · , q.
Conditions (a) and (c) imply condition (d), So, it suffices to show that
there exist β0 , · · · , βq satisfying (a)–(c).
4.2. Nonlinear mapping estimates in Besov spaces 81
If q = 1, then we can take β0 = v, β1 = s − s0 − v. Due to δ 6 s0 , we

have β1 6 [s − δ] and s0 + β1 > [s − δ]. So, conditions (a)–(c) are satisfied.
We consider the case q > 2 which is divided into the following three
subcases.
First, we assume that |αq | > s − s0 − v. Let β0 = v, β1 = · · · = βq−1 = 0
and βq = s − s0 − v. Obviously, conditions (a) and (c) hold. On the other
hand, it is easy to see that s0 +βq > |αq | and s0 +β0 > v. If |αq−1 | > s0 , then
|αq |+ |αq−1 | > s− v > [s− δ], a contraction. So we have s0 + βq−1 > |αq−1 |,
which implies the condition (b) holds.
Secondly, we consider the case s0 6 |αq | < s − s0 − v. Put β0 = v,
βi = |αi | for i = 1, · · · , q − 1 and βq = |αq | + δ − s0 . A straightforward
Pq
calculation yields i=0 βi = s − s0 . In view of δ 6 s0 and |αq | > s0 , we
see that 0 6 βq 6 |αq |, which leads the condition (a) holds. Noticing that
s0 + βq = |αq | + δ, we easily see that condition (b) holds.
Thirdly, we consider the case |αq | 6 (s − s0 − v) ∧ s0 . In the current
Pq
case, we easily see that i=1 |αi | = [s − δ] = s − δ − v > s − s0 − v. Hence,
Pq
we can choose βi ∈ [0, |αi |] (i = 1, · · · , q) satisfying i=1 βi = s − s0 − v.
Put β0 = v. Obviously, conditions (a) and (c) hold. From |αi | 6 |αq | 6 s0
it follows that the condition (b) holds.
Therefore, we have chosen β0 , · · · , βq satisfying conditions (a)–(d). We
have,
Xq 1 s 1 s + β − v + Pq (β − |α |)
0 0 0 i=1 i i
a00 + ai = p − + −
i=0
r n r n
1 s 1 δ 1
0
=p − + − = 0.
r n r n r
Applying the modified Hölder inequality,
q
Y
kIq kLr0 6 Ckukp−q
s0 kuh
Ḃr,2
− ukL1/a00 kukḂ s0 +βi . (4.23)
r,2
i=1
By (4.23), we have
q
Y
Aq 6 Ckukp−q
s kukḂ v
Ḃ 0
kukḂ s0 +βi
r,2 1/a00 ,2 r,2
i=1
q
Y
6 Ckukp−q
s
Ḃ 0
kukḂ s0+βi . (4.24)
r,2 r,2
i=0
One can choose θi (i = 0, 1, . . . , q) satisfying s0 + βi = θi s0 + (1 − θi )s. It

Pq Pq
is easy to see that 0 6 θi 6 1, i=0 θi = q and i=0 (1 − θi ) = 1. Using
the convexity Hölder inequality, we have

q
Y
kukḂ s0+βi 6 CkukqḂ s0 kukḂ s . (4.25)
r,2 r,2 r,2
i=1
So, we obtain the estimate of Aq .
Step 2. We estimate
q q Z ∞ !1/2
X X dt
Bq := Bqi := t−2v sup kIIqi k2Lr0 . (4.26)
i=1 i=1 0 |h|6t t
The estimates of Bq are similar to those of Aq and we only sketch the proof.
(I) We consider the case s 6 s0 + 1. If q = 1, the estimate of Bq is very
easy and we omit the details. We now consider the estimates of Bqi , i 6= q,
q > 2. Put
1 s
0
a0 = (p + 1 − q) − ,
r n
1 s0 − |αj |
aj = − , j 6= i, j = 1, · · · , q − 1,
r n
1 s0 − v − |αi | 1 s − |αq |
ai = − , aq = − .
r n r n
It is easy to see that s0 > v + |αj |. If not, then s − δ > |αq | + |αj | + v >
s0 + 1 > s, which is impossible. So, aj > 0, j = 1, · · · , q − 1. By s 6 s0 + 1
we see that aq > 0. Using the modified Hölder inequality, we can get the
estimate of kIIqi kLr0 . Thus, we obtain the estimates of Bqi for i 6= q. Put
1 s 1 s0 − |αj |
0
a0 = (p + 1 − q) − , aj = − , j = 1, · · · , q − 1,
r n r n
1 s0 − v − |αq |
aq = − ,
r n
then the estimate of Bqq follows.
(II) We consider the case s > s0 + 1. Put
1 s 1 s0 + βj − |αj |
0
a0 = (p + 1 − q) − , aj = − , j = 1, · · · , q,
r n r n
where β1 , · · · , βq can be chosen as in Step 1. So,
q
X
(a) 0 6 βi 6 |αi |; (b) s0 + βi > |αi |; (c) βj = s − s0 − v.
j=1
Applying the modified Hölder inequality,

Y
Bqi 6 Ckukp+1−q
Ḃ
s0 kuk s0 +βj kukḂ s0+βi +v . (4.27)
r,2 Ḃr,2 r,2
j6=i
4.3. Critical and subcritical NLS in H s 83
Since s0 + βj + v 6 s, 1 6 j 6 q, we claim that there exist θi (i = 1, . . . , q)

satisfying
s0 + βj = θj s0 + (1 − θj )s, j 6= i, j = 1, ..., q,
s0 + βi + v = θi s0 + (1 − θi )s.
Pq Pq
Since i=1 θi = q −1 and i=1 (1−θi ) = 1, in view of the convexity Hölder
inequality we obtain the estimates of Bqi for i = 1, · · · , q.
Remark 4.1.
(i) If δ = s0 = 0 and s > 0, then the result of (4.18) can be slightly
improved by
kf (u)kḂ s0 6 CkukpLr kukḂ s . (4.28)
r ,2 r,2
(ii) Lemma 4.1 covers the nonlinear estimates for NLS and NLKG as
in [19; 26; 41; 173; 191; 236; 237]), where Cazenave and Weissler
[26] considered the case δ = 0, s0 = s; Wang [236; 237] discussed
the nonlinear estimates for NLKG.
2
(iii) If the nonlinearity has an exponential growth, say sinh u, (e|u| −
1)u, the nonlinear mapping estimates can not be covered by Lemma
4.1, one can refer to Nakamura and Ozawa [173], and Wang [239].
4.3 Critical and subcritical NLS in H s
4.3.1 Critical NLS in H s

We consider the initial value problem for NLS,
iut + ∆u = f (u), u(0, x) = u0 (x), (4.29)
σ
where f (u) = c|u| u. If u is a solution of (4.29), then uλ (t, x) =
λ2/σ u(λ2 t, λx) also solves (4.29) with initial data λ2/σ u0 (λ·) at t = 0. Let
us observe that
kuλ kḢ s (Rn ) = λ2/σ+s−n/2 kukḢ s (Rn ) .
This implies that σ = 4/(n−2s) is the unique index such that the norm of uλ
in Ḣ s is invariant for all λ > 0. From this point of view, σ = 4/(n − 2s) > 0
is said to be the critical power in Ḣ s (or in H s ) for NLS. For s = n/2 − 2/σ,
Ḣ s (H s ) is said to be the critical space for NLS.
When s < n/2, corresponding to the critical case, σ < 4/(n − 2s) is
said to be a subcritical power in H s . If s > n/2, σ < ∞ is said to be a
subcritical power in H s . σ > 4/(n − 2s) is said to be a supercritical power
in H s for NLS.
4.3.2 Wellposedness in H s
We consider the following equivalent integral equation with respect to
(4.29),
Z t
u(t) = S(t)u0 − i S(t − τ )f (u(τ ))dτ, (4.30)
0
where S(t) = eit∆ . According to the Strichartz inequalities, if u0 ∈ H s , then
(3.75) and (3.76) indicate that the solution should belong to Lγ(p) (I, Bp,2 s
).
Assume that
|f (k) (u)| 6 C|u|σ+1−k , k = 0, 1, ..., [s] + 1, (4.31)
where 0 6 s < n/2, [s] denotes the integer part of s, 0 < σ 6 4/(n − 2s).
We have (see [26])
Theorem 4.1. Let 0 6 s < n/2, 0 < σ < 4/(n − 2s). Assume that
f ∈ C [s]+1 satisfies (4.31), [s] 6 σ. If u0 ∈ H s , then there exists a T ∗ :=
T ∗ (ku0 kH s ) > 0 such that (4.30) has a unique solution
γ(r)
u ∈ Lloc ([0, T ∗ ); Br,2
s
), (4.32)
where
n(2 + σ)
r= . (4.33)
n + sσ
∗
Moreover, for any 2 6 p 6 2 (p 6= ∞), we have
γ(p)
u ∈ Lloc ([0, T ∗ ); Bp,2
s
), (4.34)
∗
and if T < ∞, then
kukLγ(r) ([0,T ∗ );Br,2
s ) = ∞, (4.35)
ku(t)kH s & (T ∗ − t)1−σ(n−2s)/4 , 0 < t < T ∗ . (4.36)
In Theorem 4.1, we obtain the local wellposedness in H s , where σ is
a subcritical power in H s . If σ is a critical power in H s , we can get the
global wellposedness for NLS with small data.
Theorem 4.2. Let 0 6 s < n/2 and σ = 4/(n − 2s). Assume that f ∈
C [s]+1 satisfies (4.31), [s] 6 σ. If u0 ∈ H s , then there exists a T ∗ :=
T ∗ (u0 ) > 0 such that (4.30) has a unique solution u verifying (4.32), where
r = n(2 + σ)/(n + sσ) and γ(r) = 2 + σ. Moreover, for any 2 6 p 6 2∗
(p 6= ∞), the solution satisfies (4.34). If T ∗ < ∞, then (4.35) holds.
If ku0 kḢ s is sufficiently small, then the above solution is a global one,
i.e., T ∗ = ∞ and
kuk∩26p62∗, p6=∞ Lγ(p) (0,∞;Ḃ s ) . Cku0 kḢ s , (4.37)
p,2
kuk∩26p62∗, p6=∞ Lγ(p) (0,∞;Bp,2

s ) . Cku0 kH s . (4.38)
4.3. Critical and subcritical NLS in H s 85
Remark 4.2. If T ∗ < ∞ in Theorem 4.1, then T ∗ has a lower bound

which depends on ku0 kH s , i.e., T ∗ > δ(ku0 kH s ) > 0. However, if T ∗ < ∞
in Theorem 4.2, then we can not get the lower bound of T ∗ which only
depends on ku0 kH s , namely, T ∗ may depend on not only ku0 kH s , but also
the choice of u0 in H s , which is easily seen from the scaling u(t, x) →
uλ (t, x) := λ2/σ u(λ2 t, λx).
Proof. Now we prove Theorems 4.1 and 4.2. Let T > 0, M > 0 and
δ > 0 which will be chosen later. Put
n o
D = u ∈ Lγ(r) (0, T ; Br,2
s
) : kukLγ(r) (0,T ;Ḃ s ) 6 δ, kukLγ(r) (0,T ;Br,2
s ) 6 M ,
r,2
(4.39)
which is equipped with the metric
d(u, v) = ku − vkLγ(r) (0,T ;Lr ) . (4.40)
Z t
T : u(t) → S(t)u0 − i S(t − τ )f (u(τ ))dτ, (4.41)
0
we show that T : (D, d) → (D, d) is a contraction mapping for some

T, δ, M > 0. Since
1 s 1 1
σ − + = 0, (4.42)
r n r r
by Lemma 3.1 we have
σ+1
kf (u)kḂ s0 . kukḂ s , (4.43)
r ,2 r,2
kf (u)kBrs0 ,2 . kukσḂ s kukBr,2

s . (4.44)
r,2
So,
kf (u)kLγ(r)0 (0,T ;Ḃ s0 ) . T 1−σ(n−2s)/4 kukσ+1
Lγ(r) (0,T ;Ḃ s )
. (4.45)
r ,2 r,2
In view of Theorem 3.1, one has that

kT ukLγ(r) (0,T ;Ḃ s . ku0 kḢ s + T 1−σ(n−2s)/4 kukσ+1 . (4.46)
r,2 ) Lγ(r) (0,T ;Ḃ s ) r,2
Similarly, we have
kT ukLγ(r) (0,T ;Br,2
s )
. ku0 kH s + T 1−σ(n−2s)/4 kukσLγ(r) (0,T ;Ḃ s ) kukLγ(r) (0,T ;Br,2

s ) (4.47)
r,2
kT u − T vkLγ(r) (0,T ;Lr ) ,

. T 1−σ(n−2s)/4 (kukσ + kvkσ )Lγ(r) (0,T ;Ḃ s ) ku − vkLγ(r) (0,T ;Lr ) . (4.48)
r,2
Case 1. 0 < σ < 4/(n − 2s). Put

δ = 2Cku0 kḢ s , M = 2Cku0 kH s . (4.49)
We take T > 0 satisfying
2CT 1−σ(n−2s)/4 M σ 6 1/2. (4.50)
(4.46)–(4.48) imply that
kT ukLγ(r) (0,T ;Ḃ s . δ, (4.51)
r,2 )
kT ukLγ(r) (0,T ;Ḃ s . M, (4.52)

r,2 )
1
kT u − T vkLγ(r) (0,T ;Lr ) ku − vkLγ(r) (0,T ;Lr ) . (4.53)
2
So, T : (D, d) → (D, d) is a contraction mapping. So, there exists a u ∈ D
satisfying (4.30). Again, in view of the Strichartz inequalities (3.75) and
(3.76), we have
kukLγ(p) (0,T ;Ḃ s . ku0 kḢ s + T 1−σ(n−2s)/4 kukσ+1 , (4.54)
p,2 ) Lγ(r) (0,T ;Ḃ s )
r,2
from which we obtain that u ∈ Lγ(p) (0, T ; Bp,2

s
). By a standard argument,
we can extend the solution above. Considering the mapping
Z t
T : u(t) → S(t − T )u(T ) − i S(t − τ )f (u(τ ))dτ, (4.55)
T
and noticing that u(T ) ∈ H s , we can use the same way as in the above
to solve (4.55). Repeating this argument step by step, we find a maximal
T ∗ > 0 satisfying (4.33)–(4.36). The uniqueness of the solution can be
shown by following the same way as that of (4.48) and we omit the details.
Case 2. σ = 4/(n − 2s). By Strichartz estimate (3.75), we see that
kS(t)u0 kLγ(r) (0,T ;Br,2
s ) → 0, as T → 0. (4.56)
Put
n o
D0 = u ∈ Lγ(r) (0, T ; Br,2
s
) : kukLγ(r) (0,T ;Br,2
s ) 6 M , (4.57)
which is equipped with the metric as in (4.40). Take M > 0 satisfying

2CM σ 6 1/2 (4.58)
and T > 0 such that
kS(t)u0 kLγ(r) (0,T ;Br,2
s ) 6 M/2. (4.59)
4.4. Global wellposedness of NLS in L2 and H 1 87
So, analogous to the above, one can show the local wellposedness of The-
orem 4.2. For the global wellposedness with small data, it suffices to take
T = ∞ in (D, d) as in Case 1. In fact, analogous to (4.46)–(4.48), we have
kT ukLγ(r) (0,∞;Ḃ s . ku0 kḢ s + kukσ+1 , (4.60)
r,2 ) Lγ(r) (0,∞;Ḃ s )
r,2
kT ukLγ(r) (0,∞;Br,2
s )
. ku0 kH s + kukσLγ(r) (0,∞;Ḃ s ) kukLγ(r) (0,∞;Br,2

s ), (4.61)
r,2
kT u − T vkLγ(r) (0,∞;Lr ) ,
. (kukσ + kvkσ )Lγ(r) (0,∞;Ḃ s ) ku − vkLγ(r) (0,∞;Lr ) . (4.62)
r,2
Put
δ = 2Cku0 kḢ s 1, M = 2Cku0 kH s . (4.63)
Using the contraction mapping argument, we can get the result.
4.4 Global wellposedness of NLS in L2 and H 1
We consider the initial value problem for NLS:

iut + ∆u = λ|u|σ u, u(0, x) = u0 (x), (4.64)
where λ ∈ R. λ > 0 is the defocusing case and λ < 0 is the focusing case.
Recall that the solution of NLS formally satisfies
ku(t)k22 = ku0 k22 , E(u(t)) = E(u0 ), (4.65)
1 2λ
E(u(t)) = k∇u(t)k22 + ku(t)k2+α
2+α . (4.66)
2 2+α
From the conservation of energy we see that for the focusing NLS, the
kinetic energy 21 k∇uk22 and the potential energy 2+λ
2λ
kuk2+σ
2+σ have opposite
signs and for the defocusing NLS, the kinetic energy and the potential
energy are both positive. So, the defocusing case is better than the focusing
case. Roughly speaking, the defocusing NLS has global solutions and the
focusing NLS has blowup phenomena.
Theorem 4.3. Let 0 < σ < 4/n. If u0 ∈ L2 , then (4.64) has a unique
solution

γ(r)
u ∈ C([0, ∞); L2 ) ∩ ∩2<r62∗ , r6=∞ Lloc (0, ∞; Lr ) (4.67)
and ku(t)k2 = ku0 k2 .

Theorem 4.4. Let 2∗ be as in (3.73), λ > 0 and 0 < σ < 2∗ − 2. If

u0 ∈ H 1 , then (4.64) has a unique solution

γ(r)
u ∈ C([0, ∞); H 1 ) ∩ ∩2<r62∗ , r6=∞ Lloc (0, ∞; Hr1 ) , (4.68)
and (4.65) holds.
Theorems 4.3 and 4.4 are corollaries of Theorem 4.1 and the conserva-
tion laws in (4.65). Since we have gotten the local wellposedness in Theorem
4.1, using the standard argument, one can get that the local solution of NLS
(4.64) satisfies (4.65) at any local time. By (4.36), we have T ∗ = ∞.
For the energy critical case σ = 4/(n − 2), n > 3, Theorem 4.4 also
holds, see Chapter 7.
4.5 Critical and subcritical NLKG in H s
We study the Cauchy problem for the nonlinear Klein-Gordon equation

(NLKG):
utt + (m2 − ∆)u = f (u), u(0, x) = u0 (x), ut (0, x) = u1 (x), (4.69)
where m2 > 0. The main results in this section are the following
Theorem 4.5. Let n > 2, 1/2 6 s < n/2 and 0 < σ < 4/(n − 2s). Assume
that f ∈ C [s+1/2] satisfies
|f (k) (u)| 6 C|u|σ+1−k , k = 0, 1, ..., [s + 1/2], (4.70)

[s + 1/2] 6 σ + 1. If (u0 , u1 ) ∈ H s × H s−1 , then there exists a T ∗ :=
T ∗ (ku0 kH s , ku1 kH s−1 ) > 0 such that (4.69) has a unique solution

γ(r) s−β(r)
u ∈ C([0, T ∗ ); H s ) ∩ ∩r∈(2,2∗ ) Lloc ([0, T ∗ ); Br,2 ) , (4.71)
where 2β(r) = (n + 1)(1/2 − 1/r), 2/γ(r) = (n − 1)(1/2 − 1/r), and

∗ 2(n − 1)/(n − 3), n > 3,
2 =
∞, n = 2, 3.
Moreover, if T ∗ < ∞, then
ku(t)kH s & (T ∗ − t)−δ/p , 0 < t < T ∗ , (4.72)
where δ = 1 for 0 < σ < 2/(n − 2s), δ = (4 − σ(n − 2s))/2 for 2/(n − 2s) 6
σ < 4/(n − 2s).
4.5. Critical and subcritical NLKG in H s 89
In Theorem 4.5 we get the local wellposedness in H s , where σ is a

subcritical power in H s . If σ is a critical power in H s , we have the following
global wellposedness result with small data:
Theorem 4.6. Let n > 2, 1/2 6 s < n/2 and σ = 4/(n− 2s). Assume that
f ∈ C [s+1/2] satisfies (4.70), [s + 1/2] 6 σ + 1. If (u0 , u1 ) ∈ H s × H s−1 ,
then there exists a T ∗ := T ∗ (u0 , u1 ) > 0 such that (4.69) has a unique
solution u verifying (4.71). Moreover, if k(u0 , u1 )kH s ×H s−1 is sufficiently
small, then T ∗ = ∞.
If the initial data belong to energy spaces, we have
Theorem 4.7. Let n > 2, 0 < σ < 4/(n − 2) and f (u) = |u|σ u. If
(u0 , u1 ) ∈ H 1 × L2 , then (4.69) has a unique solution

γ(r) s−β(r)
u ∈ C([0, ∞); H 1 ) ∩ ∩r∈(2,2∗ ) Lloc ([0, ∞); Br,2 ) , (4.73)
and
1 m2 1 2
E(u, ut ) := k∇uk22 + kuk22 + kut k22 + kuk2+σ
2+σ = E(u0 , u1 ).
2 2 2 2+σ
(4.74)
Theorems 4.5 and 4.6 can be developed to the nonlinear wave equation
(NLW), i.e., m2 = 0 in (4.69). Indeed, substituting H s , Bp,2 a
by relevant
s a
Ḣ , Ḃp,2 , we see that Theorems 4.5 and 4.6 also hold for the following NLW
utt − ∆u = f (u), u(0, x) = u0 (x), ut (0, x) = u1 (x). (4.75)

We have similar results to Theorems 4.5–4.7 for NLKG in one spatial
dimension, whose proofs are slightly different from the case n > 2. For
NLW, one can use the L2 estimates to get the local wellposedness in one
spatial dimension. However, the Strichartz estimates fail in 1D and Theo-
rem 4.6 can not be generalized to one spatial dimension. If σ = 4/(n − 2),
the global wellposedness result will be discussed in Chapter 7.
Proof. [Sketch Proof of Theorem 4.5] We only consider the case 2/(n −
2s) 6 σ 6 4/(n − 2s). Put
2(n − 1)(2 + σ)
ρ= .
(n − 1)(2 + σ) + 4 − 2σ(n − 2s)
ρ satisfies the following identity,

1 s − β(ρ) 1 1 − 2β(ρ) 1
σ − + − = 0.
ρ n ρ n ρ
By Lemma 4.1, we have

kf (u)kḂ s−1+β(ρ) . kukσ+1
s−β(ρ) . (4.76)
ρ0 ,2 Ḃρ,2
Consider the equivalent integral equation of (4.69),

Z t
u(t) = K 0 (t)u0 + K(t)u1 + K(t − τ )f (u(τ ))dτ, (4.77)
0
where K(t) = (m2 − ∆)−1/2 sin(m2 − ∆)1/2 and K 0 (t) = cos(m2 − ∆)1/2 .
Put
s−β(ρ)
D = {u ∈ Lγ(ρ) (0, T ; Bρ,2 ) : kukLγ(ρ) (0,T ;B s−β(ρ) ) 6 M },
ρ,2
and
d(u, v) = ku − vkLγ(ρ) (0,T ;Lρ ) .
Using the Strichartz inequalities together with (4.76), we can show that
Z t
0
T : u(t) → K (t)u0 + K(t)u1 + K(t − τ )f (u(τ ))dτ (4.78)
0
is a contraction mapping from (D, d) into itself. For the details, see Wang
[236; 237].
Chapter 5
The low regularity theory for the

nonlinear dispersive equations
From PDE point of view, one of the main reasons that we like Fourier
transform is that it makes the differential operation into fundamental
algebraic multiplication.
This chapter is devoted to introduce some methods in studying the low

regularity theory for the Cauchy problems of the nonlinear dispersive equa-
tions: X s,b method and I-method. By the low regularity theory we mean
to study the Cauchy problems assuming rough initial data. For instance,
whether the Cauchy problems are locally well-posed in H s and how low can
s be? How about global well-posedness?
The low regularity theory is important for several reasons: First, for
Sobolev H s , if s is smaller, then the space H s is larger; thus the scope of
initial data is wide, for example, it can include Dirac measure–δ function
which belongs to H −1/2− , > 0. Secondly, for some equations, conser-
vation laws of L2 and H 1 can be easily obtained, so the task of extending
a local existence result to a global existence result can be easier if one is
working at low regularities than high regularities.
In this chapter, we take the Korteweg de-Vries equation and the
Schrödinger equation with derivative as examples to study the low reg-
ularity solution for the nonlinear dispersive equations.
5.1 Bourgain space
In this section, we introduce Bourgain space and its properties. The Cauchy
problem for a general dispersive equation usually has the following form:
(
∂t u − iφ(D)u = f (u, ∂ α u), (x, t) ∈ Rn × R
(5.1)
u(x, 0) = u0 (x) ∈ H s (Rn ).
91
92 The low regularity theory for the nonlinear dispersive equations
where u(x, t) : Rn → C is a unknown function, u0 (x) is a given data,

∂ α = (∂xα11 , ..., ∂xαnn ), and φ(D) is a Fourier multiplier:
Z
φ(D)u = (2π)−n/2 eixξ φ(ξ)b
u(ξ)dξ.
Rn
Here φ(ξ) is a real-valued continuous function1 , which we refer as dispersion

relation of the equation or phase function. Nonlinear term f is a function
of multi-variables, for example f is a polynomials or usually of the form
|u|p u.
Many dispersive equations can be reduced to the form (5.1), for in-
stance, Schrödinger equation (with dispersion relation φ(ξ) = |ξ|2 ), Ko-
rteweg de-Vries equation (with dispersion relation φ(ξ) = ξ 3 ), and so on.
In the previous several chapters we have discussed some space-time struc-
tures, such as Strichartz-type space Lqt Lrx , smoothing effect-type Lqx Lrt . In
this chapter we will apply a new class of space-time structure X s,b . To
briefly introduce the motivation, we view (5.1) as a perturbation of a linear
equation as before, hence consider first the corresponding linear equation:
∂t u − iφ(D)u = 0. (5.2)
Denote S(t)f = Fx−1 eitφ(ξ) Fx f . Taking Fourier transform with respect to

both space and time on both side of (5.2), then it becomes
(τ − φ(ξ))b
u(ξ, τ ) = 0. (5.3)
b or F u will always denote the Fourier trans-

Unless particularly specified, u
form of the function u(x, t) with respect to x and t, Fx u denotes the
one only on x.2 From (5.3) we see that u b is supported on the surface
{(ξ, τ ) : τ = φ(ξ)}. One can imagine that this surface is intimately related
to the equation (5.1). By viewing x and t equally, we then introduce the
Sobolev-type space for (5.1) as Sobolev space for the Laplacian equation.
Definition 5.1. Assume φ : Rd → R is a continuous function. Let s, b ∈ R.

The space Xτs,b d
=φ(ξ) (R ×R), simply denoted as X
s,b
(Rd ×R) or X s,b , is then
defined to be the closure of the Schwartz functions S(Rd+1 ) under the norm
s b
kukX s,b = khξi hτ − φ(ξ)i F ukL2ξ L2τ . (5.4)
1 We emphasize that φ must be real-valued.
2 Ifthere is no confusion, we also denote by u
b the Fourier transform of u for function
u : Rn → C.
5.1. Bourgain space 93
These modern forms of the spaces were first used to systematically study
nonlinear dispersive wave problems by Bourgain [15; 14]. Klainerman and
Machedon [146] used similar ideas in their study of the nonlinear wave
equation. The spaces appeared earlier in a dierent setting in the works
[194], [7] of Rauch, Reed, and M. Beals. Then the bilinear estimates in
these space were deeply studied by Kenig, Ponce and Vega [132], and by
Tao [216] in an abstract setting. Bourgain spaces X s,b can exploit deeply
dispersive effect of the equation, and analyze the frequency interactions of
waves evolving from the equation. Thus it is usually a powerful tool in the
well-posedness study, see [132; 218]. We take the nonlinear term uux as an
example, then Bourgain space methods reduce to the following two crucial
estimates:
Z t

ψ(t) S(t − s)f (s)ds 6Ckf kX s,b−1 , (5.5)

0 X s,b
k∂x (uv)kX s,b−1 6CkukX s,b kvkX s,b , (5.6)
where ψ(t) is a smooth cut-off function. Intuitively, we can view these two
inequalities in this way: s is the degree of regularity for the space variable,
R t for the operator ∂t − iφ(D). From (5.5) we see the nonlinear
and b is that
evolution 0 S(t − s) · ds gains one derivative with respect to ∂t − iφ(D),
and (5.6) shows this gain can compensate the loss of derivative on the
space variable. How much regularity can be obtained through this process
depends on the strength of dispersive effect of the equation. More precisely,
we will show (5.5) always hold for any φ when 1/2 < b < 1, hence the main
task is to show the second inequality which is the main topics of the next
section.
This section devotes to prove some basic properties of Bourgain X s,b .
Assume ψ ∈ C0∞ (R), suppψ ⊂ [−2, 2], and equals to 1 on [−1, 1]. Let
ψδ (·) = ψ(·/δ) for any δ > 0. We will use the following lemma, which is
proved in Chapter 3.
Lemma 5.1. If 0 < α < 1, 1 < p < ∞, then
k(−∆)α/2 (f g)kLp . kf kL∞ k(−∆)α/2 gkLp + kgkL∞ k(−∆)α/2 f kLp . (5.7)
From the definition of X s,b , it is easy to see that
Lemma 5.2. Let s, b ∈ R. Then

kf kX s,b = hξis ke−itφ(ξ) (Fx f )(ξ, t)kHtb L2 .
ξ
Proposition 5.1. Assume s ∈ R, 1/2 < b < b0 < 1, 0 < δ < 1. Then
0
kψδ (t)f kX s,b .δ 1/2−b kf kX s,b , kψδ (t)f kX s,b−1 .δ b −b kf kX s,b0 −1 .
Proof. We may assume s = 0. It follows from Lemma 5.2 and Lemma

5.1 that

kψδ (t)f kX 0,b = ke−itφ(ξ) ψδ (t)Fx (f )kHtb L2
ξ

. kψδ (t)kHtb ke−itφ(ξ) Fx (f )(ξ, t)kL∞ t
2
Lξ

+ kψδ (t)kL∞ t
ke −itφ(ξ)
Fx (f )(ξ, t)kHtb L2 .
ξ
By simple calculations one easily get for any λ > 0

kf (λt)kḢ b = λb−1/2 kf (t)kḢ b , kf (λt)kH b .(λ1/2 + λb−1/2 )kf (t)kH b . (5.8)
The first inequality follows from (5.8) and Sobolev’s inequality H b ,→ L∞ .
Next we show the second inequality. For simplicity, let c = 1 − b and
d = 1 − b0 , and then 0 6 d < c < 1/2. From Lemma 5.2 it suffices to prove
kψδ (t)hkH −c .δ c−d khkH −d .
t t
By duality it suffices to prove

kψδ (t)gkHtd .δ c−d kgkHtc , ∀ g ∈ Htc . (5.9)
It follows from Hölder’s inequality and Sobolev’s inequality that
kψδ (t)gkL2t .δ c kgkHtc . (5.10)
Then it suffices to prove
kψδ (t)gkḢ d .δ c−d kgkHtc . (5.11)
t
From (5.10) and Gagliardo-Nirenberg inequality we get

kψδ (t)gkḢ d .kψδ (t)gk1−θ
Ḣ c
kψδ (t)gkθL2 .δ c−d kψδ (t)gk1−θ
Ḣ c
kgkθHtc ,
t t t t
where θ = (c − d)/c, thus it suffices to show

kψδ (t)gkḢ c .kgkHtc . (5.12)
t
Indeed,
kψδ (t)gkḢ c .(kψδ kL∞
t
kgkHtc + k((−∆)c/2 ψδ )gkL2t )
t
.(kgkHtc + k((−∆)c/2 ψδ )gkL2t ).

By Hölder’s inequality and Sobolev’s inequality one get
k((−∆)c/2 ψδ )gkL2t .k(−∆)c/2 ψδ kL1/c kgkHtc .kgkHtc . (5.13)
In conclusion, the proof of the proposition is completed.
Proposition 5.2. (a) Let s ∈ R, u0 ∈ H s (Rn ), 1/2 < b < 1, 0 < δ < 1.
Then
kψδ (t)S(t)u0 kX s,b .δ 1/2−b ku0 kH s . (5.14)
s,b−1
(b) Let s ∈ R, f ∈ X and 1/2 < b < 1. Then
Z t

ψδ (t) S(t − t 0
)f (t 0
) .δ 1/2−b kf kX s,b−1 . (5.15)

0 X s,b
Proof. First we show (a). It follows from Lemma 5.2 that

kψδ (t)S(t)u0 kX s,b = kψδ (t)kHtb ku0 kHxs .
Using (5.8), we have
kψδ (t)kHtb 6 (δ 1/2−b + δ 1/2 )kψ(t)kHtb .δ 1/2−b , (5.16)
then we get (5.14).
For (b), It follows from Lemma 5.2 that the left-hand side of (5.15)
equals to
Z t
s −it0 φ(ξ)

hξi ψδ (t) e 0 0
Fx (f )(ξ, t )dt .
2 b
0 Lξ Ht
Thus to prove (b) it suffices to prove for any g ∈ Htb−1

Z t

ψδ (t) g(t 0
)dt 0 1/2−b
kgkH b−1 .
6 Cδ t
0 Htb
Z t
Let h(t) = ψδ (t) g(t0 )dt0 . Simple calculation implies that
0
Z tZ Z
0 eitτ − 1
h(t) = ψδ (t) eit τ b
g(τ )dτ dt0 = ψδ (t) gb(τ )dτ.
0 R R iτ
To control kh(t)kHtb , we divide h into two parts h(t) = h1 (t) + h2 (t), where
Z Z
eitτ − 1 eitτ − 1
h1 (t) = ψδ (t) gb(τ )dτ, h2 (t) = ψδ (t) gb(τ )dτ.
|τ |61 iτ |τ |>1 iτ
Since ktn ψδ (t)kH b = δ n k(t/δ)n ψ(t/δ)kH b 6 δ n δ 1/2−b 4n , then from Taylor’s
expansion we get
Z X (itτ )n

kh1 kHtb = ψδ (t) b
g(τ )dτ b
|τ |61 n>1 n!(iτ ) Ht
X ktn ψδ (t)kH b Z
g(τ )
(iτ )n b
. dτ .kgkH b−1 . (5.17)
n! |τ |61 (iτ ) t
n>1
Next we need to control kh2 kHtb . Divide it into two parts

Z Z
g(τ )dτ
b eitτ gb(τ )

kh2 kHtb . ψδ (t) + ψδ (t) dτ := I + II.
|τ |>1 iτ b
|τ |>1 iτ
Ht Htb
For the term I we have

Z hτ i
2(1−b) 1/2
I.kψδ kH b kgkH b−1 dτ .δ 1/2−b kgkH b−1 . (5.18)
t
|τ |>1 |τ |2 t
For the term II we have

−1 χ|τ |>1 ĝ(τ )

II = ψδ (t)Ft (t)
iτ b
Ht

−1 χ|τ |>1 ĝ(τ ) −1 χ|τ |>1 ĝ(τ )

.kψδ kH b Ft
iτ ∞ + kψδ kL Ft
∞
iτ
L Htb
.δ 1/2−b kgkH b−1 . (5.19)

t
Combining (5.17), (5.18)(5.19), then we obtain (5.15) if 1/2 < b < 1.
Lemma 5.3. Assume Y is a space-time Banach space such that for all
u0 ∈ L2 (Rn ) and τ0 ∈ R
keitτ0 S(t)u0 kY .ku0 kL2 (Rn ) .
Then when 1/2 < b < 1, for any u ∈ X 0,b
kukY .kukX 0,b .
Proof. From the inverse Fourier transform and change of variables we

get
Z
u(x, t) = (2π)−(n+1)/2 b(ξ, τ )eixξ+itτ dξdτ
u
Rn+1
Z Z
−(n+1)/2 itτ
= (2π) e b(ξ, τ + φ(ξ))eixξ eitφ(ξ) dξdτ.
u
R Rn
Thus from Minkowski’s inequality and the hypothesis on Y we get

Z Z
kukY . kb u(ξ, τ + φ(ξ))kL2ξ dτ . hτ i−b khτ ib u
b(ξ, τ + φ(ξ))kL2ξ dτ.
R R
The lemma follows from the condition b > 1/2 and Cauchy-Schwartz in-
equality.
If b = 1/2, there are counter-examples to show both Proposition 5.2

(b) and Lemma 5.3 fail. We leave this to the readers as an exercise. When
b = 1/2, one has a good substitute for X s,b : Besov-type Bourgain space F s .
We will use this space combined with a special low frequency structure to
solve the H −3/4 GWP problem for the KdV equation in Section 5.4. Now
we introduce the space F s . For convenience, for k ∈ N ∪ {0}, denote
I0 = [−2, 2]; Ik = [2k−1 , 2k+1 ], k > 1. (5.20)
Assume {ϕk }∞ k=0 is the sequence of functions for the non-homogeneous
dyadic decomposition which are constructed in Section 1.3, and {4k }∞ k=0
are the corresponding Littlewood-Paley projector operators. For k ∈ N∪{0}
we define the dyadic X s,b -type space Xk (R2 ):
( )
f ∈ L2 (R2 ) : f (ξ, τ ) is supported in Ik × R and
Xk = P∞ . (5.21)
kf kXk = j=0 2j/2 kϕj (τ − φ(ξ)) · f (ξ, τ )kL2ξ,τ < ∞
Then we resemble these space in a Littlewood-Paley manner

X
kuk2F s = 22sk kϕk (ξ)F uk2Xk , (5.22)
k>0
X
kuk2N s = 22sk khτ − φ(ξ)i−1 ϕk (ξ)F uk2Xk . (5.23)
k>0
These Besov-type X s,b space was first introduced by Tataru [221], and then
widely used by Ionescu and Kenig [113; 115], Tao [219] and the authors
[96; 90; 91; 92]. F s is a good substitute of X s,1/2 for the following two
reasons: Firstly, it is easy to see X s,1/2+ ⊂ F s ⊂ X s,1/2 , and moreover,
F s space can control many space-time norm such as Strichartz-type space
Lpt Lqx (while X s,1/2 fails logarithmicly), secondly, F s has the same scale in
time as X s,1/2 . This is very similar as the comparison between Besov space
n/2
B2,1 and Sobolev space H n/2 , H n/2+ .
First we prove the linear estimates in the space F s .
Proposition 5.3. (a) Let s ∈ R and u0 ∈ H s . Then
kψ(t)S(t)u0 kF s .ku0 kH s .
(b) Assume s ∈ R, k ∈ N ∪ {0} and hτ − φ(ξ)i−1 F u ∈ Xk . Then there
exists C > 0 independent of k such that
Z t

F ψ(t) S(t − s)u(s)ds 6 Ckhτ − φ(ξ)i−1 F ukX .
k
0 Xk
Proof. To prove (a), from the definition of F s it suffices to show that for
any k ∈ N ∪ {0}
kϕk (ξ)F (ψ(t)S(t)u0 )kXk .kϕk (ξ)b
u0 (ξ)kL2 .
By the definition of Xk we get
X
kϕk (ξ)F (ψ(t)S(t)u0 )kXk = b )c
2j/2 kϕk (ξ)ϕj (τ )ψ(τ u0 (ξ)kL2
j>0
.kψkB 1/2 kϕk (ξ)b

u0 (ξ)kL2 ,
2,1
which implies (a).

Next we show (b). It follows from simple calculation that
Z t
F ψ(t) S(t − s)u(s)ds (ξ, τ )
0
Z b − τ 0 ) − ψ(τ
b − φ(ξ))
ψ(τ
=C F u(ξ, τ 0 ) 0
dτ 0 .
R τ − φ(ξ)
Let f (ξ, τ ) = F u(ξ, τ 0 )hτ 0 − φ(ξ)i−1 . For fk ∈ Xk we define the operator
0
T:
Z b − τ 0 ) − ψ(τ
b − φ(ξ))
ψ(τ
T (fk )(ξ, τ ) = fk (ξ, τ 0 ) 0
hτ 0 − φ(ξ)idτ 0 .
R τ − φ(ξ)
Then it suffices to prove that
kT fk kXk 6 Ckfk kXk
hold uniformly for k > 0.
Denote fk] (ξ, τ ) = fk (ξ, τ + φ(ξ)), (T fk )] (ξ, τ ) = (T fk )(ξ, τ + φ(ξ)). By
change of variables we get
Z b − τ 0 ) − ψ(τ
b )
ψ(τ
(T fk )] (ξ, τ ) = fk] (ξ, τ 0 ) 0
hτ 0 idτ 0 .
R τ

ψ(τ
b − τ 0 ) − ψ(τ
b )
0
0
hτ i 6 C[(1 + |τ |)−4 + (1 + |τ − τ 0 |)−4 ].
τ
]
For j ∈ N∪{0}, define fk,j = fk (ξ, τ )ϕj (τ −φ(ξ)) and fk,j = fk,j (ξ, τ +φ(ξ)).
Then we have
"Z #1/2
] −4 j/2 ] 0 2 0
(T fk,j ) (ξ, τ )|.(1 + |τ |) 2 |fk,j (ξ, τ )| dτ
Ij
5.2. Local smoothing effect and maximal function estimates 99
2
X Z
]
+ ϕj+l (τ ) |fk,j (ξ, τ 0 )|(1 + |τ − τ 0 |)−4 dτ 0
l=−2 Ij
:=I + II.
Hence
X 0
kT fk kXk . 2j /2
kϕj 0 (τ )(T fk )] (ξ, τ )kL2
j 0 >0
X 0
. 2j /2 kϕj 0 (τ )(T fk,j )] (ξ, τ )kL2 .
j 0 ,j>0
For the term I, it is obvious that

X 0
2j /2 kϕj 0 (τ )IkL2 .kfk kXk .
j 0 ,j>0
For the term II, it follows from Young’s inequality that

X 0
2j /2 kϕj 0 (τ )IIkL2 .kfk kXk .
|j 0 −j|.1
Therefore, the proposition is proved.

We give an analogue of Lemma 5.3 in the case b = 1/2 in the following
lemma which can be proved similarly.
Lemma 5.4. Assume Y is space-time Banach space such that for all u0 ∈
L2 (Rd ) and τ0 ∈ R
keitτ0 S(t)u0 kY 6 Cku0 kL2 (Rn ) .
Then for any k ∈ N ∪ {0} and u ∈ F 0
\
k4k (u)kY .k4 k (u)kXk .
5.2 Local smoothing effect and maximal function estimates
In the last Section, we already see that the elements of Bourgain space X s,b
are very close to the linear solutions u = S(t)u0 with initial data u0 ∈ H s .
Thus to use X s,b we need to get the estimates for the linear solutions. In
studying the dispersive equations with derivatives in the nonlinear terms,
there are two important kinds of estimates known as Local smoothing effect
and maximal function estimates. In this section, we introduce these two
α
estimates using S α (t) = Fx−1 eit|ξ| ξ Fx , 1 6 α 6 2 as examples. The
results were due to Kenig-Ponce-Vega, for example see [127; 126].
Theorem 5.1. Assume 1 6 α 6 2. Then

Z 1/2 Z 1/2
α 2 u0 (ξ)|2
|c
sup |S (t)u0 | dt . dξ , (5.24)
x R R |ξ|α
Z 1/4 Z 1/2
sup |S α (t)u0 |4 dx . u0 (ξ)|2 |ξ|1/2 dξ
|c . (5.25)
R t∈R R
Proof. Let ω(ξ) = |ξ|α ξ. Then ω is invertible and the inverse is denoted
by ω −1 . By change of variable η = ω(ξ), we have
Z Z
α ixξ itω(ξ) −1 b0 (ω −1 (η))
u
S (t)u0 = C e e û0 (ξ)dξ = C eixω (η) eitη 0 −1 dη.
R R ω (ω (η))
Using Plancherel’s equality and making change of varible η = ω(ξ) (noting
ω 0 (ξ) = (α + 1)|ξ|α ), we get
Z Z
α 2 u0 (ω −1 (η))|2
|c u0 (ξ)|2
|c
kS (t)u0 kL2 . 0 −1 (η))|2
dη. dξ.
t
R |ω (ω R |ξ|α
Therefore, (5.24) is proved.
Next we prove (5.25). The main idea is to permute the position of
x, t, namely, view t as space variable and x is time variable, then L4x L∞ t
becomes Strichartz-type estimate, and (4, ∞) is usually admissible pair.
From the proof of the first inequality we see S α (t)u0 can be viewed as the
linear solution to the dispersive equation with dispersive relation ω −1 (η),
−1
and with the initial data whose Fourier transform is uω c0 (ω (η))
0 (ω −1 (η)) . It is easy
to check that ω −1 (ξ) satisfies (H1), (H2), (H3), (H4) in Section 2.1, and
m1 = m2 = α1 = α2 = 1/(α + 1). Then we get the Strichartz estimate
−1
kFx−1 eitω (ξ)
Fx f kL4t L∞
x
.kf k 1− 1 .
Ḣ 2 4(α+1)
Thus we have
Z !1/2
1 − 1 û0 (ω −1 (η) 2
α
kS (t)u0 kL4x L∞
. |η| dη .
2 4(α+1)
t
R ω 0 (ω −1 (η))
By change of variable η = ω(ξ) we get
kS α (t)u0 kL4x L∞
t
.ku0 kḢ 1/4 ,
which is (5.25) as desired.

The proof we gave here is a simplified argument for the concrete exam-
ple. In [126], Kenig, Ponce and Vega proved a more precise estimate for a
general class of dispersive equation
Z 41 Z 0 1/2 !1/2
4 2 ω (ξ)
sup |S(t)u0 (x)| dx . |c
u0 (ξ)| dξ .
R t∈R R ω 00 (ξ)
Next we give another maximal function estimate for S α (t), which was
proved in [127].
Theorem 5.2. Assume s > α+1 s

4 , α > 1 and u0 ∈ H . Then
Z
∞ 1/2
sup |S α (t)u0 |2 dx 6 Cku0 kH s , (5.26)
−∞ t∈[−T,T ]
where the constant C > 0 depends on T .
To prove Theorem 5.2, we need the following lemma.
Lemma 5.5. Assume ϕ is a C ∞ function supported in [2k−1 , 2k+1 ] and

k ∈ N. For α > 1, define the function Hkα (·)
 k
 2 , |x| 6 1,
Hkα (x) = 2k/2 |x|−1/2 , 1 6 |x| 6 C2αk ,

1/(1 + x2 ), |x| > C2αk .
Then, if |t| 6 2 there exists C > 0 independent of x, t and k such that
Z

ei(t|ξ|α ξ+xξ) ϕ(ξ)dξ 6 CHkα (x). (5.27)

R
Moreover, if k = 0 and ϕ is smooth function supported in [−2, 2], then the

same conclusion holds.
R α
Proof. Let u(x, t) = R ei(t|ξ| ξ+xξ) ϕ(ξ)dξ. First we have the trivial esti-
mate for any x, t ∈ R
|u(x, t)| 6 C2k .

Thus it remains to consider the case |x| > 1.
Assume k = 0. It follows from integration by parts and |t| 6 2 that
Z 2
−2 ixξ d i(t|ξ|α ξ)

|u(x, t)|. x e e ϕ(ξ) dξ .|x|−2 .
R dξ 2
Thus it remains to consider the case k > 1.
Define the set Ω = {ξ ∈ supp ϕ : |(α + 1)t|ξ|α + x| 6 |x|/2}. Choose

a cut-off function ζ ∈ C ∞ supported in Ω such that ζ = 1 when |(α +
1)t|ξ|α + x| 6 |x|/3, for example we may set

(α + 1)t|ξ|α + x
ζ =η ,
|x|
where η is a fixed smooth cut-off function.
Since |t| 6 2, if ξ ∈ Ω, then |x| ∼ αt|ξ|α 6 C2αk , hence h(ξ, x) =
tξ|ξ|α + xξ satifies
|h00 (ξ)| = Ct|ξ|α−1 > C2−k |x|.
It follows from Van der Corput’s lemma that

Z

eih(ξ,x) ϕ(ξ)ζdξ 6 C2k/2 |x|−1/2 .

If ξ ∈ supp (1 − ζ), then |h0 (ξ)| = |(α + 1)t|ξ|α + x| > |x|/3. Integrating
by parts, we get for |x| > 1
Z

eih(ξ,x) (1 − ζ)ϕ(ξ)dξ 6 C .
1 + x2
In view of the estimates above, the lemma is proved.

Using the lemma above, we can prove the following lemma which implies
Theorem 5.2 immediately.
1+α
Lemma 5.6. If s > 4 , α > 1, then
X
∞ 1/2
sup sup |S α (t)u0 |2 6 Cku0 kH s . (5.28)
j<x<j+1
j=−∞ |t|61
α
Proof. Define Skα (t)u0 = Fx−1 eitξ|ξ| ϕk (ξ)Fx u0 , where {ϕk }∞
k=0 is non-
homogeneous dyadic functions. It suffices to prove
X
∞ 1/2 (1+α)k
sup sup |Skα (t)u0 |2 .2 4 ku0 kL2 .
j=−∞ |t|61 j<x<j+1
By duality we need to prove

Z 1 ∞ Z
h X 1 Z j+1 2 i1/2
(1+α)k
S α
(t)g(·, t)dt .2 4 |g(x, t)|dxdt .
k
−1 L2 j=−∞ −1 j
(5.29)
By duality (see [223]), (5.29) follows from the following:

X
∞ Z 1 2 1/2

sup sup Skα (t − t0 )g(·, t0 )dt0
j<x<j+1
j=−∞ |t|61 −1
∞ Z
h X 1 Z j+1 2 i1/2
(1+α)k
.2 2 |g(x, t)|dxdt . (5.30)
j=−∞ −1 j
To prove (5.30), we get from Lemma 5.5 that

Z 1 Z Z 1
α 0 0 0 α
Sk (t − t )g(x, t )dt . Hk (y) |g(x − y, t0 )|dt0 dy
−1 −1
∞
X Z l+1 Z 1
. Hkα (l) |g(x − y, t0 )|dt0 dy.
l=−∞ l −1
Thus the left-hand side of (5.30) is bounded by

h X
∞ ∞
X Z l+1 Z 1 2 i1/2
sup Hkα (l) |g(x − y, t0 )|dt0 dy (5.31)
j=−∞ j6x<j+1 l −1
l=−∞
By change of variables we get

∞ X
h X ∞ Z l−j+2 Z 1 2 i1/2
(5.31). Hkα (l) |g(z, t0 )|dt0 dz .
j=−∞ l=−∞ l−j−1 −1
It follows from Minkowski’s inequality that

X∞ h X ∞ Z l−j+2 Z 1 2 i1/2
α
(5.31). Hk (l) |g(z, t0 )|dt0 dz
l=−∞ j=−∞ l−j−1 −1
X∞ ∞ Z
h X j+1 Z 1 2 i1/2
. Hkα (l) |g(x, t0 )|dt0 dx
l=−∞ j=−∞ j −1
∞ Z
h X 1 Z j+1 2 i1/2
(α+1)k
.2 2 |g(x, t0 )|dxdt0 ,
j=−∞ −1 j
where in the last inequality we use the following simple fact:

[2αk ] [2αk ] Z l+1
X X
−1/2
l . x−1/2 dx.2αk/2 .
l=1 l=1 l
Thus we complete the proof of the lemma.

The maximal function associated to dispersive equation was first pro-

posed by Carleson [24]. We take Schrödinger equation as an example,
(
iut + ∆u = 0,
u(x, 0) = u0 (x).
From Plancherel’s equality we know as t → 0 the solution u = eit∆ u0 → u0

in H s . Carleson asked the following question: under what condition one
has u(x, t) → u0 (x) almost everywhere? To answer the question, he asked
further: what is least regularity on u0 ∈ H s so that sup|t|61 |eit∆ u0 (x)| is
locally integrable?
This problem was completely solved in one dimension, but remains un-
clear in high dimension. A general conjecture states as following: in any
dimension u(x, t) → u0 (x) almost everywhere if and only if s > 1/4. In one
dimension we already show if s > 1/2
keit∂xx u0 kL2x L∞
|t|61
.ku0 kH s .
A conjecture of Kenig and Vega states that it still hold when s = 1/2.
Actually we show the inequality holds when s = 1/2 and kukH s is replaced
s
by B2,1 . For these problems we refer the readers to [126; 127].
5.3 Bilinear estimates for KdV and local well-posedness
In this section we consider Korteweg-de Vries (KdV) equation, and prove

the local well-posedness by showing the bilinear estimates in X s,b . KdV
equation is given by the following form:
(
∂t u + ∂x3 u + 6u∂x u = 0, (x, t) ∈ R2 ,
(5.32)
u(x, 0) = u0 (x) ∈ H s (R).
Then the dispersion relation for KdV equation is φ(ξ) = ξ 3 . KdV equation
is the fundamental equation in shallow water theory, which was founded
in 1895 by Korteweg and de Vries. From the discussion in Section 5.1 we
know the main task is to prove the bilinear estimate. First we prove
Theorem 5.3. If −3/4 < s < −1/2, then there exists b ∈ (1/2, 1) such
that for any b0 ∈ (1/2, b] with b − b0 6 min{−s − 1/2, 1/4 + s/3} we have
k∂x (u1 u2 )kX s,b−1 6 Cku1 kX s,b0 ku2 kX s,b0 . (5.33)

5.3. Bilinear estimates for KdV and local well-posedness 105
We will prove Theorem 5.3 by adopting the [k; Z]-multiplier ideas. The
method is quite standard now. From the definition of X s,b , (5.33) is equal
to
0 0
khξis hτ −ξ 3 ib−1 ξ(b v )(ξ, τ )kL2ξ,τ .khξis hτ −ξ 3 ib u
u ∗b bkL2ξ,τ khξis hτ −ξ 3 ib vbkL2ξ,τ .
For the convenience, we denote
Λs,b (ξ, τ ) = hξis hτ − ξ 3 ib .
Then (5.33) is equal to

ξΛs,b−1 (ξ, τ )[(Λ−1 −1
s,b u) ∗ (Λs,b v)](ξ, τ ) .kukL2(R2 ) kvkL2 (R2 ) . (5.34)
L2ξ,τ
Assume {ϕk }∞ ∞
k=0 ({χk }k=−∞ ) is the nonhomogeneous (homogeneous)
dyadic functions, the intervals {Ik }∞ k=0 are given by (5.20). Decompose
ξi , τi − ξi3 into dyadic pieces, then by duality we get (5.34) is equivalent to
X h2k3 is 2j3 (b−1) 2k3
h2k1 is 2j1 b0 h2k2 is 2j2 b0
ki ∈Z,ji ∈Z+ ,
Z
× [1Dk3 ,j3 (uk1 ,j1 ∗ vk2 ,j2 )f ](ξ, τ )dξdτ .kukL2 kvkL2 kf kL2 , (5.35)
where L2 = L2 (R2 ), for k ∈ Z, j ∈ Z+ we denote

Dk,j = {(ξ, τ ) : ξ ∈ [2k−1 , 2k+1 ], τ − ξ 3 ∈ Ij },
and
uk1 ,j1 = χk1 (ξ)ϕj1 (τ − ξ 3 )u.
Thus to prove Theorem 5.3, we need to prove first
Z

1 (ξ, τ )(u ∗ v ) · f dξdτ .
2 Dk3 ,j3 k1 ,j1 k2 ,j2
R
So far, the issues reduce to some elementary calculus. We rewrite it in a
symmetric form, then the bilinear estimates will follow from the trilinear
estimates below
Z 3
Y
f1 (ξ1 , τ1 )f2 (ξ2 , τ2 )f3 (ξ3 , τ3 ).C(ki , ji ) kfi kL2 , (5.36)
Γ3 i=1
where fi are non-negative functions supported in Dki ,ji , and
Γ3 = {ξ1 + ξ2 + ξ3 = 0, τ1 + τ2 + τ3 = 0},
endowed with induced measure 3 . The estimates of the form (5.36) were
first studied by Bourgain [14; 15] using the relations between X s,b and
3
R R
Γ3
Πi=1,2,3 fi (ξi , τi ) = R4
f1 (ξ1 , τ1 )f2 (ξ2 , τ2 )f3 (−ξ1 − ξ2 , −τ1 − τ2 )dξ1 dξ2 dτ1 dτ2 .
other space-time norm, then by Kenig, Ponce, Vega [132] using elementary
calculus. Later, Tao [216] systematically studied that in an abstract setting
by using dyadic decomposition.
We will study the estimate of (5.36) in the next lemma, while for a
general class of dispersion relation φ(ξ) = |ξ|α ξ we refer the readers to [92].
Our proofs are simplified version of the [k; Z]-multiplier, quite elementary
but a little complicated. We suggest the readers to see the crucial ideas in
the calculation. For convenience, if a1 , a2 , a3 ∈ R, we use amin 6 amed 6
amax to denote the maximum, medium, minimum among a1 , a2 , a3 . We
also denote
(f1 ∗ f2 ) · f3 = f1 ∗ f2 · f3 .
Lemma 5.7. Assume ki ∈ Z, ji ∈ Z+ , and fki ,ji ∈ L2 (R2 ) are non-

negative function supported in ∪jl=0 i
Dki ,l and kfki ,ji k2 6 1, i = 1, 2, 3.
Then
(a) for any k1 , k2 , k3 ∈ Z and j1 , j2 , j3 ∈ Z+
Z
fk1 ,j1 ∗ fk2 ,j2 · fk3 ,j3 dξdτ 6 C2jmin /2 2kmin /2 ; (5.37)
R2
(b) if kmin 6 kmax − 10, then for i = 1, 2, 3

Z
fk1 ,j1 ∗ fk2 ,j2 · fk3 ,j3 dξdτ 6 C2(j1 +j2 +j3 )/2 2−kmax /2 2−(ji +ki )/2 ; (5.38)
R2
(c) if kmin > kmax − 10 > 10, then
Z
fk1 ,j1 ∗ fk2 ,j2 · fk3 ,j3 dξdτ 6 C2jmin /2 2jmed /4 2−kmax /4 . (5.39)
R2
R
Proof. For f, g, h ∈ L2 (R2 ), let J(f, g, h) = R2 f ∗ g · hdξdτ . By change
of variables we get
|J(f, g, h)| = |J(g, f, h)| = |J(fe, h, g)| = |J(f, e
g , h)|,
where fe(ξ, µ) = f (−ξ, −µ). Note that φ(ξ) = ξ 3 is an odd function4 , Thus
feki ,ji and fki ,ji have the same support. By Cauchy-Schwartz inequality and
support properties of functions fki ,ji we get
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 )
Z
jmin /2
.2 kfk1 ,j1 (ξ1 , ·)kL2τ kfk2 ,j2 (ξ2 , ·)kL2τ kfk3 ,j3 (ξ1 + ξ2 , ·)kL2τ dξ1 dξ2
R2
4 If not odd, there are less symmetries, but the methods still work.
3
Y
.2kmin /2 2jmin /2 kfki ,ji kL2 ,
i=1
which is (a).
For (b), from the support properties we know J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 ) = 0
unless
|kmax − kmed | 6 5. (5.40)
From symmetry we may assume k1 6 k2 6 k3 , hence |k2 − k3 | 6 5. Then

we divide it into three cases: j1 = jmax , j2 = jmax , j3 = jmax . If j3 = jmax ,
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 ) can be rewritten into the following form:
Z
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 ) = fk]1 ,j1 (ξ1 , τ1 )fk]2 ,j2 (ξ2 , τ2 )
R4
× fk]3 ,j3 (ξ1 + ξ2 , τ1 + τ2 + Ω(ξ1 , ξ2 ))dξ1 dξ2 dτ1 dτ2 ,
where fk]i ,ji (ξ, τ ) = fki ,ji (ξ, τ + ξ 3 ), i = 1, 2, 3, and
Ω(ξ1 , ξ2 ) = ξ13 + ξ23 − (ξ1 + ξ2 )3 = −3ξ1 ξ2 (ξ1 + ξ2 ).
By change of variable ξ20 = ξ1 + ξ2 and Hölder’s inequality we get
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 )

Z
= fk]1 ,j1 (ξ1 , τ1 )fk]2 ,j2 (ξ2 − ξ1 , τ2 )
R4
× fk]3 ,j3 (ξ2 , τ1 + τ2 + Ω(ξ1 , ξ2 − ξ1 ))dξ1 dξ2 dτ1 dτ2
Z
. kfk]1 ,j1 (·, τ1 )kL2 (R) kfk]2 ,j2 (·, τ2 )kL2 (R)
R2
× kfk]3 ,j3 (ξ2 , τ1 + τ2 + Ω(ξ1 , ξ2 − ξ1 ))kL2 (|ξi |∼Ni , i=1,2) dτ1 dτ2 .
(5.41)
By
change of variable µ2 = τ1 + τ2 + Ω(ξ1 , ξ2 − ξ1 ) and noting
∂ξ1 [Ω(ξ1 , ξ2 − ξ1 )] ∼ N22 , we get
kfk]3 ,j3 (ξ2 , τ1 + τ2 + Ω(ξ1 , ξ2 − ξ1 ))kL2 (|ξi |∼Ni , i=1,2) . N2−1 kfk]3 ,j3 kL2 (R2 ) .
(5.42)
It follows from (5.41), (5.42) and Hölder’s inequality that
3
Y
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 ) . 2−k3 kfki ,ji kL2 (R2 ) .
i=1
If j2 = jmax , from symmetry it is identical to the case j3 = jmax . If

j1 = jmax , it follows from symmetry that
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 )
Z
= fk]1 ,j1 (ξ1 , τ2 + τ3 + Ω(ξ2 , ξ1 − ξ2 ))fk]2 ,j2 (ξ2 , τ2 )fk]3 ,j3 (ξ1 − ξ2 , τ3 ).
R 4

Note that ∂ξ2 [Ω(ξ2 , ξ1 − ξ2 )] ∼ N2 N1 , then repeating the argument we
get
3
Y
J(fk1 ,j1 , fk2 ,j2 , fk3 ,j3 ) . 2−(k1 +k3 )/2 kfki ,ji kL2 (R2 ) ,
i=1
hence (b) is proved.
Now we prove (c). For the simplicity of notations, we denote fi =
fk]i ,ji , i = 1, 2, 3. We may assume j1 6 j2 6 j3 and rewrite J(f1 , f2 , f3 )
into the following form
Z
f1 (ξ1 , µ1 )f2 (ξ2 − ξ1 , µ2 − µ1 − ξ13 − (ξ2 − ξ1 )3 )
R4
× f3 (ξ2 , µ2 − ξ23 )dξ1 dµ1 dξ2 dµ2 .
It follows from Cauchy-Schwartz inequality that J(f1 , f2 , f3 ) is bounded by
Z

3 3
2 f1 (ξ1 , µ1 )f2 (ξ2 − ξ1 , µ2 − µ1 − ξ1 − (ξ2 − ξ1 ) )dξ1 dµ1 2 · kf3 kL2 .
R Lξ
2 ,µ2
The integral area in L2ξ2 ,µ2 is

E = {(ξ1 , µ1 ) : |µ1 |.2j1 , ξ13 + (ξ2 − ξ1 )3 = µ2 − µ1 + O(2j2 )},
where |ξ2 | ∼ 2k2 . Since ξ13 + (ξ2 − ξ1 )3 = 3ξ2 (ξ1 − ξ2 /2)2 + ξ23 /4, then
3ξ2 (ξ1 − ξ2 /2)2 + ξ23 /4 − µ2 + µ1 = O(2j2 ),
moreover, if ξ23 /4 − µ2 + µ1 > 0 then
|ξ1 − ξ2 /2|.2(j2 −k2 )/2 ,
or, if ξ23 /4 − µ2 + µ1 > 0, for some θ > 0,
(ξ1 − ξ2 /2 + θ)(ξ1 − ξ2 /2 − θ) = O(2j2 −k2 ),
and then |ξ1 − ξ2 /2 + θ| . 2(j2 −k2 )/2 or |ξ1 − ξ2 /2 − θ| . 2(j2 −k2 )/2 . In
any case we always have |E|.2j1 2(j2 −k2 )/2 . Thus from Cauchy-Schwartz
inequality we have
J(f1 , f2 , f3 )
.2j1 /2 2(j2 −k2 )/4 kf1 (ξ1 , µ1 )f2 (ξ2 − ξ1 , µ2 − µ1 − ξ13 − (ξ2 − ξ1 )3 )kL2ξ
2 ,µ2 ,ξ1 ,µ1
j1 /2 (j2 −k2 )/4
.2 2 ,
which is (c), and therefore, the lemma is proved.
Now we prove Theorem 5.3 using the lemma. In order to better under-
stand how the conditions are imposed, we divide the proof into many cases.
For each case one has a condition, then intersection of all conditions would
ensure the theorem.
Proof. [Proof of Theorem 5.3] From the previous discussion it suffices to

prove (5.35). We may assume kuk2 = kvk2 = 1. Then we need to show
X h2k3 is 2j3 (b−1) 2k3
sup 0 0
1D · u k

1 ,j1 ∗ vk2 ,j2 L2
k s j b k s
h2 1 i 2 1 h2 2 i 2 2 j b k3 ,j3
kmax >1 ξ,τ
kmin 6kmax ,ji ∈Z+ ,
(5.43)
is bounded. Actually, from the support properties of uk1 ,j1 , vk2 ,j2 we see
1Dk3 ,j3 · uk1 ,j1 ∗ vk2 ,j2 = 0 unless
|kmax − kmed | 6 5, 2jmax ∼ max(2jmed , 2k1 +k2 +k3 ). (5.44)
Thus one may assume (5.44) in (5.43). Now we prove (5.35) using (5.43).
Divide the left-hand side of (5.35) into several parts, we may assume
(k2 , k3 ) = (kmed , kmax ), and k2 , k3 > 1, then from (5.43) we get
X Z
h2k3 is 2j3 (b−1) 2k3
[1Dk3 ,j3 (uk1 ,j1 ∗ vk2 ,j2 )f ](ξ, τ )dξdτ
|k2 −k3 |65,ji ∈Z+ ,
X
. kukL2 k4k2 vkL2 k4k3 f kL2 .kukL2 kvkL2 kf kL2 ,
|k2 −k3 |65
which is (5.35). Next we study (5.43) according to the frequency interac-

P P
tions. Fix kmax , then the summation below ki ∈Z means kmin ∈Z .
Case 1 (low-low interactions): kmax 6 100. It follows from Lemma 5.7
(a) that
X h2k3 is 2j3 (b−1) 2k3
(5.43). 2jmin /2 2kmin /2 .1.
ki ∈Z,ji ∈Z+
Case 2 (high-low interaction): k2 > 100, |k3 − k2 | 6 5, k1 6 k2 − 10 (or

k1 , k2 exchange). If j3 = jmax , then we get from Lemma 5.7 (b) that
X 2j3 (b−1) 2k3
(5.43). min(2(j1 +j2 )/2 2−k2 , 2j1 /2 2k1 /2 ).
h2k1 is 2j1 b0 2j2 b0
ki ∈Z,ji ∈Z+
Divide the summation into three parts, the first part is k1 6 −2k2 , denoted
by I, the second part is k1 > −2k2 and 2k1 /2 6 2j2 /2−k2 , denoted by II, the
third part is k1 > −2k2 and 2k1 /2 > 2j2 /2−k2 , denoted by III. It is easy to
see that I.1. For II, summing on j1 , j2 , j3 we get
X 2(k1 +2k2 )(b−1)
II. .
h2k1 is
−2k2 6k1 6k2
Then it is easy to get that II.1 if 1 + s 1

and k1 6 1. The term III can be similarly handled and we get
X 2(k1 +2k2 )(b−1)
III. .
h2k1 is
−2k2 6k1 6k2
Thus if 1 + s < b 6 1 + s/3 then III.1.

If j2 = jmax , then from a similar argument and using Lemma 5.7 (b)
we can get if 0 6 b − b0 6 1/2 + s/3,
X 2j3 (b−1) 2k3
(5.43). min(2(j1 +j3 )/2 2−k2 , 2j1 /2 2k1 /2 ).1.
h2k1 is 2j1 b0 2j2 b0
ki ∈Z,ji ∈Z+
If j1 = jmax , it follows from Lemma 5.7 (b) that if 0 6 b − b0 6 1/4,

X 2j3 (b−1) 2k3
(5.43). min(2(j2 +j3 )/2 2−(k2 +k1 )/2 , 2j2 /2 2k1 /2 ).1.
h2k1 is 2j1 b0 2j2 b0
ki ∈Z,ji ∈Z+
Case 3 (high-low interaction II): k2 > 100, |k3 − k2 | 6 10, k1 > k2 − 10.
Similarly, it follows from Lemma 5.7 (c) that if s > −3/4, 1/2 1
Case 4 (high-high interactions): k2 > 100, |k1 − k2 | 6 5, k3 6 k2 − 10.

If j3 = jmax , from Lemma 5.7 (c) we get 1/2 < b < 5/4 + s,
X h2k3 is 2j3 (b−1) 2k3
(5.43) . 2(j1 +j2 )/2 2−k1 /2 2−k3 /2
ki ∈Z,ji ∈Z+
X h2k3 is max(1, 2k3 +2k1 )b−1 2k3 /2
. .1. (5.45)
22sk1 2k1 /2
ki ∈Z
If j1 = jmax , from Lemma 5.7 (c) we get if b − b0 6 −1/2 − s,

X h2k3 is 2j3 (b−1) 2k3
(5.43) . 2(j2 +j3 )/2 2−k1
ki ∈Z,ji ∈Z+
X h2k3 is max(1, 2k3 +2k1 )b−b0 −1/2 2k3 −k1

. .1. (5.46)
22sk1
ki ∈Z
The case j2 = jmax is identical to the case j1 = jmax .

Taking all the conditions, we proved that if s ∈ (−3/4, −1/2), 1/2 <
b < 3/4 + s/3, 1/2 < b0 6 b with b − b0 6 −1/2 − s, then (5.33) holds.
Therefore, Theorem 5.3 is proved.
If s is large, one has similar estimates. Actually, the same methods
above also shows the following theorem. For a ∈ R, a+ denotes a + for a
fixed 0 < 1.
Theorem 5.4. If −3/4 < s 6 0, then there exists b ∈ (1/2, 1) such that
k∂x (u1 u2 )kX s,b−1 6 Cku1 kX s,b ku2 kX s,b (5.47)
and
k∂x (u1 u2 )kX s,b−1 .ku1 kX −3/4+,b ku2 kX s,b + ku1 kX s,b ku2 kX −3/4+,b . (5.48)
Remark 5.1. The condition s > −3/4 in Theorem 5.4 is necessary. When
s 6 −3/4, (5.47) fails for any b ∈ R. The case s < −3/4 was due to
Kenig, Ponce and Vega [132], and the case s = −3/4 was due to Nakanishi,
Takaoka and Tsutsumi [180].
Now we prove the local well-posedness for KdV equation using the bi-
linear estimates. First assume u0 ∈ H s , −3/4 < s < −1/2, and define the
operator and set:
Z t
Φu0 (u) = ψ1 (t)S(t)u0 − ψ1 (t) S(t − t0 )ψT (t0 )∂x (u2 )(t0 )dt0 ,
0
B = {u ∈ X s,b : kukX s,b 6 2Cku0 kH s }.

We will show if T is sufficiently small, the map Φu0 is a contraction mapping
in B.
From Proposition 5.2, Proposition 5.1 and Theorem 5.3, we have for
1/2 < b < b0 < 1
0
kΦu0 (u)kX s,b 6Cku0 kH s + CT b −b kuk2X s,b
0
6Cku0 kH s + 4C 2 T b −b ku0 k2H s .
Thus choose T such that
0
4CT b −b ku0 kH s 6 1/2,
then we have Φu0 (B) ⊂ B. Assume (u1 , u2 ) ∈ B, and we have

kΦu0 (u1 ) − Φu0 (u2 )kX s,b 6 1/2ku1 − u2 kX s,b .
Thus, Φu0 is a contraction mapping in B and there exists a unique u ∈ B
such that Φu0 (u) = u:
Z t
u(t) = ψ1 (t)S(t)u0 − ψ1 (t) S(t − t0 )ψT (t0 )∂x (u2 )(t0 )dt0 .
0
It is easy to see when t ∈ [−T, T ]
Z t
u(t) = S(t)u0 − S(t − t0 )∂x (u2 )(t0 )dt0 ,
0
hence u is a solution to KdV equation (5.32) in [−T, T ] and u ∈ XTs,b , where

XTs,b is defined as following:
kukX s,b = inf{ke
ukX s,b : u
e(t) = u(t)t ∈ [−T, T ]}.
T
Next we will show u is the unique solution in XTs,b by using the ideas
in [165]. Assume u1 , u2 ∈ XTs,b are two solutions to KdV equation with
the same initial data u0 , we will prove u1 (t) = u2 (t), t ∈ [−T, T ]. From
symmetry it suffices to prove u1 (t) = u2 (t), t ∈ [0, T ]. For δ > 0 which will
be determined later, i = 1, 2, define uei


ui (t), t ∈ [0, δ],

uei = ui (2δ − t), t ∈ [δ, 2δ], (5.49)


u , otherwise.
0
Thus t → uei (t) is continuous, and ψ(t)e ui (t) ∈ X s,b , u

e1 (t) − u
e2 (t) = 0 if
t ∈ R \ [0, 2δ].
Since u1 , u2 are solutions to KdV equation, then when t ∈ [0, δ]
Z t
u1 (t) − u2 (t) = −ψ1 (t) S(t − t0 )ψδ (t0 )∂x [(e e2 )(u1 + u2 )](t0 )dt0 .
u1 − u
0
For T > 0 define
kukX s,b = inf{ke
ukX s,b : u
e(t) = u(t) if t ∈ [0, T ]}.
T+
Thus we get from Proposition 5.2, Proposition 5.1 and Theorem 5.3 that
0
ku1 − u2 kX s,b 6 δ b −b (kukX s,b + kvkX s,b )ke
u1 − u
e2 kX s,b .
δ+ T T
From the constructing we know

ke
u1 − u
e2 kX s,b 6 2ku1 − u2 kX s,b ,
δ+
5.4. Local well-posedness for KdV in H −3/4 113
then we have
0
ku1 − u2 kX s,b 6 Cδ b −b (kukX s,b + kvkX s,b )ku1 − u2 kX s,b .
δ+ T T δ+
b0 −b
Choose δ such that Cδ (kukX s,b + kvkX s,b ) < 1/2, then we get u1 (t) =
T T
u2 (t) if t ∈ [0, δ]. Repeating this procedure, we obtained the uniqueness in
[0, T ].
If s is large, then we use the scaling invariance. It is easy to see that
KdV equation has the following scaling invariance: for any λ > 0
u(x, t) → λ2 u(λx, λ3 t), u0 (x) → λ2 u0 (λx). (5.50)
Then Ḣ −3/2 is critical in the sense that: kλ2 u0 (λ·)kḢ −3/2 = ku0 kḢ −3/2 .
From the fact
3
kλ2 u0 (λx)kH −3/4+ .λ 2 + ku0 kH −3/4+ + λ3/4+ ku0 kH −3/4+ ,
thus choosing λ sufficiently small, we may assume
kφkH −3/4+ 6 0 1. (5.51)
The rest argument is similar to the case −3/4 < s < −1/2, and we leave it
to the readers. Therefore, we prove
Theorem 5.5. Assume s ∈ (−3/4, 0], u0 ∈ H s , Then there exists T =

T (ku0 kH −3/4+ ) > 0 and b > 1/2 such that KdV equation (5.32) has a
unique solution u(x, t) ∈ XTs,b ⊂ C([−T, T ]; H s ). Moreover, ∀ R > 0, the
map u0 → u(t) is Lipschitz continuous from {φ ∈ H s , kφkH s 6 R} to
C([−T, T ]; H s ).
5.4 Local well-posedness for KdV in H −3/4
In Section 5.3, we consider the local well-posedness of Korteweg-de Vries

equation in H s with s > −3/4. It is known that the bilinear estimate (5.47)
is invalid if s 6 −3/4, one can refer to [132]. Moreover, KdV equation is
not locally well-posed in H s for s < −3/4, the solution operator fails to be
uniformly continuous with respect to the H s norm [47; 134]. Thus there
exists a natural question: whether is the Cauchy problem of KdV equation
locally well-posed in H −3/4 ? In this section, we give the answer, one can
refer to [90].
From the results in Section 5.3, it follows that the space X s,b with
b > 1/2 is unsuitable to be chosen as the work space for the local well-
popsedness in H −3/4 . We try to choose a new space instead of X s,b with
b > 1/2. From the definition of Bourgain’s spaces X s,b , it is expected to

1/2
obtain more elaborate estimates if Htb with b > 1/2 is replaced by B2,1
in spaces X s,b . In Section 5.1, we have introduced this space F s , where
φ(ξ) = ξ 3 . The key approach is prove whether the following bilinear in
spaces F s holds
k∂x (uv)kN s 6 C(kukF s kvkF s + kvkF s kukF s ). (5.52)
By the definitions of F s and N s , in order to obtain the bilinear estimate

(5.52), it suffices to show that the following dyadic bilinear estimate
[
khτ − ξ 3 i−1 ϕk3 (ξ)ξ(4 [ [ [
k1 u ∗ 4k2 v)kXk3 6C(k1 , k2 , k3 )k4k1 ukXk1 k4k2 vkXk2 .
[
By examining the supports of the functions, ϕk3 (ξ)(4 [
k1 u∗ 4k2 v) ≡ 0 unless
| max(k1 , k2 , k3 ) − med(k1 , k2 , k3 )| 6 5.
Thus we have the following several cases depending on the relative sizes
k1 , k2 , k3 : high × low → high, low × high → high, high × high → low,
high × high → high, low × low → low. For each case, we have the corre-
sponding estimtaes.
Using Lemma 5.4, Theorem 5.1 and the proof of Lemma 5.6, we have
the following proposition.
Proposition 5.4. If k ∈ Z+ , j ∈ N, I ⊂ R and |I|.1, then
k4k (u)kL∞ 2 .kF [4k (u)]kX ,

t Lx k
k4k (u)kL2x L∞
t∈I
.23k/4 kF [4k (u)]kXk ,
k4k (u)kL4x L∞
t
.2k/4 kF [4k (u)]kXk ,
−j
k4j (u)kL∞ 2 .2
x Lt
kF [4j (u)]kXj .
Now we prove dyadic bilinear estimates according to frequency interac-

tions.
Proposition 5.5. (a) If k > 10, |k − k2 | 6 5, then

d
khτ − ξ 3 i−1 ϕk (ξ)ξ(4 [ d [
0 u ∗ 4k2 v)kXk .k40 vkX0 k4k2 vkXk2 . (5.53)
(b) If k > 10, |k − k2 | 6 51 6 k1 6 k − 9, then

[
khτ − ξ 3 i−1 ϕk (ξ)ξ(4 [ 3 −k/2−k1 [
k1 u ∗ 4k2 v)kXk .k 2
[
k4k1 ukXk1 k4 k2 vkXk2 .
(5.54)
Proof. For the simplicity of notations, we assume k = k2 . We always

denote
uk,j = ϕk (ξ)ϕj (τ − ξ 3 )b
u.
First we show (a). From the definition of Xk we get
X
d
khτ − ξ 3 i−1 ϕk (ξ)ξ(4 d
0 u ∗ 4k v)kXk .2
k
2−j/2 k1Dk,j (u0,j1 ∗ vk,j2 )k2 .
j,j1 ,j2 >0
Similar to the proof of Lemma 5.7 (b), we get

k1Dk,j (u0,j1 ∗ vk,j2 )k2 .2−k 2(j1 +j2 )/2 ku0,j1 k2 kvk,j2 k2 .
Thus
X
d
0 u ∗ 4k v)kXk .2
k
2−j/2 k1Dk,j (u0,j1 ∗ vk,j2 )k2
j,j1 ,j2 >0
d
.k4 [
0 vkX0 k4k2 vkXk2 .
Then (a) is proved.

For (b), from the definition of Xk we have
X
[
k1 u ∗ 4k v)kXk .2
k
2−j3 /2 k1Dk,j3 (uk1 ,j1 ∗ vk,j2 )k2 .
ji >0
(5.55)
From the support properties we may assume jmax > 2k + k1 − 10 in the
right-hand side of (5.55). We may also assume j1 , j2 , j3 6 10k, otherwise
(b) follows from Lemma 5.7 (a) immediately. It follows from Lemma 5.7
(b) that
X
2k 2−j3 /2 k1Dk,j (uk1 ,j1 ∗ vk,j2 )k2
j3 ,j1 ,j2 >0
X
. 2k 2−j/2 2jmin /2 2−k/2 2−k1 /2 2jmed /2 kuk1 ,j1 k2 kvk,j2 k2
j3 ,j1 ,j2 >0
X
.2 k [
k 3 2−k/2 2−k1 /2 2−jmax /2 k4 d
k1 ukXk1 k4k vkXk
jmax >2k+k1 −10
.k 2 3 −k/2 −k1
2 [
k4 d
k1 ukXk1 k4k vkXk .
Proposition 5.6. If k > 10, |k − k2 | 6 5 and k − 9 6 k1 6 k + 10, then

d
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(4 [
k u ∗ 4k2 v)kXk1 . 2
−3k/4 d [
k4k ukXk k4 k2 vkXk2 .
(5.56)
Proof. We assume k = k2 . From the definition of Xk1 we get
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(4d d

k u ∗ 4k v)kXk1
X
. 2k1 2−j1 /2 k1Dk1 ,j1 (uk,j2 ∗ vk,j3 )k2 . (5.57)
ji >0
As before, we may assume jmax > 3k − 20 and j1 , j2 , j3 6 10k. It follows

from Lemma 5.7 (c) that
X
2k1 2−j1 /2 k1Dk1 ,j1 (uk,j2 ∗ vk,j3 )k2
j1 ,j2 ,j3 >0
X X X −j1 /2 3k/4 jmin /2 j /4
. + + 2 2 2 2 med kuk,j2 k2 kvk,j3 k2
j1 =jmax j2 =jmax j3 =jmax
:= I + II + III.
The term I is easy to control, we omit the details. By symmetry we only

need to control II, and dividing it into two parts we get
X X
II. + 2−j1 /2 23k/4 2jmin /2 2jmed /4 kuk,j2 k2 kvk,j3 k2
j2 =jmax ,j1 6j3 j2 =jmax ,j1 >j3
:=II1 + II2 .
For II1 , by summing on j1 we get

X
II1 . 2−j1 /2 23k/4 2j1 /2 2j3 /4 kuk,j2 k2 kvk,j3 k2
j2 =jmax ,j1 6j3
X
. 23k/4 2j3 /2 kuk,j2 k2 kvk,j3 k2
j2 >3k−20,j3 >0
d
. 2−3k/4 k4 [
k ukXk k4k2 vkXk2 .
For II2 we have

X
II2 . 2−j1 /2 23k/4 2j3 /2 2j1 /4 kuk,j2 k2 kvk,j3 k2
j2 =jmax ,j1 >j3
d
. 2−3k/4 k4 [
k ukXk k4k2 vkXk2 .
Proposition 5.7. If 0 6 k1 , k2 , k3 6 100, then

[
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(4 [ [ [
k2 u ∗ 4k3 v)kXk1 .k4k2 ukXk2 k4k3 vkXk3 . (5.58)
Proof. By the definition we get

X
khτ −ξ i3 −1 [
ϕk1 (ξ)ξ(4 [
k2 u∗4k3 v)kXk1 . 2
k1
2−j1 /2 k1Dk1 ,j1 (uk2 ,j2 ∗vk3 ,j3 )k2 .
ji >0
From the support properties we have 1Dk1 ,j (uk2 ,j1 ∗vk3 ,j2 ) = 0 unless |jmax −
jmed | 6 10 or jmax 6 1000. It follows from Lemma 5.7 (a) that
[
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(4 [ [ [
k2 u ∗ 4k3 v)kXk1 .k4k2 ukXk2 k4k3 vkXk3 .

Actually, for the low-low interactions, we can prove a stronger result.
Proposition 5.8. If 0 6 k1 , k2 , k3 6 100, then

[
khτ − ξ 3 i−1 ϕk1 (ξ)ξ F [ψ(t)4k2 u] ∗ 4 k3 v kXk1 .k4k2 ukL∞ 2 k4k vkL∞ L2 .
t Lx 3 t x
Proof. It follows from the definition of Xk1 , Plancherel’s equality and

Bernstein’s inequality that
[
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(F [ψ(t)4k2 u] ∗ 4 k3 v)kXk1
X
. 2k1 2−j3 /2 kψ(t)4k2 u · 4k3 vkL2t L2x .k4k2 ukL∞ 2 k4k vkL∞ L2 ,
t Lx 3 t x
j3 >0
which implies the proposition.

The last one is the high-high to low interactions.
Proposition 5.9. (a) If k > 10|k − k2 | 6 5, then

d
khτ − ξ 3 i−1 ϕ0 (ξ)ξ(4 [
k u ∗ 4k2 v)kX0 . k2
−3k/2 d [
k4k ukXk k4 k2 vkXk2 .
(b) If k > 10, |k − k2 | 6 5, 1 6 k1 6 k − 9, then

d
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(4 [
k u ∗ 4k2 v)kXk1
d
.(2−3k/2 + k2−2k+k1 /2 )k4 [
k ukXk k4k2 vkXk2 .
Proof. First we show (a), assuming k = k2 . The left-hand side of the

inequality in (a) is bounded by
0
X X
2k3 2−j3 /2 k1Dk3 ,j3 · (uk,j1 ∗ vk,j2 )k2 . (5.59)
k3 =−∞ j1 ,j2 ,j3 >0
We may assume k3 > −10k and j1 , j2 , j3 6 10k. We only consider the

worst case |j3 − 2k − k3 | 6 10. It follows from Lemma 5.7 (b) that
d
khτ − ξ 3 i−1 ϕ0 (ξ)ξ(4 d
k u ∗ 4k v)kX0
0
X X
. 2−k 2−k3 /2 2k3 2−k/2 2−k3 /2 2j1 /2 2j2 /2 kuk,j1 k2 kvk,j2 k2
k3 =−10k j1 ,j2 >0
d
.k2−3k/2 k4 d
k ukXk k4k vkXk . (5.60)
Thus, (a) is proved.

Now we prove (b), assuming k = k2 . From definition we have
X
d
khτ − ξ 3 i−1 ϕk1 (ξ)ξ(4 d
k u ∗ 4k v)kXk1 .2
k1
2−j1 /2 k1Dk1 ,j1 (uk,j2 ∗ vk,j3 )k2 .
ji >0
(5.61)
We may assume jmax > 2k + k1 − 10 and j1 , j2 , j3 6 10k. We will control

the right-hand side of (5.61) case by case. If j1 = jmax , it follows from
Lemma 5.7 (b) that
X
2k1 2−j1 /2 k1Dk1 ,j1 · (uk,j2 ∗ vk,j3 )k2
j1 ,j2 ,j3 >0
X X
. 2k1 2−j1 /2 2−k/2 2−k1 /2 2(j2 +j3 )/2 kuk,j2 k2 kvk,j3 k2
j1 >2k+k1 −10 j2 ,j3 >0
.2 −3k/2 d
k4 [
k ukXk k4k2 vkXk2 .
If j2 = jmax , we have better estimate for the characterization multiplier. It

follows from Lemma 5.7 (b) that
X
2k1 2−j1 /2 k1Dk1 ,j1 · (uk,j2 ∗ vk,j3 )k2
j1 ,j2 ,j3 >0
X X
. 2k1 2−j1 /2 2−k 2(j1 +j3 )/2 kuk,j2 k2 kvk,j3 k2
j2 >2k+k1 −10 j1 ,j3 >0
.k2 −2k k1 /2
2 d
k4 [
k ukXk k4k2 vkXk2 ,
where in the last inequality we use j1 6 10k. By symmetry, the case

j3 = jmax is identical to the case j2 = jmax . Thus the proposition is
proved.
Next we prove the bilinear estimates in F s .
Theorem 5.6. Fix s ∈ (−3/4, 0]. Then ∀ s 6 σ 6 0, there exists C > 0

such that ∀ u, v ∈ F σ
k∂x (uv)kN σ 6 C(kukF s kvkF σ + kvkF s kukF σ ). (5.62)

Proof. From definition we have

X
k∂x (uv)k2N σ = 22σk3 khτ − ξ 3 i−1 ϕk3 (ξ)ξ(b
u ∗ vb)k2Xk .
3
k3 ∈Z+
Applying Littlewood-Paley decomposition to u, v we get
khτ − ξ 3 i−1 ϕk3 (ξ)ξ(b

u ∗ vb)kXk3
X
. 3 −1
khτ − ξ i ϕk3 (ξ)ξ(4 [ [
k1 u ∗ 4k2 v)kXk3 . (5.63)
k1 ,k2 ∈Z+
We may assume |kmax − kmed | 6 5. By symmetry we may also assume

k1 6 k2 . Then the right-hand side of (5.63) is bounded
4
X X
[
khτ − ξ 3 i−1 ϕk3 (ξ)ξ(4 [
k1 u ∗ 4k2 v)kXk3 , (5.64)
j=1 k1 ,k2 ∈Aj
where Aj , j = 1, 2, 3, 4, is defined as following
A1 = {k2 > 10, |k2 − k3 | 6 5, k1 6 k2 − 10};

A2 = {k2 > 10, |k2 − k3 | 6 5, k2 − 9 6 k1 6 k2 + 10};
A3 = {k2 > 10, |k2 − k1 | 6 5, k3 6 k1 − 10};
A4 = {k1 , k2 , k3 6 100}.
Thus, (5.62) follows from Propositions 5.5-5.9, the condition −3/4 < s 6 0
and discrete Young’s inequality.
With this bilinear estimates and the scaling (5.50), we can also obtain
the LWP of KdV equation in H s for s > −3/4. From the proof of Theorem
5.6 we see the condition s > −3/4 was only needed in Proposition 5.9 (a).
Thus to study the endpoint case s = −3/4, one naturally asks whether
Theorem 5.6 holds at s = −3/4, more precisely, whether the bound in
Proposition 5.9 (a) can be improved to 2−3k/2 ? The answer is negative by
the following counter-example which was given by N. Kishimoto [141] using
the one in [180].
Proposition 5.10. If k > 200 and |k − k2 | 6 5, then there exists u, v ∈

F −3/4 such that
d
khτ − ξ 3 i−1 ϕ0 (ξ)ξ(4 [ − 3k d [
k u ∗ 4k2 v)kX0 & log(k)2 2 k4k ukXk k4k2 vkXk2 .
(5.65)
Proof. Let N = 2k , choose m ∈ N sufficiently large such that 2m

N 1/2 . For j = 0, 1, · · · , m, define Rj0 ⊂ R2 to be the parallelogram with
vortex
(τ, ξ) = (0, 0), (1, 0), (3 · 2−j N 3/2 + 1, 2−j N −1/2 ), (3 · 2−j N 3/2 , 2−j N −1/2 ).
Then let Rj := ((N + 2j N −1/2 )3 , N + 2j N −1/2 ) + Rj0 . Choose
m
X
b=b
u v=N 2j/2 aj 1Rj ∪(−Rj ) ,
j=0
where aj > 0. It is easy to see that ∪m 3

j=0 Rj ⊂ {(τ, ξ) : |τ − ξ | 6 10}, thus
we get
m
X m
X
d
N −3/2 k4 [
k ukXk k4k2 vkXk2 ∼ N
2
[N −3/4 2j/2 aj |Rj |1/2 ]2 ∼ a2j .
j=0 j=0
On the other hand, if 1 6 j 6 m

1Rj ∗ 1−R0 &|Rj |1(τ (j) ,ξ(j) )−1/2R00 ∼ 2−j N −1/2 1(τ (j) ,ξ(j) )−1/2R00 , (5.66)
where
(τ (j) , ξ (j) ) = ((N + 2j N −1/2 )3 , N + 2j N −1/2 ) − ((N + N −1/2 )3 , N + N −1/2 ).
d
Thus khτ − ξ 3 i−1 ϕ0 (ξ)ξ 4 [
k u ∗ 4k2 vkX0 is larger than
∞
X m
X
−j 0 /2 3
2 kϕ (τ − ξ )ϕ0 (ξ)ξ(N
j0 2j/2 aj 1Rj ) ∗ (N a0 1−R0 )kL2 . (5.67)
j 0 =0 j=1
From (5.66) we get (5.67) is larger than

∞
X m
X
0
N 3/2 a0 2−j /2 kϕj 0 (τ − ξ 3 )ϕ0 (ξ)ξ 2−j/2 aj 1(τ (j) ,ξ(j) )−1/2R00 kL2 .
j 0 =0 j=1
(5.68)
It is easy to see that if (τ, ξ) ∈ (τ (j) , ξ (j) ) − 1/2R0 (j > 1), then |ξ| ∼
2j N −1/2 |τ | ∼ 2j N 3/2 , hence (5.68) is equivalent to
X 0
N 3/2 a0 2−j /2 kϕj 0 (τ − ξ 3 )ϕ0 (ξ)ξ2−j/2 aj 1(τ (j) ,ξ(j) )−1/2R00 kL2
j 0 >0,2j N 3/2 ∼2j0
Xm m
X
∼ N 3/2 a0 (2j N 3/2 )−1/2 2j N −1/2 2−j/2 aj |R00 |1/2 ∼ a0 aj .
j=1 j=1
Taking aj = 1/(j + 1), we complete the proof of the proposition.

The proposition above shows that if s = −3/4, one can not obtain local
well-posedness by using F s or X s,b . From the discussion we see the only
problem is the low frequency structure P60 u. Fortunately, by using a new
low frequency structure, we can overcome the logarithmic divergence in
Proposition 5.9 (a). We will use the following structure:
kukX̄0 = kukL2xL∞
t
.
From Proposition 5.4 we have
kϕ0 (t)40 ukX̄0 .k40 ukX0 . (5.69)
Thus it is weaker than X0 , but on the other hand, we have for any 1 6 q 6
∞ and 2 6 r 6 ∞
k40 ukLq|t|6T Lrx ∩Lrx Lq|t|6T .T k40 ukL2x L∞
|t|6T
. (5.70)
For −3/4 6 s 6 0, define
X
F̄ s = {u ∈ S 0 (R2 ) : kuk2F̄ s = 22sk kϕk (ξ)F uk2Xk + k40 uk2X̄0 < ∞}.
k>1
Assume T > 0, define the local space F̄ s (T ):

kukF̄ s (T ) = inf {k40 ukL2xL∞
|t|6T
+k(I−40 )wkF̄ s , w(t) = u(t), t ∈ [−T, T ]}.
w∈F̄ s
Now we will show local well-posedness for KdV equation at s = −3/4.
Theorem 5.7. Assume s > −3/4 and φ ∈ H s . Then

(a) Existence. There exists T = T (ku0kH −3/4 ) > 0 and a solution u to
the Cauchy problem (5.32) such that
u ∈ F̄ s (T ) ⊂ C([−T, T ] : H s ).
(b) Uniqueness. The solution map ST : u0 → u is the unique continuous
extension of the smooth solution map H ∞ → C([−T, T ] : H ∞ ).
(c) Lipschitz continuity. For any R > 0, the map u0 → u is Lipschitz
continuous from {u0 ∈ H s : ku0 kH s < R} to C([−T, T ] : H s ).
(d) Persistence. If u0 ∈ H s for σ > s, then u ∈ H σ .
Now we indicate our ideas in constructing F¯s . The starting point is the
bilinear estimates in F s :
k∂x (uv)kN s 6 CkukF s kvkF s .
This estimate fails at s = −3/4. On the other hand, we expect contraction
principle still work at s = −3/4. Thus we need a new space F̄ −3/4 , which
of course satisfies the following
F̄ −3/4 ⊂ C(R : H −3/4 ).
From the equivalent integral equation of (5.32):

Z t
u = S(t)u0 + C S(t − s)∂x (u2 )(s)ds.
0
Localizing in time, we get
Z t
u = Tu0 (u) = ψ(t)S(t)u0 + Cψ(t) S(t − s)∂x (u2 )(s)ds, (5.71)
0
where ψ(t) = ϕ0 (t). By an iteration,

u(0) = 0; · · · ; u(n+1) = Tu0 (u(n) ); · · ·
we obtain a sequence {u(n) }. If contraction principle works in F̄ −3/4 and u0
satisfies ku0 kH −3/4 1, then {u(n) } is a Cauchy sequence in C(R : H −3/4 ).
Then a basic requirement is
u(n) ∈ C(R : H −3/4 ), ∀ n ∈ N. (5.72)
Next we check (5.72) for {u(n) }. For n = 0, 1, it is obvious that
u , u(1) ∈ C(R; H −3/4 ). The case n = 2 is nontrivial. From the defi-
(0)
nition we have
Z t
u(2) = ψ(t)S(t)u0 + Cψ(t) S(t − s)∂x (S(s)u0 · S(s)u0 )ds.
0
It suffices to prove (I − 40 )(u ) ∈ C(R; H −3/4 )40 (u(2) ) ∈ C(R; L2 ). For

(2)
the high frequency part

Z t
ψ(t) S(t − s)∂x (S(s)u0 · S(s)u0 )ds
0
Z t
= ψ(t) S(t − s)∂x [ψ(s/2)S(s)u0 · ψ(s/2)S(s)u0 ]ds.
0
From the linear estimate
kψ(t/2)S(t)u0 kF s .ku0 kH s ,
then by dyadic bilinear estimates and Proposition 5.3, we get for k ∈ N,
4k (u(2) ) ∈ Xk ⊂ C(R; H −3/4 ).
For the low frequency part, we can not obtain 40 (u(2) ) ∈ X0 , due to
the logarithmic divergence. By calculation we get
Z t

Fx ψ(t) S(t − s)40 ∂x [4k1 u(s)4k2 v(s)]ds (ξ)
0
Z t Z
3 3
\ isξ 3 \
= ψ(t)η0 (ξ)iξ ei(t−s)ξ eisξ1 4 k1 u0 (ξ1 )e 2 4k2 v0 (ξ2 )ds
0 ξ=ξ1 +ξ2
Z 3 3 3
3 1 − e−it(ξ −ξ1 −ξ2 ) \ \
= ψ(t)η0 (ξ)eitξ ξ 4k1 u0 (ξ1 )4 k2 v0 (ξ2 )
ξ=ξ1 +ξ2 ξ 3 − ξ13 − ξ23
:= Fx (I) + Fx (II).
Since in the hyperplane ξ = ξ1 + ξ2 one has ξ 3 − ξ13 − ξ23 = 3ξξ1 ξ2 , then we
get
Z \ \
itξ 3 4 k1 u0 (ξ1 )4k2 v0 (ξ2 )
Fx (I) = ψ(t)ϕ0 (ξ)e .
ξ=ξ1 +ξ2 3ξ1 ξ2
Applying Theorem 5.2 we get
Z
\
4 u (ξ )4\ v (ξ )
k1 0 1 k2 0 2
kIkL2x L∞ 6 C 6 C2−3k1 /2 ku0 kL2 kv0 kL2 .
t
ξ=ξ1 +ξ2 3ξ1 ξ2
L2ξ
For the term II we have

Z 3 3
−eit(ξ1 +ξ2 ) \ \
Fx (II) = ψ(t)ϕ0 (ξ) 4k1 u0 (ξ1 )4 k2 v0 (ξ2 ).
ξ=ξ1 +ξ2 3ξ1 ξ2
Applying Theorem 5.2 we get
3 3
kIIkL2xL∞
t
6Cket∂x ∂x−1 4k1 u0 · et∂x ∂x−1 4k2 v0 kL2x L∞
t
3 3
6Cket∂x ∂x−1 4k1 u0 kL4x L∞
t
ket∂x ∂x−1 4k2 v0 kL4x L∞
t
6C2−3k1 /2 ku0 kL2 kv0 kL2 .

Thus, we proved 40 (u(2) ) ∈ L2x L∞
t .
Next we consider the case n > 3. If n > 3, do we have (I − 40 )(u(n) ) ∈
F s and 40 (u(n) ) ∈ L2x L∞
t ? To answer this, we have the following proposi-
tion:
Proposition 5.11 (X̄0 estimate). Assume |k1 − k2 | 6 5 and k1 > 10.

Then u, v ∈ F̄ 0
Z t

ψ(t) S(t − s)4 ∂ [4 u(s)4 v(s)]ds
0 x k1 k2
0 L2x L∞
t
3k
.2 − 21 [
k4 [
k1 ukXk1 k4k2 ukXk2 .
Rt
Proof. Denote Q(u, v) = ψ(t) 0 S(t − s)40 ∂x [4k1 u(s)4k2 v(s)]ds. Di-
rect computations show that
Z b b − ξ3 )
ψ(τ − τ 0 ) − ψ(τ
F [Q(u, v)] (ξ, τ ) =c ϕ0 (ξ)iξ
R τ 0 − ξ3
Z
× dτ 0 [
4 [
k1 u(ξ1 , τ1 )4k2 v(ξ2 , τ2 ).
ξ=ξ1 +ξ2 ,τ 0 =τ1 +τ2
Fix ξ ∈ R, divide Γ := {ξ = ξ1 + ξ2 , τ 0 = τ1 + τ2 } as following

Γ1 ={|ξ|.2−2k1 } ∩ Γ;
Γ2 ={|ξ| 2−2k1 , |τi − ξi3 | 3 · 22k1 |ξ|, i = 1, 2} ∩ Γ;
Γ3 ={|ξ| 2−2k1 , |τ1 − ξ13 |&3 · 22k1 |ξ|} ∩ Γ;
Γ4 ={|ξ| 2−2k1 , |τ2 − ξ23 |&3 · 22k1 |ξ|} ∩ Γ.
Then we get
Z t
F ψ(t) · S(t − s)40 ∂x [4k1 u(s)4k2 v(s)]ds (ξ, τ ) = A1 + ... + A4 ,
0
where
Z b − τ 0 ) − ψ(τ
b − ξ3 ) Z
ψ(τ [ [ 0
Ai =C 40 (ξ)iξ 4 k1 u(ξ1 , τ1 )4k2 v(ξ2 , τ2 )dτ .
R τ 0 − ξ3 Γi
First we consider the estimate of A1 . It follows from Proposition 5.4
and Proposition 5.3 (b) that
Z
0
kF −1 (A1 )kL2x L∞ . hτ − ξ 3 i−1 ϕ0 (ξ)ξ [
4 u(ξ , τ )4[ v(ξ , τ ) .
t k1 1 1 k2 2 2
A1 X0
−2k1
Since in Γ1 one has |ξ|.2 , then we get
Z
0
hτ − ξ 3 i−1 ϕ0 (ξ)iξ [
4 u(ξ , τ )P[ v(ξ , τ )
k1 1 1 k2 2 2
Γ1 X0
X X X
−j3 /2 k3
. 2 2 k1Dk3 ,j3 · (uk1 ,j1 ∗ vk2 ,j2 )kL2 .
k3 6−2k1 +10 j3 >0 j1 >0,j2 >0
Using (5.37) we get

X X
kF −1 (A1 )kL2x L∞
t
. 2−j3 /2 2k3 2jmin /2 2k3 /2 kuk1 ,j1 kL2 kvk2 ,j2 kL2
k3 6−2k1 +10 ji >0
[
.2−3k1 k4 [
k1 ukXk1 k4k2 ukXk2 ,
which suffices for the bound of A1 .

Next we consider A3 . As for A1 , it follows from Proposition 5.4 and
Proposition 5.3 (b) that
Z
0
−1
kF (A3 )kL2x L∞ 3 −1
. hτ − ξ i ϕ0 (ξ)ξ [ [
4k1 u(ξ1 , τ1 )4k2 v(ξ2 , τ2 )
t
A3 ∪A4 X0
X X X
−j3 /2 k3
. 2 2 k1Dk3 ,j3 · (uk1 ,j1 ∗ vk2 ,j2 )kL .
2
k3 60 j3 >0 j1 ,j2 >0

We may assume j3 6 10k1 and |τ1 − ξ13 |&3|ξξ1 ξ2 |. It follows from Lemma
5.7 (b) that
X X
kF −1 (A3 )kL2x L∞
t
. 2k3 2j2 /2 2−k1 kuk1 ,j1 kL2 kvk2 ,j2 kL2
k3 60 j1 >k3 +2k1 −10,j2 ,j3 >0
[
.k1 2−2k1 k4 [
which gives the bound for A3 . Similarly, we can control A4 .

Now we consider A2 , which is the main term. Simple calculations show
that
Z t Z
3
Ft−1 (A2 ) =ψ(t) ei(t−s)ξ ϕ0 (ξ)iξ eis(τ1 +τ2 )
0 R2
Z
× uk1 (ξ1 , τ1 )vk2 (ξ2 , τ2 ) dτ1 dτ2 ds
ξ=ξ1 +ξ2
where
b(ξ1 , τ1 ),
uk1 (ξ1 , τ1 ) =ϕk1 (ξ1 )1{|τ1 −ξ13 |3·22k1 |ξ|} u
vk2 (ξ2 , τ2 ) =ϕk2 (ξ2 )1{|τ2 −ξ23 |3·22k1 |ξ|} vb(ξ2 , τ2 ).
By change of variable τ10 = τ1 − ξ13 , τ20 = τ2 − ξ23 , we get
Z t Z
−1 itξ 3 −isξ 3
Ft (A2 ) =ψ(t)e ϕ0 (ξ)iξ e eis(τ1 +τ2 )
0 R2
Z
3 3
× eisξ1 uk1 (ξ1 , τ1 + ξ13 )eisξ2 vk2 (ξ2 , τ2 + ξ23 ) dτ1 dτ2 ds.
ξ=ξ1 +ξ2
By interchange of the integral we get it is equal to

Z Z 3 3 3
itξ 3 it(τ1 +τ2 ) eit(ξ1 +ξ2 −ξ ) − e−it(τ1 +τ2 )
ψ(t)e ϕ0 (ξ)ξ e
R2 ξ=ξ1 +ξ2 τ1 + τ2 − ξ 3 + ξ13 + ξ23
× uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 ) dτ1 dτ2
:= Ft−1 (II1 ) − Ft−1 (II2 ).
For the term II2 ,
Ft−1 (II2 )
Z Z
itξ 3 uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 )
= ψ(t)e ϕ0 (ξ)ξ dτ1 dτ2 .
R2 ξ=ξ1 +ξ2 τ1 + τ2 − ξ 3 + ξ13 + ξ23
Since in the integral area |τ1 +τ2 −ξ 3 +ξ13 +ξ23 | ∼ |ξξ1 ξ2 |, then from Theorem
5.2 we get
kF −1 (II2 )kL2x L∞
t
Z Z
uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 )
. ξ dτ1 dτ2
2
τ1 + τ2 − ξ 3 + ξ13 + ξ23
R ξ=ξ1 +ξ2 L2 ξ
3k1
.2 − 2 [
k4 [
k1 ukXk1 k4k2 ukXk2 .
To prove 5.11, it suffices to prove

kF −1 (II1 )kL2x L∞ [
.2−3k1 /2 kP [
k1 ukXk1 kPk2 ukXk2 .
t
Compare II1 with II10 defined as following:

Z Z 3 3 3
−1 0 itξ 3 it(τ1 +τ2 ) eit(ξ1 +ξ2 −ξ )
Ft (II1 ) =ψ(t)e ϕ0 (ξ)ξ e
R2 ξ=ξ1 +ξ2 −ξ 3 + ξ13 + ξ23
× uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 ) dτ1 dτ2 .
For II10 we have
Z
Ft−1 (II10 ) = ψ(t)ϕ0 (ξ)eit(τ1 +τ2 ) 1{|ξ||τ1 |2−2k1 } 1{|ξ||τ2 |2−2k1 }
R2
Z 3 3
eit(ξ1 +ξ2 )
× F (fτ1 )(ξ1 )F (gτ2 )(ξ2 ) dτ1 dτ2 ,
ξ=ξ1 +ξ2 −3ξ1 ξ2
where for τ1 , τ2 ∈ R we denote
[
F (fτ1 )(ξ) = 4 3 [ 3
k1 u(ξ, τ1 + ξ ), F (gτ2 )(ξ) = 4k2 v(ξ, τ2 + ξ ).
Since (We may need a smooth cut-off 1{|ξ|λ} ) ∀ λ > 0

kFx−1 1{|ξ|λ} Fx ukL2x L∞
t
.kukL2xL∞
t
,
From Theorem 5.2 we get
Z
kF −1 (II10 )kL2x L∞
t
. kS(t)∂x−1 fτ1 S(t)∂x−1 fτ2 kL2x L∞
t
dτ1 dτ2
R 2
Z
. kS(t)∂x−1 fτ1 )kL4x L∞
t
kS(t)∂x−1 fτ2 kL4x L∞
t
dτ1 dτ2
R2
3k1
.2 − 2 [
k4 [
which gives the bound of II10 .

To finish the proof of Proposition 5.11, it remains to prove
kF −1 (II1 − II10 )kL2x L∞ [
.2−3k1 /2 kP [
t
Since in the integral area |τi | 22k1 |ξ|, i = 1, 2, thus in the hyperplane
X∞ n
1 1 1 τ1 + τ2
− = .
τ1 + τ2 − ξ 3 + ξ13 + ξ23 −ξ 3 + ξ13 + ξ23 n=1
3ξξ1 ξ2 3ξξ1 ξ2
then we get
Ft−1 (II1 − II10 )
Z Z ∞
X n
1 τ1 + τ2
= ψ(t)ϕ0 (ξ)ξ eit(τ1 +τ2 )
R2 ξ=ξ1 +ξ2 n=1 3ξξ1 ξ2 3ξξ1 ξ2
it(ξ13 +ξ23 )
× e uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 ) dτ1 dτ2 .
Decomposing dyadically on the low frequency, {χk (ξ)}∞
k=−∞ denote homo-
geneous decomposition, we get
Ft−1 (II1 − II10 )
∞ Z
X X 3 3
= eit(τ1 +τ2 ) ψ(t)eit(ξ1 +ξ2 ) χk3 (ξ)
n=1 R2
2k3 2−2k1 max(|τ1 |,|τ2 |)
Z n
τ1 + τ2 uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 )
× dτ1 dτ2 .
ξ=ξ1 +ξ2 3ξξ1 ξ2 3ξ1 ξ2
We rewrite it into the following form
Ft−1 (II1 − II10 )
∞ Z
X X 3 3
= eit(τ1 +τ2 ) ψ(t)eit(ξ1 +ξ2 ) χk3 (ξ)(ξ/2k3 )−n
n=1 R2
2k3 2−2k1 max(|τ1 |,|τ2 |)
Z n
τ1 + τ2 uk1 (ξ1 , τ1 + ξ13 )vk2 (ξ2 , τ2 + ξ23 )
× 2−nk3 dτ1 dτ2 .
ξ=ξ1 +ξ2 3ξ1 ξ2 3ξ1 ξ2
Since χk3 (ξ)(ξ/2k3 )−n is the multiplier for the L2x L∞ 0
t , then as for II1 we
get
kF −1 (II1 − II10 )kL2x L∞
t
∞ Z
X X
. C n |τ1 + τ2 |n 2−nk3
n=1 R2
2k3 2−2k1 max(|τ1 |,|τ2 |)
Z
it(ξ13 +ξ23 ) F (fτ1 )(ξ1 )F (gτ2 )(ξ2 )

× ψ(t)e dτ1 dτ2 .
3ξ1n+1 ξ2n+1
ξ=ξ1 +ξ2 L2 L∞ x t
By Theorem 5.2 and summing on k3 we get for some M 1

kF −1 (II1 − II10 )kL2x L∞
t
∞ Z
X X
. C n |τ1 + τ2 |n 2−nk3 2−2nk1
n=1 R2
2k3 2−2k1 max(|τ1 |,|τ2 |)
−3k1 /2
×2 kF (fτ1 )kL2 kF (gτ2 )kL2 dτ1 dτ2
∞
X Z
. (C/M )n 2−3k1 /2 kF (fτ1 )kL2 kF (gτ2 )kL2 dτ1 dτ2
n=1 R2
[
.2−3k1 /2 kP [
Therefore, we complete the proof of the proposition.
Inspired by the proposition, then we may use the structure L2x L∞ t for
the low frequency part. It remains to control the high-low interaction.
Proposition 5.12. If k > 10, |k − k2 | 6 5, then for any u, v ∈ F̄ 0

d
khτ − ξ 3 i−1 ϕk (ξ)ξ(4 [
0 u ∗ 4k2 v)kXk .k40 ukL2x L∞
[
k4 k2 vkXk2 . (5.73)
t
Proof. We may assume k = k2 . From the definition of Xk we get
khτ − ξ 3 i−1 ϕk (ξ)ξ(4d d

0 u ∗ 4k v)kXk
X
.2 k
2 −j/2 d [
k 40 u ∗ 4 k2 vkL2ξ,τ . (5.74)
j>0
It follows from Plancherel’s equality and Proposition 5.4 that

d
2k k4 [ k
0 u ∗ 4k2 vkL2ξ,τ .2 k40 ukL2x L∞ k4k ukL∞ d
2 .k40 ukL2 L∞ k4 k vkXk ,
t x Lt x t
thus, the proposition is proved.
Next we prove a crucial bilinear estimate which is key to prove Theorem

5.7. For u, v ∈ F̄ s we define bilinear operator
Z t

B(u, v) = ψ(t/4) S(t − τ )∂x ψ 2 (τ )u(τ ) · v(τ ) dτ. (5.75)
0
Consider the following integral equation

Z t

u = ψ(t)S(t)u0 + ψ(t/4) S(t − τ )∂x ψ 2 (τ )u(τ ) · u(τ ) dτ. (5.76)
0
In order to apply the contraction principle, the rest task is to show the
boundedness B : F̄ s × F̄ s → F̄ s .
Proposition 5.13. Assume −3/4 6 s 6 0. Then there exists C > 0 such

that
kB(u, v)kF̄ s 6 C(kukF̄ s kvkF̄ −3/4 + kukF̄ −3/4 kvkF̄ s ) (5.77)
hold for all u, v ∈ F̄ s .

Proof. From the definition of F s we have

X
kB(u, v)k2F̄ s = k40 B(u, v)k2X̄0 + 22k1 s kϕk1 (ξ)F [B(u, v)]k2Xk1 . (5.78)
k1 >1
First we consider the second term on the right-hand side of (5.78). By

applying Littlewood-Paley decomposition to u, v we get
X
kϕk1 (ξ)F [B(u, v)]kXk1 . kϕk1 (ξ)F [B(4k2 (u), 4k3 (v))]kXk1 .
k2 ,k3 >0
(5.79)
It follows from Proposition 5.3 (b) that the right-hand side of (5.79) is
bounded by
X
\
khτ − ξ 3 i−1 ϕk1 (ξ)ξ ψ(t)4 \
k2 u ∗ ψ(t)4k3 v)kXk1 . (5.80)
k2 ,k3 >0
From symmetry we may assume k2 6 k3 in (5.80). It suffices to prove

X
X 2 1/2
22k1 s \
khτ − ξ 3 i−1 ϕk1 (ξ)ξ ψ(t)4 \
k2 u ∗ ψ(t)4k3 v)kXk1
k1 >1 k2 ,k3 >0
.kukF̄ −3/4 kvkF̄ s . (5.81)

If kmax 6 20, then from Proposition 5.8 we get (5.80) is bounded by
X
k4k2 ukL∞ 2 k4k vkL∞ L2 ,
t Lx 3 t x
kmax 620
which is sufficient to give (5.81), since in this case k4k ukL∞ 2 .k4k ukX
t Lx k
for k > 1, and k4k ukL∞ 2

t Lx
.k4 k uk X̄k for k = 0. Now we assume kmax > 20
in (5.81). There are three cases. If |k1 − k3 | 6 5, k2 6 k1 − 10, then we use
Proposition 5.5 (a) for k2 = 0, and (b) for k2 > 1; if |k1 − k3 | 6 5, k1 − 9 6
k2 6 k3 , then we use Proposition 5.6; if |k2 − k3 | 6 5, 1 6 k1 6 k2 − 5, then
we use Proposition 5.9 (b). Therefore, we get (5.81).
It remains to prove
kB(u, v)kX̄0 6 C(kukF̄ s kvkF̄ −3/4 + kukF̄ −3/4 kvkF̄ s ). (5.82)
By applying Littlewood-Paley decomposition to u, v we get
X
kB(u, v)kX̄0 6 kB(4k2 u, 4k3 v)kX̄0 .
k2 ,k3 >0
If max(k2 , k3 ) 6 10, then from Proposition 5.14 and Proposition 5.8 we get
kB(4k2 u, 4k3 v)kX̄0 .k4k2 ukL∞ 2 k4k vkL∞ L2 ,
t Lx 3 t x
which is sufficient to give (5.82). If max(k2 , k3 ) > 10, then |k2 − k3 | 6 5. It

follows from Proposition 5.11 that
X
kB(u, v)kX̄0 6 2−3k2 /2 kF (4k2 u)kXk2 kF (4k3 v)kXk3
|k2 −k3 |65, k2 ,k3 >10
.kukF̄ −3/4 kvkF̄ −3/4

which gives (5.82). Therefore, the proposition is proved.
Similarly, as the proof of 5.3 (a), it is easy to prove the following propo-
sition by Lemma 5.4.
Proposition 5.14. Assume s ∈ R and φ ∈ H s . Then there exists C > 0

such that
kψ(t)S(t)φkF̄ s 6 CkφkH s .
To prove Theorem 5.7, we need the bilinear form contraction principle

which is easy to prove. The proof is left to the readers as an exercise.
Lemma 5.8. Assume (X, k · k) is a Banach space with the norm k · k. Let
B : X × X→X be a bilinear operator satisfying
kB(x1 , x2 )k 6 ηkx1 kkx2 k, ∀ x1 , x2 ∈ X.
Then for all y ∈ X with 4ηkyk < 1, the equation x = y + B(x, x) has unique
1
solution x ∈ X such that kxk < 2η .
Using Lemma 5.8, Proposition 5.13 and Proposition 5.14, we can prove
Theorem 5.7.
5.5 I-method
In the last two sections, we studied the local well-posedness for the KdV
equation. In this section, we study the global well-posedness, namely ex-
tend the local solution to a global one. KdV equation (5.32) is completely
integrable, and hence has infinite conservation laws. With these conser-
vation laws, one can derive some a priori estimates of the solution. For
example, if u is a smooth solution to the KdV equation (5.32), then for
k ∈ N ∪ {0} and T > 0
ku(t)kH k 6 C(T, ku0 kH k ), ∀ t ∈ [−T, T ]. (5.83)
With these a priori estimates and local well-posedness, one can easily obtain
the following: KdV equation is globally well-posed in H k , where k ∈ N ∪
5.5. I-method 131
{0}. In view of the local well-posedness results, it is a natural question:

what about global well-posedness in H s for s ∈ [−3/4, 0)? For the regularity
below L2 , there is no conservation law, and hence the global well-posedness
do not hold automatically. I-method is an effective method to study
this kind of problem. It was introduced by J. Colliander, M. Keel, G.
Staffilani, H. Takaoka, and T. Tao [42], inspired by Bourgain’s frequency
decomposition techniques [17].
We explain briefly the basic ideas of I-method. We recall the proof of
(5.83) in the case k = 0, 1. Multiplying u on both sides of equation (5.32),
and then integrating on x, we get
d
ku(t)k2L2 = 0,
dt
which is (5.83) at k = 0. For k = 1, it is easy to see that
d
ku(t)k2H 1 = 0
dt
fails, the reason is the quantity ku(t)kH 1 does not exploit the structure of
the equation. A natural idea is to find some other quantity Q1 (u(t)) such
that
d
Q1 (u(t)) ∼ ku(t)k2H 1 , [Q1 (u(t))] = 0.
dt
If such Q1 (u(t)) exists, then we get ∀ t ∈ R
ku(t)kH 1 ∼ kQ1 (u(t))k = kQ1 (u0 )k.ku0 kH 1 ,
which is (5.83) at k = 1. Now we know such quantity exists, for instance,
we can take
Z
1
Q1 (u(t)) = u2x − u3 + Cu2 dx.
R 3
d
Actually, obviously dt [Q(u(t))] = 0, it suffices to show Q1 (u(t)) ∼ ku(t)k2H 1 .
By Gagliardo-Nirenberg’s inequality we have
5/2 1/2
kuk3L3 .kukL2 kux kL2 .
Then using
ap bq
ab 6 + , 1 < p 6 q < ∞, 1/p + 1/q = 1,
p q
and taking q = 4, we get
10/3 4/3
3kukL2 kux k2L2 3ku0 kL2 kuk2L2 kux k2L2
kuk3L3 6 + 6 + .
4 4 4 4
4/3
Finally take C = ku0 kL2 .
From this, for general H s norm, we also need to find quantity Qs (u(t)),
they have similar properties as Q1 (u(t)). However, since such quantity does
not exist for s < 0, then we relax the condition to Qs (u(t)) increase much
slower than the norm kukH s during the evolution. The issues that how
to construct Qs (u(t)) and control the increasing rate are systematically
studied in ’I-method’.
Given m : Rk → C, m is said to be symmetric if
m(ξ1 , · · · , ξk ) = m(σ(ξ1 , · · · , ξk ))
for all σ ∈ Sk , where Sk is the permutation group for k elements. The

symmetrization of m is defined by
1 X
[m]sym (ξ1 , ξ2 , · · · , ξk ) = m(σ(ξ1 , ξ2 , · · · , ξk )).
k!
σ∈Sk
For each m, we define a k-linear functional acting on k functions

u1 , · · · , uk (m is said to be k-multiplier) as following
Z
Λk (m; u1 , · · · , uk ) = m(ξ1 , · · · , ξk )c
u1 (ξ1 ) · · · u
ck (ξk ).
ξ1 +···+ξk =0
Usually we apply Λk on k functions which are all u. For convenience,

Λk (m; u, · · · , u) is simply denoted by Λk (m). From symmetries we see
Λk (m) = Λk ([m]sym ). By using the KdV equation (5.32) we can get the
following proposition.
Proposition 5.15. Assume u satisfies (5.32) and m is a symmetric func-

tion. Then
dΛk (m) ik
= Λk (mhk ) − Λk+1 m(ξ1 , · · · , ξk−1 , ξk + ξk+1 )(ξk + ξk+1 ) ,
dt 2
where
hk = i(ξ13 + ξ23 + · · · + ξk3 ).
Next we introduce the I-operator. Assume m : R → R is a real-valued

even function, then we define Fourier multiplier operator Iu as following
c
Iu(ξ) = m(ξ)b
u. (5.84)
Define the modified energy EI2 (t)

EI2 (t) = kIuk2L2 .
5.5. I-method 133
Since m is real-valued even function and u is real-valued, then by

Plancherel’s equality we get
EI2 (t) = Λ2 (m(ξ1 )m(ξ2 )).
It follows from Proposition 5.15 that
d 2
E (t) = Λ2 (m(ξ1 )m(ξ2 )h2 ) − iΛ3 (m(ξ1 )m(ξ2 + ξ3 )(ξ2 + ξ3 )).
dt I
Since ξ13 +ξ23 = 0 for ξ1 +ξ2 = 0, then the first term vanishes. Symmetrizing
the second term we get
d 2
E (t) = Λ3 (−i[m(ξ1 )m(ξ2 + ξ3 )(ξ2 + ξ3 )]sym ).
dt I
Denote
M3 (ξ1 , ξ2 , ξ3 ) = −i[m(ξ1 )m(ξ2 + ξ3 )(ξ2 + ξ3 )]sym .
Define a new modified energy
EI3 (t) = EI2 (t) + Λ3 (σ3 ),
where symmetric function σ3 will be set momentarily. The role of σ3 is to
make a cancelation. It follows from Proposition 5.15 that
d 3 3
EI (t) = Λ3 (M3 ) + Λ3 (σ3 h3 ) − iΛ4 (σ3 (ξ1 , ξ2 , ξ3 + ξ4 )(ξ3 + ξ4 )).
dt 2
(5.85)
Take σ3 = −M3 /h3 such that the two trilinear terms in (5.85) cancels.
Denote
3
M4 (ξ1 , ξ2 , ξ3 , ξ4 ) = −i [σ3 (ξ1 , ξ2 , ξ3 + ξ4 )(ξ3 + ξ4 )]sym ,
2
then we have
d 3
E (t) = Λ4 (M4 ).
dt I
Similarly we define a new modified energy
EI4 (t) = EI3 (t) + Λ4 (σ4 ),
where
M4
σ4 = − .
h4
It is easy to get
d 4
E (t) = Λ5 (M5 )
dt I
where
M5 (ξ1 , . . . , ξ5 ) = −2i[σ4 (ξ1 , ξ2 , ξ3 , ξ4 + ξ5 )(ξ4 + ξ5 )]sym .
This procedure can be made for a finite times, but we stop here since it
suffices for our purposes.
In order to control the norm kukH s , it seems to convenient to take
m(ξ) = hξis . However, it is difficult to control its increase. Basing on the
L2 conservation, we take m as following: m is a smooth even function which
has the form

1, |ξ| < N,
m(ξ) = (5.86)
N −s |ξ|s , |ξ| > 2N.
It is easy to see that if N → ∞ then Iu → u, and hence the increase rate
of the modified energy EIk (t) tends to 0 as N → ∞. The basic question is
how fast it tends to 0. If m is of the form (5.86), then it is easy to see m2
satisfies

 2 2 0 0
m (ξ) ∼ m (ξ ) |ξ| ∼ |ξ |,

2
(m2 )0 (ξ) = O( m|ξ|(ξ) ), (5.87)


(m2 )00 (ξ) = O( m2 (ξ) ).
|ξ|2
Next we need to estimate the multipliers M3 , M4 , M5 . We will use two

mean value formulas: if |η|, |λ| |ξ|
|a(ξ + η) − a(ξ)|.|η| sup |a0 (ξ 0 )|, (5.88)
|ξ 0 |∼|ξ|
and
|a(ξ + η + λ) − a(ξ + η) − a(ξ + λ) + a(ξ)|.|η||λ| sup |a00 (ξ 0 )|. (5.89)
|ξ 0 |∼|ξ|
First we give the estimate of M3 .
Proposition 5.16. Let m be given by (5.86). In the set {ξ1 + ξ2 + ξ3 =

0, |ξi | ∼ Ni } where Ni is dyadic, we have
|M3 (ξ1 , ξ2 , ξ3 )|. max(m2 (ξ1 ), m2 (ξ2 ), m2 (ξ3 )) min(N1 , N2 , N3 ).
Proof. From symmetry we may assume N1 = N2 > N3 . If N3 &N1 ,

then it follows directly from the definition of m. If N3 N1 , from the
mean-value formula we get
m2 (ξ1 )ξ1 − m2 (ξ1 + ξ3 )(ξ1 + ξ3 ) + m2 (ξ3 )ξ3 6 max(m2 (ξ1 ), m2 (ξ2 ))N3 .
Thus the proposition is proved.
5.5. I-method 135
To estimate M4 , first we extend the multiplier σ3 from the hyperplane

to the entire space.
Proposition 5.17. If m is of the form (5.86), then for any dyadic num-
bers λ 6 µ there exists an extension of σ3 from {ξ1 + ξ2 + ξ3 = 0, |ξ1 | ∼
λ, |ξ2 |, |ξ3 | ∼ µ} to the entire space {(ξ1 , ξ2 , ξ3 ) ∈ R3 , |ξ1 | ∼ λ, |ξ2 |, |ξ3 | ∼
µ} such that
|∂1β1 ∂2β2 ∂3β3 σ3 (ξ1 , ξ2 , ξ3 )| 6 Cm2 (λ)µ−2 λ−β1 µ−β2 −β3 , (5.90)
where the constant C is independent of λ, µ.
Proof. Since on the hyperplane {(ξ1 , ξ2 , ξ3 ) : ξ1 + ξ2 + ξ3 = 0} we have
h3 = i(ξ13 + ξ23 + ξ33 ) = 3iξ1 ξ2 ξ3 ,
then |h| ∼ λµ2 . From
M3 (ξ1 , ξ2 , ξ3 ) = − i[m(ξ1 )m(ξ2 + ξ3 )(ξ2 + ξ3 )]sym

=i(m2 (ξ1 )ξ1 + m2 (ξ2 )ξ2 + m2 (ξ3 )ξ3 ),
we will discuss case by case. If λ ∼ µ, then we extend σ3 as following

i(m2 (ξ1 )ξ1 + m2 (ξ2 )ξ2 + m2 (ξ3 )ξ3 )
σ3 (ξ1 , ξ2 , ξ3 ) = − ,
3iξ1 ξ2 ξ3
which suffices for the purpose. If λ µ, then we extend σ3 as following
i(m2 (ξ1 )ξ1 + m2 (ξ2 )ξ2 − m2 (ξ1 + ξ2 )(ξ1 + ξ2 ))
σ3 (ξ1 , ξ2 , ξ3 ) = − .
3iξ1 ξ2 ξ3
Then (5.90) follows from (5.88) and (5.87).
Next we give the estimate of σ4 . The proof here was due to [96]. Com-
pared to the proof given in [42], the proof here is more flexible to a general
class of dispersive equations. But the ideas are the same.
Proposition 5.18. Let m be of the form (5.86). Then on the area |ξi | ∼
Ni , |ξj + ξk | ∼ Njk , where Ni , Njk are dyadic numbers, we have
|M4 (ξ1 , ξ2 , ξ3 , ξ4 )| m2 (min(Ni , Njk ))
. . (5.91)
|h4 | (N + N1 )(N + N2 )(N + N3 )(N + N4 )
Proof. From symmetry we may assume N1 > N2 > N3 > N4 . Since ξ1 +
ξ2 +ξ3 +ξ4 = 0, then N1 ∼ N2 . We may also assume N1 ∼ N2 &N , otherwise
M4 vanishes, since m2 (ξ) = 1 if |ξ| 6 N , and if max(N12 , N13 , N14 ) N1 ,
then ξ3 ≈ −ξ1 , ξ4 ≈ −ξ1 , which contradicts to ξ1 + ξ2 + ξ3 + ξ4 = 0. The

right-hand side of (5.91) is actually
m2 (min(Ni , Njk ))
. (5.92)
N1 2 (N + N3 )(N + N4 )
Since on the hyperplane ξ1 + ξ2 + ξ3 + ξ4 = 0
h4 = ξ13 + ξ23 + ξ33 + ξ43 = 3(ξ1 + ξ2 )(ξ1 + ξ3 )(ξ1 + ξ4 ),
thus we get
CM4 (ξ1 , ξ2 , ξ3 , ξ4 )
= [σ3 (ξ1 , ξ2 , ξ3 + ξ4 )(ξ3 + ξ4 )]sym
= σ3 (ξ1 , ξ2 , ξ3 + ξ4 )(ξ3 + ξ4 ) + σ3 (ξ1 , ξ3 , ξ2 + ξ4 )(ξ2 + ξ4 )
+ σ3 (ξ1 , ξ4 , ξ2 + ξ3 )(ξ2 + ξ3 ) + σ3 (ξ2 , ξ3 , ξ1 + ξ4 )(ξ1 + ξ4 )
+ σ3 (ξ2 , ξ4 , ξ1 + ξ3 )(ξ1 + ξ3 ) + σ3 (ξ3 , ξ4 , ξ1 + ξ2 )(ξ1 + ξ2 )
= [σ3 (ξ1 , ξ2 , ξ3 + ξ4 ) − σ3 (−ξ3 , −ξ4 , ξ3 + ξ4 )](ξ3 + ξ4 )
+ [σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 )](ξ2 + ξ4 )
+ [σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (−ξ2 , −ξ3 , ξ2 + ξ3 )](ξ2 + ξ3 )
:= I + II + III. (5.93)
We will get (5.91) by studying different cases.
Case 1. |N4 |& N2 .
Case 1a. N12 , N13 , N14 &N1 .
For this case, it follows directly from (5.90) that
|M4 (ξ1 , ξ2 , ξ3 , ξ4 )| m2 (N4 )
. . (5.94)
|h4 | N1 N2 N3 N4
Case 1b. N12 N1 , N13 &N1 , N14 &N1 .
First we consider the contribution of I. From (5.90) we get
|I| m2 (min(N4 , N12 ))
. . (5.95)
|h4 | N1 N2 N3 N4
For the contribution of II, if N12 &N3 , then we use (5.88) and (5.90),
else if N12 N3 , then we use (5.88)(5.90). Thus we get
II m2 (N4 )
. . (5.96)
h4 N 1 N 1 N 1 N 3
From symmetry, the contribution of III is identical to that of II.
Case 1c. N12 N1 , N13 N1 , N14 &N1 .
Since N12 N1 , N13 N1 , then N1 ∼ N2 ∼ N3 ∼ N4 .
5.5. I-method 137
First we consider the contribution of I. Since

I =[σ3 (ξ1 , ξ2 , ξ3 + ξ4 ) − σ3 (−ξ3 , ξ2 , ξ3 + ξ4 )](ξ3 + ξ4 )
+ [σ3 (−ξ3 , ξ2 , ξ3 + ξ4 ) − σ3 (−ξ3 , −ξ4 , ξ3 + ξ4 )](ξ3 + ξ4 )
=I1 + I2 .
Using (5.90) and (5.88) we get
I I1 I2 m2 (N12 )
. + . .
|h4 | |h4 | |h4 | N14
The contribution of II is identical to that of I.
Finally we consider III. Since
III =1/2[σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (−ξ2 , −ξ3 , ξ2 + ξ3 )
− σ3 (−ξ3 , −ξ2 , ξ2 + ξ3 ) + σ3 (ξ4 , ξ1 , ξ2 + ξ3 )](ξ2 + ξ3 ),
then using (5.89) four times we get
III m2 (N1 )
. .
|h4 | N14
Case 1d. N12 N1 , N13 &N1 , N14 N1 .
This case is identical to Case 1c.
Case 2. N4 N/2.
It is obvious that in this case m2 (min(Ni , Njk )) = 1, and N13 ∼ |ξ1 +
ξ3 | = |ξ2 + ξ4 | ∼ N1 .
Case 2a. N1 /4 > N12 &N/2.
Since N4 N/2 and |ξ3 + ξ4 | = |ξ1 + ξ2 |&N/2, then N3 &N/2. From
|h4 | ∼ N12 N12 , we control the six terms in (5.93) respectively and get
|M4 | 1
. . (5.97)
|h4 | N12 N3 N
Case 2b. N12 N/2.
Since N12 = N34 N/2 and N4 N/2, then N3 N/2, and N13 ∼
N14 ∼ N1 .
For the contribution of I, since N3 , N4 , N34 N/2, then from the
definition of m we have σ3 (−ξ3 , −ξ4 , ξ3 + ξ4 ) = 0. Using (5.90) we get
|I| |σ3 (ξ1 , ξ2 , ξ3 + ξ4 )| 1
. . 4. (5.98)
|h4 | N12 N1
Next we consider the contribution of II and III. We can not deal with
the two terms separately, but need to exploit a cancelation between the two
terms. Rewrite II + III as following
II + III =[σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 )](ξ2 + ξ4 )
+ [σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (−ξ2 , −ξ3 , ξ2 + ξ3 )](ξ2 + ξ3 )

=[σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 )]ξ4
+ [σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (−ξ2 , −ξ3 , ξ2 + ξ3 )]ξ3
+ [σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 )
+ σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (−ξ2 , −ξ3 , ξ2 + ξ3 )]ξ2
=J1 + J2 + J3 . (5.99)
For J1 , since
|J1 | |[σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 )]ξ4 |
6 (5.100)
|h4 | |h4 |
if N12 N3 (in this case N3 ∼ N4 ), then use (5.88) twice, otherwise use
(5.88) and (5.90), thus we get
|J1 | 1
. .
|h4 | N14
J2 is identical to J1 . Now consider J3 . First we assume N12 &N3 . From the
symmetry of σ3 we get
J3 =[σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 − ξ3 , ξ3 , ξ2 )
+ σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (−ξ2 − ξ4 , ξ4 , ξ2 )]ξ2 .
From (5.88) and that N12 &N3 we get
|J3 | 1
. .
|h4 | N14
If N12 N3 , then N3 ∼ N4 . Rewrite J3 as following
J3 =[σ3 (−ξ2 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 )
+ σ3 (ξ1 , ξ4 , ξ2 + ξ3 ) − σ3 (ξ1 , −ξ3 , ξ2 + ξ3 )]ξ2
+ [σ3 (ξ1 , ξ3 , ξ2 + ξ4 ) − σ3 (−ξ2 , ξ3 , ξ2 + ξ4 )
+ σ3 (ξ1 , −ξ3 , ξ2 + ξ3 ) − σ3 (−ξ2 , −ξ3 , ξ2 + ξ3 )]ξ2
=J31 + J32 .
It follows from (5.88) that
|J32 | 1
. . (5.101)
|h4 | N14
It suffices to control J31 . Since in this case m2 (ξ3 ) = m2 (ξ4 ) = 1, then we
get
J31 =[σ3 (−ξ2 , ξ3 , ξ2 + ξ4 ) − σ3 (ξ1 , −ξ3 , ξ2 + ξ3 )
5.5. I-method 139
− σ3 (−ξ2 , −ξ4 , ξ2 + ξ4 ) + σ3 (ξ1 , ξ4 , ξ2 + ξ3 )]ξ2

−m2 (ξ2 )ξ2 + ξ3 + m2 (ξ2 + ξ4 )(ξ2 + ξ4 )
= ξ2
−ξ2 ξ3 (ξ2 + ξ4 )
−m2 (ξ2 )ξ2 − ξ4 + m2 (ξ2 + ξ4 )(ξ2 + ξ4 )
− ξ2
ξ2 ξ4 (ξ2 + ξ4 )
m2 (ξ1 )ξ1 + ξ4 + m2 (ξ2 + ξ3 )(ξ2 + ξ3 )
+ ξ2
ξ1 ξ4 (ξ2 + ξ3 )
m2 (ξ1 )ξ1 − ξ3 + m2 (ξ2 + ξ3 )(ξ2 + ξ3 )
− ξ2 .
−ξ1 ξ3 (ξ2 + ξ3 )
Noting that there is a cancelation, then we get
ξ3 + ξ4 −m2 (ξ2 )ξ2 + m2 (ξ2 + ξ4 )(ξ2 + ξ4 )
J31 = − ξ2
ξ3 ξ4 ξ2 (ξ2 + ξ4 )
ξ3 + ξ4 m2 (ξ1 )ξ1 + m2 (ξ2 + ξ3 )(ξ2 + ξ3 )
+ ξ2 . (5.102)
ξ3 ξ4 ξ1 (ξ2 + ξ3 )
Rewrite (5.102) as following
−m2 (ξ2 )ξ2 + m2 (ξ2 + ξ4 )(ξ2 + ξ4 ) + m2 (ξ1 )ξ1 + m2 (ξ2 + ξ3 )(ξ2 + ξ3 )
− ξ2
ξ2 (ξ2 + ξ4 )
ξ3 + ξ4 ξ3 + ξ4 2
× + [m (ξ1 )ξ1 + m2 (ξ2 + ξ3 )(ξ2 + ξ3 )]
ξ3 ξ4 ξ3 ξ4

1 1
× + ξ2 .
ξ1 (ξ2 + ξ3 ) ξ2 (ξ2 + ξ4 )
Thus, for the first term we use (5.89), and (5.88) for the second term, finally
we get
|J31 | 1
. ,
|h4 | N14
The estimate of M5 follows immediately from the estimate of σ4 .
Proposition 5.19. If m is of the form (5.86), then

m2 (N∗45 )N45
|M5 (ξ1 , . . . , ξ5 )|. ,
(N + N1 )(N + N2 )(N + N3 )(N + N45 ) sym
where
N∗45 = min(N1 , N2 , N3 , N45 , N12 , N13 , N23 ).

Next we will prove two properties. One is to show the modified energy
EI4 (t) increase slowly, the other is to prove it is close to EI2 (t). To control
the increase rate of EI4 (t), it suffices to control its derivative
d 4
E (t) = Λ5 (M5 ).
dt I
Proposition 5.20. Assume I ⊂ R and |I|.1. Let 0 6 k1 6 . . . 6 k5 and
k4 > 10. Then
Z Z
Y5 5
Y
5(k1 +k2 +k3 )
P ki (wi )(x, t)dxdt .2 12 2 −k4 −k5
k4\
kj (wj )kXkj ,
I
i=1 j=1
(5.103)
where on the right-hand side if kj = 0 then replace Xkj by X̄0 .
Proof. It follows from Hölder inequality that the left-hand side of (5.103)
is bounded by
3
Y
k4ki (wi )kL3x L∞
t∈I
· k4k4 (w4 )kL∞ 2 · k4k (w5 )kL∞ L2 .
x Lt 5 x t
i=1
For the term k4k4 (w4 )kL∞x Lt

2 and k4k (w5 )kL∞ L2 , we use Proposition 5.4.
5 x t
For the term k4ki (wi )kL3x L∞ t∈I
, we use interpolation between
k4ki (wi )kL2x L∞
t∈I
and k4ki (wi )kL4x L∞t∈I
, then use Proposition 5.4.
Proposition 5.21. Assume δ.1. Let m be given by (5.86) with s = −3/4.

Then
Z
δ 5
Y

Λ5 (M5 ; u1 , · · · , u5 )dt .N −15/4 kI(uj )kF̄ 0 (δ) . (5.104)
0
j=1
Proof. It follows from Proposition 5.19 that the left-hand side of (5.104)
is bounded by
X Z δ N45 m2 (N∗45 )
Λ5 Q ;
ki >0 0 (N + N1 )(N + N2 )(N + N3 )(N + N45 ) 5i=1 m(Ni )
5
Y
15
Pk1 u1 , · · · , Pk5 u5 dt.N − 4 kuj kF̄ 0 (δ) ,
i=1
ki N∗45
where Ni = 2 . Cancel (N +N45 ) 6 1 and consider only the worst case
m2 (N∗45 ) = 1. it suffices to prove
P R Q
δ 3 1 1 1
k1 ,··· ,k5 >0 0 Λ5 i=1 (N +Ni )m(Ni ) m(N4 ) m(N5 ) ; Pk1 u1 , · · · , Pk5 u5 dt
5.5. I-method 141
15 Q5
.N − 4 i=1 kuj kF̄ 0 (δ) .
From symmetry we may assume N1 > N2 > N3 N4 > N5 and then

at least two of Ni satisfy Ni &N . Fix the extension u ei of ui such that
ke
ui kF̄ 0 .2kui kF̄ 0 (δ) , still denoted by ui .
From (5.86) and s = −3/4, it is easy to see (N +Ni1)m(Ni ) .N −3/4 hNi i−1/4
and
1 3/4 3/4
.N −3/2 N4 N5 .
m(N4 )m(N5 )
Hence we need to control
!
15
XZ δ Y3
3/4 3/4
N− 4 Λ5 hNi i−1/4 N4 N5 ; u1 , · · · , u5 dt. (5.105)
ki 0 i=1
If N2 ∼ N1 &N , N4 .N2 , then consider the worst case N1 > N2 > N4 >
N5 > N3 . We get from (5.103) that
X 5
Y
(5.105).N − 4
15
hN1 i−5/4 hN2 i−5/4 hN3 i1/6 N4 N5
7/6 7/6 [
kPki ukXki
Ni i=1
5
Y
15
.N − 4 kIuj kF̄ 0 (δ) . (5.106)
j=1
The rest cases N4 ∼ N5 &N , N1 .N5 or N1 ∼ N4 &N can be handled

similarly.
The following proposition shows that EI4 (t) is very close to EI2 (t).
Proposition 5.22. Assume I the Fourier multiplier with symbol m given

by (5.86) and s = −3/4. Then
|EI4 (t) − EI2 (t)|.kIu(t)k3L2 + kIu(t)k4L2 .
Proof. Since EI4 (t) = EI2 + Λ3 (σ3 ) + Λ4 (σ4 ), it suffices to prove

3
Y
|Λ3 (σ3 ; u1 , u2 , u3 )|. kIuj (t)kL2 , (5.107)
j=1
and
4
Y
|Λ4 (σ4 ; u1 , u2 , u3 , u4 )|. kIuj (t)kL2 . (5.108)
j=1
We may assume ubj is non-negative. To prove (5.107), it suffices to prove

2 Y 3
2 2
Λ3 m (ξ1 )ξ1 + m (ξ2 )ξ2 + m (ξ3 )ξ3 ; u1 , u2 , u3 . kuj k2 . (5.109)
ξ1 ξ2 ξ3 m(ξ1 )m(ξ2 )m(ξ3 )
i=1
Making dyadic frequency decomposition, we get the left-hand side of (5.109)

is bounded by

X m2 (ξ1 )ξ1 + m2 (ξ2 )ξ2 + m2 (ξ3 )ξ3

Λ3 ; 4k1 u1 , 4k2 u2 , 4k3 u3 .
ξ1 ξ2 ξ3 m(ξ1 )m(ξ2 )m(ξ3 )
ki >0
(5.110)
Denote Ni = 2ki . From symmetry we may assume N1 > N2 > N3 , and
hence N1 ∼ N2 &N .
Case 1. N3 N .
Now m3 (N3 ) = 1, hence we have
X N s N s

(5.112). Λ3 ; 4 u , 4 u , 4 u
1+s 1+s
N1 N1
k1 1 k2 2 k3 3
ki >0
X −1/4 −1/4

. Λ3 N1 N2 ; 4k1 u1 , 4k2 u2 , 4k3 u3 .
ki >0
It suffices to prove
XZ −1/2
3
Y 3
Y
N1 ηki (ξi )ubi (ξi ). kui kL2 .
ki >0 ξ1 +ξ2 +ξ3 =0,|ξi |∼Ni i=1 i=1
Define vi (x), whose Fourier transform is

−1/6
vbi (ξ) = Ni ubi (ξ)χ{|ξ|∼Ni } (ξ).
From Sobolev’s embedding theorem we get kvi kL3 .kui kL2 , thus use
Hölder’s inequality we get
XZ −1/2
3
Y
N1 ηki (ξi )ubi (ξi )
ki >0 ξ1 +ξ2 +ξ3 =0,|ξi |∼Ni i=1
X 3
Y 3
Y
−1/6 1/6
. N1 N3 kvi kL3 . kui kL2 .
ki >0 i=1 i=1
Case 2. N3 &N . It is easy to see that

!
X N3
−3/4 −3/4
N Y

3
(5.112). Λ3 ; Pk1 u1 , Pk2 u2 , Pk3 u3 . kui kL2 .
N
1/2
ki >0 1 i=1
5.5. I-method 143
Thus (5.107) is proved.

Next show (5.108). It suffices to prove
Y 4
σ4
Λ4 ; u , u , u , u . kuj k2 . (5.111)
m(ξ1 )m(ξ2 )m(ξ3 )m(ξ4 )
1 2 3 4
i=1
By dyadic frequency decomposition, we get the left-hand side of (5.111) is

bounded by

X σ4

Λ4 ; 4k1 u1 , 4k2 u2 , 4k3 u3 , 4k4 u4 .
m(ξ1 )m(ξ2 )m(ξ3 )m(ξ4 )
ki >0
(5.112)
Denote Ni = 2ki . From symmetry we may assume N1 > N2 > N3 > N4 ,
and hence N1 ∼ N2 &N . In the integration area

σ4 1 N −3
.Q . Q .
m(ξ1 )m(ξ2 )m(ξ3 )m(ξ4 ) 4 4 1/4
i=1 (N + Ni )m(Ni ) N i=1 i
We get from Hölder’s inequality that

X N −3
(5.112). Q4 1/4
k4k1 u1 kL2 k4k2 u2 kL2 k4k3 u3 kL∞ k4k4 u4 kL∞
ki >0 N
i=1 i
4
Y
. kuj k2 .
i=1
Thus we proved the proposition.

Now we are ready to extend the local solution to a global one. First we
need a variant local well-posedness which can be proved similarly.
Proposition 5.23. Let −3/4 6 s 6 0. Assume φ satisfies kIφkL2 (R) 6

20 1. Then there exists a unique solution to (5.32) on [−1, 1] such that
kIukF̄ 0 (1) 6 C0 , (5.113)
where C is independent of N .
For any given u0 ∈ H −3/4 and time T > 0, our purpose is to construct
a solution on [0, T ]. If u is a solution to KdV equation with initial data u0 ,
then for any λ > 0, uλ (x, t) = λ−2 u(x/λ, t/λ3 ) is also a solution to KdV
equation with initial data u0,λ = λ−2 u0 (x/λ). Simple calculations show
that
3
kIu0,λ kL2 .λ− 2 −s N −s ku0 kH s .
2s
For fixed N (N will be determined later), we take λ ∼ N − 3+2s such that
3
λ− 2 −s N −s kφkH s = 0 < 1.
For simplicity of notations, we still denote uλ by u, u0,λ by u0 , and assume
kIu0 kL2 6 0 . The purpose is then to construct solutions on [0, λ3 T ]. By
Proposition 5.23 we get a solution on [0, 1], and we need to extend the
solution. It suffices to control the modified energy EI2 (t) = kIuk2L2 .
First we control EI2 (t) for t ∈ [0, 1], we will prove EI2 (t) < 420 . By
standard bootstrap, we may assume EI2 (t) < 520 . Then from Proposition
5.22 we get
EI4 (0) = EI2 (0) + O(30 )
and
EI4 (t) = EI2 (t) + O(30 ).
From Proposition 5.21 we get for all t ∈ [0, 1]
EI4 (t) 6 EI4 (0) + C50 N −15/4 .
Thus
kIu(1)k2L2 = EI4 (1) + O(30 ) 6EI4 (0) + C50 N −15/4 + O(30 )
=20 + C50 N −15/4 + O(30 ) < 420 .
Thus the solution u can be extended to t ∈ [0, 2]. Iterating this procedure
M steps, then we get for t ∈ [0, M + 1]
EI4 (t) 6 EI4 (0) + CM 50 N −15/4
as long as M N −15/4 .1. Thus we get
EI2 (M ) = EI4 (t) + O(30 ) = 20 + O(30 ) + CM 50 N −15/4 < 420 ,
hence the solution can be extended to t ∈ [0, N 15/4 ]. Take N (T ) sufficiently
large such that
N 15/4 > λ3 T ∼ N 3 T.
Therefore, the u is extended to [0, λ3 T ].
In the end of this section, we prove some properties of the global solu-
tion. By the scaling we get
sup ku(t)kH s ∼λ3/2+s sup kuλ (t)kH s 6 λ3/2+s sup kIuλ (t)kL2 ,
t∈[0,T ] t∈[0,λ3 T ] t∈[0,λ3 T ]
kIφλ kL2 .N −s kφλ kH s ∼ N −s λ−3/2−s kφkH s .

5.6. Schrödinger equation with derivative 145
From the proof we have

sup kIuλ (t)kL2 .kIφλ kL2 ,
t∈[0,λ3 T ]
and hence
sup ku(t)kH s .N −s kφkH s .
t∈[0,T ]
Take λ such that kIφλ kL2 ∼ 0 1, thus we get

2
kφkH s 3+2s − 3+2s
2s
λ = λ(N, 0 , kφkH s ) ∼ N .
0
15 6s
Take N such that N 4 > λ3 T ∼ckφkH s ,0 N − 3+2s T , then N ∼ T 4/3 . Thus
the global solution u(x, t) satisfies
ku(t)kH −3/4 . (1 + |t|)kφkH −3/4 .
For the case s ∈ (−3/4, 0], we left the proof to the readers.
5.6 Schrödinger equation with derivative
In above several sections, we use the Bourgain’s spaces to prove some bi-
linear estimates with respect to KdV equation. However, in some cases,
for example: local well-posedness of the KdV equation in H −3/4 , it is not
enough to get the local result directly using the Bourgain’s spaces. In this
section, we consider the following Cauchy problem of the Schrödinger equa-
tion with derivative such that the reader can understand more the method
of the Bourgain’s spaces
(
iut + uxx = iλ(|u|2 u)x , (x, t) ∈ R2
(5.114)
u(x, 0) = u0 (x).
From Section 5.1 in this chapter, it follows that dispersion relation of (5.114)
φ(ξ) = −ξ 2 . Thus we can obtain the following the definition of the norm
of the Bourgain’s spaces X s,b with respect to Schrödinger equation with
derivative:
kukX s,b = khξi2 hτ + ξ 2 ib u
b(ξ, τ )kL2 .
In order to prove the well-posedness of the Cauchy problem (5.114),
from the standard method, it suffices to show
k∂x (uv̄w)kX s,b−1 .kukX s,b kvkX s,b kwkX s,b . (5.115)
But the above inequality is invalid for any s, b ∈ R, one can refer to [91].
In this section, we will use gauge transform to improve the derivative
nonlinearity, one can refer to [105; 184; 213]. The result in this section was
first given by Takaoka[213]. Via the gauge transformation which was first
used by Hayashi [102] (see also [106])
Rx
|u(y,t)|2 dy
v(x, t) = Gλ (u)(x, t) = e−iλ −∞ u(x, t), (5.116)
(5.114) is formally rewritten as the Cauchy problem
( 2
i∂t v + ∂x2 v = −iλv 2 ∂x v̄ − |λ|2 |v|4 v,
(5.117)
v(x, 0) = v0 (x),
Rx
|u0 (y)|2 dy
where v0 (x) = e−iλ −∞ u0 (x). The corresponding inverse gauge
transformation,
Rx
|v(y,t)|2 dy
u(x, t) = eiλ −∞ v(x, t) = G−λ (v).
Since the map Gλ is a bicontinuous map from H s to H s , we can show that
the global well-posedness of (5.114) in H s is equivalent to that of (5.117).
in the following, we only consider the Cauchy problem (5.117).
Notice that the nonlinear term of the original equation iλ(|u|2 u)x =
2iλ|u|2 ux + iλu2 ūx , via the gauge transformation, the derivative in the
nonlinearity |u|2 ux has been replaced by the quintic nonlinearity |v|4 v. The
Strichartz estimate can control the nonlinearity |v|4 v easy. But, at first
sight, there still exists a derivative in the nonlinearity u2 ūx in equation
(5.117). However a derivative of the complex conjugate of the solution u
can be handled while a derivative of u cannot since ω(ξ) = −ξ 2 is a even
function.
From the definition of Bourgain’s spaces, it follows that for complex
conjugate of u
s b s b
kūkX s,b = khξi hτ − ξ 2 i F ukL2ξ L2τ , kukX s,b = khξi hτ + ξ 2 i F ukL2ξ L2τ .
For the nonlinearityu2 ūx ,we can prove that following estimate holds for
s > 1/2 and b > 1/2
kuv w̄x kXs,b−1 .kukXs,b kvkXs,b kwkXs,b . (5.118)
But for |u|2 ux , the above inequality is invalid.
Now we mainly estimate the nonlinear term v 2 ∂x v̄. Similarly with the
KdV equation, we also need some space-time estimates (Strichartz esti-
mate, local smoothing effect and maximal function estimate) of Schrödinger
eqution, the proofs of these estimates are similar with one of the Benjamin-
Ono equation (the dispersion relation ω(ξ) = −|ξ|ξ) in Section 5.2.
2
Lemma 5.9. The group of Schrödinger equation S(t) = Fx−1 eitξ Fx sat-
isfies :
kS(t)v0 kL6x L6t .kv0 kL2 ,
1/2
kD S(t)v0 kL∞ 2 .kv0 kL2 ,
x Lt
kD−1/4 S(t)v0 kL4x L∞

t
.kv0 kL2 .
Using Lemma 5.3 and Lemma 5.9, we can easily obtain the following
lemma:
Lemma 5.10. Let u ∈ X 0,b with 1/2 1/2, 1/2 1/2. Then
kv1 v2 ∂x v¯3 kXs,b−1 6 Ckv1 kXs1 ,b0 kv2 kXs2 ,b0 kv3 kXs3 ,b0 . (5.119)
Proof. By duality and the Plancherel identity, it suffices to show that

Z Q Y4
hξ4 is hτ4 − ξ42 ib−1 ξ3 4i=1 fi (ξi , τi )
2 0 2 0 2 0
Q 3 . kfi kL2 , (5.120)
Γ4 hτ1 + ξ1 ib hτ2 + ξ2 ib hτ3 − ξ3 ib j=1 hξi i
s
i=1
for all fi ∈ L2 (R2 ) > 0, i = 1, 2, 3, 4; where the hyperplane Γ4 is defined by

Γ4 = {(ξ, τ ) ∈ R4 × R4 : ξ1 + ξ2 + ξ3 + ξ4 = 0, τ1 + τ2 + τ3 + τ4 = 0}
which we endow with the standard measure: θi = (ξi , τi )
Z Z
h(θ1 , θ2 , θ3 , θ4 ) = h(θ1 , θ2 , θ3 , −θ1 − θ2 − θ3 )dθ1 dθ2 dθ3 .
Γ4 R6
From Plancherel identity, it follows that

Z Y 4 Z Y
4
fi (ξi , τi ) = F −1 (fi )(x, t)dxdt. (5.121)
Γ4 i=1 R2 i=1
By symmetry it suffices to estimate the integral in the domain |ξ1 | 6 |ξ2 |.

We define |ξ|max , |ξ|sub , |ξ|thd , |ξ|min be the maximum, second largest,
third largest and minimum of |ξ1 |, |ξ2 |, |ξ3 |, |ξ4 |; and define σj± = τj ± ξj2 ,
j = 1, 2, 3, 4. Let

fi fk
Fi = F −1 , i = 1, 2; Fk = F −1
, k = 3, 4.
hτ + ξ 2 ib0 hτ − ξ 2 ib0
|ξ3 |hξ4 is
K(ξ1 , ξ2 , ξ3 , ξ4 ) = .
hξ1 is hξ2 is hξ3 is
In order to obtain the boundedness of the integral in the left side of (5.120),
we split the domain of integration in several pieces as in [91].
Case 1. If max(|ξ1 |, |ξ2 |, |ξ3 |, |ξ4 |) 6 10, then
K(ξ1 , ξ2 , ξ3 , ξ4 ) 6 C.
By Lemma 5.10 and (5.121), the left side of (5.120) restricted to this domain
is bounded by
Z Q4 Y3 Y4
i=1 fi (ξi , τi )
2 b 0 2 b 0 2 b0 .kf 4 k 2 kFi k L 6. kfi kL2 .
Γ4 hτ1 + ξ1 i hτ2 + ξ2 i hτ3 − ξ3 i i=1 i=1
Case 2. If |ξ|max 1 and |ξ|thd |ξ|sub , then from ξ1 + ξ2 + ξ3 + ξ4 = 0

it follows that |ξ|max ∼ |ξ|sub . We split this case into the following several
subcases.
Subcase 2a. If |ξ2 | |ξ3 | ∼ |ξ4 |, then recall the following identity on
the hyperplane Γ4
h(ξ1 , ξ2 , ξ3 , ξ4 ) = σ1+ + σ2+ + σ3− + σ4− = 2(ξ1 + ξ3 )(ξ2 + ξ3 ). (5.122)
This implies that
max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |)&|ξ3 |2 .
By symmetry, we only consider |σ4− | = max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |) and
|σ2+ | = max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |).
If |σ4− | = max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |), then for b 6 3/4 and s > 1/4, we
have
K(ξ1 , ξ2 , ξ3 , ξ4 ) |ξ3 ||ξ3 |2(b−1) |ξ3 |1/2
− 1−b . s s
. .
|σ4 | hξ1 i hξ2 i hξ1 is hξ2 is
By Lemma 5.10, (5.121) and Hölder inequality, the left side of (5.120)
restricted to this domain is bounded by
Z Q4
|ξ3 ||ξ3 |1/2 i=1 fi (ξi , τi )
Q2 s 2 b0 2 b0 2 b0
Γ4 i=1 hξi i hτ1 + ξ1 i hτ2 + ξ2 i hτ3 − ξ3 i
4
Y
.kJ −s F1 kL4x L∞
t
kJ −s F2 kL4x L∞
t
kf4 kL2x,t kΛ1/2 F3 kL∞ 2.
x Lt
kfi kL2 .
i=1
0
If |σ2+ | = max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |), then from & hσ4− i1−b hσ2+ ib
− b0
hσ4 i hσ2+ i1−b ,similarly with above, for b 6 3/4 and s > 1/4, the left
side of (5.120) restricted to this domain is bounded by
Z Q
|ξ3 ||ξ3 |2(b−1) 4i=1 fi (ξi , τi )
Q2 s 2 b0 2 b0 2 b0
Γ4 i=1 hξi i hτ1 + ξ1 i hτ3 − ξ3 i hτ4 − ξ4 i
4
Y
.kJ −s F1 kL4x L∞
t
kF4 kL4x L∞
t
kf2 kL2x,t kΛ1/2 F3 kL∞ 2.
x Lt
kfi kL2 ,
i=1
which follows from Lemma 5.10, (5.121) and Hölder inequality.
Subcase 2b. Assume |ξ2 | ∼ |ξ3 | ∼ |ξ|max or |ξ2 | ∼ |ξ4 | ∼ |ξ|max . By
symmetry, we can assume |ξ2 | ∼ |ξ3 | ∼ |ξ|max .
If |ξ2 + ξ3 | 6 1 or |ξ4 | . |ξ1 |, then hξ1 i ∼ hξ4 i. For s > 1/2, we have
K(ξ1 , ξ2 , ξ3 , ξ4 ) 6 C.
Similarly with Case 1, the left side of (5.120) restricted to this domain is
bounded by
Z Q4 Y4
i=1 fi (ξi , τi )
2 b 0 2 b 0 2 b0 . kfi kL2 .
Γ4 hτ1 + ξ1 i hτ2 + ξ2 i hτ3 − ξ3 i i=1
If |ξ2 + ξ3 | > 1 and |ξ4 | |ξ1 |, then
max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |)&|ξ1 + ξ3 ||ξ2 + ξ3 | ∼ hξ3 ihξ4 i.
By symmetry, we only consider |σ4− | = max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |). Then
for s > 1/2 and b 6 3/4
K(ξ1 , ξ2 , ξ3 , ξ4 ) |ξ3 |1/2 |ξ3 |1/2
− 1−b . s 1−b 1−b
. .
|σ4 | hξ1 i hξ3 i hξ4 i hξ1 i1/4 hξ2 i1/4
Similarly with Subcase 2a, we can obtain the result.
Subcase 2c. If |ξ1 | ∼ |ξ3 | ∼ |ξ|max or |ξ1 | ∼ |ξ4 | ∼ |ξ|max , then similarly
with Subcase 2b, we can obtain the result.
Subcase 2d. If |ξ1 | ∼ |ξ2 | ∼ |ξ|max , then max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |)
&|ξ3 |2 , similarly with Subcase 2a, we can obtain the result.
Case 3. If |ξ|min |ξ|thd ∼ |ξ|sub ∼ |ξ|max , then we have
max(|σ1+ |, |σ2+ |, |σ3− |, |σ4− |)&|ξ3 |2 .
Similarly with Subcase 2a, we can obtain the result.
Case 4. If |ξ|min ∼ |ξ|thd ∼ |ξ|sub ∼ |ξ|max , similarly with Case 1, we
can obtain the result.
This completes the proof of theorem.
Next we estimate the nonlinearity |u|4 u. From the above standard ar-
gument, it suffices to show that
5
Y
ku1 u2 u3 ū4 ū5 kX s,b−1 . kui kX s,b . (5.123)
i=1
Notice that there exists the condition s > 1/2 when we estimate the non-
linearity u2 ūx , so we do not need to obtain the sharp index of s for (5.123),
the index s > 1/2 is enough for the local well-posedness. But the author
believe that the sharp index of s for (5.123) is s > 0.
Lemma 5.11. Let s ∈ R, 1/2 < b 6 1. Then
kψ(t/T )ukX s,b−1 .kukL2/(3−2b) H s .

T x
Proof. Without loss of generality, we assume s = 0. Using the defi-

2/(3−2b)
nition of X s,b , imbedding inequality Lt ,→ Htb−1 and Minkowski’s
inequality, we have

kψ(t/T )ukX 0,b−1 = ke−itω(ξ) ψ(t/T )(Fx u)(ξ, t)kH b−1 L2
t ξ
−itω(ξ)

. ke ψ(t/T )(Fx u)(ξ, t)kL2/(3−2b) L2
t ξ
.kukL2/(3−2b) L2 .
T x
This completes the proof of lemma.
Theorem 5.9. Let s > 1/2 and 1/2 0
such that
5
Y
kψ(t/T )u1u2 u3 ū4 ū5 kX s,b−1 .T θ kui kX s,b . (5.124)
i=1
Proof. Assume u1 = · · · = u5 = u, it suffices to show that
kψ(t/T )|u|4ukX s,b−1 .T θ kuk5X s,b . (5.125)
Using Lemma 5.11, Lemma 5.10 and the Leibniz rule for fractional deriva-
tives, we have
kψ(t/T )|u|4 ukX s,b−1 .k|u|4 ukL2/(3−2b) H s .ku4 kL∞ 2 kuk 2/(3−2b) s
T Lx L H
T x T ∞
.T θ kuk4L∞ L8x kukL4T H∞ θ 5

s .T kuk s,b .
X
T
This completes the proof.

Using the proof of Theorem 5.5 in Section 5.3, we have
Proposition 5.24. Let 1/2 6 s < 1. For v0 ∈ H s , then there exists

b > 1/2 and T = T (kv0 kH 1/2 ) > 0 such that the Cauchy problem (5.117)
admits a unique local solution v ∈ XTs,b .
For the original equation (5.114), we have
Theorem 5.10. Let λ ∈ R and s > 1/2. For u0 ∈ H s (R), then there exists
b > 1/2 and T = T (ku0 kH 1/2 ) > 0 such that the Cauchy problem (5.114)
admits a unique local solution
u ∈ C([−T, T ] : H s (R)), Gλ (u) ∈ XTs,b .
Moreover, given t ∈ (0, T ), the map u0 → u(t) is Lipschitz continuous from
H s to C(0, T ; H s ).
Proof. For u0 ∈ H s and 1/2 6 s < 1, we define v0 ∈ H s

Rx
|u0 (y)|2 dy
v0 (x) = e−iλ/2 −∞ u0 (x).
(n) (n)
Assume v0is a sequence in H ∞ satisfying v0 → v0 in H s . Let vn
be a solution of the Cauchy problem (5.117) obtained in Proposition 5.24.
Define
un (x, t) =G−λ (vn )(x, t),
Rx (n)
(n) (y)|2 dy (n)
u0 (x) =eiλ/2 −∞
|v0
v0 (x).
From the definition of un , it follows that
kun kL∞ 2 = kvn kL∞ L2 .
T Lx T x
Using the Leibniz rule for fractional derivatives and Sovolev inequality, for
1/2 6 s < 1, we have

s iλ/2 R−∞ x (n)
|v0 (y)|2 dy
kDs un kL∞
T L x
s
2 .kD vn kL∞ L2 + D
T x
e vn ∞ 2
LT Lx
Rx
s s iλ/2 −∞ |v0 (y)|2 dy
(n)
.kD vn kL∞
T Lx
2 + D e ∞ p kvn kL∞ q
T Lx
LT Lx
s
.kD vn kL∞ 2 +
T Lx
kvn k2L∞ L2p
x
1 kvn kL∞ H 1/2
T T x
.(1 + kvn k2L∞ s )kvn kL∞ H s ,

T Hx T x
where 2 < p < ∞, 1/p + 1/q = 1/2 and 1/p1 = 1/p + 1 − s. Similarly with
above, we have
2
kun − um kL∞ s .(1 + kvn kL∞ H s + kvm kL∞ H s ) kvn − vm kL∞ H s .
T Hx T x T x T x
It suffices to show that the map Gλ is Lipschitz continuous from H s to Lp

for any 2 6 p < ∞. In fact, we have

iλ/2 R−∞
x
|f (y)|2 dy
Rx 2
e f (x) − eiλ/2 −∞ |g(y)| dy g(x)
Lp

iλ/2 R−∞
x 2
|f (y)| dy
Rx 2
.kf − gkH s + e − eiλ/2 −∞ |g(y)| dy ∞ (kf kH s + kgkH s ).
L
0
Using the fact that |f (x) − f (y)| 6 kf kL1 , we have

iλ/2 R−∞
x
|f (y)|2 dy
Rx 2
e − eiλ/2 −∞ |g(y)| dy
∞
L
iλ/2 R−∞
x
(|f (y)|2 −|g(y)|2 )dy
. e − 1 ∞
L
.k|f |2 − |g|2 kL1 .kf − gkH s (kf kH s + kgkH s ).
Let n → ∞, we can show that u is a solution of the Cauchy problem (5.114).
This completes the proof of theorem.
5.7 Some other dispersive equations
Except for KdV equation and Schrödinger equation, there exist some other
important dispersive equations which have been considered by lots of math-
ematicer recently. In this section, the authors give some methods for some
other dispersive equations such that the readers can understand more the
method of Bourgain’s spaces.
From the bilinear estimates of KdV equation and trilinear estimates of
Schrödinger equation with derivative, it follows that we can use not only
the method of Tao’s [k; Z] multiplier [216], but also the method of the
local smoothing effects and maximal function estimates [213] to consider
the well-posedness of dispersive equations. We take the bilinear estimates
of KdV equation for example to show the difference of the two methods.
First, we can use the method of Tao’s [k; Z] multiplier to prove Theorem
5.3. That is
k∂x (u1 u2 )kX s,b−1 6 Cku1 kX s,b0 ku2 kX s,b0 , s > −3/4, b, b0 > 1/2. (5.126)
However, (5.126) holds with the condition s > −5/8 using only Strichartz
estimate, the local smoothing effects and the maximal function estimates
of KdV equation, one can refer to [128]. If −5/8 > s > −3/4, Kenig, Ponce
and Vega [132] used the following inequalities to obtain (5.126).
Z ∞
dx C 1
2b (1 + |x − β|)2b
6 2b
, b> (5.127)
−∞ (1 + |x − α|) (1 + |α − β|) 2
5.7. Some other dispersive equations 153
Z ∞
dx C 1
√ 6 1 , b > (5.128)
−∞ (1 + |x|)2b a − x (1 + |a|) 2 2
Z ∞
dx C 1
6 , b>
−∞ (1 + |x − α|)2(1−b) (1 + |x − β|)2b (1 + |α − β|)2(1−b) 2
(5.129)
Z 1
dx C(1 + β)2(b− 2 ) 1
√ 6 1 , b> . (5.130)
|x|6β (1 + |x|)2(1−b) a−x (1 + |a|) 2 2
As in Section 5.3 of this chapter, we first define ξ = ξ1 + ξ2 , τ = τ1 + τ2
σ = τ − ξ 3 , σ1 = τ1 − ξ13 , σ2 = τ2 − ξ23 . We split the the hyperplane
{(ξ, ξ1 , ξ2 ) × (τ, τ1 , τ2 ) ∈ R3 × R3 : ξ = ξ1 + ξ2 , τ = τ1 + τ2 } into suitable
several parts. For the case: |ξ1 | ∼ |ξ2 | |ξ| and |σ| > |σ1 |, |σ2 |, we only use
the inequality (5.127); for other cases, similarly with the paper [128], we
use the Strichartz estimates, the local smoothing effects and the maximal
function estimates, thus we can also obtain (5.126). This can simplify the
proof of the paper [132], the readers can try it. From this, it follows that the
interaction: high × high → low is worst when one consider low regularity
solution in Sobolev spaces of negative indice. Moreover, corresponding
to the inequality (5.127), there exists the following inequality, which is
also important when one consider the local well-posedness of the dispersive
equations
Z
dx C 1 1
2b 2b
6 4b−1
, <b< . (5.131)
R hx − αi hx − βi hα − βi 4 2
Next we take the fourth-order nonlinear Schrödinger equation for exam-
ple, and show how to consider the well-posedness for the dispersive equation
with complicated phase function (the dispersion relations of KdV equation
φ(ξ) = ξ 3 and Schrödinger equation φ(ξ) = |ξ|2 do not have non-zero sin-
gular points, that is, the solutions of the phase function φ(ξ) = 0, φ0 (ξ) = 0
and φ00 (ξ) = 0 are only zero; the complicated phase functions are generally
assumed to have non-zero singular points). The Cauchy problem for the
fourth-order Schrödinger equation which describes the motion of the vortex
filament as follows:
i∂t u + ∂x2 u + ν∂x4 u = F (u, ū, ∂x u, ∂x ū, ∂x2 u, ∂x2 ū), (x, t) ∈ R × R, (5.132)
where ν is a non-zero real constant, the nonlinear term F is given by
F (u, ū, ∂x u, ∂x ū, ∂x2 u, ∂x2 ū)
1
= − |u|2 u + λ1 |u|4 u + λ2 (∂x u)2 ū + λ3 |∂x u|2 u + λ4 u2 ∂x2 ū + λ5 |u|2 ∂x2 u
2
with λ1 = − 3µ ν
4 , λ2 = −2µ+ 2 , λ3 = −4µ−ν, λ5 = −2µ+ν, where µ is also a
real constant. The equation (5.132) describes the three-dimensional motion
of an isolated vortex filament, which is embedded in inviscid incompressible
fluid fulfilled in an infinite region. Note that the dispersion relation of the
equation (5.132) φ(ξ) = νξ 4 − ξ 2 has non-zero singular points. we can not
directly use the local smoothing effects and the maximal function estimates
of Section 5.2 in this chapter to consider it. So we need use the following
Fourier restriction operators
Z Z
PNf = eixξ fˆ(ξ)dξ, PN f = eixξ fˆ(ξ)dξ, ∀N > 0, (5.133)
|ξ|>N |ξ|6N
to eliminate the singularity of the phase function. Using the Fourier re-
striction operators and the results of Section 5.2, we have
3
kDx2 P 2a S(t)ϕkL∞ 2 6 CkϕkL2 ,
x Lt
(5.134)
− 41
kDx P a S(t)ϕkL4x L∞ t
6 CkϕkL2 , (5.135)
Z ∞ 1/2
sup |P a S(t)u0 |2 dx 6 CT,s ku0 kH s , s > 1, (5.136)
−∞ [−T,T ]
where a depends on ν. Then using (5.134), (5.135), (5.136) and the

Strichartz estimates, and the Bourgain’s spaces, we can obtain the local
well-posedness of the Cauchy problem (5.132). One can refer to [109;
110].
For some other dispersive equation, in order to obtain the well-
posedness, we can not directly use the Bourgain’s spaces X s,b to construct
the contraction mapping. So we need new method to consider them. In
this book, we do not introduce them in details, one can refer to Guo’s thesis
[93]. For example: the generalized Benjamin-Ono (BO) equation:
∂t u + H∂xx u = µ∂x (uk ), u(x, 0) = u0 (x), (5.137)
whereR k ∈ Z, µ ∈ R, and H is the Hilbert transform H(f )(x) =
1 f (y) d b
π p.v. R x−y dy. From Hf = −isgn(ξ)f (ξ), it follows that the dispersion
relation of BO equation ω(ξ) = −|ξ|ξ. This model was first introduced
by Benjamin [10] and Ono [182], and it describe one-dimensional internal
waves in deep water. Unlike Schrödinger equation, if initial data u0 is a
real number, then the solution u of (5.137) is also a real number. The real
and complex BO equation are different.
Compared with KdV equation, BO equation has weaker dispersive ef-
fect. If k = 2, the Cauchy problem of BO equation (5.137) is badly behaved
with respect to Picard iterative methods in standard Sobolev spaces. But
5.7. Some other dispersive equations 155
for the definition of some weaker well-posedness (for example: the solution
mapping is only continuous), some local results of the real BO equation
were obtained. For now, the best local result was obtained by Kenig and
Ionescu[113], who obtained local well-posedness in L2 . Their method based
on the gauge transform and the Besov-type Bourgain’s spaces X s,b . If
k = 3, the best result of the well-posedness was obtained by Kenig and
Takaoka [136], who used the dyadic gauge transform to prove that the
Cauchy problem is globally well-posed in H 1/2 . Moreover,they showed
that the index s = 1/2 is sharp in the sense: the solution map u0 → u
as mapping from H s to C([−T, T ]; H s ) is no longer uniformly continuous
when s < 1/2. Recently, in the paper [91], the author followed the idea in
[113], used the contraction mapping to obtain the global well-posedness in
H 1/2 with small data in L2 norm, here the author did not use the gauge
transform. So this method can apply to complex BO equation.
KdV equation and BO equation can be viewed as the special cases of
the following dispersion generalized BO equation:
∂t u + |∂x |1+α ∂x u = µ∂x (uk ), u(x, 0) = u0 (x), (5.138)
2
where 0 6 α 6 1, u : R →R and |∂x | is the Fourier multiplier operator
with symbol |ξ|. This model is very interesting from mathematical view, it
let us understand the relation between the dispersion effect and the well-
posedness. From the intensity of dispersion effect, it lies between KdV
equation and BO equation, for example: if 0 < α < 1, then it is similar
with BO equation, H s assumption on the initial data is insufficient for the
local well-posedness via Picard iteration by showing the solution mapping
fails to be C 2 smooth from H s to C([−T, T ]; H s ) at the origin for any s,
one can refer to [167]. For the methods to obtained the well-posedness for
(5.138), the method of refined energy was used in [127], the method of short
time X s,b in [115] was used in [92], and a para-differential renormalization
technique was used in [108].
Notice that the spatial dimension of the dispersive equations considered
above in this chapter is one. How about the application of the Bourgain’s
spaces to dispersive equations in higher dimension? For example: the Bour-
gain’s spaces can be applied to the higher dimensional Schrödinger equation.
Compared with one dimension case, cases in higher dimension are more del-
icate. For instance, a quadratic non-linear Schrödinger equation(n = 1, 2)
iut + ∆u = N (u, ū), u(x, 0) = u0 (x),
where (x, t) ∈ R × R, N (u, ū) = c1 |u|2 + c2 u2 + c3 ū2 . Using only the
n
Strichartz estimates, one can obtain the well-posedness in H s with s > 0.

However, by the Bourgain’s spaces, it was showed that it is well-posed in

H s with s > −3/4(c1 = 0) and H s with s > −1/4(c1 6= 0) respectively, one
can refer to [131; 41]. These results can be improved, one can refer to [9;
142]. For other nonlinearity |u|p u to non-linear Schrödinger equation, one
can not obtain the better result of low regularity solution using Bourgain’s
spaces, but one can obtain the better result of global well-posedness using
I-method, one can refer to [45]. Finally, we introduce KP equation, which
can be viewed as two dimensional KdV equation. The KP equation is also
an important shallow water model. The Cauchy problem of KP-I equation
as follows:
∂t u + ∂x3 u − ∂x−1 ∂y2 u + ∂x (u2 /2) = 0; u(x, y, 0) = φ(x, y), (5.139)
3
where u(x, y, t) : R → R. The KP-I equation and The KP-II equation,
in which the sign of the term ∂x−1 ∂y2 u in (5.139) is + instead of −, arise
in physical contexts as models for the propagation of dispersive long waves
with weak transverse effects. The KP-II equation is well understood from
the point of view of well-posedness, one can refer to [16; 214; 212; 97;
98]. For the KP-I equation, Molinet, Saut and Tzvetkov [169] showed that
it is badly behaved with respect to Picard iterative methods in standard
Sobolev spaces H s1 ,s2 with any s1 , s2 ∈ R. On the positive side, it is
known that the KP-I initial value problem is globally well-posed in the
second energy spaces [122; 168]; these global well-posedness results rely
on refined energy methods. Recently, Ionescu, Kenig and Tataru [115]
obtained the global well-posedness in the natural energy space E1 = {φ ∈
L2 (R2 ), ∂x φ ∈ L2 (R2 ), ∂x−1 ∂y φ ∈ L2 (R2 )}, they introduce a new method,
which can be looked as the blend of Bourgain’s spaces method and energy
method. Recently, the authors in [95] remove the condition ∂x−1 ∂y φ ∈
L2 (R2 ). Compared with the result of KP-II equation, the global well-
posedness of KP-I equation in L2 is a open problem.
Chapter 6
Frequency-uniform decomposition
techniques
W. Orlicz had a small apartment and he once applied to the city ad-
ministration for a bigger one. The answer of an employee was: “Your
apartment is really small but we cannot accept your claim since we
know that you have your own spaces!” —— Lech Maligranda2
In this chapter we study the NLS and the NLKG by using the frequency-
uniform decomposition techniques, see [243; 245; 246; 247].
Let Qk be the unit cube with the center at k, {Qk }k∈Zn constitutes a
decomposition of Rn . Such a kind of decomposition goes back to the work
of N. Wiener [252], and we say that it is the Wiener decomposition of Rn .
We can roughly write
k ∼ F −1 χQk F , k ∈ Zn , (6.1)
where χE denotes the characteristic function on the set E. Since Qk is a
translation of Q0 , k (k ∈ Zn ) have the same localized structures in the
frequency space, which are said to be the frequency-uniform decomposition
operators. Similar to Besov spaces, one can use {k }k∈Zn and `q (Lp ) to
generate a class of function spaces, so called modulation spaces for which
the norm is defined by
!1/q
X
sq q
kf kMp,q
s = hki k k f kp .
k∈Zn
6.1 Why does the frequency-uniform decomposition work
Comparing with the dyadic decomposition, the frequency-uniform decom-

position has at least two advantages for the Schrödinger semi-group:
2 W. Orlicz (1903-1990), Polish mathematician and he is known for his Orlicz spaces,
L. Maligranda is one of his students, see http://www.sm.luth.se/l̃ech.
157
158 Frequency-uniform decomposition techniques
0
(1) k eit∆ : Lp → Lp satisfies a uniform truncated decay;
(2) k eit∆ is uniformly bounded in Lp .
It is known that S(t) = eit∆ : Lp → Lp if and only if p = 2. This is one
of the main reasons that we can not solve NLS in Lp (p 6= 2). However, if
we consider the frequency-uniform decomposition, there holds
kk S(t)f kp . (1 + |t|)n|1/2−1/p| kk f kp . (6.2)
n
Taking the summation to the above inequality over all k ∈ Z , we can get
the relevant estimate in modulation spaces, which enable us to solve NLS
0
in modulation spaces Mp,1 , 1 6 p 6 ∞.
Now we give a comparison between frequency-uniform and dyadic de-
compositions. Recalling that
4k ∼ F −1 χ{|ξ|∼2k } F
and noticing that |{ξ : |ξ| ∼ 2k }| = O(2nk ) and |Qk | = 1, we see that their
Bernstein’s estimates are quite different:
k4k f kq . 2n(1/p−1/q)k k4k f kp , kk f kq . kk f kp , p 6 q.
We find that there is no regularity increasement for the Bernstein estimate
of k , which leads to the Schrödinger semi-group S(t) = eit∆ satisfies the
following truncated decay estimate,
kk S(t)f kp . (1 + |t|)−n(1/2−1/p) kk f kp0 , p > 2, 1/p + 1/p0 = 1. (6.3)
0
Comparing it with the classical Lp → Lp estimate
k4k S(t)f kp . |t|−n(1/2−1/p) k4k f kp0 , (6.4)
we see that the singularity at t = 0 disappears in (6.3).
Recalling that in the classical Strichartz inequalities, in order to handle
the singularity of |t|−n(1/2−1/p) at t = 0, we need condition n(1/2−1/p) 6 1,
which arises from the Hardy-Littlewood-Sobolev inequality and it is an
essential condition. To solve NLS with the nonlinearity |u|σ u, one need to
improve the spatial regularity at least at the Ḣ n/2−2/σ level for σ > 4/n.
Considering the Strichartz inequalities with the frequency-uniform de-
composition, due to the truncated decay, we can remove the condition
n(1/2 − 1/p) 6 1. As a result, in solving NLS with the nonlinearity |u|σ u,
it is not necessary to improve the spatial derivative regularity if σ is large.
The frequency-uniform decomposition is more delicate than the dyadic
decomposition and it is easier to handle the derivatives in the nonlinearity,
say for k = (k1 , ..., kn ) with k1 1 and j 1,
k∂x1 k f kp ∼ k1 kk f kp , k∂x1 4j f kp . 2j k4j f kp .
6.2. Frequency-uniform decomposition, modulation spaces 159
Using the frequency-uniform decomposition techniques, the initial data

s s
should belong to modulation spaces Mp,q . The modulation space Mp,q
s
coincides with the Sobolev space H in the case p = q = 2. Re-
cently, this technique was also developed to the Navier-Stokes equa-
tion and the dissipative nonlinear electrohydrodynamic system [247; 116;
56].
6.2 Frequency-uniform decomposition, modulation spaces
Roughly speaking, dyadic decomposition operators combined with function

spaces `q (Lp ) generate Besov spaces, frequency-uniform decomposition op-
erators joint with function spaces `q (Lp ) produce modulation spaces. The
modulation space was introduced by Feichtinger [73] in 1983, from the
history point of view, it was defined by the short-time Fourier transform3 .
During the past twenty years, the frequency-uniform decomposition had not
been attached the importance to applications and it is even not mentioned
in Gröchenig’s book [86]. However, from PDE point of view, the combi-
nation of frequency-uniform decomposition operators and Banach function
spaces `q (X(Rn ))4 seems to be important in making nonlinear estimates,
which contains an automatic decomposition on high-low frequencies.
We now give an exact definition on frequency-uniform decomposition
operators. Since χQk can not make differential operations, one needs to
replace χQk in (6.1) by a smooth cut-off function. Let ρ ∈ S (Rn ), ρ :
Rn → [0, 1] be a smooth function verifying ρ(ξ) = 1 for |ξ|∞ 6 1/2 and
ρ(ξ) = 0 for |ξ|∞ > 15 . Let ρk be a translation of ρ,
ρk (ξ) = ρ(ξ − k), k ∈ Zn . (6.6)

3 Let g ∈ S (Rn ),
Z
Vg f (x, ω) = e−itω g(t − x)f (t)dt.
Rn
Vg f is said to be the short-time Fourier transform of f . The norm on modulation spaces

is given by
Z Z q/p !1/q
kf koM s = p
|Vg f (x, ω)| dx sq
hωi dω . (6.5)
p,q
Rn Rn
One can prove that k · koM s and k · kMp,q

s are equivalent norms; cf. [73] for a proof on
p,q
modulation spaces defined in an Abel group and [246] for a straightforward proof.
4 X is a Banach function space defined in Rn .
5 For ξ = (ξ , ..., ξ ), |ξ|
1 n ∞ := maxi=1,...,n |ξi |.
P n
We see that ρk (ξ) = 1 in Qk and so, k∈Zn ρk (ξ) > 1 for all ξ ∈ R .
Denote
!−1
X
σk (ξ) = ρk (ξ) ρk (ξ) , k ∈ Zn . (6.7)
k∈Zn
Then we have


 |σk (ξ)| > c, ∀ ξ ∈ Qk ,

supp σk ⊂ {ξ : |ξ − k|∞ 6 1},
P (6.8)

 n σk (ξ) ≡ 1, ∀ ξ ∈ Rn ,
 k∈Z
|D σk (ξ)| 6 C|α| , ∀ ξ ∈ Rn , α ∈ (N ∪ {0})n .
α
Hence, the set

Υn = {{σk }k∈Zn : {σk }k∈Zn satisfies (6.8)} (6.9)
is non-void. If there is no confusion, in the sequel we will write Υ = Υn .
Let {σk }k∈Zn ∈ Υ. Denote
k := F −1 σk F , k ∈ Zn . (6.10)
{k }k∈Zn are said to be frequency-uniform decomposition operators. For
k ∈ Zn , we denote |k| = |k1 | + ... + |kn |, hki = 1 + |k|. Let s ∈ R,
0 < p, q 6 ∞,
 !1/q 
 X 
s
Mp,q (Rn ) = f ∈ S 0 (Rn ) : kf kMp,q
s = hkisq k k f kqp <∞ .
 n

k∈Z
(6.11)
0 s
For simplicity, we write Mp,q = Mp,q . Mp,q is said to be the modulation
space.
6.2.1 Basic properties on modulation spaces

As indicated in Proposition 1.16, if we consider a function f with a compact
support set in the frequency space, then we can compare kf kp with kf kq ,
which is one of the advantages of the frequency localizations. The following
is a refinement of Proposition 1.16.
Lemma 6.1. Let Ω be a compact subset of Rn , diam Ω < 2R, 0 0 which only depends on p, q and R
such that
kf kq 6 Ckf kp , ∀ f ∈ LpΩ , (6.12)
where LpΩ = {f ∈ L : suppfˆ ⊂ Ω}.
p
Proof. We emphasize that the constant in (6.12) is independent of the

position of Ω. Let ψ ∈ S (Rn ) satisfy supp ψb ⊂ B(0, 2R) and ψ| b B(0,R) = 1.
Due to diam Ω < 2R, we can take ξ0 satisfting Ω ⊂ B(ξ0 , R). For any
f ∈ S (Rn ) ∩ LpΩ , we have fb = fb · ψ(·
b − ξ0 ), which implies that
Z
f (x) = c f (x − y)eiξ0 y ψ(y)dy.
Rn
If 1 6 p 6 ∞, using Young’s inequality, we can get the conclusion. If

0 < p < 1,
Z
1−p
kf k∞ 6 Cψ sup |f (x − y)| |f (x − y)|p dy,
y Rn
we obtain that for q = ∞ and 0 < p < 1, the result holds. For general
0 < p 6 q < ∞, by Hölder’s inequality, Lq norm can be controlled by Lp
and L∞ norms, from which we get the result, as desired.
Lemma 6.2 (LpΩ -multiplier). Let Ω ⊂ Rn be a compact subset, 0 < r 6

∞ and σr = n(1/(r ∧ 1) − 1/2). If s > σr , then there exists a C > 0 such
that
kF −1 ϕF f kr 6 CkϕkH s kf kr (6.13)
holds for all f ∈ LrΩ and ϕ ∈ H s .
Proof. If r > 1, in view of Bernstein’s multiplier estimates, we have the

result, see Proposition 1.11. If r < 1, the proof is similar to the case r > 1,
cf. [224].
Proposition 6.1 (Completeness). Let 0 < p, q 6 ∞ and s ∈ R.

s
(1) Mp,q is a (quasi-) Banach space. Moreover, if 1 6 p, q 6 ∞, then
s
Mp,q is a Banach space.
(2) S (Rn ) ⊂ Mp,q
s
⊂ S 0 (Rn ).
(3) Let 0 < p, q < ∞, then S (Rn ) is dense in Mp,q
s
.
Proof. Analogous to Besov spaces, we can prove the consequence and

the details are omitted.
Proposition 6.2 (Equivalent norm). Let {σk }k∈Zn , {ϕk }k∈Zn ∈ Υ.

s
Then {σk }k∈Zn and {ϕk }k∈Zn generate equivalent norms on Mp,q .
Proof. We have the translation identity,

(F −1 mF f )(x) = eixk F −1 m(· + k)F (e−iky f (y)) (x).
For convenience, we denote
σk := F −1 σk F , ϕ
k := F
−1
ϕk F .
Noticing the almost orthogonality of σk
X
σk = σk ϕ
k+` , (6.14)
|`|∞ 61
we have
X
kσk f kp 6 kϕ σ
k+` (k f )kp .
|`|∞ 61
Using the multiplier estimate

kϕ σ ϕ
k+` (k f )kp . kσk kH s kk+` f kp , s > n(1/(1 ∧ p) − 1/2),
and kσk kH s 6 C, we immediately have
X
kσk f kp . kϕ
k+` f kp .
|`|∞ 61
{σ }
k k {ϕ }
So, kf kMp,q
s 6 kf kMp,q
s .
Proposition 6.2 indicates that one can choose {σk }k∈Zn ∈ Υn according
to our requirement. In applications of PDE, it is convenient to us the
following {σk }k∈Zn ∈ Υn . Let {ηk }k∈Z ∈ Υ1 , we denote
σk (ξ) := ηk1 (ξ1 )...ηkn (ξn ), (6.15)
then we have {σk }k∈Zn ∈ Υn . the above σk (ξ) realizes the separation of
different variables.
Proposition 6.3 (Embedding). Let s1 , s2 ∈ R and 0 < p1 , p2 , q1 , q2 6

∞.
(1) If s2 6 s1 , p1 6 p2 and q1 6 q2 , then Mps11,q1 ⊂ Mps22,q2 .
s1 s2
(2) If q2 < q1 and s1 − s2 > n/q2 − n/q1 , then Mp,q 1
⊂ Mp,q 2
.
Proof. Recall that

X
kk f kp2 6 kF −1 σk F (k+` f )kp2 .
|`|∞ 61
Using Lemma 6.1, we

Xhave X
kk f kp2 . kF −1 σk F (k+` f )kp1 . kk+` f kp1 .
|`|∞ 61 |`|∞ 61
q1 q2
In view of ` ⊂ ` , we can get the conclusion of (1).
Now we prove (2). Using Hölder’s inequality,
!1/q2
X
s q q
kf kMp,q
s2 = hki 2 2
k k f kp2
2
k∈Zn
!(q1 −q2 )/q1 q2
X
(s2 −s1 )q1 q2 /(q1 −q2 )
6 kf kMp,q
s1 hki . (6.16)
1
k∈Zn
Noticing that
X ∞
X
hki(s2 −s1 )q1 q2 /(q1 −q2 ) . hiin−1+(s2 −s1 )q1 q2 /(q1 −q2 ) , (6.17)
k∈Zn i=0
s1 − s2 > n/q2 − n/q1 implies that the right hand side of (6.17) is a con-
vergent series. It follows that (2) holds.
Proposition 6.4 (Dual space). Let s ∈ R and 0 < p, q < ∞. If p > 1,

we denote 1/p + 1/p0 = 1; If 0 1, Proposition 6.4 is similar to that of Besov spaces, however,

if 0 < p < 1, the result is quite different from that of Besov spaces. The
details of the proof of Proposition 6.4 can be found in [246] by following
the proof of the relevant result in Besov spaces.
Remark 6.1. If p, q ∈ [1, ∞], Propositions 6.1 and 6.4 were obtained by
Feichtinger [73]. In [246; 247], the cases 0 < p < 1 and 0 < q < 1 were
considered. The proofs of the results in this section are due to [246; 247].
Soon after the work [247], Kobayashi [148] independently defined Mp,q
for all 0 < p, q 6 ∞ and obtained Proposition 6.1. Almost at the same
time as [246], Kobayashi [149] discussed the dual space of Mp,q and ob-
tained partial results of Proposition 6.4: if 0 < p < 1 or 1 < q < ∞,
he obtained Mp0 ,q0 ⊂ (Mp,q )∗ ⊂ M∞,∞ . For the other cases, he showed
(Mp,q )∗ = Mp0 ,q0 . Recently, by using the molecular decomposition tech-
niques of modulation spaces, Kobayashi and Sawano [150] reconsidered the
s
dual space of Mp,q and they also obtained the result of Proposition 6.4.
It is worth to mention that Triebel [225] introduced a class of generalized
modulation spaces for all indices 0 < p, q 6 ∞, however, those spaces have
no complete norms, which seems harder to use in the study of PDEs.
6.3 Inclusions between Besov and modulation spaces
From the definitions, we see that Besov spaces and modulation spaces are
rather similar, both of them are the combinations of frequency decomposi-
tion operators and function spaces `q (Lp ). In fact, we have the following
inclusion results.
Theorem 6.1 (Embedding). Let 0 < p, q 6 ∞ and s1 , s2 ∈ R. We have

the following results.
s1 s2
(1) Bp,q ⊂ Mp,q if and only if s1 > s2 + τ (p, q), where

1 1 1 1
τ (p, q) = max 0, n − , n + −1 ;
q p q p
s1 s2
(2) Mp,q ⊂ Bp,q if and only if s1 > s2 + σ(p, q), where

1 1 1 1
σ(p, q) = max 0, n − , n 1− − .
p q p q
1
q
p=2
S1
S3
( 12 , 12 )
S2
(0, 0) (1, 0) 1
p
Fig. 6.1 The distribution of τ (p, q) R2+ : τ (p, q) = n( q1 − 1

p
) in S1 ; τ (p, q) = 0 in S2 ;
τ (p, q) = n( p1 + 1
q
− 1) in S3 .
6.3. Inclusions between Besov and modulation spaces 165
1
q
R1
1
q
=
1
p
(0, 1)
1
p
+
1
q
=
1
R2
( 12 , 12 )
R3
(0, 0) (1, 0) 1
p
Fig. 6.2 The distribution of σ(p, q): σ(p, q) = 0 in R1 ; σ(p, q) = n( p1 − 1

q
) in R2 ;
1 1
σ(p, q) = n(1 − p
− q
) in R3 .
The inclusions between Besov and modulation spaces in the cases

(1/p, 1/q) ∈ [0, 1]2 were first discussed by Gröbner [85] and he has never
published his results. When (1/p, 1/q) is in the vertices of the square [0, 1]2 ,
Gröbner’s results are optimal. Afterwards, Toft [222] obtained the suf-
ficiency of Theorem 6.1 in the cases (1/p, 1/q) ∈ [0, 1]2 . Sugimoto and
Tomita [210] showed the necessity of the first inclusion of Theorem 6.1 in
the cases (1/p, 1/q) ∈ [0, 1]2 , and by duality they obtained the second in-
clusion is also sharp if (1/p, 1/q) ∈ [0, 1]2 and p, q 6= ∞. Sugimoto and
Tomita’s idea is to use Feichtinger’s norm and the dilation property of
modulation spaces. In [245; 246; 247] the authors proved the conclusions
of Theorem 6.1 by using frequency-uniform decomposition techniques.
Corollary 6.1. We have the following inclusions.

s+n/2 s s s+n s s
B2,1 ⊂ M2,1 ⊂ B2,1 , B∞,1 ⊂ M∞,1 ⊂ B∞,1 .
The above embedding theorem is of importance for the study of non-

linear PDEs and we give the details of the proof. We can first prove
Theorem 6.1 for some special p, q and then by the interpolation to carry
out the other cases. The following proof is a collection of [85; 247; 246;
245], which is different from those of [210; 222].
Lemma 6.3. Let 0 < p, q 6 ∞. We have

0
Mp,q ⊂ Bp,q , ∀ q 6 p ∧ 1,
n(1/(p∧1)−1/q) 0
Mp,q ⊂ Bp,q , ∀ q > 1 ∧ p.
Proof. First, we show the first inclusion

X∞
kf kqB 0 = k4k f kqp . (6.19)
p,q
k=0
√
Case 1. q < 1 and p > 1. Let ak = max(0, 2k−1 − n) and bk =
√
2k+1 + n. Noticing that for |i| ∈
6 [ak , bk ], 4k i f = 0, we obtain that
 q
X X
k4k f kqp 6  k4k (i f )kp  . ki f kqp .
i∈Zn , |i|∈[ak ,bk ] i∈Zn , |i|∈[ak ,bk ]
(6.20)
From (6.20) we deduce the result.
Case 2. q < 1 and p < 1. From p < 1 and q/p 6 1 it follows that
 q/p
X Z
k4k f kqp 6  |4k (i f )(x)|p dx
i∈Zn , |i|∈[ak ,bk ]
X
6 k4k (i f )kqp . (6.21)
i∈Zn , |i|∈[ak ,bk ]
In view of the multiplier estimate on LpΩ , 4k i : LpB(i,√n) → LpB(i,√n) . By

(6.21) we get the conclusion.
Next, we show the second inclusion. We consider the following two
cases.
Case 1. p > 1. Denote
√
Λ0 = {k ∈ Zn : B(k, n) ∩ {ξ : |ξ| ∈ [0, 2)} 6= ∅}, (6.22)
n
√ j−1 j+1
Λj = {k ∈ Z : B(k, n) ∩ {ξ : |ξ| ∈ [2 , 2 )} 6= ∅}, j > 1. (6.23)
In view of the multiplier estimate on LpΩ ,
X X X
k4j f kp . k4j k+` k f kp . kk f kp . (6.24)
k∈Λj |`|∞ 61 k∈Λj
Since q > p ∧ 1, we have

(a1 + ... + am )q 6 mq−1 (aq1 + ... + aqm ). (6.25)
For k ∈ Λj , one has that |k| ∼ 2j and Λj overlaps at most O(2nj ) cubes,
which follows that
 q
X∞ X
kf kqB 0 .  kk f kp 
p,q
j=0 k∈Λj
∞
X X
. 2jn(q−1) kk f kqp
j=0 k∈Λj
∞
X X
. hkin(q−1) kk f kqp
j=0 k∈Λj
. kf kq n(1−1/q) . (6.26)
Mp,q
Case 2. p < 1. In view of q/p > 1, we have

q/p
(a1 + ... + am )q/p 6 mq/p−1 (a1 + ... + aq/p
m ). (6.27)
By the multiplier estimate on LpΩ ,
 q/p
X XZ
q
k4j f kp .  |4j k+` k f |p dx
k∈Λj `∈Λ Rn
X
jn(q/p−1)
.2 kk f kqp . (6.28)
k∈Λj
Similar to Case 1, we have the conclusion, as desired.
Lemma 6.4. We have

n(1/q−1/2)
B2,q ⊂ M2,q , ∀ 0 < q 6 2.
Proof. By Plancherel’s identity,

X
kf kqM2,q ∼ kχQk fbkq2 . (6.29)
k∈Zn
Denote Λj = {k ∈ Zn : |k| ∈ [2j−1 , 2j )}. Let L 1. One has that

X ∞ X
X
kf kqM2,q . kχQk fbkq2 + kχQk fbkq2 . (6.30)
|k|62L j=L k∈Λj
It is easy to see that Λj has at most O(2nj ) elements. So,

q
X∞ X

q q
kf kM2,q . kf k2 + 2 jn(1−q/2) b
χQk f

j=L k∈Λj
2
∞
X
. kf kq2 + 2jn(1−q/2) kχ|·|∈(2j−2 ,2j+1 ) fbkq2
j=L
. kf kq n(1/q−1/2) , (6.31)
B2,q
which is the result, as desired.
Lemma 6.5. Let 1 6 p, q 6 ∞. We have

σ(p,q) 0 1 1
Mp,q ⊂ Bp,q , σ(p, q) = max 0, n − . (6.32)
p ∧ p0 q
Proof. By the dual versions of Lemmas 6.3 and 6.4, we have
n 0
Mp,∞ ⊂ Bp,∞ , 1 6 p 6 ∞, (6.33)
n(1/2−1/q) 0
M2,q ⊂ B2,q , 2 6 q 6 ∞. (6.34)
Taking p = 1, ∞ and q = ∞, we have
n 0 0 n/2
n 0
M1,∞ ⊂ B1,∞ , M2,∞ ⊂ B2,∞ , M∞,∞ ⊂ B∞,∞ . (6.35)
By the complex interpolation on (6.35) (see Appendix),
0
max(n/p, n/p ) 0
Mp,∞ ⊂ Bp,∞ , 1 6 p 6 ∞. (6.36)
Lemma 6.3 also implies that
0
Mp,1 ⊂ Bp,1 , 1 6 p 6 ∞. (6.37)
Recall that
0
M2,2 = B2,2 . (6.38)
Making a complex interpolation on (6.36)–(6.38), we obtain the result, as
desirted.
When 0 < p < 1, we need the following multiplier estimate (cf. Peetre
[192] and Triebel [224]).
Proposition 6.5. Let Ω ⊂ Rn be a compact set, and 0 0 and 0 0 is independent
of b > 0.
Proof. If f ∈ LpB(0,b) , then f (b−1 ·) ∈ LpB(0,1) . In view of
(F −1 M F f )(x) = [F −1 M (b ·)f\
(b−1 ·)](bx),
it follows from Proposition 6.5 that Corollary 6.2 holds.
Lemma 6.6. Let 0 < p 6 ∞. Then we have

n(1/(p∧1)−1)
Bp,∞ ⊂ Mp,∞ . (6.39)
Proof. First, we consider the case 0 < p < 1. By Corollary 6.2, for any
|k| 1, |k| ∈ [2j−1 , 2j ),

4
X
−1
kk f kp = F σk ϕj+` F f

`=−4 p
4
X
. kσk (2j+5 ·)kB n(1/p−1/2) k4j+` f kp . (6.40)
2,p
`=−4
Using the scaling in Besov spaces (cf. [224])
kg(λ ·)kBp,q
s 6 λs−n/p kgkBp,q
s , λ & 1,
we obtain that6
kσk (2j+5 ·)kB n(1/p−1/2) . 2jn(1/p−1) kσk kB n(1/p−1/2) . 2jn(1/p−1) . (6.41)

2,p 2,p
Inserting (6.41) into (6.40), we immediately have

4
X
jn(1/p−1)
kk f kp . 2 k4j+` f kp . (6.42)
`=−4
By (6.42), we get (6.39).

Now we consider the case p > 1. By Young’s inequality, for |k| 1,

4
X
−1
kk f kp = F σk ϕj+` F f

`=−4 p
4
X 4
X
. kF −1 σk k1 k4j+` f kp . k4j+` f kp . (6.43)
`=−4 `=−4
This implies the result.

6 We can assume that σk = σ0 (· − k) in the definition of s .
Mp,q
Proof. [Proof of Theorem 6.1] (Sufficiency) First, we show that

s+σ(p,q) s
Mp,q ⊂ Bp,q . By Lemma 6.5, we see that the conclusion holds if
1 6 p, q 6 ∞. By Lemma 6.3 we have the result if 0 0} and (see Fig. 6.1),
S1 = {(1/p, 1/q) ∈ R2+ : 1/q > 1/p, 1/p 6 1/2};
S2 = {(1/p, 1/q) ∈ R2+ : 1/q 6 1/p, 1/p + 1/q 6 1};
S3 = R2+ \ (S1 ∪ S2 ).
We first consider the case (1/p, 1/q) ∈ S3 . We have τ (p, q) = n(1/p +
1/q − 1). Take (p0 , q0 ) and (p1 , q1 ) satisfying
1 1 1 1
= + , = 0;
p0 p q q0
1 1 1 1 1 1
= , = + − .
p1 2 q1 p q 2
Let θ = q1 ( p1 + 1
q − 12 )−1 , we have
1 1−θ θ 1 1−θ θ
= + , = + ;
p p0 p1 q q0 q1

1 1 1 1 1
+ − 1 = (1 − θ) −1 + − θ.
p q p0 q1 2
By Lemmas 6.4 and 6.6,
n(1/q1 −1/2)
B2,q1 ⊂ M2,q1 , Bpn(1/p
0 ,∞
0 −1)
⊂ Mp0 ,∞ .
A complex interpolation yields
n(1/p+1/q−1)
Bp,q ⊂ Mp,q .
Secondly, we consider the case (1/p, 1/q) ∈ S1 . If (1/p, 1/q) ∈ Ṡ1 (Ṡ1
denotes the set of all inner points of S1 ), then (1/p, 1/q) can be lying in
the segment by connecting (1/∞, 1/∞) and a point (1/p1 , 1/q1 ) in the line
{(1/p, 1/q) : p = 2, q < 2}. By a complex interpolation, we have
n(1/q−1/p)
Bp,q ⊂ Mp,q . (6.44)
Thirdly, if (1/p, 1/q) ∈ S2 , the dual version of the first inclusion implies
the result, as desired.
(Necessity) We need to show that for any 0 < η 1,
τ (p,q)−η
Bp,q 6⊂ Mp,q . (6.45)
Case 1.1. (1/p, 1/q) ∈ S3 . Let f = F −1 ϕj , j 1. We have

1
X
kf kq τ (p,q)−η = 2q(τ (p,q)−η)(j+`) kF −1 ϕj+` ϕj kqp
Bp,q
`=−1
. 2q(n/q−η)j . (6.46)
Assume without loss of generality that

5 j−1 3 j+1
ϕj (ξ) = 1, if ξ ∈ Dj := ξ : ·2 6 |ξ| 6 · 2 .
4 4
Noticing that
√
Λj = k ∈ Zn : B(k, n) ⊂ Dj
contains at least O(2jn ) elements, we have
X X
kf kqMp,q = kk F −1 ϕj kqp > kF −1 σk ϕj kqp & 2nj . (6.47)
k∈Zn k∈Λj
By (6.46) and (6.47),

kf kMp,q & 2ηj kf kB τ (p,q)−η ,
p,q
which implies that (6.45) holds.

Case 1.2. (1/p, 1/q) ∈ S2 . We consider the case q = ∞. Taking
k(j) = (2j , 0, ..., 0) and f = F −1 σk(j) , we see that
kf kMp,∞ & 1 & 2ηj kf kBp,∞
−η .
If q < ∞, we show that

ε ε
Mp,q 6⊂ Bp,r ∪ B∞,∞ . (6.48)
Let f ∈ S (Rn ) be a Schwartz function satisfying suppfb ⊂ {ξ : |ξi | <
1/2, i = 1, ..., n}. Let N 1, 0 < ε 1,
k(j) = (2N j , 0, ..., 0) ∈ Zn , (6.49)
∞
X
b
F (ξ) = 2−εN j fb(ξ − k(j)). (6.50)
j=1
Noticing that suppFb ⊂ ∪∞

j=1 Qk(j) , we have
k F = 0 if k 6= k(j) + `, |`|∞ 6 1. (6.51)

Since N 1, we see that
∞ X
X
kF kqMp,q 6 k k(j)+` F kqp
j=1 |`|∞ 61
∞ X
X
6 2−εN qj kF −1 σk(j)+` fb(· − k(j))kqp
j=1 |`|∞ 61
∞ X
X
= 2−εN qj k` f kqp . 1, (6.52)
j=1 |`|∞ 61
which leads to F ∈ Mp,q . On the other hand, letting s > ε, we have

∞
X
kF krBp,r
s > 2srj k 4j F krp
j=1
∞
X
& 2srN j k 4N j F krp
j=1
X∞
& 2(s−ε)rN j k F −1 ϕN j χQk(j) fb(· − k(j))krp , (6.53)
j=1
where ϕj = ϕ(2−j ·). We can assume that ϕ(ξ) = 1 if |ξ| ∈ [3/4, 5/4].
Hence,
F −1 ϕN j χQk(j) fb(· − k(j)) = F −1 χQk(j) fb(· − k(j)). (6.54)
By (6.53) and (6.54),
∞
X
kF krBp,r
s & 2(s−ε)rN j = ∞. (6.55)
j=1
s
The above discussion also implies that F 6∈ B∞,∞ .
Case 1.3. (1/p, 1/q) ∈ S1 . We can assume that σk (ξ) = 0 if ξ ∈ 6 Q̃k :=
{ξ : |ξi − ki | 6 5/8, 1 6 i 6 n}, and the dyadic decomposition function
sequence satisfying ϕj (ξ) = 1 if ξ ∈ Dj (= {ξ : 45 · 2j−1 6 |ξ| 6 43 · 2j+1 }).
Let
Aj = {k ∈ Zn : Q̃k ⊂ Dj }, j 1. (6.56)
It is easy to see that Aj has at most O(2 ) elements. Let f ∈ S (Rn ) be
nj
a radial Schwartz function satisfying suppfˆ ⊂ B(0, 1/8),

X
g(x) = eixk (τk f )(x), τk f = f (· − k). (6.57)
k∈Aj
Taking notice of supp τd d

k f ⊂ B(0, 1/8), we see that supp τk (τk f )∩supp σ` =
∅, k 6= `. So,
 1/q
X
kgkMp,q >  kF −1 σk F gkqp 
k∈Aj
 1/q
X
= kF −1 σ0 (τd
k f )kp 
q
& 2jn/q . (6.58)
k∈Aj
On the other hand, supp ĝ ⊂ {ξ : 2j−1 6 |ξ| 6 2j+1 }. Hence,

1
!1/q
X
nj(1−q/p)−ηjq −1 q
kgkB n(1/q−1/p)−η 6 2 kF ϕj+` F gkp . (6.59)
p,q
`=−1
Using the multiplier estimates and Hölder’s inequality, we have

2/p
kF −1 ϕj+` F gkp . kgkp 6 kgk1−2/p
∞ kgk2 . (6.60)
By Plancherel’s identity,
 1/2
Z X
kgk2 = kĝk2 =  |τk (e−ikξ fˆ(ξ))|2 dξ  . 2nj/2 . (6.61)
Rn k∈A
j
We can further assume that f (x) = f (|x|) is a decreasing function on |x|.

In view of f ∈ S (Rn ), we have
|f (x − k)| . (1 + |x − k|)−N , N 1. (6.62)
Denote
B0 = {k ∈ Aj : |x − k| 6 2}, Bi = {k ∈ Aj : 2i < |x − k| 6 2i+1 }.
Bi contains at most O(2n i ) elements. It follows from (6.62) that
XX X X
|g(x)| 6 |f (x − k)| . f (0) + 2ni |f (2i )| . 2(n−N )i . 1.
i>0 k∈Bi i>1 i>0
(6.63)
Collecting (6.60), (6.61) and (6.63), we have
kF −1 ϕj+` F gkp . 2nj/p . (6.64)
Inserting (6.64) into (6.59) and using (6.58), we immediately have
kgkB n(1/q−1/p)−η . 2nj/q−ηj . 2−ηj kgkMp,q . (6.65)
p,q
This implies the conclusion.

Now we show the necessity of the second inclusion. It suffices to show
σ(p,q)−η 0
Mp,q 6⊂ Bp,q , ∀η > 0. (6.66)
Notice that (see Fig. 6.2)
R1 = {(1/p, 1/q) ∈ R2+ : 1/q > 1/p, 1/p + 1/q > 1},
R2 = {(1/p, 1/q) ∈ R2+ : 1/q 6 1/p, 1/p > 1/2},

R3 = R2+ \ (R1 ∪ R2 ).
We consider the following three cases.
Case 2.1. (1/p, 1/q) ∈ R1 . We have discussed it in Case 1.1.
Case 2.2. (1/p, 1/q) ∈ R3 . Assume that there exists an η > 0 such that
n(1−1/p−1/q)−η 0
Mp,q ⊂ Bp,q , we will get a contradiction. If 1 6 p, q < ∞, by
−n(1/p0 +1/q0 −1)+η
duality, Bp00 ,q0 ⊂ Mp0 ,q0 , which contradicts the first inclusion.
If p = ∞ or q = ∞, one can use the same way as in Case 2.1 to get the
result. In fact, letting f = F −1 ϕj , we have
kf kBp,q
0 > kF −1 ϕj ϕj kp & 2nj(1−1/p) . (6.67)
On the other hand,
kf kM n/p0 6 suphkin(1−1/p) kF −1 σk ϕj kp . 2nj(1−1/p) , (6.68)
p,∞
k
 1/q
X
kf kM n/q0 6  2nj(q−1) kF −1 σk ϕj kq∞  . 2nj . (6.69)
∞,q
|k|∈[2j−1 ,2j+1 ]
From (6.67)–(6.69) it follows that (6.66) holds in the cases p = ∞ or q = ∞.

Case 2.3. (1/p, 1/q) ∈ R2 . Take f ∈ S (Rn ) satisfying f (0) = 1,
supp fˆ ⊂ Q0 . Choose 0 < a 1 (which will be fixed in (6.75) below).
Denote fa (x) = f (x/a). We easily see that supp fâ ⊂ Q0,a := {ξ : |ξi | 6
1/2a, 1 6 i 6 n}. Recall that Dj = {ξ : 45 · 2j−1 6 |ξ| 6 43 · 2j+1 }
contains at most O(an 2jn ) pairwise disjoint cubes Qk(i),a := k(i) + Q0,a ,
i = 1, ..., O(an 2jn ). We write Aj = {k(i) : i = 1, ..., O(an 2jn )},
X
g(x) = eixk (τk fa )(x). (6.70)
k∈Aj
For any N 1,
|f (x)| 6 CN (1 + |x|)−N , (6.71)
it follows that
|fa (x)| 6 CN aN |x|−N . (6.72)
By the continuity of f (x) and f (0) = 1, we deduce that there exists a % > 0
such that7
|fa (x)| > 1/2, x ∈ B(0, %). (6.73)
7 In fact, % can be chosen as % = a%0 , %0 > 0 is independent of a.
Xx ∈ B(k(i), %), that

In view of (6.70) and (6.73), we get, for any
|g(x)| > |fa (x − k(i))| − |fa (x − k)|
k∈Aj \{k(i)}
1 X
> − |fa (x − k)|. (6.74)
2
k∈Aj \{k(i)}
Denote Aj,` := {k ∈ Aj : 2 6 |k − k(i)| < 2`+1 }. We can further assume
`
that f (x) is a decreasing function on |x|. Since Aj,` has at most O(an 2`n )
elements, weX have for any x ∈ B(k(i),
X X %),
|fa (x − k)| 6 |fa (x − k)|
k∈Aj \{k(i)} `>1 k∈Aj,`
X
6C an 2n` |fa (2` − %)|
`>1
X
. CN an+N 2(n−N )` 6 1/4, (6.75)
`>1
where N > n + 1 and CCN an+N 6 1/4. Hence, it follows from (6.74) and
(6.75) that
|g(x)| > 1/4, x ∈ B(k(i), %). (6.76)
By (6.76), we have n nj
O(aX 2 )
1
kg(x)kp >
χ
B(k(i),%) & (a%)
n/p nj/p
2 , (6.77)
i=1 4
p
where % and a are independent of j 1. We can assume that ϕj (ξ) = 1 for
ξ ∈ Dj . Since supp ĝ ⊂ Dj , we have F −1 ϕj F g = g. Hence, (6.77) implies
that
kgkBp,q
0 > kF −1 ϕj F gkp & (a%)n/p 2nj/p . (6.78)
On the other hand,
 1/q
X
kgk n(1/p−1/q) = 
Mp,q
2nj(q/p−1) kF −1 σk F gkqp 
|k|∈[2j−1 ,2j+1 ]
nj/p
62 sup kF −1 σk F gkp . (6.79)
|k|∈[2j−1 ,2j+1 ]
Since supp σk overlaps at most finite many supp τ` (τd
` fa ), using multiplier
estimates, we have
kF −1 σk F gkp . kf kp . (6.80)
Hence, (6.79) and (6.80) imply that
kgkM n(1/p−1/q) . 2nj/p . (6.81)
p,q
By (6.78) and (6.81) we immediately have (6.66). We finish the proof of
Theorem 6.1.
6.4 NLS and NLKG in modulation spaces
As indicated in §6.1, the dispersive semi-group combined with the

frequency-uniform decomposition operator has some advantages and we
discuss them in this section. The results of this section can be found in [11;
12; 50; 247; 246].
6.4.1 Schrödinger and Klein-Gordon semigroup in modula-

tion spaces
Let S(t) = eit4 denote the Schrödinger semi-group. In [247], Wang, Zhao
and Guo obtained the uniform boundedness for the Ginzburg-Landau semi-
group L(t) = e(a+i)t4 (a > 0) in modulation spaces and their proof is also
adapted to the Schrödinger semi-group (a = 0 in L(t)). First, we show that
k S(t) : Lp → Lp is uniformly bounded on k ∈ Zn by following the proof
in [247].
X 2
kk S(t)f kp 6 kF −1 σk+` eit|ξ| σk fˆkp
|`|∞ 61
X 2
6 kF −1 (σk+` eit|ξ| )k1 kk f kp . (6.82)
|`|∞ 61
2
Hence, it suffices to estimate kF −1 (σk eit|ξ| )k1 . Using the multiplier esti-
mate, we have
2 2
kF −1 (σk eit|ξ| )k1 = kF −1 (σ0 eit|ξ| )k1
1−n/2L
X 2 n/2L
. kσ0 k2 kDα (σ0 eit|ξ| )k2
|α|=L
n/2
. (1 + |t| ). (6.83)
Noticing that
kk S(t)f k2 = kk f k2 , (6.84)
by a complex interpolation, we immediately have
kk S(t)f kp . (1 + |t|)n|1/2−1/p| kk f kp . (6.85)

s
Proposition 6.6 (Uniform boundedness of S(t) in Mp,q ). Let s ∈
R, 1 6 p 6 ∞ and 0 < q < ∞. Then we have
kS(t)f kMp,q
s . (1 + |t|)n|1/2−1/p| kf kMp,q
s . (6.86)
6.4. NLS and NLKG in modulation spaces 177
Shortly after the work [247], Proposition 6.6 is independently obtained

by Bényi, Gröchenig, Okoudjou and Rogers in [12] and their result contains
α
more general semi-group eit(−∆) with α 6 1, whose proof is based on the
short-time frequency analysis technique. Miyachi, Nicola, Riveti, Taracco
and Tomita [164] were able to consider the case α > 1, Chen, Fan and Sun
α
[38] obtained some refined estimates for eit(−∆) with any α > 0 by using
the oscillatory integral estimates in higher spatial dimensions.
Now we consider the truncated decay of S(t); cf. [246]. Let us recall
0
the Lp − Lp estimate of S(t),
kS(t)f kp . |t|−n(1/2−1/p) kf kp0 , 2 6 p 6 ∞, (6.87)
where 1/p + 1/p0 = 1. We also have
X
k k S(t)f kp . k k+` f kp0 , 2 6 p 6 ∞. (6.88)
|`|∞ 61
Indeed, by Young’s and Hölder’s inequalities, we have

X
kk S(t)f kp 6 kF −1 σk σk+` exp(−it|ξ|2 )F f kp .
`∈Λ
X
6 kσk+` exp(−it|ξ|2 )σk F f kp0 . kk f kp .
`∈Λ
Combing (6.87) with (6.88), one has that

kk S(t)f kp . (1 + |t|)−n(1/2−1/p) kk f kp0 , 2 6 p 6 ∞. (6.89)
Multiplying both sides of (6.89) by hkis and then taking `q norm, we have
kS(t)f kMp,q
s . (1 + |t|)−n(1/2−1/p) kf kMps0 ,q .
Proposition 6.7. Let s ∈ R, 2 6 p < ∞, 1/p + 1/p0 = 1 and 0 < q < ∞.

Then we have
kS(t)f kMp,q
s . (1 + |t|)−n(1/2−1/p) kf kMps0 ,q . (6.90)
Propositions 6.6 and 6.7 are optimal in the sense that the powers of time
variable are sharp, cf. [51]. Now we consider the truncated decay estimate
1/2
for the Klein-Gordon semi-group G(t) = eitω where ω = I − ∆.
X
kk G(t)f kp 6 kF −1 σk+` eithξi σk fˆkp
|`|∞ 61
X
6 kF −1 (σk+` eithξi )k1 kk f kp . (6.91)
|`|∞ 61
So, it suffices to estimate kF −1 (σk eithξi )k1 . Using the Fourier multiplier
estimate, we have
kF −1 (σk eithξi )k1 = kF −1 (σ0 eithξ+ki )k1
1−n/2L
X n/2L
. kσ0 k2 kDα (σ0 eithξ+ki )k2
|α|=L
n/2
. (1 + |t| ). (6.92)
Noticing that
kk G(t)f k2 = kk f k2 , (6.93)
by a complex interpolation, we immediately have
kk G(t)f kp . (1 + |t|)n|1/2−1/p| kk f kp . (6.94)
Proposition 6.8. Let s ∈ R, 1 6 p 6 ∞ and 0 < q < ∞. Then we have

kG(t)f kMp,q
s . (1 + |t|)n|1/2−1/p| kf kMp,q
s . (6.95)
0
It is known that G(t) satisfies the following Lp − Lp estimate
kG(t)f kH −2σ(p) . |t|−n(1/2−1/p) kf kp0 , (6.96)
p
where

1 1
2 6 p < ∞, 2σ(p) = (n + 2) − . (6.97)
2 p
From (6.96) it follows that
k k G(t)f kH −2σ(p) . |t|−n(1/2−1/p) k k f kp0 . (6.98)
p
Applying the multiplier estimate,

k k (I − 4)δ/2 gkp . hkiδ kgkp . (6.99)
In view of (6.98) and (6.99), we have
X
k k G(t)f kp . hki2σ(p) k k+` G(t)f kH −2σ(p)
p
`∈Λ
X
. hki2σ(p) |t|−n(1/2−1/p) k k+` f kp0 . (6.100)
`∈Λ
On the other hand, by Hölder’s and Young’s inequality,

2 1/2
k k G(t)f kp . kσk eit(1+|ξ| ) fbkp0
X 2 1/2
. kσk eit(1+|ξ| ) F k+` f kp0
`∈Λ
X
. kF k+` f kp
`∈Λ
X
. k k+` f kp0 . (6.101)
`∈Λ
So, for any θ ∈ [0, 1], it follows from (6.100) and (6.101) that
X
k k G(t)f kp . hki2σ(p)θ |t|−nθ(1/2−1/p) k k+` f kp0 . (6.102)
`∈Λ
Noticing that σ(p) > 0, by (6.101) we have

X
k k G(t)f kp . hki2σ(p)θ k k+` f kp0 . (6.103)
`∈Λ
Combining (6.102) and (6.103), we get

X
k k G(t)f kp . hki2σ(p)θ (1 + |t|)−nθ(1/2−1/p) k k+` f kp0 . (6.104)
`∈Λ
Multiplying both sides of (6.104) by hkis and then taking `q norm, we

immediately obtain that
Proposition 6.9. Let s ∈ R, 2 6 p < ∞, 1/p + 1/p0 = 1, 0 < q < ∞,

θ ∈ [0, 1] and σ(p) is as in (6.97). Then we have
kG(t)f kMp,q
s . (1 + |t|)−nθ(1/2−1/p) kf kM s+2σ(p)θ . (6.105)
p0 ,q
6.4.2 Strichartz estimates in modulation spaces

For convenience, we write
!1/q
X
kf k`s,q (Lγ (I,Lp )) = hkisq
kk f kqLγ (I,Lp ) , (6.106)

k∈Zn
`q (Lγ (I, Lp )) := `0,q γ p q p q p p

(L (I, L )), ` (Lx,t∈I ) := ` (L (I, L )). Recall that
the truncated decay can be generalized to the following estimate
kU (t)f kMp,q
α . (1 + |t|)−δ kf kMp0 ,q , (6.107)
where 2 6 p < ∞, 1 6 q < ∞, α = α(p) ∈ R, δ = δ(p) > 0, α and δ are
independent of t ∈ R, U (t) is a dispersive semi-group,
U (t) = F −1 eitP (ξ) F , (6.108)
and P (·) : Rn → R is a real valued function. In the sequel we will assume
that U (t) satisfies conditions (6.107) and (6.108), from which we can get
some Strichartz inequalities for U (t) in modulation spaces.
Generally speaking, all of the Strichartz estimates in Sec. 3.2 hold if we

replace X ∗ and X α by Mp0 ,2 and Mp,2 α
in Sec. 3.2, respectively. However,
do not forget that the estimate in (6.107) contains no singularity at t = 0,
which leads that we can remove the restriction condition δ 6 1 in the
Strichartz estimates on modulation spaces and the results are much better
than those in Sec. 3.2.
Proposition 6.10 (Strichartz inequalities). Let U (t) satisfy (6.107)

and (6.108). For any γ > 2 ∨ (2/δ), we have
kU (t)f k`α/2,q (Lγ (R,Lp )) . kf kM2,q . (6.109)

In addition, if γ > q, then we have
kU (t)f kLγ (R,M α/2 ) . kf kM2,q . (6.110)

p,q
Proof. The proof is similar to that as in Sec. 3.2. However, we need

to carefully handle the indices α, p, q and γ. First, we consider the case
1 < q < ∞ to show that
Z
(U (t)f, ψ(t))dt . kf kM2,q kψk`−α/2,q0 (Lγ 0 (R,Lp0 )) (6.111)
R
holds for all f ∈S (R ) and ψ ∈ C0∞ (R, S (Rn )). Noticing that S (Rn ) and
n
−α/2,q0 0 0
C0∞ (R, S (Rn )) are dense in M2,q and ` (Lγ (R, Lp )), respectively,
(6.111) implies (6.109). By duality,
Z Z

(U (t)f, ψ(t))dt . kf kM2,q
U (−t)ψ(t)dt
. (6.112)
R R M2,q0
For any k ∈ Zn ,
Z 2

k U (−t)ψ(t)dt

R
2 Z

. k k ψkLγ 0 (R,Lp0 )
k U (t − s)ψ(s)ds
. (6.113)
R Lγ (R,Lp )
Recall that {k }k∈Zn are almost orthogonal. By (6.107), the definition of
k · kMp0 ,q together with the multiplier estimate, we have
X
k k U (t)f kp . hti−δ hki−α k k k+` f kMp0 ,q
|`|∞ 61
−δ −α
. hti hki k k f kp0 . (6.114)
If δ 6= 1, applying (6.114), Young’s and Hardy-Littlewood-Sobolev’s in-

equalities, we obtain that8
Z

k U (t − s)ψ(s)ds . hki−α k k ψkLγ 0 (R,Lp0 ) . (6.115)
γ
R p L (R,L )
If δ = 1 and γ > 2, we can use Young’s inequality to get that (6.115) holds.
If γ = 2 and δ = 1, applying the same way as in Sec. 3.3, we can obtain
(6.115). Hence, in view of (6.113) and (6.115), we have
Z

k U (−t)ψ(t)dt −α/2
. hki k k ψkLγ 0 (R,Lp0 ) . (6.116)
R 2
0
Both sides in (6.116) are taken the norm in `q ,
Z

U (−s)ψ(s)ds . kψk`−α/2,q0 (Lγ 0 (R,Lp0 )) . (6.117)

R M2,q0
(6.112) and (6.117) imply (6.111).

If γ > q, by Minkowski’s inequality, one sees that the left hand side of
(6.110) can be bounded by the left hand side of (6.109).
Now we consider the case q = 1. It suffices to show that
Z
(U (t)f, ψ(t))dt . kf kM2,q kψkc−α/2(Lγ 0 (R,Lp0 )) (6.118)
R
n
holds for all f ∈ S (R ) and ψ ∈ C0∞ (R, S (Rn )). Repeating the above
procedure, we can prove our result.
Denote
Z t
(U f )(t) = U (t − s)f (s, ·)ds. (6.119)
0
Proposition 6.11. Let U (t) satisfy (6.107) and (6.108). For any γ >
2 ∨ (2/δ), we have
kU f k`q (L∞ (R,L2 )) . kf k`−α/2,q (Lγ 0 (R,Lp0 )) . (6.120)

In addition, if γ 0 6 q, then
kU f kL∞ (R,M2,q ) . kf kLγ 0 (R,M −α/2 ) . (6.121)
p0 ,q
Proof. We only sketch the proof. Applying the same way as in (6.113),
(6.115) and (6.116), we see that
kk U f k22 . hki−α k k f k2Lγ 0 (R,Lp0 ) . (6.122)
(6.122) implies (6.120). By Minkowski’ s inequality, it follows from (6.120)
that (6.121) holds.
8 Due to ht − si−δ < |t − s|−δ , if δ < 1, γ = 2/δ, one can apply Hardy-Littlewood-
Sobolev’s inequality.
Proposition 6.12. Let U (t) satisfy (6.107) and (6.108). For any γ >
2 ∨ (2/δ), we have
kU f k`α/2,q (Lγ (R,Lp )) . kf k`−α/2,q (Lγ 0 (R,Lp0 )) . (6.123)

In addition, we assume that q 6 2 if δ = 1 and γ = 2, then we have for any

γ > 2 ∨ (2/δ),
kU f kLγ (R,M α/2 ) . kf kLγ 0 (R,M −α/2 ) . (6.124)
p,q p0 ,q
Proof. We only give an outline of the proof of (6.124). By (6.107), we

have
Z t
kU f kM α/2 . ht − si−δ kf (s)kM −α/2 ds. (6.125)
p,q p0 ,q
0
If δ 6= 1, or δ = 1 and γ > 2, then we can use the same way as in (6.115)
to get the conclusion. If δ = 1 and γ = 2, it follows from (6.115) and
Minkowski’s inequality that (6.124) holds.
Proposition 6.13. Assume that U (t) satisfies (6.107) and (6.108), γ >
max(2/δ, 2). Then we have
kU f k`α/2,q (Lγ (R,Lp )) . kf k`q (L1 (R,L2 )) . (6.126)

In addition, if γ > q, then

kU f kLγ (R,M α/2 ) . kf kL1 (R,M2,q ) . (6.127)
p,q
Proof. Assume that f, ψ ∈ C0∞ (R, S (Rn )). By Proposition 6.11, we

have Z Z
t

U (t − τ )f (τ )dτ, ψ(t) dt
R+ 0
Z ∞

. kf kL1 (R,M2,q )
U (· − t)ψ(t)dt

· L∞ (R,M2,q0 )
. kf kL1 (R,M2,q) kψk`−α/2,q0 (Lγ 0 (R,Lp0 )) . (6.128)

0 0 0
−α/2,q
Since ψ ∈ C0∞ (R, S (Rn )) is dense in ` (Lγ (R, Lp )) and
−α/2 γ0 p0
c (L (R, L )), by duality, we get the result, as desired.
The Schrödinger semi-group corresponds to the cases α = 0, δ =
n(1/2 − 1/p) and 2 6 p < ∞. Taking q = 1 in Propositions 6.10–6.12,
we immediately have
Corollary 6.3. Let 2 6 p < ∞, γ > 2 ∨ γ(p), and
2 1 1
=n − . (6.129)
γ(p) 2 p
Rt
Let S(t) = eit∆ , A = 0 S(t − s) · ds. Then
kS(t)ϕk`1 (Lγ (R,Lp )) . kϕkM2,1 , (6.130)

kA f k`1 (Lγ (R,Lp ))∩`1 (L∞ (R,L2 )) . kf k`1 (Lγ 0 (R,Lp0 )) . (6.131)

Similar to Corollary 6.3, we have

Corollary 6.4. Let 2 6 p < ∞, θ ∈ (0, 1], 1 6 q < ∞,
2 1 1 1 1
= nθ − , 2σ = (n + 2)θ − . (6.132)
γθ (p) 2 p 2 p
Rt
Let G(t) be as in (6.96), G = 0 G(t − s) · ds. Then for any γ > 2 ∨ γθ (p),
we have
kG(t)ϕk`−σ,q (Lγ (R,Lp )) . kϕkM2,q , (6.133)

kG f k`−σ,q (Lγ (R,Lp ))∩`q (L∞ (R,L2 )) . kf k`σ,q (Lγ 0 (R,Lp0 )) . (6.134)

Related Strichartz estimates in Wiener amalgam spaces for the Schrödinger

equation were obtained by Cordero and Nicola [49].
6.4.3 Wellposedness for NLS and NLKG

We consider the Cauchy problem for NLS,
iut + ∆u = f (u), u(0, x) = u0 (x). (6.135)
n 0 ∞
Noticing that B∞,1 ⊂ M∞,1 ⊂ B∞,1 ⊂ L are sharp embeddings, up to
now, we can not get the wellposedness of NLS in L∞ or in B∞,1
0
. However,
we can obtain the local wellposedness of NLS in M∞,1 . We have (see [11;
50])
Theorem 6.2. Let n > 1, f (u) = λ|u|κ u, κ ∈ 2N, λ ∈ R, u0 ∈ Mp,1
and 1 6 p 6 ∞. Then there exists a T > 0 such that (6.135) has
a unique solution u ∈ C([0, T ), Mp,1 ). Moreover, if T < ∞, then
lim supt%T ku(t)kMp,1 = ∞.
2
If the nonlinearity has an exponential growth, say f (u) = λ(e|u| − 1)u,
n/2 0
the result in Theorem 6.2 also holds. Noticing that B2,1 ⊂ M2,1 ⊂ B2,1 ∩
n
C(R ) are sharp embeddings, we can get that NLS is global wellposed in
M2,1 if initial data are sufficiently small.
Theorem 6.3. Let n > 1, f (u) = λ|u|κ u, κ ∈ 2N, λ ∈ R, κ > 4/n,
u0 ∈ M2,1 and there exists a sufficiently small δ > 0 such that ku0 kM2,1 6 δ.
Then (6.135) has a unique solution
u ∈ C(R, M2,1 ) ∩ `1 (Lpx,t∈R ), (6.136)
where p ∈ [2 + 4/n, 2 + κ] ∩ N, `1 (Lpx,t∈R ) is as in (6.106).

2
Theorem 6.4. Let n > 2, f (u) = λ(e%|u| − 1)u, λ ∈ C and % > 0.
Assume that u0 ∈ M2,1 and there exists a sufficiently small δ > 0 such that
ku0 kM2,1 6 δ. Then (6.135) has a unique solution
u ∈ C(R, M2,1 ) ∩ `1 (L4x,t∈R ). (6.137)
The proof of Theorem 6.2 relies upon the algebra structure of Mp,1 .
Lemma 6.7. Let 1 6 p 6 ∞, then Mp,1 is a Banach algebra.
The proof of Lemma 6.7 can be shown by following the same technique as
s
that of E2,1 and we omit the details.
Proof. [Proof of Theorem 6.2] Denote

Z t
A f (t, x) = S(t − τ )f (τ, x)dτ.
0
Consider the mapping
T : u(t) → S(t)u0 − iA f (u). (6.138)
We write
D = {u : kukC([0,T ]:Mp,1) 6 M }, d(u, v) = ku − vkC([0,T ]:Mp,1 ) ,
where M = 2Cku0 kMp,1 . If u ∈ D, by Proposition 6.6 and Lemma 6.7,

kT ukC([0,T ]:Mp,1) . (1 + T )n/2 ku0 kMp,1 + T kukκ+1
C([0,T ]:Mp,1) . (6.139)
We can choose a sufficiently small 0 < T < 1 such that CT M κ 6 1/2. It

follows that T : (D, d) → (D, d) is a contraction mapping. The left part of
the proof is standard and we omit the details of the proof.
Lemma 6.8. Let 1 6 p, pi , γ, γi 6 ∞ satisfy

1 1 1 1 1 1
= + ... + , = + ... + . (6.140)
p p1 pN γ γ1 γN
Then we have
N
Y
ku1 u2 ...uN k`1 (Lγ (R,Lp )) 6 C N kui k`1 (Lγi (R,Lpi )) . (6.141)

i=1
Proof. It suffices to consider the case N = 2. We have

X
k k (u1 u2 )kp 6 kk (i u1 j u2 )kp . (6.142)
i,j∈Zn
Noticing that k (i u1 j u2 ) = 0 if |k−i−j| > k0 (n), where k0 (n) depends

only on n, we have from (6.142) that
X
k k (u1 u2 )kp 6 kk (i u1 j u2 )kp χ|k−i−j|6k0 (n) . (6.143)
i,j∈Zn
Applying Bernstein’s estimate and Hölder’s inequality, by (6.143) we have

X
k k (u1 u2 )kp 6 k i u1 kp1 k j u2 kp2 χ|k−i−j|6k0 (n) . (6.144)
i,j∈Zn
So, by (6.144), Hölder’s and Minkowski’s inequalities, we get that

ku1 u2 k`1 (Lγ (R,Lp ))

 1/γ
X Z X γ
.  k i u1 (t)kp1 k j u2 (t)kp2 χ|k−i−j|6k0 (n) dt
k∈Z R i,j∈Zn
X X
. k i u1 (t)kLγ1 (R,Lp1 ) k j u2 (t)kLγ2 (R,Lp2 ) χ|k−i−j|6k0 (n) .
k∈Z i,j∈Zn
(6.145)
By Young’s inequality, it follows from (6.145) that
ku1 u2 k`1 (Lγ (R,Lp )) . ku1 k`1 (Lγ1 (R,Lp1 )) ku2 k`1 (Lγ2 (R,Lp2 )) . (6.146)

By the induction and (6.146), we can get the result, as desired.

Taking p ∈ N, p ∈ [2 + 4/n, 2 + κ] and
X = `1 (L∞ (R, L2 )) ∩ `1 (Lp (R, Lp )), (6.147)
we easily see that

2 1 1 n(p − 2)
6n − = (6.148)
p 2 p 2p
and “ = ” in (6.148) holds if and only if p = 2 + 4/n. Using Corollary 6.3,
we have
kS(t)u0 kX . ku0 kM2,1 , (6.149)
kA f kX . kf (u)k`1 (Lp0 . (6.150)
x,t∈R )
Proof. [Proof of Theorem 6.3] Let X be as in (6.147) and T be as in

(6.138). By (6.149) and (6.150),
kT ukX . ku0 kM2,1 + kf (u)k`1 (Lp0 . (6.151)
x,t∈R )
Since p ∈ [2 + 4/n, 2 + κ], we have

1 p−1 κ+2−p
= + . (6.152)
p0 p ∞
By Lemma 6.8,

π(u1+κ ) 1 p0 . kukp−1 kuk2+κ−p . (6.153)
` (L x,t∈R )
`1 (Lp
x,t∈R ) `1 (L∞ x,t∈R )
Since
k i uk∞ . k i uk2 , i ∈ Zn , (6.154)
in view of (6.153) and (6.154), we have

π(u1+κ ) 1 p0 . kuk1+κ
` (Lx,t∈R ) X . (6.155)
It follows from (6.151) and (6.155) that

1+κ
kT ukX . ku0 kM2,1 + kukX . (6.156)
Putting
D = {u : kukX 6 M }, d(u, v) = ku − vkX , (6.157)
we easily see that, if M > 0 is sufficiently small, ku0 kM2,1 . M/2, then
T : (D, d) → (D, d) is a contraction mapping, which leads that (6.135) has
a solution u ∈ X. The left part of the proof is standard and we omit the
details.
The proof of Theorem 6.4 is similar to that of Theorem 6.3. Let
Y = `1 (L∞ (R, L2 )) ∩ `1 (L4x,t∈R ). (6.158)
Then we have
∞
X %k
kT ukY . ku0 kM2,1 + k|u|2k uk`1 (L4/3 . (6.159)
k! x,t∈R )
k=1
Using Lemma 6.8, one has that

1+2k
u 1 4/3 . C 2k+1 kuk31 4 kuk2k−2 . C 2k+1 kuk2k+1 .
` (L

)
x,t∈R
` (L x,t∈R ) `1 (L∞
x,t∈R )
Y
(6.160)
6.5. Derivative nonlinear Schrödinger equations 187
Hence,
∞
X C 2k+1
kT ukY . ku0 kM2,1 + kuk2k+1
Y . (6.161)
k!
k=1
Applying (6.161) and combining the proof of Theorem 6.3, we can get the
conclusion of Theorem 6.4.
We now consider the initial value problem for NLKG,
utt + (I − ∆)u + f (u) = 0, u(0) = u0 , ut (0) = u1 . (6.162)
Analogous to NLS, we have
Theorem 6.5. Let n > 1, f (u) = λ|u|κ u, κ ∈ 2N, λ ∈ R, (u0 , u1 ) ∈

−1
Mp,1 × Mp,1 and 1 6 p 6 ∞. Then there exists a T > 0 such that (6.135)
−1
has a unique solution (u, ut ) ∈ C([0, T ), Mp,1 ) × C([0, T ), Mp,1 ). Moreover,
if T < ∞, then lim supt%T (ku(t)kMp,1 + kut (t)kM −1 ) = ∞.
p,1
If the nonlinearity has an exponential growth, the corresponding results

as in Theorem 6.5 also hold.
Theorem 6.6. Let n > 1, f (u) = λu1+κ , κ ∈ N and κ > 4/n. Put
n+2
σ= . (6.163)
n(2 + κ)
σ σ−1
Assume that (u0 , u1 ) ∈ M2,1 × M2,1 and there exists a sufficiently small
δ > 0 such that ku0 kM2,1 + ku1 kM σ−1 6 δ. Then (6.162) has a unique
σ
2,1
solution
σ
u ∈ C(R, M2,1 ) ∩ `1 (L2+κ
x,t∈R ). (6.164)
Theorem 6.7. Let n > 2, f (u) = sinh u − u and σ = (n + 2)/4n. Assume

σ σ−1
that (u0 , u1 ) ∈ M2,1 × M2,1 and there exists a sufficiently small δ > 0 such
that ku0 kM2,1σ + ku1 k
M σ−1 6 δ. Then (6.162) has a unique solution
2,1
σ
u ∈ C(R, M2,1 ) ∩ `1 (L4x,t∈R ). (6.165)
6.5 Derivative nonlinear Schrödinger equations
We study the initial value problem for the derivative nonlinear Schrödinger
equation (gDNLS)
iut + ∆± u = F (u, ū, ∇u, ∇ū), u(0, x) = u0 (x), (6.166)
where u is a complex valued function of (t, x) ∈ R × Rn ,

n
X
∆± u = εi ∂x2i , εi ∈ {1, −1}, i = 1, ..., n, (6.167)
i=1
∇ = (∂x1 , ..., ∂xn ), F : C2n+2 → C is a series of z ∈ C2n+2 ,

X
F (z) = F (z1 , ..., z2n+2 ) = cβ z β , cβ ∈ C, (6.168)
m+16|β|<∞
2 6 m < ∞, m ∈ N, supβ |cβ | < ∞9 . The typical nonlinear term is
F (u, ū, ∇u, ∇ū) = |u|2~λ · ∇u + u2 ~µ · ∇ū + |u|2 u,

see [48; 63; 227]. Another model is
∞
X
F (u, ū, ∇u, ∇ū) = (1 + |u|2 )−1 (∇u)2 ū = (−1)k |u|2k (∇u)2 ū, |u| < 1,
k=0
which is an equivalent version of the Schrödinger flow [61; 88; 114;

261]. The non-elliptic gDNLS arises in the strongly interacting many-
body systems near the criticality, where anisotropic interactions are mani-
fested by the presence of the non-elliptic case, as well as additional resid-
ual terms which involve cross derivatives of the independent variables [48;
63; 227]. Some water wave and completely integrable system mod-
els in higher spatial dimensions are also non-elliptic, cf. [1; 152; 258;
259]. A large amount of work has been devoted to the study of gDNLS,
see [100; 101; 114; 125; 126; 130; 133; 135; 143; 147; 107; 186; 209;
199].
Since the nonlinearity in gDNLS contains derivative terms and the
Strichartz inequalities can not absorb any derivatives, gDNLS can not be
solved if we use only the Strichartz estimate. One needs to look for some
other ways to handle the derivative terms in the nonlinearity. Up to now,
three kinds of methods seem to be very useful for gDNLS. One is to use
the energy estimate to deal with the derivatives in the nonlinearity, the sec-
ond way is to use Bourgain’s space X s,b and the third technique is Kato’s
smooth effect estimates. Of course, there are some connections between
these methods.
We will use the smooth effect estimates together with the frequency-
uniform decomposition techniques to study gDNLS and we show that it is
globally wellposed and scattering in a class of modulation spaces.
9 In fact, we only need the condition |cβ | 6 C |β| .
For convenience to the readers, we sketch the ideas. Denote

P
Z t
it∆± −1 it n εj ξj2
S(t) = e =F e j=1 F , A f (t, x) = S(t − τ )f (τ, x)dτ.
0
For the sake of brevity, we simplify the nonlinearity as F (u, ū, ∇u, ∇ū) =
~λ·∇(|u|2 u). According to the equivalent integral form of gDNLS, one needs
to solve
u(t) = S(t)u0 − iA ~λ · ∇(|u|2 u).
On the basis of the smooth effect of the Schrödinger semi-group in 1D, we
easily get its smooth effect estimates in all dimensions,

1/2
Dxi S(t)u0 ∞ 2 2
. ku0 k2 , (6.169)
1+n Lx L(x Lt (R )
i j )j6=i
k∂xi A f kL∞ L2 L2t (R1+n ) . kf kL1x L2(x L2t , (6.170)

xi (xj )j6=i i j )j6=i
where

kf kLpx1 Lp2 p
Lt 2 = kf kLpx2 ,...,x p
L 2 (R×Rn−1 ) p
. (6.171)
i (xj )j6=i 1 i−1 ,xi+1 ,...,xn t Lx1i (R)
(6.169) and (6.170) are scaling-invariant and so, they are optimal estimates.
In view of (6.170), one should choose L∞ 2 2
xi L(xj )j6=i Lt as a framework to
handle the partial derivative ∂xi in the nonlinearity. According to the
integral equation, we have
n
X
k∂x1 ukL∞ 2
x L(x L2 . kDx1/2
1
u0 k2 + k∂xi (|u|2 u)kL1x L2(x L2t . (6.172)
1 j )j6=1 t 1 j )j6=1
i=1
So, one needs to make the following two kinds of nonlinear estimates,
I = k∂x1 (|u|2 u)kL1x L2(x L2t (R1+n ) , II = k∂x2 (|u|2 u)kL1x L2(x L2t .
1 j )j6=1 1 j )j6=1
Let us consider the estimates of I. Using Hölder’s inequality, we have

2
I 6 k∂x1 ukL∞ 2
x L(x L2t kukL2x L∞ L∞
t
. (6.173)
1 j )j6=1 1 (x j )j6=1
Hence, we need to estimate kukL2x L∞

(x
L∞
t
. By the integral equation,
1 j )j6=1
kukL2x L∞
(x
L∞
t
6 kS(t)u0 kL2x L∞ L∞
t
+ k∇A (|u|2 u)kL2x L∞ L∞
t
.
1 j )j6=1 1 (x j )j6=1 1 (x j )j6=1
Unfortunately, we can not get a straightforward global estimates. So, we

consider the frequency-localized version of kS(t)u0 kL2x L∞
(x )
L∞
t
:
1 j j6=1
1/2
kk S(t)u0 kL2x L∞
(x
L∞
t
. hk1 i kk u0 k2 . (6.174)
1 j )j6=1
Since (6.174) is only a frequency-local version, one needs to localize (6.172)

and introduce the following
X
kuk`1,s (L∞ L2 L2 ) := hk1 is kk ukL∞ 2
x L(x ) L2t , (6.175)
hi x1 (xj )j6=1 t 1 j j6=1
k∈Zn , |k1 |>4
X
kuk`1 (L2x L∞ L∞
t )
:= kk ukL2x L∞
(x
L∞
t
. (6.176)
1 (x j )j6=1 1 j )j6=1
k∈Zn
Noticing that the smooth effect estimate in the lower frequency part is
worse than that of the Strichartz estimate, we have thrown away the lower
frequency on the ξ1 orientation in (6.175). If we do not consider the sharp
estimate, (6.174) implies that
X
kuk`1 (L2x L∞ L ∞ ) . kuk
t M
1/2 + hk1 i1/2 kk ∇(|u|2 u)kL1t L2x .
(x ) 1 j j6=1 2,1
k∈Zn
Making the nonlinear estimate, we have

!3
X
3/2
kuk`1 (L2x L∞ L∞
t )
. kukM 1/2 + hki kk ukL3t L6x , (6.177)
1 (x j )j6=1 2,1
k∈Zn
which means that we need the following norm

X
kuk`1,3/2 (L3 L6 ) := hki3/2 kk ukL3t L6x . (6.178)
t x
k∈Zn
The local version of (6.172) is

k∂x1 uk`1,2 (L∞ 2 L2t )
hi x1 L(x j )j6=1
n
X
. ku0 kM 5/2 + hk1 i2 kk ∂xi (|u|2 u)kL1x L2(x L2t . (6.179)
2,1 1 j )j6=1
i=1
After making nonlinear estimates, we find that the right hand side of (6.179)
can be estimated by (6.175), (6.176) and (6.178). The estimate of II is more
complicated and one needs to consider the interaction between the partial
derivative ∂x2 and the space `1,2 ∞ 2 2
hi (Lx1 L(xj )j6=1 Lt ), see below for details.
We now state the global wellposedness and scattering results for gDNLS.
Let us recall the anisotropic Lebesgue space Lpx1i Lp(x2j )j6=i Lpt 2 is defined by
(6.171). For k = (k1 , ..., kn ), we write
n
X X
kukXαs = hki is−1/2 ∂xα` k u L∞ L2 L2t
xi (xj )j6=i
i, `=1 k∈Zn , |ki |>4
n
X X
+ ∂xα k u m ∞ , (6.180)
` L L xi (xj )j6=i
L∞
t
i, `=1 k∈Zn
n X
X
kukSαs = hkis−1 ∂xα` k u L∞ L2 T L3 L6 , (6.181)
t x t x
`=1 k∈Zn
X X
kukX s = kukXαs , kukS s = kukSαs . (6.182)
α=0,1 α=0,1
5/2
Theorem 6.8. Let n > 3, m = 2, u0 ∈ M2,1 and there exists a suitably
small δ > 0 such that ku0 kM 5/2 6 δ. Then (6.166) has a unique solution
2,1
5/2
u ∈ C(R, M2,1 ) ∩ X 5/2 ∩ S 5/2 , kukX 5/2 ∩S 5/2 . δ. Moreover, the scattering
5/2
operator S of (6.166) carries a whole zero neighborhood in C(R, M2,1 ) into
5/2
C(R, M2,1 ).
6.5.1 Global linear estimates

Proposition 6.14 (Smooth effect). For any i = 1, ..., n, we have the
following estimates

1/2
Dxi S(t)u0 ∞ 2 2
. ku0 k2 , (6.183)
Lx L(x Lt
i j )j6=i
k∂xi A f kL∞
x L
2 L2t . kf kL1x L2(x L2t , (6.184)
i (xj )j6=i i j )j6=i
k∂xi A f kL∞ L2x . kDx1/2

i
f kL1x L2(x L2t . (6.185)
t i j )j6=i
Proof. By standard dual estimates, (6.183) implies (6.185). So, it suffices

to show the first two inequalities. Denote x̄ = (x2 , ..., xn ). By Plancherel’s
identity and Minkowski’s inequality,

−1 itε1 ξ12
kS(t)u0 kL∞ 2 2
x Lx̄ Lt
6 Fξ1 e Fx1 (Fx̄ 0 2 ∞ 2 .
u ) (6.186)
1 Lξ̄ Lx1 Lt
Recall the smooth effect of S(t) in one spatial dimension,

−1 itξ2
Fξ e Fx u0 ∞ 2 1+1 . kDx−1/2 u0 kL2 (R) . (6.187)
Lx Lt (R )
So, (6.186), (6.187) together with Plancherel’s equality yield (6.183).

Denote
−1 ξ1
u = cFt,x 2 Ft,x f. (6.188)
|ξ|± − τ
¯ 2 . By Plancherel’s
We can assume that |ξ|2 = ξ 2 +ε2 ξ 2 +...+εn ξ n := ξ 2 +|ξ|
± 1 2 n 1 ±
identity,

−1 ξ1
kukL∞ L2x̄ L2 6
Fξ 2 ¯ 2
F f
t,x
. (6.189)
x1 t 1
ξ1 + |ξ|± − τ L2 L∞ L2 ξ̄ x1 τ
Making the change of the variable τ → µ + |ξ| ¯ 2 , the right hand side of
±
(6.189) becomes

−1 ξ1 −it|ξ̄|2±

F (e f )
F
ξ1 ξ 2 − µ t,x1 Fx2 ,...,xn 2 ∞ 2. (6.190)
1 Lξ̄ Lx1 Lµ
Recalling the smooth effect in one spatial dimension

−1 ξ
F
τ,ξ ξ 2 − τ Ft,x f ∞ 2 1+1 . kf kL1xL2t (R1+1 ) , (6.191)
Lx Lt (R )
in view of (6.189), (6.190) and (6.191) we get that

−it|ξ̄|2±
kukL∞ 2 2
x Lx̄ Lt
. e Fx 2 ,...,x f
n . (6.192)
1 L2ξ̄ L1x1 L2t
Using Minkowski’s inequality and Plancherel’s identity, we immediately

have
kukL∞ 2 2 . kf kL1 L2 L2 .
x1 Lx̄ Lt x x̄ t
(6.193)
R∞ 1
Noticing that ∂x1 A f = u − ∂x1 S(t) −∞ S(s)sgn(s)f (s)ds, we easily see
that in (6.193), substituting u by ∂x1 A f , the result also holds.
6.5.2 Frequency-localized linear estimates

We consider the frequency-localized versions of the smooth effect estimate,
the maximal function estimate and their relations to the Strichartz estimate
for the Schrödinger semi-group. Let {ηk }k∈Z ∈ Υ1 and
σk (ξ) := ηk1 (ξ1 )...ηkn (ξn ). (6.194)
We have {σk }k∈Zn ∈ Υn . In the sequel we will always use the expression
of σk in (6.194). Recall that k = F −1 σk F . For convenience, we denote
P
ek =
|`|∞ 61 k+` .
Lemma 6.9. Let Dxσi = (−∂x2i )σ/2 . For any σ ∈ R, k = (k1 , ..., kn ) ∈ Zn
with |ki | > 4, we have
kk Dxσi ukLpx1 Lpx2 ,...,xn Lpt 2 . hki iσ kk ukLpx1 Lpx2 ,...,xn Lpt 2 .
1 2 1 2
If σ ∈ N, replacing Dxσi by ∂xσi , the above estimate holds for all k ∈ Zn .
Proof. Using (6.194), we have
X1 Z
σ
k Dxi u = Fξ−1
i
(ηki +` (ξi )|ξi | σ
) (yi )(k u)(xi − yi )dyi .
`=−1 R
By Young’s inequality and

kFξ−1
i
(ηki +` (ξi )|ξi |σ )kL1 (R) . hki iσ ,
we immediately have the result, as desired.
Proposition 6.15 (Maximal function estimate). Let 4/n < q 6 ∞

and q > 2. Then we have
kk S(t)u0 kLqxi L∞
(x
L∞
t
. hki i1/q kk u0 kL2 (Rn ) . (6.195)
j )j6=i
Proof. For convenience, we denote x̄ = (x2 , ..., xn ). By duality, it suffices

to show that
2
¯ q/2 ∞ n . hk1 i2/q .
kF −1 eit|ξ|± ηk1 (ξ1 )η (ξ)k k̄ Lx1 Lx̄,t (R )
k S(t) satisfies the following decay

2
¯ L∞ (Rn−1 ) . (1 + |t|)−(n−1)/2 ,
kFξ̄−1 eit|ξ̄|± ηk̄ (ξ)k x̄
2
kFξ−1
1
eitξ1 ηk1 (ξ1 )kL∞
x (R)
. (1 + |t|)−1/2 .
1
On the other hand, integrating by parts, we can get that for |x1 | > 4|t|hk1 i,
2
|Fξ−1
1
eitξ1 ηk1 (ξ1 )| . |x1 |−2 .
So, for any |x1 | > 1,
2
¯ . (1 + |x1 |)−2 + hk1 in/2 (hk1 i + |x1 |)−n/2 .
|F −1 eit|ξ|± ηk1 (ξ1 )ηk̄ (ξ)|
This implies that
2
¯ q/2 ∞ n
kF −1 eit|ξ|± ηk1 (ξ1 )ηk̄ (ξ)k L L (R ) x1 x̄,t
n/2 −n/2
. 1 + hk1 i k(hk1 i + |x1 |) kLq/2 (R)
x1
2/q
. hk1 i .
It follows that (6.195) holds.
By Proposition 6.14, we have
Proposition 6.16 (Frequency-localized smooth effect). For any
k = (k1 , ..., kn ) ∈ Zn , we have
kk A ∂xi f kL∞ L2 L2t . kk f kL1x L2(x L2t , (6.196)
xi (xj )j6=i i j )j6=i
kk A ∂xi f kL∞ L2 . hki i1/2 kk f kL1x L2(x L2t . (6.197)
t x i j )j6=i
Proof. By Proposition 6.14, we have (6.196). In view of Proposition 6.14

and Lemma 6.9, we obtain that (6.197) holds for all |ki | > 3. If |ki | 6 2,
from Proposition 6.14 it follows that

kk A ∂xi f kL∞ L2 . Dx−1/2
i
k A ∂x i f ∞ 2
t x Lt Lx
. kk f kL1x L2 L2 ,
i (xj )j6=i t

Proposition 6.17. (Relations of Strichartz, smooth effect and max-

imal function estimates). Let 2 6 r < ∞, 2/γ(r) = n(1/2 − 1/r) and
γ > γ(r) ∨ 2. We have
kk S(t)u0 kLγt Lrx . kk u0 kL2 (Rn ) , (6.198)
kk A f kL∞ L2 ∩ Lγ Lr . kk f kLγ 0 Lr0 , (6.199)
t x t x t x
kk A ∂xi f kLγ Lrx . hki i1/2 kk f kL1x L2(x L2t , (6.200)
t i j )j6=i
1/2
kk A ∂xi f kL∞ L2 2 . hki i kk f kLγ 0 Lr0 , (6.201)
xi (xj )j6=i Lt t x

k A ∂ α f 2 ∞ . hki iα+1/2
kk f kL1t L2x . (6.202)
xi L L L∞
xi t
(xj )j6=i
Proof. (6.198) and (6.199) are corollaries of the Strichartz estimate. By

(6.195) we can get (6.202). For convenience, we write
Z Z

Lk (f, ψ) := k S(t − τ )f (τ )dτ, ψ(t) dt . (6.203)
R R
Now we prove (6.200). Applying the Strichartz inequality, Lemma 6.9 and
Proposition 6.16, we have

Lk (∂x1 f, ψ) . hki i1/2 kk f kL1 L2 ˜ γ 0 r0
L2 k ψ x1 x2 ,...,xn t Lt Lx
1/2
. hki i kk f kL1x L2x2 ,...,xn L2t kψkLγ 0 Lr0 . (6.204)
1 t x
By the duality, (6.210) and Christ-Kiselev’s lemma (see Appendix), we can

get (6.200). Changing the role of f and ψ, we immediately get that (6.201)
holds for r > 2. If r = 2, (6.201) is a straightforward consequence of the
smooth effect estimate of S(t).
Corollary 6.5. Let 2 6 q < ∞, q > 4/n and 4/n 6 p < ∞. We have the
following results.

1/2
Dx1 k S(t)u0 2
. kk u0 kL2 (Rn ) , (6.205)
L∞ 2
x1 Lx2 ,...,xn Lt
kk S(t)u0 kLqx L∞ ∞

x2 ,...,xn Lt
. hki i1/q kk u0 kL2 (Rn ) , (6.206)
1
kk S(t)u0 kL2+p ∩ L∞ L2 . kk u0 kL2 (Rn ) , (6.207)

t,x t x
kk A ∂x1 f kL∞ L2 2 . kk f kL1x L2x2 ,...,xn L2t , (6.208)

x1 x2 ,...,xn Lt 1
1/2
kk A f kL∞ L2 ∩ L2+p . hk1 i kk f kL1x L2x2 ,...,xn L2t , (6.209)
t x t,x 1
kk A ∂x1 f kL∞ L2 2 . hk1 i1/2 kk f kL(2+p)/(1+p) , (6.210)

x1 x2 ,...,xn Lt t,x
kk A ∂x1 f kLqx L∞ ∞ . hk1 i1+1/q kk f kL1t L2 , (6.211)

1 x2 ,...,xn Lt
kk A f kL∞ L2 ∩ L2+p . kk f kL(2+p)/(1+p) . (6.212)

t x t,x t,x
Moreover, if L2+p 3 6
x,t is replaced by Lt Lx in (6.207), (6.209) and (6.212), then
corresponding results also hold.
By (6.208) in Corollary 6.5, the operator k ∂x1 A : L1x1 L2x2 ,...,xn L2t →
L∞ 2 2
x1 Lx2 ,...,xn Lt
and the partial derivative ∂x1 is successfully absorbed. How-
ever, A in the space L∞ 2 2
x1 Lx2 ,...,xn Lt can not handle the partial derivative
∂x2 . So, one needs to use another way to deal with ∂x2 A when it appears
in the space L∞ 2 2
x1 Lx2 ,...,xn Lt .
Proposition 6.18. Let i = 2, ..., n, 2 6 q 6 ∞, q > 4/n, 2 6 r < ∞,

2/γ(r) = n(1/2 − 1/r), γ > γ(r), γ > 2. Then we have
kk ∂xi A f kL∞ L2 2 . k∂xi ∂x−1 k f kL1x L2x2 ,...,xn L2t , (6.213)
x1 x2 ,...,xn Lt 1 1
kk ∂xi A f kL∞ L2 2 . k∂xi Dx−1/2 k f kLγ 0 Lrx0 , (6.214)

x1 x2 ,...,xn Lt 1
kk ∂xi A f kLqx L∞ ∞ . hki ihk1 i1/q kk f kL1t L2x . (6.215)
1 x2 ,...,xn Lt
Proof. (6.213) is a corollary of Proposition 6.14. We have

Z Z

L(∂x2 f, ψ) := S(t − τ )∂x2 f (τ )dτ, ψ(t) dt
Z R R

6 S(−τ )∂x2 Dx1 f (τ )dτ
−1/2

R L2 (Rn )
Z
1/2

× Dx1 S(−t)ψ(t)dt . (6.216)

R L2 (Rn )
In view of the Strichartz inequality and Proposition 6.14,
L(∂x2 f, ψ) . k∂x2 Dx−1/2
1
f kLγ 0 Lr0 kψkL1x L2x2 ,...,xn L2t . (6.217)
t x 1
If r > 2, by the duality, Christ-Kiselev’s lemma and (6.217), we have

(6.214). If r = 2, in view of the smooth effect estimate of S(t), we see
that (6.214) holds. By Proposition 6.15, we have (6.215).
Lemma 6.10. Let ψ : [0, ∞) → [0, 1] be a smooth cut-off function satisfying

ψ(x) = 1 for |x| 6 1; and ψ(x) = 0 for |x| > 2. Let ψ1 (ξ) = ψ(ξ2 /2ξ1 ),
ψ2 (ξ) = 1 − ψ(ξ2 /2ξ1 ), ξ ∈ Rn . For any σ > 0, we have
X

hk1 iσ Fξ−1
1 ,ξ2
ψ1 Fx 1 ,x 2 k ∂x 2 A f ∞ 2 2 Lx1 Lx2 ,...,xn Lt
k∈Zn , |k1 |>4
X
. hk1 iσ kk f kL1 2 2 . (6.218)
x1 Lx2 ,...,xn Lt
k∈Zn , |k1 |>4
For σ > 1,
X

hk1 iσ Fξ−1
1 ,ξ2
ψ2 Fx1 ,x2 k ∂x2 A f 2
L∞ 2
x1 Lx2 ,...,xn Lt
k∈Zn , |k1 |>4
X
. hk2 iσ kk f kL1x L2x2 ,...,xn L2t . (6.219)
1
k∈Zn , |k2 |>4
Proof. For simplicity, we denote x̄ = (x2 , ..., xn ) and

I = Fξ−11 ,ξ2
ψ1 Fx1 ,x2 k ∂x2 A f ∞ 2 2 ,
Lx1 Lx̄ Lt

II = Fξ−1
1 ,ξ2
ψ2 Fx1 ,x2 k ∂x2 A f 2 2
.
L∞
x1 Lx̄ Lt
Let ηk be as in Lemma (6.194). For k ∈ Zn , |k1 | > 4, using the almost

orthogonality of k , we have

X Y
I. F −1 ψ ξ2 ξ2 η (ξ )F ∂ A f .
ξ1 ,ξ2 2ξ1 ξ1 i=1,2
ki +ì i x1 ,x2 k x1
|`1 |,|`2 |61 ∞ 2 2 Lx1 Lx̄ Lt
(6.220)
Denote
Z
(f ~12 g)(x) = f (t, x1 − y1 , x2 − y2 , x3 , ..., xn )g(t, y1 , y2 )dy1 dy2 .
R2
(6.221)
For any Banach function space X defined in R1+n , we have
kf ~12 gkX 6 kgkL1y (R2 ) sup kf (·, · − y1 , · − y2 , ·, ..., ·)kX . (6.222)
1 ,y2
y1 ,y2
Hence, in view of (6.220) and (6.222),

X ξ2 ξ2
Y
I. −1
ηki +ì (ξi ) kk ∂x1 A f kL∞
Fξ1 ,ξ2 ψ 2ξ1 ξ1 2 2 .
x Lx̄ L t
|`1 |,|`2 |61 1
i=1,2
L1 (R2 )
(6.223)
Using the multiplier estimates, for any |k1 | > 4, we have

Y
−1 ξ2 ξ2
F ψ η (ξ )
ξ1 ,ξ2 2ξ1 ξ1 i=1,2
k i +` i i
1 2
L (R )
 

X ξ2 ξ2
Y
α
ηki +ì (ξi )

. D ψ 2ξ1 ξ1 . 1. (6.224)
|α|62 i=1,2
L2 (R2 )
By Proposition 6.16, (6.223) and (6.224), one has that

I . kk f kL1 2 2 , |k1 | > 4. (6.225)
x1 Lx̄ Lt
Now we consider the estimate of II. Applying Proposition 6.18, we have

II . Fξ−1
1 ,ξ2
(ξ2 /ξ1 )ψ2 Fx 1 ,x 2 k f 1 2 2
Lx1 Lx̄ Lt

X Y
ξ2 ξ2
−1
ηki +ì (ξi )
. Fξ1 ,ξ2 1 − ψ 2ξ1 ξ1 i=1,2

|`1 |,|`2 |61
L1 (R2 )
× kk f kL1 2 2 . (6.226)

x1 Lx̄ Lt
Notice that suppψ2 ⊂ {ξ : |ξ2 | > 2|ξ1 |}. If |k1 | > 4, then |k2 | > 6 and in
the summation of the left hand side of (6.219), one has that |k2 | > |k1 |. So,
X X
hk1 iσ II 6 hk2 iσ−1 hk1 iII.
k∈Zn , |k1 |>4 k∈Zn , |k1 |>4
We have

Y
−1 ξ2 ξ2
F 1 − ψ η (ξ )
ξ1 ,ξ2 2ξ1 ξ1 i=1,2
ki +ì i
1 2
L (R )
 
X Y
ξ2 ξ 2
α  −1
ηki +ì (ξi )

. D Fξ1 ,ξ2 1 − ψ 2ξ1 ξ1 i=1,2
|α|62
L2 (R2 )
−1
. hk2 ihk1 i . (6.227)
Combining (6.226) and (6.227), we can get the estimate of II.
6.5.3 Proof of global wellposedness for small rough data

Denote
(i)
X
%1 (u) = hki i2 kk ukL∞ 2
x L(x L2t ,
i j )j6=i
k∈Zn , |ki |>4
(i)
X
%2 (u) = kk ukLm ∞
x L(x L∞
t
,
i j )j6=i
k∈Zn
(i)
X
%3 (u) = hki i3/2 kk ukL3t L6x ∩ L∞ 2.
t Lx
k∈Zn
Let  
 3 X X
X n 
(i)
X := u ∈ S 0 : kukX := %` (∂xαj u) 6 δ .
 
`=1 α=0,1 i,j=1

T : u(t) → S(t)u0 − iA F (u, ū, ∇u, ∇ū),
we show that T : X → X is a contraction mapping. Since kukX = kūkX ,
we can assume that
X
F (u, ū, ∇u, ∇ū) = F (u, ∇u) := cκν uκ (∇u)ν ,
m+16κ+|ν|<∞
where (∇u)ν = uνx11 ...uνxnn . For convenience, we write

v1 = ... = vκ = u,
vκ+1 = ... = vκ+ν1 = ux1 , ..., vκ+|ν|−νn +1 = ... = vκ+|ν| = uxn .
By (6.183), for α = 0, 1, we have
(i)
X
%1 (∂xαj S(t)u0 ) . hki i1/2 hkj i2 kk u0 kL2 (Rn ) 6 ku0 kM 5/2 .
2,1
k∈Zn , |ki |>4
In view of (6.206) and (6.207), for α = 0, 1, we have

(i) (i)
%2 (∂xαj S(t)u0 ) + %3 (∂xαj S(t)u0 ) . ku0 kM 5/2 .
2,1
So,
kS(t)u0 kX . ku0 kM 5/2 .
2,1
(i)
In order to estimate %1 (A ∂xαj (v1 ...vκ+|ν| )), i, j = 1, ..., n, it suffices to con-
(1) (1)
sider the estimates of %1 (A ∂xα1 (v1 ...vκ+|ν| )) and %1 (A ∂xα2 (v1 ...vκ+|ν| )).
Applying the frequency-uniform decomposition, we have
X
k (v1 ...vκ+|ν| ) = k k(1) v1 ...k(κ+|ν|) vκ+|ν|
(i)
S1
X
+ k k(1) v1 ...k(κ+|ν|) vκ+|ν| , (6.228)
(i)
S2
where
(i) (1) (κ+|ν|)
S1 := {(k (1) , ..., k (κ+|ν|) ) : |ki | ∨ ... ∨ |ki | > 4},
(i) (1) (κ+|ν|) (1) (κ+|ν|)
S2 := {(k , ..., k ): |ki | ∨ ... ∨ |ki | 6 4}.
We will frequently use the almost orthogonality of k ,

k k(1) v1 ...k(κ+|ν|) vκ+|ν| = 0,
if |k − k (1) − ... − k (κ+|ν|) |∞ > κ + |ν| + 1. (6.229)
By (6.196) and (6.201),

(1)
%1 (A ∂xα1 (v1 ...vκ+|ν| ))
X X
. hk1 i2 kk k(1) v1 ...k(κ+|ν|) vκ+|ν| kL1x L2x̄ L2t
1
k∈Zn , |k1 |>4 S1
(1)
X X
+ hk1 i5/2 kk k(1) v1 ...k(κ+|ν|) vκ+|ν| k κ+|ν|+1
κ+|ν|
k∈Zn , |k1 |>4 (1) Lt,x
S2
=: I + II. (6.230)
Applying the almost orthogonality of k in (6.229), we have

X (1)
I . C κ+|ν| hk1 i2 kk(1) v1 kL∞ 2 2
x Lx̄ Lt 1
(1)
k(1) ∈Zn , |k1 |>2
κ+|ν|
X Y
× kk(i) vi kLκ+|ν|−1 L∞ L∞ . (6.231)
x1 x̄ t
k(2) ,...,k(κ+|ν|) ∈Zn i=2
By Hölder’s inequality and Lemma 6.1,
kk(i) vi kLκ+|ν|−1 L∞ L∞
x1 x̄ t
m m
κ+|ν|−1
1−κ+|ν|−1
6 kk(i) vi k Lm ∞ ∞
x1 Lx̄ Lt
kk(i) vi kL∞
x,t
m m
1−
. kk(i) vi kLκ+|ν|−1
x x̄ t
κ+|ν|−1
m L∞ L∞ kk(i) vi k ∞ 2
L L . (6.232)
1 t x
Hence, in view of vi = u or vi = uxj , it follows from (6.231) and (6.232)

that
I . (CkukX )κ+|ν| . (6.233)

(1)
By the definition of S2 , one has that |k1 | 6 C(κ + |ν|) in the summation
of II. Again, by Hölder’s inequality and Lemma 6.1,
κ+|ν|
Y
kk(1) v1 ...k(κ+|ν|) vκ+|ν| k κ+|ν|+1 6 kk(i) vi kLκ+|ν|+1
κ+|ν| x,t
Lx,t i=1
κ+|ν|
Y
. kk(i) vi kL2+m T
L∞ 2, (6.234)
x,t t Lx
i=1
which implies that
II . (CkukX )κ+|ν| . (6.235)

(1)
Now we estimate %1 (A ∂xα2 (v1 ...vκ+|ν| )). The case α = 0 has been dis-
cussed in the above, it suffices to consider the case α = 1. Let ψi (i = 1, 2)
be as in Lemma 6.10 and Pi = F −1 ψi F . We have
(1)
%1 (A ∂x2 (v1 ...vκ+|ν| ))
X
6 hk1 i2 kP1 k (A ∂x2 (v1 ...vκ+|ν| ))kL∞ 2 2
x Lx̄ Lt 1
k∈Zn , |k1 |>4
X
+ hk1 i2 kP2 k (A ∂x2 (v1 ...vκ+|ν| ))kL∞ 2 2
x Lx̄ Lt 1
k∈Zn , |k1 |>4
=: III + IV. (6.236)

Using the decomposition in (6.228), we have
X X
III 6 hk1 i2 kP1 k (A ∂x2 (k(1) v1 ...k(κ+|ν|) vκ+|ν| ))kL∞ 2 2
x Lx̄ Lt 1
k∈Zn , |k1 |>4 (1)
S1
X X
+ hk1 i2 kP1 k (A ∂x2 (k(1) v1 ...k(κ+|ν|) vκ+|ν| ))kL∞ 2 2
x Lx̄ Lt 1
k∈Zn , |k1 |>4 S2
(1)
=: III1 + III2 . (6.237)

By Lemma 6.10,
X X
III1 . hk1 i2 kk (k(1) v1 ...k(κ+|ν|) vκ+|ν| )kL1x L2x̄ L2t .
1
n
S1 k∈Z , |k1 |>4
(1)
(6.238)
(1) (1) (κ+|ν|)
By the symmetry, we can assume that |k1 | = max(|k1 |, ..., |k1 |) in
(1)
S1 . So,
III1
κ+|ν|
X (1)
Y
. C κ+|ν| hk1 i2 kk(1) v1 kL∞ 2 2
x Lx̄ Lt
kk(i) vi kLκ+|ν|−1 L∞ L∞
1 x1 x̄ t
(1)
S1 ,
(1)
|k1 |>4 i=2
κ+|ν|
(1)
Y (1) (1)
. C κ+|ν| %1 (v1 ) (%2 (vi ) + %3 (vi )) . (CkukX )κ+|ν| . (6.239)
i=2
Applying (6.214) and noticing that |k1 | 6 C in III2 , we obtain that

III2
X X
. hk1 i5/2 kk (k(1) v1 ...k(κ+|ν|) vκ+|ν| )kL(2+m)/(1+m)
t,x
k∈Zn , |k1 |>4, |k2 |.|k1 | (1)
S2
κ+|ν|
Y (1)
. C κ+|ν| %3 (vi ) 6 (CkukX )κ+|ν| . (6.240)
i=1
We have shown that
III . (CkukX )κ+|ν| . (6.241)
Now we estimate IV . Using (6.228), we have
X X
IV 6 hk1 i2 kP2 k (A ∂x2 (k(1) v1 ...k(κ+|ν|) vκ+|ν| ))kL∞ 2 2
x Lx̄ Lt 1
k∈Zn , |k1 |>4 (2)
S1
X X
+ hk1 i2 kP2 k (A ∂x2 (k(1) v1 ...k(κ+|ν|) vκ+|ν| ))kL∞ 2 2
x Lx̄ Lt 1
k∈Zn , |k1 |>4 (2)
S2
=: IV1 + IV2 . (6.242)

By Lemma 6.10,
X X
IV1 . hk2 i2 kk (k(1) v1 ...k(κ+|ν|) vκ+|ν| )kL1x L2x̄ L2t . (6.243)
1
(2)
S1 k∈Zn , |k2 |>4
(κ+|ν|)
We can choose some k (i) , say k (κ+|ν|) such that |k2 | does not attain
(i)
or does not uniquely attain max16i6κ+|ν| |k2 |, then we can take another
(1) (i)
k (i) , say k (1) such that k2 = max16i6κ+|ν| |k2 |. By Hölder’s inequality,
kk(1) v1 ...k(κ+|ν|) vκ+|ν| kL1x L2x̄ L2t
1
6 kk(1) v1 ...k(κ+|ν|−1) vκ+|ν|−1 kL2x,t kk(κ+|ν|) vκ+|ν| kL2x L∞ ∞

x̄ Lt
1
κ+|ν|
Y
6 kk(1) v1 kL∞ 2 2
x Lx̄ Lt
kk(i) vi k(L∞ 2 2
t Lx )∩(Lx L∞
x̄,t )
. (6.244)
1 1
i=2
Combining (6.243) with (6.244), we get

IV1 . (CkukX )κ+|ν| . (6.245)
In order to estimate IV2 , we apply (6.214) to obtain that
X X
IV2 . hk2 i5/2 kP2 k (k(1) v1 ...k(κ+|ν|) vκ+|ν| )kL(2+m)/(1+m)
t,x
k∈Zn , |k1 |>4 S2
(2)
X κ+|ν|
. C κ+|ν| kk(1) v1 ...k(κ+|ν|) vκ+|ν| kL(2+m)/(1+m) . kukX .
t,x
(2)
S2
(6.246)
Hence, in view of (6.245) and (6.246), we have
IV . (CkukX )κ+|ν| . (6.247)
Collecting (6.233), (6.235), (6.241) and (6.247), we have shown that

n
X X (i)
%1 (A ∂xαj (uκ (∇u)ν )) . (CkukX )κ+|ν| . (6.248)
α=0,1 i,j=1
(i)
Now we estimate %3 (A (uκ (∇u)ν )). By (6.212),
n
X X
(i)
%3 (A (uκ (∇u)ν )) . hki3/2 kk (uκ (∇u)ν )k 2+m . (6.249)
1+m
Lt,x
i=1 k∈Zn
Using Lemma 6.8, one can control the right hand side of (6.249),
X
hki3/2 kk (v1 ...vκ+|ν| )k 2+m
1+m
Lt,x
k∈Zn
m+1
! κ+|ν|
!
Y X Y X
κ+|ν| 3/2 3/2
.C hki kk vi kL2+m hki kk vi kL∞
t,x
t,x
i=1 k∈Zn i=m+2 k∈Zn
m+1
! κ+|ν|
!
Y X Y X
. C κ+|ν| hki3/2 kk vi kL2+m hki3/2 kk vi kL∞ 2
t Lx
t,x
i=1 k∈Zn i=m+2 k∈Zn
κ+|ν| n
!
Y X (1)
. C κ+|ν| %3 (vi ) 6 (CkukX )κ+|ν| . (6.250)
i=1 i=1
(1)
Now we estimate %2 (A ∂xα1 (uκ (∇u)ν )).
(1)
X
%2 (A ∂xα1 (v1 ...vκ+|ν| )) . hki3/2 kk v1 ...vκ+|ν| kL1t L2x . (6.251)
k∈Zn
Analogous to (6.250), using Lemma 6.8, we can control (6.251),

κ+|ν|
!
(1)
Y X
%2 (A ∂xα1 (v1 ...vκ+|ν| )) .C κ+|ν| 3/2
hki kk vi kL∞ 2 3 6
t Lx ∩Lt Lx
i=1 k∈Zn
. (CkukX )κ+|ν| . (6.252)

Now we estimate
X n
(i)
ρ3 (A ∂x1 (v1 ...vκ+|ν| ))
i=1
X
. kk (v1 ...vκ+|ν| )kL1t L2x
|k|64
 
X X
+ +... +  hki3/2
k∈Zn , |k1 |=kmax >4 k∈Zn , |kn |=kmax >4

× kk A ∂x1 v1 ...vκ+|ν| kL∞ 2 3 6
t Lx ∩ Lt Lx
=: Υ0 (u) + Υ1 (u) + ... + Υn (u). (6.253)

Υ0 (u) has been estimated as above. It suffices to estimate Υi (u), i = 1, 2.
Using (6.228) and Corollary 6.5,
X X
Υ1 (u) . hk1 i2 kk k(1) v1 ...k(κ+|ν|) vκ+|ν| kL1x L2x̄ L2t
1
k∈Zn , |k1 |>4 (1)
S1
X X
+ hk1 i5/2 kk k(1) v1 ...k(κ+|ν|) vκ+|ν| k κ+|ν|+1 ,
κ+|ν|
k∈Zn , |k1 |>4 (1) Lt,x
S2
(6.254)
which reduces to (6.230). Hence,
Υ1 (u) . (CkukX )κ+|ν| . (6.255)
We now estimate Υ2 (u). Taking notice of |k1 | 6 |k2 | in the summation of
Υ2 (u) and using (6.209) and (6.212), we have
X X
Υ2 (u) . hk2 i2 kk k(1) v1 ...k(κ+|ν|) vκ+|ν| kL1x L2x̄ L2t
1
k∈Zn , |k2 |>4 (2)
S1
X X
+ hk2 i5/2 kk k(1) v1 ...k(κ+|ν|) vκ+|ν| k κ+|ν|+1 .
κ+|ν|
k∈Zn , |k2 |>4 (2) Lt,x
S2
(6.256)
The first term in the right hand side has been estimated in (6.244). The sec-
ond term is the same as in II of (6.230). Summarizing the above estimates,
we get that10
X
kT ukX 6 Cku0 kM 3/2 + `2n+2 C ` kuk`X . (6.257)
m+16`<∞
We can use the contraction mapping argument to finish the proof and omit
the details.
Remark 6.2.
(1) There are some recent works which have been devoted to the study
of the derivative NLKG [54; 55; 76; 144; 185; 200]:
utt + u − ∆u = F (u, ut , uxx , uxt ),
where the methods are quite different from those of derivative NLS.
We do not know whether the frequncy-uniform decomposition tech-
niques are applicable to the derivative NLKG.
10 Notice that |cβ | 6 C |β| .
(2) We now state some results on the Zakharov system

i∂t u + ∆u = nu,
∂t2 n − ∆n = ∆|u|2 .
In 2D, Bourgain and Colliander [20] used the Fourier restriction
norm to prove the local well-posedness and global well-posedness
for initial data in the spaces H 1 × L2 × H −1 . Fang, Pecher and
Zhong [71] applied the I-method to obtain the global well-posedness
in H s × L2 × H −1 with 3/4 < s < 1. Recently, Bejenaru, Herr,
Holmer, and Tataru [8] showed the local well-posedness for the
Schrödinger data in L2 and the wave data in H −1/2 × H −3/2 . In
higher spatial dimensions, the local well posed result can be found
in Ginibre, Tsutsumi and Velo [81].
Chapter 7
Conservations, Morawetz’ estimates

of nonlinear Schrödinger equations
This result is too beautiful to be false; it is more important to have

beauty in one’s equations than to have them fit experiment. ——
Paul Dirac
On mentioning evolution partial differential equations, we will natu-

rally think whether they satisfy some physical conservation laws, such as
the mass conservation, the energy conservation and the momentum conser-
vation, and so on. For example, as indicated in (4.2), we see that NLS (4.1)
satisfies the conservations of mass and energy.
In this chapter, we will derive some conservation laws by using Nöther’s
theorem and Morawetz’ inequalities by the virial identity. It is known that
Morawetz’ estimates are fundamental tools in the study of the scattering
theory for NLS, see for instance, [27; 40; 43; 218].
7.1 Nöther’s theorem
The energy conservation law means that it does not vary with time, in
other word, it has a certain symmetry with respect to the time. This
fundamental connection between conservation laws and symmetries was
first discovered in 1915 (published in 1918) by Emmy Nöther1 and was
called the “Nöther’s theorem”. In short, Nöther’s theorem states informally
that any differentiable symmetry of the action of a physical system has
a corresponding conservation law. Nöther’s theorem is important, both
because of the insight it gives into conservation laws, and also as a practical
calculational tool. It allows us to determine the conserved quantities from
1 Amalie Emmy Nöther (1882-1935) was a German-born mathematician known for her
ground-breaking contributions to abstract algebra and theoretical physics. She revolu-

tionized the theories of rings, fields, and algebras. In physics, Nöther’s theorem explains
the fundamental connection between symmetry and conservation laws.
205
July 7, 2011 10:54 World Scientific Book - 9in x 6in booken-main
206 Conservations, Morawetz’ estimates of nonlinear Schrödinger equations
the observed symmetries of a physical system. The latter turns out to be

easier to find than the former.
Application of Nöther’s theorem allows physicists to gain powerful in-
sights into any general theories in physics, by just analyzing the various
transformations that would make the form of the laws involved invariant.
For example, the invariance of physical systems with respect to the spatial
translation (in other words, the laws of physics do not vary with locations
in space) gives the conservation law of the linear momentum; the invari-
ance with respect to the rotation gives the conservation law of the angular
momentum; the invariance with respect to the time translation gives the
well-known conservation law of energy. But how to derive the associated
conserved quantities of its if we know that an equation has a symmetry
group? Let us begin by recalling some concepts of the calculus of variations.
Suppose the interval I ⊂ R, the generalized coordinate q is dependent
on the time t. Assume that L : Rnq × Rnq̇ × I 7→ R is smooth and convex with
respect to the first slot q̇ (The dot indicates the derivative with respect to
the time.), then we call the function L to be a Lagrangian 2 of the system.
For some q satisfying the related boundary conditions, we define the action
functional
Z t1
I[q] = L(q(t), q̇(t), t)dt.
t0
Hamilton’s principle states that the true evolution of a system described
by generalized coordinates q between two specified states q(t0 ) and q(t1 )
at two specified times t0 and t1 is an extremum (i.e., a stationary point, a
minimum, maximum or saddle point) of the action functional I[q], in other
word, the first variation of the action functional I[q] is zero.
Let q(t) represent the true evolution of the system between two specified
states q(t0 ) and q(t1 ) at two specified times t0 and t1 . And let ε(t) be a
small perturbation that is zero at the endpoints of the trajectory
def
ε(t0 ) = ε(t1 ) == 0. (7.1)
To the first order in the perturbation ε(t), the change in the action func-
tional δIZ would be Z t1
t1
∂L ∂L
δI = (L(q + ε, q̇ + ε̇) − L(q, q̇)) dt = ε· + ε̇ · dt,
t0 t0 ∂q ∂ q̇
2 It is named after Joseph Louis Lagrange (1736-1813), who was an Italian-born math-
ematician and astronomer, making significant contributions to all fields of analysis, to

number theory, and to classical and celestial mechanics. He is one of the founders of
the calculus of variations. Roughly speaking, the Lagrangian of a dynamical system
is a function that summarizes the dynamics of the system. In classical mechanics, the
Lagrangian is defined as the kinetic energy of the system minus its potential energy.
7.1. Nöther’s theorem 207
where we have expanded the Lagrangian L to the first order in the per-
turbation ε(t). Applying the integration by parts to the last term results
in
t Z t1
∂L 1 ∂L d ∂L
δI = ε · + ε· −ε· dt.
∂ q̇ t0 t0 ∂q dt ∂ q̇
By the boundary condition (7.1), the first term vanishes, and then
Z t1
∂L d ∂L
δI = ε· − dt = 0.
t0 ∂q dt ∂ q̇
Noticing that we do not make any assumptions to the generalized coordi-
nates q except that all of the generalized coordinates are independent of
each other. Thus, we can apply the fundamental theorem of variational
calculus to obtain an Euler-Lagrange equation
d ∂L ∂L
− = 0, (7.2)
dt ∂ q̇ ∂q
which is a system involving n second-order equations with respect to q(t).
The Legendre transformation 3 of L is
H(q(t), p(t), t) = max[p(t) · q̇ − L(q(t), q̇(t), t)].
q̇
Let M (p, q̇) = p(t) · q̇ − L(q(t), q̇(t), t). In order to reach the maximum
of M with respect to q̇, we have to require that the first-order partial
derivative of M (p, q̇) with respect to q̇ is zero and that its second partial
derivative is less than zero (this can be guaranteed by the convexity of L,
2
i.e. ∂∂ q̇L2 > 0). Thus, if p(t) = ∂L∂ q̇ (q(t), q̇(t), t), then M (p, q̇) reaches the
maximum, i.e.

 p(t) = ∂L (q(t), q̇(t), t),
∂ q̇
 H(q(t), p(t), t) = p(t)q̇ − L(q(t), q̇(t), t),
where H is called the Hamiltonian. The second equation yields

∂H ∂L ∂H
=− , = q̇.
∂q ∂q ∂p
Thus, the Euler-Lagrange equation (7.2) can be rewritten as

 ∂H

 ṗ = − ∂q ,
(7.3)

 ∂H
 q̇ = ,
∂p
3 Legendre transformation is an operation that transforms one real-valued function of
a real variable into another. Specifically, the Legendre
transformation of a function f is
the function f ? defined by f ? (p) = maxx px − f (x) .
which are called Hamilton’s equation. This is a system of 2n first-order

equations in (p(t), q(t)). Denote u(t) = (p(t), q(t)) ∈ R2n , Hu = (Hp , Hq )
and define the matrix

0 −I
J= ,
I 0
where I denotes the unit matrix, then Hamilton’s equation (7.3) can be
written as4
u̇ = J Hu . (7.4)
2
Notice that J = −I and Hamilton’s equations are defined on an even
dimensional space. It suggests that there may be a nice connection with
complex numbers.
For (x, y) ∈ R2 , define z = x + iy, z̄ = x − iy, i2 = −1. Define
∂z = 21 (∂x − i∂y ) and ∂z̄ = 12 (∂x + i∂y ). It is clear that ∂z z = ∂z̄ z̄ = 1
and ∂z z̄ = ∂z̄ z = 0, that is, z̄ and z are independent of each other. Let
z = q + ip, then (7.3) yields
ż = q̇ + iṗ = −i(Hq + iHp ) = −2iHz̄ .
We give an example. Let I be a time interval, u : Rn × I 7→ C, u and
its derivatives are smooth and vanish at the infinity. Assume that the
Hamiltonian is defined by
Z Z
1 1
H = H[u, ū] = |∇u|2 dx = ∇u∇ūdx.
2 R 2 R
The usual calculus of variations argument shows
Z
H[u, ū + τ v̄] − H[u, ū] 1 1
lim = (−∆u)v̄dx = h−∆u, vi.
τ →0 τ 2 R 2
Therefore, Hū [u, ū] = − 21 ∆u and the associated Hamilton’s equation is
u̇ = −2iHū = i∆u,
which is precisely the linear Schrödinger equation.
Now, we consider the n-dimensional nonlinear Schrödinger equation
iut + ∆u = F 0 (|u|2 )u. (7.5)
Assume that u and its derivatives are smooth and vanish as |x| → +∞.
The nonlinearity F 0 (·) is a smooth real-valued function of its argument and
define
Z λ
F (λ) = F 0 (s)ds.
0
4 It is convenient to write u and Hu as column vectors in the computation.
Define the Lagrangian

i
(ūut − uūt ) − [|∇u|2 + F (|u|2 )],
L = L(u, ū, ut , ūt , ∇u, ∇ū) = (7.6)
2
and the associated action functional
Z t1 Z
I[u] = L dxdt, ∀u ∈ A, (7.7)
t0 Rn
where A is some appropriate class of admissible functions. The usual cal-
culus of variations argument shows that if u is a smooth critical point of
I(·), u satisfies the associated Euler-Lagrange equation
X ∂ n
∂L ∂ ∂L ∂L
− − = 0. (7.8)
∂u ∂t ∂(∂t u) j=1 ∂xj ∂(∂xj u)
Substituting (7.6) into the above equation and taking the complex conju-
gate, we can obtain the nonlinear Schrödinger equation (7.5).
Now, we give another statement for Nöther’s theorem:
Theorem 7.1 (Nöther’s theorem). If the action functional is invari-

ant under a family of transformations, then the solutions of the associated
Euler-Lagrange equation satisfy a conservation law.
We mainly consider the case of the one parameter family of transfor-

mations and apply Nöther’s theorem to the nonlinear Schrödinger equa-
tion (7.5). For convenience, denote ξ = (t, x) = (ξ0 , ξ1 , · · · , ξn ), ∂0 = ∂t ,
∂ = (∂t , ∇x ) = (∂0 , ∂1 , · · · , ∂n ), and u = (u1 , u2 ) = (u, ū), and denote the
integral domain D := [t0 , t1 ] × Rn . A one parameter group of transforma-
tions T ε is defined by
˜ u, ε),
ξ 7→ ξ(ξ, u 7→ ũ(ξ, u, ε), (7.9)
where we assume ξ˜ and ũ are differentiable with respect to ε, and it degen-
erates into an identical transformation for the case ε = 0. For infinitesimal
ε, denote
ξ̃ = ξ + δξ, ũ = u + δu, (7.10)
where δξ and δu are functions of (ξ, u, ε), both δξ = O(ε) and δu =
∂(ξ̃0 ,··· ,ξ̃n )
O(ε) tend to zero as ε → 0, and Jacobi’s determinant ∂(ξ 0 ,··· ,ξn )
=
Pn ∂(δξ)j ˜ the domain D
1 + j=0 ∂ξj + o(ε). Applied T ε , u(ξ) changes into ũ(ξ),
becomes into D̃, and the action functional
Z
I[u] = L(u, ∂u)dξ (7.11)
D
turns into
 
Z Z n
X ∂(δξ)j
Ĩ[ũ] = L(ũ, ∂˜ũ)dξ˜ = L(ũ, ∂˜ũ) 1 + + o(ε) dξ
D̃ D j=0
∂ξj
Z Z n
X ∂(δξ)j
= L(ũ, ∂˜ũ)dξ + L(u, ∂u) dξ + o(ε), (7.12)
D D j=0
∂ξj
where ∂ and ∂˜ denote the differential with respect to ξ and ξ,

˜ respectively.
The last two terms can be obtained by expanding with the help of (7.6)
and subsuming the higher order terms of ε in o(ε) in the expansion. We
only consider the transformations under which I is invariant. Denote
δI :=Ĩ[ũ] − I[u]
Z Z Xn
∂(δξ)j
= L(ũ, ∂˜ũ) − L(u, ∂u) dξ + L(u, ∂u) dξ + o(ε),
D D j=0
∂ξj
(7.13)
where the first integrand can be written as
L(ũ, ∂˜ũ) − L(u, ∂u)
 
X2
∂L Xn
∂L
=  ˜ − uk (ξ)) +
(ũk (ξ) ∂˜j ũk (ξ)
˜ − ∂j uk (ξ)  . (7.14)
∂uk j=0
∂(∂j u k )
k=1
Denote
n
X
˜ − uk (ξ) =
δũk ≡ ũk (ξ) ∂j uk (δξ)j + δuk (ξ). (7.15)
j=0
Since
n
X n
X
˜ =
∂j ũk (ξ) ˜ ∂ ξ̃l =
∂˜l ũk (ξ) (δjl +
∂(δξ)l ˜ ˜
)∂l ũk (ξ)
∂ξj ∂ξj
l=0 l=0
n
!
X ∂(δξ)l ˜
= ∂˜j + ∂l ˜
ũk (ξ), (7.16)
∂ξj
l=0
we have
∂˜j ũk (ξ)
˜ − ∂j uk (ξ) =(∂˜j − ∂j )ũk (ξ)
˜ + ∂j (ũk (ξ)
˜ − uk (ξ))
n n
!
X ∂(δξ)l X
=− ∂˜l ũk (ξ)
˜ + ∂j ∂l uk (δξ)l + δuk .
∂ξj
l=0 l=0
(7.17)
Hence,
 
2
X Xn
∂L
L(ũ, ∂˜ũ) − L(u, ∂u) =  ∂j uk (δξ)j + δuk 
∂uk j=0
k=1
2 X
n n n
!!
X ∂L X ∂(δξ)l X
+ − ∂˜l ũk (ξ)
˜ + ∂j ∂l uk (δξ)l + δuk .
∂(∂j uk ) ∂ξj
k=1 j=0 l=0 l=0
(7.18)
Noticed that
2 n
!
∂ ∂(δξ)j X ∂L X ∂L
(L(δξ)j ) = L + ∂j uk + ∂ 2 uk (δξ)j ,
∂ξj ∂ξj ∂uk ∂(∂l uk ) lj
k=1 l=0
(7.19)
and

∂L ∂ ∂L ∂ ∂L
∂j δuk = δuk − δuk , (7.20)
∂(∂j uk ) ∂ξj ∂(∂j uk ) ∂ξj ∂(∂j uk )
it is easily to obtain
L(ũ, ∂˜ũ) − L(u, ∂u)
Xn Xn 2
∂ ∂(δξ)j X ∂L
= (L(δξ)j ) − L + δuk
j=0
∂ξj j=0
∂ξj ∂uk
k=1
2 X
X n
∂ ∂L ∂ ∂L
+ δuk − δuk
j=0
∂ξj ∂(∂j uk ) ∂ξj ∂(∂j uk )
k=1
2 X
X n n
∂L X ∂(δξ)l ˜ ˜ − ∂l uk ).
− (∂l ũk (ξ) (7.21)
∂(∂j uk ) ∂ξj
k=1 j=0 l=0
Therefore, it yields
 
Z X 2 n
X
 ∂L ∂ ∂L  δuk dξ
δI = −
D k=1 ∂uk j=0 ∂ξj ∂(∂j uk )
Z X n 2
!
∂ X ∂L
+ L(δξ)j + δuk dξ + o(ε). (7.22)
D j=0 ∂ξj ∂(∂j uk )
k=1
(7.8) implies
X ∂ n
∂L ∂L
− = 0, k = 1, 2. (7.23)
∂uk j=0 ∂ξj ∂(∂j uk )
Since D is arbitrary, we have the following result in order to guarantee that

I remains invariant under the infinitesimal transformation T ε .
Theorem 7.2. Let ξ = (t, x1 , · · · , xn ), u = (u1 , u2 ). If the action

functional (7.11) is invariant under the infinitesimal transformation T ε :
˜ u, ε), u 7→ ũ(ξ, u, ε), then the following conservation law holds:
ξ 7→ ξ(ξ,
n 2
!
X ∂ X ∂L
L(δξ)j + δuk = 0, (7.24)
j=0
∂ξj ∂(∂j uk )
k=1
˜ − uk (ξ) − ∇ξ uk · δξ.
where δuk (ξ) = ũk (ξ)
Integrating with respect to the spatial variables yields the following result.
Theorem 7.3. If the action functional (7.11) is invariant under the in-
finitesimal transformation
t 7→ t̃ = t + δt(t, x, u), (7.25)
x 7→ x̃ = x + δx(t, x, u), (7.26)
u(t, x) 7→ ũ(t̃, x̃) = u(t, x) + δu(t, x), (7.27)
then
Z
∂L ∂L
(ut δt + ∇u · δx − δu) + (ūt δt + ∇ū · δx − δū) − Lδt dx
Rn ∂ut ∂ ūt
is a conserved quantity. In particular, in the case of nonlinear Schrödinger
∂L
equations, since ∂u t
= 2i ū, ∂∂L i
u¯t = − 2 u, it implies that
Z
i i
ū(ut δt + ∇u · δx − δu) − u(ūt δt + ∇ū · δx − δū) − Lδt dx = const.
Rn 2 2
(7.28)
7.2 Invariance and conservation law
In this subsection, we apply the formalism (7.28) to identify some invariant

quantities of the action functional for nonlinear Schrödinger equations and
thereby infer some conservation laws for nonlinear Schrödinger equations.
(i) Invariance by the phase shift : ũ = eiε u. For the infinitesimal quan-
tity ε, let δu = iεu, δt = 0, and δx = 0. From (7.28), we get the mass
conservation law (or the charge conservation law, or the L2 -norm conser-
vation law)
Z
N (t) := |u(t, x)|2 dx = const. (7.29)
Rn
7.2. Invariance and conservation law 213
(ii) Invariance by the time translation: t 7→ t+δt, δx = 0 and δu = δū =

0, where δt is an infinitesimal quantity independent of (t, x, u, ū). (7.28)
and (7.6) yield the energy conservation law
Z
H(t) := |∇u(t, x)|2 + F (|u(t, x)|2 )dx = const. (7.30)
Rn
(iii) Invariance by the spatial translation: x 7→ x + δx and δt = δu =

δū = 0, where δx is an infinitesimal quantity independent of (t, x, u, ū).
(7.28) implies the momentum conservation law
Z
P~ (t) := i (u(t, x)∇ū(t, x) − ū(t, x)∇u(t, x))dx = constant vector.
Rn
(7.31)
(iv) Invariance by the spatial rotation: δx = δθ(~a × x) and δt = δu =
δū = 0, which mean the spatial variables rotate anticlockwise for a angle
of δθ along the axis ~a where δθ is an infinitesimal quantity independent of
(t, x, u, ū). From (7.28), we have the angular momentum conservation law
Z
i x × (ū∇u − u∇ū)dx = constant vector. (7.32)
Rn
(v) Invariance by the Galilean transformation:



 x 7→ x̃ = x − ~ct,

t 7→ t̃ = t, (7.33)


 u 7→ ũ(t̃, x̃) = e−i[ 21 ~c·x̃+ 14 |~c|2 t̃] u(t̃, x̃ + ~ct̃),
in other words, for infinitesimal velocity ~c, let

i
δt = 0, δx = −~ct,
δu = − ~c · xu(t, x).
2
From (7.28), it is easily to gain the renormalized centroid conservation law
Z
x|u(t, x)|2 dx − tP~ (t) = constant vector. (7.34)
Rn
(vi) Invariance by the pseudo-conformal transformation:

 x

 x 7→ x̃ = ,

 `(t)

 Z t
 1
t 7→ t̃ = 2 (τ )
dτ, (7.35)

 0 `


 2
 u 7→ ũ(t̃, x̃) = `n/2 u(t, x) exp(−i `t |x| ).

` 4
Assume `tt = 0 and F 0 (|ũ|2 ) = `2 F 0 (|u|2 ), then the equation is invari-

ant under this pseudo-conformal transformation. Let ε ∈ (0, 1/|t|) be an
infinitesimal quantity and `(t) = 1 − εt, then δx = εtx, δt = εt2 and
δu = ε(− n2 t + 4i |x|2 )u. From (7.28), we have the pseudo-conformal conser-
vation law
Z

|xu + 2it∇u|2 + 4t2 F (|u|2 ) dx = const. (7.36)
Rn
Remark 7.1. 1) In fact, x + 2it∇ is a Galilean operator, which is com-

|x|2
mutable with i∂t + ∆. Let J(t) = x + 2it∇ and M (t) = e 4it , then we
have the following relations between J(t) and the Schrödinger semigroup
S(t) = eit∆ or M (t):
J(t) = S(−t)xS(t), J(t)u = 2itM (−t)∇(M (t)u).
2) For the general nonlinearity F 0 (|u|2 )u, we can differentiate (7.36)

with respect to the time t,
Z
d
|xu + 2it∇u|2 + 4t2 F (|u|2 ) dx
dt Rn
Z

=4t (n + 2)F (|u|2 ) − nF 0 (|u|2 )|u|2 dx, (7.37)
Rn
to obtain, for any time t in the time interval where the solution exists, the
following pseudo-conformal conservation law
Z

|xu + 2it∇u|2 + 4t2 F (|u|2 ) dx
Rn
Z Z t Z
2

= |xu(0, x)| dx + 4 τ (n + 2)F (|u|2 ) − nF 0 (|u|2 )|u|2 dxdτ.
Rn 0 Rn
(7.38)
7.3 Virial identity and Morawetz inequality
In the previous subsection, it involves |u|2 , x|u|2 , |x|2 |u|2 , · · · , in the con-
servation laws. Have they some kind of rules to follow? What will happen
if we replace the coefficient in front of |u|2 by a general function? Can we
obtain some other conservation laws or estimates?
7.3. Virial identity and Morawetz inequality 215
Now, let a(t, x) be an arbitrary real-valued function on R1+n , and define

the virial potential Va (t) and the Morawetz5 action Ma (t) associated to a
as
Z
Va (t) := a(t, x)|u(t, x)|2 dx, Ma (t) := ∂t Va (t).
Rn
For convenience, denote G(|u|2 ) = F 0 (|u|2 )|u|2 − F (|u|2 ). By (7.5) and the
integration by parts, we have
Z

Ma (t) = ∂t Va (t) = at |u|2 + 2Im(ū∇a · ∇u) dx, (7.39)
n
ZR
∂t Ma (t) = ∂tt Va (t) = (att − ∆2 a)|u|2 dx
Rn
Z
+4 Im(ū∇at · ∇u)dx
Rn
Z n
X
+4 ajk Re(uk ūj )dx
Rn j, k=1
Z
+2 ∆aG(|u|2 )dx. (7.40)
Rn
In the time independent case a(t, x) = a(x), we have

Z Z Xn
∂tt Va (t) = − ∆2 a|u|2 dx + 4 ajk Re(uk ūj )dx
Rn Rn j, k=1
Z
+2 ∆aG(|u|2 )dx. (7.41)
Rn
On the other hand, in the time dependent case a(t, x) 6= a(x), we get, for
any positive function Q = Q(t, x), that
4Im(ū∇at · ∇u) = |Q−1 (∇at )u − 2iQ∇u|2 − Q−2 |∇at |2 |u|2 − 4Q2 |∇u|2 .
Thus,
Z
∂tt Va (t) = (att − ∆2 a − Q−2 |∇at |2 )|u|2 dx
Rn
Z
+ |Q−1 (∇at )u − 2iQ∇u|2dx
Rn
5 Cathleen Synge Morawetz (1923- ) is a mathematician born in Toronto, Canada.
Morawetz’ research was mainly in the study of the partial differential equations gov-
erning fluid flow, particularly those of mixed type occurring in transonic flow. She is a
professor emerita at the Courant Institute of Mathematical Sciences at the New York
University, where she had also served as the director from 1984 to 1988.
Z n
X
+4 (ajk − Q2 δjk )Re(uk ūj )dx
Rn j, k=1
Z
+2 ∆aG(|u|2 )dx. (7.42)
Rn
The problem is then to choose Q and a such that as many as possible of

these terms are non-negative up to some manageable errors. Next, we show
some examples.
Example 7.1 (Virial identity). Let a(t, x) = |x|2 , then ajk = 2δjk and
∆2 a = 0, thus from (7.41), it yields the virial identity
Z Z Z
∂tt |x|2 |u|2 dx = 8n |∇u|2 dx + 4n G(|u|2 )dx.
Rn Rn Rn
2
For the defocusing case F (|u|2 ) = p+1 |u|p+1 where p > 1, i.e. G(|u|2 ) =
p−1 p+1
R
p+1 |u| , it implies that the convexity of Rn |x|2 |u|2 dx can be controlled
2
by the energy. For the focusing case F (|u|2 ) = − p+1 |u|p+1 with p > 1 + n4 ,
R
i.e. G(|u|2 ) = − p−1 p+1 |u|
p+1
, the quantity ∂tt Rn |x|2 |u|2 dx can be bounded
from
R above by a multiple of the energy, thus if the energy is negative and
2
Rn |x| |u|2 dx is finite, the solution of the equation necessarily has a blowup
in finite time.
Example 7.2 (Morawetz’ inequality). Let a(t, x) = |x|, n > 1, then

x δjk xj xk
∇a = |x| , ajk = |x| − |x| 3 and ∆a = n−1 |x| . From (7.39) and (7.41), we
have
Z
x
Ma (t) =2 Im(ū · ∇u)dx,
Rn |x|
Z Z
1 /0 u|2
|∇
∂t Ma (t) = − (n − 1) (∆ )|u|2 dx + 4 dx
Rn |x| Rn |x|
Z
G(|u|2 )
+ 2(n − 1) dx,
Rn |x|
where
x−y
/y u|2 := |∇u|2 − |
|∇ · ∇u|2 .
|x − y|
1
In the three-dimensional case n = 3, we have ∆( |x| ) = −4πδ, and so
Z
/0 u|2 + G(|u|2 )
|∇
∂t Ma (t) =8π|u(t, 0)|2 + 4 dx.
R3 |x|
Integrating over [T∗ , T ∗ ] with respect to t, we get the Morawetz identity

Z T∗ Z Z T∗
/0 u|2 + G(|u|2 )
|∇
2 dxdt + 4π |u(t, 0)|2 dt
T∗ R3 |x| T∗
Z Z
x x
= Im(ū(T ∗ , x) · ∇u(T ∗ , x))dx − Im(ū(T∗ , x) · ∇u(T∗ , x))dx.
R3 |x| R3 |x|
In the three-dimensional defocusing case, it implies the following Morawetz
inequality
Z T∗ Z Z T∗
/0 u|2 + G(|u|2 )
|∇
dxdt + 2π |u(t, 0)|2 dt
T∗ R3 |x| T∗
. sup ku(t)k2Ḣ 1/2 . (7.43)

T∗ 6t6T ∗
This inequality can be used to demonstrate scattering results for defocusing

Schrödinger equations.
1
In higher dimensional defocusing case (n > 4), we have −∆( |x| ) =
(n−3)
|x|3 > 0 and the Morawetz inequality
Z T∗ Z Z T∗ Z
|u|2 /0 u|2
|∇
(n − 1)(n − 3) 3
dxdt + 4 dxdt
T∗ Rn |x| T∗ Rn |x|
Z T∗ Z
G(|u|2 )
+ 2(n − 1) dxdt
T∗ Rn |x|
. sup ku(t)k2Ḣ 1/2 . (7.44)
T∗ 6t6T ∗
1
But in lower dimensions n = 1, 2, the distribution −∆( |x| ) is quite
nasty so that we can not obtain a simple result as in higher dimensional
cases.
Examplep7.3 (Nakanishi-Morawetz inequality). Let a(t, x) = λ :=

|(t, x)| = t2 + |x|2 , then we can compute
t xj txj |x|2
at = , aj = , atj = − 3 , att = 3 ,
λ λ λ λ
δjk t2 δjk |x|2 − xj xk n |x|2 n−1 t2
ajk = 3 + , ∆a = − = + .
λ λ3 λ λ3 λ λ3
Substituting these into (7.39) and (7.42), we obtain
Z
t 2 x
Ma (t) = |u| + 2Im(ū · ∇u) dx,
Rn λ λ
Z 2
|x| t |x|2
2
∂t Ma (t) = ( 3 − ∆2 a − 6 2 )|u|2 dx
Rn λ λ Q
Z
tx
+ |− u − 2iQ∇u|2 dx
Rn Qλ3
Z
t2 /0 u|2 |x|2
|∇
+4 ( 3 − Q2 )|∇u|2 + dx
Rn λ λ3
Z
+2 ∆aG(|u|2 )dx.
Rn
2 t2
Taking Q := λ3 , we can eliminate some terms to reduce the above identity
as
Z Z Z
|xu + 2it∇u|2 /0 u|2 |x|2
|∇
∂t Ma (t) = − ∆2 a|u|2 dx + dx + 4 dx
Rn Rn λ3 Rn λ3
Z
n−1 t2
+2 ( + 3 )G(|u|2 )dx.
Rn λ λ
Noticing ∆2 a = O(λ−3 ), we have, for G(|u|2 ) > 0, that
Z Z
|xu + 2it∇u|2 /0 u|2 |x|2
|∇
dx + 4 dx
Rn λ3 Rn λ3
Z
t2 G(|u|2 )
+2 dx
Rn λ3
Z
. ∂t Ma (t) + O(λ−3 )|u|2 dx.
Rn
Integrating over (1, T ] with respect to t (we can deal with [−T, −1) in a
similar way), we obtain
Z TZ Z TZ
|xu + 2it∇u|2 |∇/0 u|2 |x|2
dxdt + 4 dxdt
1 Rn λ3 1 Rn λ3
Z TZ
t2 G(|u|2 )
+2 dxdt
1 Rn λ3
Z TZ
. Ma (T ) − Ma (1) + O(λ−3 )|u|2 dxdt
1 Rn
Z T Z
. C(E(0), N (0)) + O(λ−3 )|u|2 dxdt,
1 Rn
where H(t) and N (t) are the energy and the L2 -norm defined before, re-
spectively. For |t| > 1, we have
Z Z Z Z
|u|2 |u|2
3
dxdt . 3
dxdt . N (0).
|t|>1 Rn λ |t|>1 Rn |t|
Taking T → ∞, we get
Z Z Z Z
|xu + 2it∇u|2 /0 u|2 |x|2
|∇
3
dxdt + 4 dxdt
|t|>1 Rn λ |t|>1 Rn λ3
Z Z
t2 G(|u|2 )
+2 dxdt
|t|>1 Rn λ3
.C(E(0), N (0)).
2
In particular, let F (|u|2 ) = p+1 |u|p+1 . Assume that p ∈ (1, ∞) if n =
4 p−1
1, 2, and p ∈ (1, 1 + n−2 ) if n > 3, then F 0 (|u|2 )|u|2 − F (|u|2 ) = p+1 |u|p+1 .
From the previous estimates, it yields
Z Z
t2 |u|p+1
3
dxdt . C(E(0), N (0)). (7.45)
|t|>1 Rn |(t, x)|
For the case |t| 6 1, we split it into two cases. In the case |x| > 1, we
have
ZZ ZZ
t2 |u|p+1
dxdt . |u|p+1 dxdt . C(E(0), N (0)). (7.46)
|t|61 |(t, x)|3 |t|61
|x|>1 |x|>1
In the case |x| 6 1, we get

ZZ ZZ
t2 |u|p+1 |u|p+1
dxdt . dxdt
|t|61 |(t, x)|3 |t|61 |(t, x)|
|x|61 |x|61
ZZ
|u|p+1
. dxdt
|t|61 |t|1−(p+1)ε |x|(p+1)ε
|x|61
Z k|x|−ε ukp+1Lp+1 (|x|61)

. 1−(p+1)ε
dt
|t|61 |t|
Z 1
1
. 1−(p+1)ε
kukp+1
H 1 (|x|61) dt
−1 |t|
. C(E(0), N (0)), (7.47)
6
where we have used the following Hardy’s inequality
k|x|−ε ukLq (|x|61) 6 CkukH 1 , ∀2 6 q < 2∗ . (7.48)

2n/(n − 2), n > 3, n/q − n/2∗ , n 6= 2,
Here 2∗ := 0<ε< In fact, by
∞, n 6 2, n/2q, n = 2.
Hölder’s7 inequality, we have, for n 6= 2, that
k|x|−ε ukLq (|x|61) 6k|x|−ε kLq2∗ /(2∗ −q) (|x|61) kukL2∗ (|x|61))
6 Godfrey Harold Hardy (1877-1947) was a prominent English mathematician, known
for his achievements in number theory and mathematical analysis.

7 Otto Ludwig Hölder (1859-1937) was a German mathematician born in Stuttgart. He
is famous for many things including: Hölder’s inequality, the Jordan-Hölder theorem,
Hölder’s theorem and the Hölder condition which is used e.g. in the theory of partial
differential equations and function spaces.
∗ −q
Z 1 ∗
2q2∗
−ε 2q2
∗ −q +n−1
6C r dr kukL2∗ (|x|61))
0
6CkukH 1 .
For the case n = 2, from Hölder’s inequality, we get
k|x|−ε ukLq (|x|61) 6k|x|−ε kL2q (|x|61) kukL2q (|x|61))
Z 1 1/2q
6C r−2qε+n−1 dr kukL2q (|x|61))
0
6CkukH 1 .
From (7.45), (7.45) and (7.45), we obtain the Nakanishi-Morawetz in-
equality (cf. [176])
Z
t2 |u|p+1
3
dxdt . C(E(0), N (0)), ∀n ∈ N. (7.49)
R1+n |(t, x)|
7.4 Morawetz’ interaction inequality
Let a(x) be an arbitrary real-valued function on Rn , and define the virial

interaction potential V a (t) and Morawetz’ interaction potential M a (t) as-
sociated to a as
Z Z
V a (t) := a(x − y)|u(t, x)|2 |u(t, y)|2 dxdy, M a (t) := ∂t V a (t).
Rn Rn
From (7.5) and the integration by parts, we have
Z Z h
Ma (t) = 2 ∇a(x − y) · Im(∇u(t, x)ū(t, x))|u(t, y)|2
Rn Rn
i
− Im(∇u(t, y)ū(t, y))|u(t, x)|2 dxdy, (7.50)
and
∂t Ma (t)
Z Z
=−2 ∆2 a(x − y)|u(t, x)|2 |u(t, y)|2 dxdy
Rn Rn
Z Z
+2 ∆a(x − y)[G(|u(x)|2 )|u(y)|2 + G(|u(y)|2 )|u(x)|2 ]dxdy
Rn Rn
Z Z n
X h
+4 ajk (x − y) Re(uk (t, x)ūj (t, x))|u(t, y)|2
Rn Rn j, k=1
+ Re(uk (t, y)ūj (t, y))|u(t, x)|2

7.4. Morawetz’ interaction inequality 221
− Im(uk (t, y)ū(t, y))Im(uj (t, x)ū(t, x))

i
− Im(uk (t, x)ū(t, x))Im(uj (t, y)ū(t, y)) dxdy. (7.51)
Denote
Aj (t, x, y) := u(t, x)ūj (t, y) + uj (t, x)ū(t, y),
Bj (t, x, y) := u(t, x)uj (t, y) + uj (t, x)u(t, y),
then, the last integral can be reduced to
Z Z Xn
2 ajk (x − y)[Aj (t, x, y)Ak (t, x, y) + Bj (t, x, y)Bk (t, x, y)]dxdy,
Rn Rn j, k=1
where we can exchange k and j since ajk = akj .

For convenience, we denote the above quantity by I. If a is a radial
0 0
(|x|) xk xj
function, i.e. a(x) = a(|x|), then ajk (x) = a |x| (|x|)
δjk +(a00 (|x|)− a |x| ) |x| |x| .
From the Cauchy-Schwartz inequality, we obtain
Z Z n
a0 (|x − y|) X
I =2 [|Aj (t, x, y)|2 + |Bj (t, x, y)|2 ]dxdy
Rn Rn |x − y| j=1
Z Z 2
a0 (|x − y|) h X xj − yj
n

+2 (a00 (|x − y|) − ) Aj (t, x, y)
R n R n |x − y| j=1
|x − y|
X n
xj − yj 2 i

+ Bj (t, x, y) dxdy
j=1
|x − y|
Z Z h Xn 2
xj − yj
>2 a00 (|x − y|) Aj (t, x, y)
Rn Rn j=1
|x − y|
X n
xj − yj 2 i

+ Bj (t, x, y) dxdy.
j=1
|x − y|
It is clear that I > 0 if a00 (λ) > 0 for λ > 0. Hence, (7.51) and (7.50) imply
Theorem 7.4 (Morawetz’ interaction inequality). Let a(x) be a real-

valued radial convex function, and u(t, x) be a solution of (7.5), then it
holds
Z Z
(−∆2 a(x − y))|u(t, x)|2 |u(t, y)|2 dxdy
Rn Rn
Z Z
+ ∆a(x − y)[G(|u(t, x)|2 )|u(t, y)|2 + G(|u(t, y)|2 )|u(t, x)|2 ]dxdy
Rn Rn
1
6 ∂t Ma (t), (7.52)
2
where Ma (t) is defined in (7.50).
Next, we consider an example of three dimensional nonlinear

Schrödinger equations.
Example 7.4 (Morawetz’ interaction inequality of 3D NLS).
x 2
Let n = 3, a(x) = |x|, then ∇a = |x| , ∆a = |x| > 0, and −∆2 a = 8πδ(x).
From (7.52), we get
Z T∗ Z
8π |u(t, x)|4 dxdt
T∗ R3
Z T∗ Z Z
G(|u(t, x)|2 )|u(t, y)|2 + G(|u(t, y)|2 )|u(t, x)|2
+2 dxdydt
T∗ R3 R3 |x − y|
1
6 (Ma (T ∗ ) − Ma (T∗ ))
Z2 Z h
x−y
= Im( · ∇u(T ∗ , x)ū(T ∗ , x))|u(T ∗ , y)|2
Rn Rn |x − y|
x−y i
− Im( · ∇u(T ∗ , y)ū(T ∗ , y))|u(T ∗ , x)|2 dxdy
|x − y|
Z Z h
x−y
− Im( · ∇u(T∗ , x)ū(T∗ , x))|u(T∗ , y)|2
Rn Rn |x − y|
x−y i
− Im( · ∇u(T∗ , y)ū(T∗ , y))|u(T∗ , x)|2 dxdy
|x − y|
. N 2 (0) sup ku(t)k2 1 . (7.53)
T∗ 6t6T ∗ Ḣ 2
7.5 Scattering results for NLS
The Morawetz estimates are time-decaying estimates, which are very useful
for us to study the existence of the scattering operators. In this section,
we only state some results on the existence of the scattering operators for
NLS
iut + ∆u = |u|p u, u(0, x) = u0 (x). (7.54)
The existence of the scattering operators in the H 1 -subcritical but L2 -
supercritical cases was shown by Nakanishi [176] in 1 and 2 spatial dimen-
sions and by Ginibre and Velo [79] in higher spatial dimensions n > 3:
Theorem 7.5. Let p ∈ (4/n, 4/(n − 2)) for n > 3, and p ∈ (4/n, ∞) for
n = 1, 2. For any initial datum u0 ∈ H 1 with E(u0 ) < ∞, (7.54) has a
p(n+2)/2
unique global solution u ∈ Ct0 (Ḣx1 ) ∩ Lt,x satisfying
Z ∞Z
|u(t, x)|p(n+2)/2 dxdt 6 C(E(u0 )),
−∞ Rn
7.5. Scattering results for NLS 223
where
Z
1 2
E(u(t)) = |∇u(t, x)|2 + |u(t, x)|2+p dx
Rn 2 2+p
is the energy conserved quantity. Moreover, there exist finite-energy solu-
tions u± of the linear Schrödinger equation (i∂t + ∆)u± = 0 such that
ku± (t) − u(t)kḢ 1 → 0, as t → ±∞,
x
and the mapping u0 7→ u± (0) is a homeomorphism from Ḣ 1 (Rn ) to itself.
Next, we introduce briefly the related results for the scattering theory
of H 1 -critical defocusing NLS
4
iut + ∆u = |u| n−2 u, u(0, x) = u0 (x). (7.55)
In 1999, J. Bourgain [19] proved the global wellposedness of the solution
with radial initial data in 3 and 4 dimensions. Then M. Grillakis [84]
gave a different proof for Bourgain’s results in 2000. In 2005, these re-
sults were generalized to any spherically symmetric initial data in any
dimensions n > 3 by T. Tao [217]. At the same time (the paper was
published in 2008), J. Colliander, M. Keel, G. Staffilani, H. Takaoka, and
T. Tao [44] obtained the global well-posedness and the scattering theory
for any general initial data in three dimensions based on the induction
arguments on the energy of Bourgain’s. They applied the induction ar-
gument not only on the frequency space but also on the physical space,
and replaced the original Morawetz’ inequality (7.43) by Morawetz’ in-
teraction inequality (7.53) in order to deal with the non-radial cases. In
addition, the Morawetz’ interaction inequality, together with the almost
conserved quantity of the momentum which controlled the mass in the fre-
quency space, described the possibilities of the energy concentration. For
the higher dimensional case, it is generalized by E. Ryckman and M. Visan
[196] for n = 4 and by M. Visan [232] for n > 5. Thus, the scattering in
the energy class is now known for (7.55) with any general initial data in
all dimensions n > 3, namely, as stated in the following results [44; 196;
232].
Theorem 7.6. Let n > 3. For any initial datum u0 ∈ H 1 with E(u0 ) < ∞,
2(n+2)
then (7.55) yields a unique global solution u ∈ Ct0 (Ḣx1 ) ∩ Lt,xn−2 such that
Z ∞ Z
2(n+2)
|u(t, x)| n−2 dxdt 6 C(E(u0 )),
−∞ Rn
where
Z
1 n−2 2n
E(u(t)) = |∇u(t, x)|2 + |u(t, x)| n−2 dx
Rn 2 2n
is the energy conserved quantity.
The above theorem implies the scattering results [44; 232; 196]:
Corollary 7.1. Let n > 3. Assume that u0 has finite energies, u is the
2(n+2)
unique global solution of (7.55) on u ∈ Ct0 (Ḣx1 ) ∩ Lt,xn−2 . Then, there exist
finite-energy solutions u± of the linear Schrödinger equation (i∂t + ∆)u± =
0 such that
ku± (t) − u(t)kḢ 1 → 0, t → ±∞,
x
and the mapping u0 7→ u± (0) is a homeomorphism from Ḣ 1 (Rn ) to itself.
Along with the perfect solution of the scattering theory of the energy-
critical defocusing NLS, most researchers transfer their interests into an-
other important problem – the global well-posedness and scattering theories
of the mass-critical defocusing NLS
4
iut + ∆u = |u| n u, u(0, x) = u0 (x). (7.56)
At present, there are some results in this aspect:
• n > 1, for sufficiently small initial datum u0 ∈ L2 (Rn ), the solution
of (7.56) is global well-posed and scatters.
• n > 2, for radial initial data u0 ∈ L2 (Rn ), the solution of (7.56) is
global well-posed and scatters, cf. [138; 140].
• n > 1, for any initial data u0 ∈ L2 (Rn ), the solution of (7.56) is
global well-posed and scatters, cf. [64; 65; 66].
For other nonlinearities, there are also some interesting results. K.
Nakanishi established the scattering theory for the mixed power nonlinear-
ity |u|p1 u+|u|p2 u for 4/n < p1 6 p2 < 4/(n−2) and 4/n 6 p1 < p2 6 2∗ −2
in the energy space H 1 , cf. [176; 179; 220]. For the nonlinearity (V ∗ |u|2 )u,
i.e. the Hartree equation, he also proved the existence of the scattering
operator in H 1 for V ∈ Lp1 + Lp2 , p1 , p2 > 1 and n/4 < p2 6 p1 < n/2,
cf. [175]. For the nonlinear Schrödinger equation with the following non-
linearity
X λ2k
f (u) = µ |u|2k u, µ, λ > 0, n = 1, 2, (7.57)
(2k)!
k>1+2/n
7.5. Scattering results for NLS 225
the scattering operator also exists under some assumptions, cf. [244].
In addition, there is a great process for focusing nonlinear Schrödinger
equations

iut + ∆u = −|u|p u,
(7.58)
u(0, x) = u0 (x).
For Ḣ 1 -critical nonlinearities in n = 3, 4, 5 dimensions, i.e. p = 4/(n −

2), the solution of (7.58) is global well-posed and scatters for any radial
initial datum u0 ∈ Ḣ 1 (Rn ) whose energy and kinetic energy are less than
those of the ground state, cf. [124]; For the n > 5 dimensional case, the
condition “radial” can be eliminated and it holds that the solution of (7.58)
is global well-posed and scatters for any initial datum u0 ∈ Ḣ 1 (Rn ) whose
energy and kinetic energy are less than those of the ground state, cf. [139].
Otherwise, in the n > 4 dimensions, for a global solution u to the focusing
mass-critical nonlinear Schrödinger equation (7.58) (where p = 4/n) with
spherically symmetric H 1 initial data and mass equal to that of the ground
state Q, u does not scatter then, up to the phase rotation and the scaling,
u is the solitary wave eitQ . And the only spherically symmetric minimal-
mass non-scattering solutions are, up to the phase rotation and the scaling,
the pseudo-conformal ground state and the ground state solitary wave, cf.
[137]. Recently, scattering for the focusing energy-subcritical NLS has been
studied by Fang, Xie and Cazenave [72].
Chapter 8
Boltzmann equation without angular

cutoff
The kinetic theory of the gas is a theory devoted to the study of evolutionary
behaviors of the gas in the one-particle phase space of position and velocity.
Recently, the kinetic theory is getting more and more recognized to be
significant both in mathematics and practical applications as a key theory
connecting the microscopic and macroscopic theory of gases and fluids. In
this sense, the kinetic theory is in-between or mesoscopic.
In the macroscopic scales where the gas and fluid are regarded as a
continuum, their motion is described by the macroscopic quantities such
as macroscopic mass density, bulk velocity, temperature, pressure, stresses,
heat flux and so on. The Euler and Navier-Stokes equations, compressible or
incompressible, are the most famous equations among governing equations
proposed so far in fluid dynamics.
The extreme contrary is the microscopic scale where the gas, fluid, and
hence any matter, are looked at as a many-body system of microscopic
particles (atom/molecule). Thus, the motion of the system is governed
by the coupled Newton equations, within the framework of the classical
mechanics. The number of the involved equations is 6N if the total number
of the microscopic particles is N .
Although the Newton equation is the first principle of the classical me-
chanics, it is not of practical use because the number of the equations is
so enormous (N ∼ the Avogadro number 6 × 1023 ) that it is hopeless to
specify all the initial data, and we must appeal to statistics. On the other
hand, the macroscopic (fluid dynamical) quantities mentioned above are
related to statistical average of quantities depending on the microscopic
state. Thus, the kinetic theory that gives the mesoscopic descriptions of
the gas and fluid is noticed to be a key theory that links the microscopic and
227
228 Boltzmann equation without angular cutoff
macroscopic scales. The Boltzmann equation, which is the subject of these

notes, is the most classical but fundamental equation in the mesoscopic
kinetic theory.
The goal of this chapter is to introduce, and make a preliminary dis-
cussion of the Boltzmann equation. We mainly use harmonic analysis to
consider the Boltzmann equation without angular cutoff. We will show
that the Boltzmann collision operator without angular cutoff behaves like
a singular integral operator or pseudo-differential operator whose leading
term is characterized by the operator (−∆)ν/2 .
This was first pointed out by Pao [188], see also [228], and was for-
mulated explicitly by Lions [160]. The optimal Sobolev exponent ν/2 is
due to Villani [231]. In the middle of 1990s, Desvillettes managed to prove
the regularity of solutions to some simplified models for the spatially ho-
mogeneous problems, cf. [57; 58]. Around 2000s, the regularity induced
by the grazing collision was analyzed in terms of the entropy production
integral, cf. the work [2] and others. In particular, [2] establishes several
elegant formulations associated with the collision operator which have been
essentially used to the study of the spatially homogeneous problem.
8.1 Models for collisions in kinetic theory
In this section, we will first give the derivation of the Boltzmann equation,
in details, one can refer to [28; 59; 229].
8.1.1 Transport model

The object of kinetic theory is the modelling of a gas (or plasma, or any
system made up of a large number of particles) by a distribution function
in the particle phase space. This phase space includes macroscopic vari-
ables, i.e. the position in physical space, but also microscopic variables,
which describe the state of the particles. In the present survey, we shall
restrict ourselves, most of the time, to systems made of a single species
of particles (no mixtures), and which obey the laws of classical mechanics
(non-relativistic, non-quantum).
Assume that the gas is contained in a (bounded or unbounded) domain
Ω ⊂ Rn (n = 3 in applications) and observed on a time interval [0; T ], or
[0; +∞). Then, under the above simplifying assumptions, the corresponding
8.1. Models for collisions in kinetic theory 229
kinetic model is a nonnegative function f (t; x; v), defined on [0; T ]× Ω× Rn.

Here the space Rn = Rnv is the space of possible velocities, and should
be thought of as the tangent space to Ω. f (t; x; v) is the phase space
density of particles which at time t and point x move with velocity v,
where x = (x1 , x2 , ...xn ) and v = (v1 , v2 , ...vn ).
When there is no interaction between the particles and their surrounding
environment (including themselves), they will move at a constant velocity
and along straight lines. In other words, for all times t and s, point x and
velocity v, a particle which at time t sits at point x and move with velocity
v will sit at time t + s at point x + vs and will keep its velocity v. This
entails that
f (t + s, x + vs, v) = f (t, x, v), for any s, (8.1)
or, after differentiation with respect to s
∂t f + v · ∇x f = 0. (8.2)
(8.2) is free transport model. v·∇x is the transport operator, where v·∇x =
Pn ∂
k=1 vk ∂xk . (8.2) implies that the change rate of density function is zero
if there is no collisions of the particles.
When a given force F (t, x) acts on the particles (such a force can also de-
pend on v in specific situations, for example when the particles are charged
and feel the action of a magnetic field, or when the force is the drag force
due to a surrounding gas), the particles will follow the trajectories of the
following system of differential equations:
ẋ(t) = v(t) (8.3)
v̇(t) = F (t, x(t)) (8.4)

and the corresponding partial differential equation satisfied by f (that is,
the PDE whose characteristic curves are exactly the solutions of (8.3) and
(8.4)), is the Vlasov equation
∂t f + v · ∇x f + F (t, x) · ∇v f = 0, (8.5)
Pn
where F (t, x) · ∇v = k=1 Fk ∂v∂k .
8.1.2 Boltzmann model

When the forces acting on the particles are mainly due to the collisions of
the particles between themselves, one is led to write down the Boltzmann
equation. (Very often, particles will be assumed to interact via a given
interaction potential φ(r), r distance between particles). We first shall

make several postulates:
(1). We assume that particles interact via binary collisions : this is a
vague term describing the process in which two particles happen to come
very close to each other, so that their respective trajectories are strongly
deviated in a very short time. Underlying this hypothesis is an implicit
assumption that the gas is dilute enough that the effect of interactions
involving more than two particles can be neglected. Typically, if we deal
with a three-dimensional gas of n hard spheres of radius r, this would mean
N r3 1, N r2 ' 1. (8.6)
(2). Moreover, we assume that these collisions are localized both in
space and time, i.e. they are brief events which occur at a given position x
and a given time t. This means that the typical duration of a collision is
very small compared to the typical time scale of the description, and also
quantities such as the impact parameter (see below) are negligible in front
of the typical space scale (say, a space scale on which variations due to the
transport operator are of order unity).
(3). Next, we further assume these collisions to be elastic : momentum
and kinetic energy are preserved in a collision process.
(4). We also assume collisions to be microreversible. This word can
be understood in a purely deterministic way : microscopic dynamics are
time-reversible; or in a probabilistic way : if let (v, v∗ ) stand for the ve-
locities before collision, and (v 0 , v∗0 ) stand for the velocities after collision;
the probability that velocities (v, v∗ )are changed into (v 0 , v∗0 ) in a collision
process, is the same as the probability that (v 0 , v∗0 ) are changed into (v, v∗ ).
(5). Finally, we make the Boltzmann chaos assumption : the velocities of
two particles which are about to collide are uncorrelated. Roughly speaking,
this means that if we randomly pick up two particles at position x, which
have not collided yet, then the joint distribution of their velocities will be
given by a tensor product (in velocity space) of f with itself. Note that
this assumption implies an asymmetry between past and future: indeed, in
general if the pre-collisional velocities are uncorrelated, then post-collisional
velocities have to be correlated!
Thus we only take into account interactions between particles of rarefied
gas, equation (8.2) can be written by a generalized form
∂t f + v · ∇x f = Q(f )(v). (8.7)
(8.7) means that change rate of density function of particles depends on
itself, is a function of f , and is not zero when there exist collision between
particles. Next we will derive the form of Q(f ).
We denote by f2 (v1 , v2 ) the joint density of two particles with respective

velocities v1 and v2 . It follows (due to the assumption that the collisions
are binary) that we must take into account only two distinct phenomena
which modify the number density of particles with velocity v.
First, because of a possible collision with a particle of velocity v∗ , a
particle which had v for velocity will end up with a velocity v 0 (its partner
in the collision will end up with velocity v∗0 ).
Secondly, some particle with a velocity w will encounter a particle with
velocity w∗ and will end up with a velocity v after the collision (its partner
in the collision will end up with velocity w∗0 ).
We now denote by p(v1 , v2 → v3 , v4 ) the (density of) probability that for
two particles sitting at the same point x at a given time t, a collision occurs
and transforms the ingoing velocities v1 and v2 in the outgoing velocities
v3 and v4 (we shall see that in the so-called non cutoff case, this quantity is
in fact far from being a probability density, since it is not integrable).
We see that Q(f ) is the sum of two terms −Q− (f ) and Q+ (f ) which
respectively correspond to the two phenomena described above. According
to their definition Q− (f ) and Q+ (f ) write
Z Z Z
Q− (f )(v) = f2 (v, v∗ )p(v, v∗ → v 0 , v∗0 )dv∗0 dv 0 dv∗ , (8.8)
v∗ v0 v∗0
Z Z Z
+
Q (f )(v) = f2 (w, w∗ )p(w, w∗ → v, w∗0 )dw∗0 dw∗ dw. (8.9)
w w∗ w∗0
From the assumption (4), it follows that

∀ v1 , v2 , v3 , v4 , p(v1 , v2 → v3 , v4 ) = p(v3 , v4 → v1 , v2 ).
Moreover, using the assumption (5), since v and v∗ are irrelative variables,
w and w∗ are irrelative variables; we have that f2 (v, v∗ ) = f (v)f (v∗ ) and
f2 (w, w∗ ) = f (w)f (w∗ ). Let Q(f ) = Q(f, f ) = Q+ (f, f ) − Q− (f, f ), where
Z Z Z
−
Q (f, f )(v) = f (v)f (v∗ )p(v, v∗ → v 0 , v∗0 )dv∗0 dv 0 dv∗ , (8.10)
v∗ v0 v∗0
Z Z Z
Q+ (f, f )(v) = f (v 0 )f (v∗0 )p(v 0 , v∗0 → v, v∗ )dv∗0 dv 0 dv∗ . (8.11)
v∗ v0 v∗0
We have
Z Z Z
Q(f, f )(v) = f (v 0 )f (v∗0 ) − f (v)f (v∗ ) p(v, v∗ → v 0 , v∗0 )dv∗0 dv 0 dv∗ .
v∗ v0 v∗0
(8.12)
Let us explain (8.12) a little bit. This operator can formally be split, in a
self-evident way, into a gain and a loss term, Q(f ; f ) = Q+ (f ; f )−Q−(f ; f ):
The loss term counts all collisions in which a given particle of velocity v will
encounter another particle, of velocity v∗ . As a result of such a collision,
this particle will in general change its velocity, and this will make less
particles with velocity v. On the other hand, each time particles collide
with respective velocities v 0 and v∗0 , then the v 0 particle may acquire v
as new velocity after the collision, and this will make more particles with
velocity v : this is the meaning of the gain term.
Next, we only need to give the expression of p(v, v∗ → v 0 , v∗0 ). Using the
conservation of momentum and kinetic energy in a collision, we have

v + v∗ = v 0 + v∗0 ,
(8.13)
|v|2 + |v∗ |2 = |v 0 |2 + |v∗0 |2 .
Since this is a system of n + 1 scalar equations for 2n scalar unknowns,

it is natural to expect that its solutions can be defined in terms of n − 1
parameters. Here is a convenient representation of all these solutions, which
we shall sometimes call the σ-representation:
v + v∗ |v − v∗ | v + v∗ |v − v∗ |
v0 = + σ, v∗0 = − σ. (8.14)
2 2 2 2
Here the parameter σ ∈ S n−1 varies in the n − 1 unit sphere. One can see
the following figure
v0
θ/2 θ
v∗ v
k
v∗0
Fig. 8.1 The geometry of collisions

The Galilean invariance which holds in the context of binary collisions

entails that the measure p(v, v∗ → v 0 , v∗0 ) can only depend on |v − v∗ | and
v−v∗ 0 0
|v−v∗ | · σ, p(v, v∗ → v , v∗ ) can be written by B(·, ·). We now can write
down the final form of Boltzmann’s collision operator :
Z Z
v − v∗
Q(f, f ) = B(|v − v∗ |, · σ){f (v∗0 )f (v 0 ) − f (v∗ )f (v)}dσdv∗ ,
Rn S n−1 |v − v∗ |
(8.15)
where function B is called the Boltzmann collision cross section (or the
collision kernel). For convenience, we shall also use the bilinear form Q(g, f )
related to the quadratic form Q(f, f ) and defined by
Z Z
Q(g, f ) = B(|v − v∗ |, σ){g(v∗0 )f (v 0 ) − g(v∗ )f (v)}dσdv∗ . (8.16)
Rn S n−1
For simplicity, we sometimes use standard abbreviations: g(v∗0 )f (v 0 ) = g∗0 f 0

and g(v∗ )f (v) = g∗ f . Finally, we write down the standard form of the
Boltzmann equation
∂t f + v · ∇x f = Q(f, f )(v), (8.17)
where Q(f, f ) is given by (8.15). The spatially homogeneous Boltzmann
equation describes the behavior of a dilute gas, in which the velocity dis-
tribution of particles is assumed to be independent of the position; it reads
∂t f = Q(f, f )(v), (8.18)
where the unknown f = f (t, v) is assumed to be nonnegative, and stands
for the density of particles at time t with velocity v. These equations are
called homogeneous since f is assumed to be independent of the position.
This model is simple, in this chapter we only consider (8.18).
8.1.3 Cross section

In most cases, the collision kernel B cannot be expressed explicitly. How-
ever, to capture its main properties, it is usually assumed to have the form:
v − v∗ π
B(|v − v∗ |, σ) = Φ(|v − v∗ |)b(cos θ), cos θ = h , σi, 0 6 θ 6 ,
|v − v∗ | 2
(8.19)
where the deviation angle θ is the angle between pre- and post-collisional
velocities. Using n-dimensional spherical coordinates, we have
Z Z π2
n−2
B(|v − v∗ |, σ)dσ = |S | B(|v − v∗ |, cos θ) sinn−2 θdθ,
S n−1 0
where |S n−2 | denotes the surface area of unit sphere in n dimension.

The collision kernel should be computed in terms of the interaction
potential φ. Here, we consider the important potential: inverse-power law
potentials. It is possible to (almost) explicitly compute the cross section
if potential φ(r) is inverse-power law potentials, that is, the inter particle
force is proportional to r−s (with r denoting the interparticle distance and
s > 2). In such a case (and in dimension 3) B writes
s−5
B(|v − v∗ |, σ) = |v − v∗ | s−1 b(cos θ) (8.20)
with b > 0 a smooth function except at point 1 and satisfying:
Z π2
K
sinn θb(cos θ)dθ < ∞ and sinn−2 θb(cos θ) ≈ s+1 (8.21)
0 θ s−1
when θ → 0, K > 0. Since s+1 s−1 > 1, b is not integrable at θ = 0. This
means
Z Z π2
n−2
b(k · σ)dσ = |S | sinn−2 θb(cos θ)dθ = ∞.
S n−1 0
Because of the difficulties entailed by this singularity, Grad has proposed
to introduce an angular cutoff near θ = 0. It means that we replace B in
(8.20) and (8.21) by a new cross section cross
s−5
B̃(|v − v∗ |, σ) = |v − v∗ | s−1 b̃(cos θ), (8.22)
where b̃ is smooth, or at least such that sinn−2 θb̃(cos θ) is integrable near
θ = 0. In our model case, this means
Z Z π2
n−2
b̃(k · σ)dσ = |S | sinn−2 θb̃(cos θ)dθ < ∞.
S n−1 0
In the sequel,we shall speak of cutoff cross sections (cutoff potentials or
Grad’s angular cut-off) when B is locally integrable, and of non cutoff
cross sections (or non cutoff potentials) when B has a singularity like in
(8.21).
In this chapter, we mainly consider the Boltzmann equation without
cutoff cross section. Note that the separation Q(f, f ) = Q+ (f, f )−Q− (f, f )
make sense if B is cutoff cross section, where
Z Z
v − v∗
Q+ (f, f ) = B(|v − v∗ |, · σ)f (v∗0 )f (v 0 )dσdv∗ , (8.23)
Rn S n−1 |v − v∗ |
Z Z
v − v∗
Q− (f, f ) = B(|v − v∗ |, · σ)f (v∗ )f (v)dσdv∗ . (8.24)
R n S n−1 |v − v∗ |
If B is non cutoff cross section, for convenience, we sometimes need to cal-
culate the gain and loss terms Q± in the collision operator separately. This
8.2. Basic surgery tools for the Boltzmann operator 235
separation is not allowed for the non-cutoff cross section but is legitimate
as an intermediate step in the calculation.
In this chapter, we mainly consider the inverse-power law potentials,
the cross section B have the form:
B(|v − v∗ |, σ) = |v − v∗ |γ b(cos θ). (8.25)
If γ > 0, it is called hard potentials; in general, we consider the case of
0 < γ 6 2.
If γ = 0, it is called Maxwellian potentials.
If −n < γ < 0, it is called soft potentials; for −2 < γ < 0, it is called
moderately soft potentials, for γ 6 −2 it is called very soft potentials.
Note that if s = 2 then φ(r) becomes Coulomb potential, which leads
to Fokker-Planck-Landau equation, we do not consider this model in this
book.
8.2 Basic surgery tools for the Boltzmann operator
In this section, we give some basic tools which one often needs for a fine
study of the Boltzmann operator. In details, one can refer to [59; 229].
1.Symmetrization of the collision kernel. In view of formulas (8.14),
the quantity f∗0 f 0 − f∗ f is invariant under the change of variables σ →
−σ. From the physical point of view this reflects the undiscernability of
particles. Thus one can replace (from the very beginning, if necessary) B
by its symmetrized version
B̄(z, σ) = {B(z, σ) + B(z, −σ)}Iz·σ>0 .
This is the reason why we assume that angle θ varies from 0 to π2 in (8.19).
In fact, we get rid of collisions with deflexion angle larger than π2 by this
symmetrization trick. B in (8.19) is indeed B̄ in the above equality.
2. Pre-postcollisional change of variables. A universal tool in the
Boltzmann theory is the involutive change of variables with unit Jacobian
v−v∗
(v, v∗ , σ) → (v 0 , v∗0 , k), where k = |v−v ∗|
is the unit vector along v − v∗ .
Indeed, denote the mapping Φ : (v, v∗ , σ) → (v 0 , v∗0 , k), then it’s easy to
verify that Φ2 = I. Thus the Jacobian of Φ is 1. Then for any function f
we have
Z Z Z
f (v, v∗ , v 0 , v∗0 , σ)dσdv∗ dv
Rn Rn S n−1
Z Z Z (8.26)
v − v∗
= f (v 0 , v∗0 , v, v∗ , )dσdv∗ dv.
Rn Rn S n−1 |v − v∗ |
3. The change of variables (v, v∗ , σ) → (v∗ , v, σ) is clearly involutive

and has unit Jacobian. For function f = f (v, v∗ , v 0 , v∗0 , σ), one has
Z Z Z
f (v, v∗ , v 0 , v∗0 , σ)dσdv∗ dv
Rn Rn S n−1
Z Z Z (8.27)
= f (v∗ , v, v 0 , v∗0 , σ)dσdv∗ dv.
Rn Rn S n−1
Next, we give some applications of the two change of variables above.
We first give the conservation laws of the Boltzmann equation. As a con-
sequence, if ϕ(v) is an arbitrary continuous function of the velocity v, then
we get the following various weak formulations for Boltzmann’s kernel Q
Z
Q(g, f )ϕ(v)dv (8.28)
Rn
Z Z
v − v∗
= B(|v − v∗ |, · σ){g(v∗0 )f (v 0 ) − g(v∗ )f (v)}ϕ(v)dσdv∗ dv
R 2n S n−1 |v − v∗ |
Z Z
v − v∗
= B(|v − v∗ |, · σ)g(v∗ )f (v)(ϕ(v 0 ) − ϕ(v))dσdv∗ dv.
R 2n S n−1 |v − v∗ |
In the special case g = f , we have that
Z
Q(f, f )ϕ(v)dv
Rn
Z Z
1
= Bf (v∗ )f (v)(ϕ(v∗0 ) + ϕ(v 0 ) − ϕ(v∗ ) − ϕ(v))dσdv∗ dv.
2 Rn ×Rn S n−1
From the mathematical point of view, it is interesting because expres-
sions like (8.28) may be well-defined in situations where Q(f, f ) is not.
From
R the physical point of view, it expresses the change in the integral
Rn
Q(f, f )(v)dv which is due to the action of collisions. Let f be a solu-
tion of the Boltzmann equation (8.17), set in the whole space Rn to simplify.
By the conservative properties of the transport operator v · ∇x ,
Z Z
d
f (t, x, v)ϕ(v)dxdv = Q(f, f )ϕ(v)dxdv (8.29)
dt Rn Rn
and the right-hand side is just the x-integral of any one of the expressions
in formulas (8.28). As an immediate consequence, whenever ϕ satisfies the
functional equation
∀(v, v∗ , σ) ∈ Rn × Rn × S n−1 , ϕ(v 0 ) + ϕ(v∗0 ) = ϕ(v) + ϕ(v∗ ), (8.30)
then, at least formally,
Z
d
f (t, x, v)ϕ(v)dxdv = 0 (8.31)
dt Rn
along solutions of the Boltzmann equation. The words at least formally of

course mean that the preceding equations must be rigorously justified with
the help of some integrability estimates on the solutions to the Boltzmann
equation.
Pluging ϕ(v) = 1, v1 , ..., vn , |v|2 /2, in formula (8.28) and (8.31), by
(8.30), we get the conservation of mass, momentum and energy at the level
of the Boltzmann operator:
Z Z Z
|v|2
Q(f, f )(v)dv = 0, Q(f, f )(v)vj dv = 0, Q(f, f )(v) dv = 0,
Rn Rn Rn 2
(8.32)
or,
Z Z Z
d d d |v|2
f (t, x, v)dxdv = f (t, x, v)vj dv = f (t, x, v)) dv = 0.
dt Rn dt Rn dt Rn 2
(8.33)
Next, we give the Boltzmann’s H theorem. Without caring about inte-
grability issues, we plug ϕ(v) = log f (v) into (8.28),
Z
Q(f, f )(v) log f (v)dv = −D(f ),
Rn
where D is the entropy dissipation functional. Using the fact that (F −

G)(log F − log G) > 0, we have
Z Z
1 v − v∗
D(f ) = B(|v − v∗ |, · σ) f (v∗0 )f (v 0 ) − f (v∗ )f (v)
4 Rn ×Rn S n−1 |v − v∗ |
(f (v∗ )f (v 0 )
0
× log dσdv∗ dv > 0. (8.34)
f (v∗ )f (v)
Now, we introduce Boltzmann’s H functional,
Z
H(f ) = f log f dxdv.
Rn n
x ×Rv
Of course, the transport operator −v ·∇x does not contribute in any change
of the H functional in time. As a consequence, if f = f (t, x, v) is a solution
of the Boltzmann equation, then H(f ) will evolve in time because of the
effects of the collision operator:
Z
d
H(f (t, ·, ·)) = − D(f (t, ·, ·))dx 6 0. (8.35)
dt Rn
x
This is the first part of Boltzmann’s H-theorem. Next we assume that

B(|v − v∗ |, σ) > 0 for almost all (v, v∗ , σ), which is always the case in
applications of interest. Then equality in Boltzmann’s H theorem occurs if

and only if for almost all x, v, v∗ , σ
f (v∗0 )f (v 0 ) = f (v∗ )f (v). (8.36)
Then, it is possible to prove (under suitable, but rather weak assumptions
on B and f ) that
D(f ) = 0 ⇐⇒ Q(f, f ) = 0 ⇐⇒ f = M (v), (8.37)
ρ − |v−u|
2
M (v) = Mρ,u,T (v) = e 2T , (8.38)
(2πT )n/2
where M is called the Maxwellian and is known to describe the velocity
distribution of a gas in an equilibrium state with the mass density ρ(x), bulk
velocity u(x) and temperature T . Here, (ρ, u, T ) are taken to be parameters.
If (ρ, u, T ) are constants,then M is called a global ( absolute) Maxwellian ;
if they are functions of (x, t), then it is called a local Maxwellian. Evidently,
the global Maxwellian is a stationary solution Boltzmann equation (8.17)
without force F . This is the second part of Boltzmann’s H-theorem.
4. Bobylev’s identities . We now turn to more intricate tools introduced
by Bobylev, who first made Fourier transform an extremely powerful tool
in the study of the Boltzmann operator with Maxwellian collision kernel.
Now we give the Fourier transform of Q(g, f ) in (8.19). We first perform the
calculation of the Fourier transform of the gain term in a general Boltzmann
collision operator Q+ (g, f ).
For any test function ϕ(v), using the pre-postcollisional change of vari-
ables in (8.26), we have
Z
Q+ (g, f )ϕ(v)dv
R n
Z Z (8.39)
v − v∗
= B(|v − v∗ |, · σ)g(v∗ )f (v)ϕ(v 0 )dσdv∗ dv.
Rn ×Rn S n−1 |v − v∗ |
Plugging ϕ(v) = e−ivξ in (8.39), we have
F (Q+ (g, f ))(ξ)
Z Z
v − v∗ v+v∗ |v−v∗ |
= B(|v − v∗ |, · σ)g(v∗ )f (v)e−i 2 ·ξ e−i 2 σ·ξ dσdv∗ dv.
R2n S n−1 |v − v∗ |
(8.40)
Using the following general equality:
Z Z
F (k · σ, l · σ)dσ = F (l · σ, k · σ)dσ, |l| = |k| = 1, (8.41)
S n−1 S n−1
(due to the existence of an isometry on S n−1 exchanging l and k) we have

Z
v − v∗ |v−v∗ |
B(|v − v∗ |, · σ)e−i 2 σ·ξ dσ
S n−1 |v − v∗ |
Z (8.42)
ξ |ξ|
= B(|v − v∗ |, · σ)e−i 2 σ·(v−v∗ ) dσ.
S n−1 |ξ|
Then, using (8.42) and the Fourier inversion formula, we have
F (Q+ (g, f ))(ξ)
Z Z
ξ v+v∗ |ξ|
= B(|v − v∗ |, · σ)g(v∗ )f (v)e−i 2 ·ξ e−i 2 σ·(v−v∗ ) dσdv∗ dv
2n n−1 |ξ|
ZR ZS
ξ + −
= B(|v − v∗ |, · σ)g(v∗ )f (v)e−ivξ e−iv∗ ξ dσdv∗ dv
2n n−1 |ξ|
ZR ZS
ξ
= B(|v − v∗ |, · σ)
R2n |ξ|
+ −

× ĝ(η∗ )fˆ(η)eivη eiv∗ η∗ e−ivξ e−iv∗ ξ dη∗ dη dσdv∗ dv
Z
= ĝ(η∗ )fˆ(η)
R2n ×S n−1
Z ξ + −

× B(|v − v∗ |, · σ)eiv(η−ξ ) eiv∗ (η∗ −ξ ) dv∗ dv dσdη∗ dη,
|ξ|
(8.43)
where
ξ + |ξ|σ ξ − |ξ|σ
ξ+ = , ξ− = . (8.44)
2 2
R
Let δ be the Dirac measure, B̂(|ξ|, cos θ) = Rn B(|q|, cos θ)e−iq·ξ denotes
the Fourier transform of B in the relative velocity variable. By the change
of variables q = v − v∗ ,
Z
ξ + −
B(|v − v∗ |, · σ)eiv(η−ξ ) eiv∗ (η∗ −ξ ) dv∗ dv
R2n |ξ|
Z
ξ −
= B(|q|, · σ)eiv(η∗ +η−ξ) e−iq(η∗ −ξ ) dqdv (8.45)
R 2n |ξ|
ξ
= B̂(|η∗ − ξ − |, · σ)δ(η = ξ − η∗ ).
|ξ|
By (8.45), we give the Fourier transform of Q+ (g, f )
Z
ξ
F (Q+ (g, f ))(ξ) = ĝ(η∗ )fˆ(ξ − η∗ )B̂(|η∗ − ξ − |, · σ))dσdη∗ .
Rn ×S n−1 |ξ|
(8.46)
Let ξ∗ = η∗ − ξ − , weZ find

ξ
F (Q+ (g, f ))(ξ) = ĝ(ξ − + ξ∗ )fˆ(ξ + − ξ∗ )B̂(|ξ∗ |, · σ))dσdξ∗ .
Rn ×S n−1 |ξ|
(8.47)
In Maxwellian potentials (Maxwellian molecules) case, that is,
B(|z|, cos θ) = b(cos θ), we have
B̂(|ξ∗ |, cos θ) = δ(ξ∗ = 0)b(cos θ). (8.48)
and as a consequence Z
ξ
F (Q+ (g, f ))(ξ) = ĝ(ξ − )fˆ(ξ + )b( · σ)dσ. (8.49)
S n−1 |ξ|
−
For the Fourier transform of Q (g, f ), if B(|z|, cos θ) = b(cos θ), then we
easily have Z
ξ
−
F (Q (g, f ))(ξ) = ĝ(0)fˆ(ξ)b( · σ)dσ, (8.50)
R S n−1 |ξ|
which follows from the fact that S n−1 b(k · σ) does not depend on k.
8.3 Properties of Boltzmann collision operator without cut-

off
In this section, we will consider the properties of the Boltzmann collision

operator Q(g, f ) without cutoff. We prove that for a given distribution
function g ∈ L1 , the Boltzmann operator Q(g, f ) behaves essentially as a
fractional power of the Laplacian:
Q(g, f ) = −Cg (−∆)ν/2 f + more regular terms.
In details, one can refer to [2]. In the following discussion, we also adopt
the notations for the weighted function spaces,
kf kLpr = kf (v)hvir kLZp , 1 6 p 6 ∞, r ∈ R,
kf kLlogL = f log(1 + f )dv.
Rn n
2 +1 (Rn ), we have
weak formulation of Q(g, f ). For f ∈ H
Wefirst give the
Q(g, f ), f
L2
Z
v − v∗
= B(|v − v∗ |, · σ)g(v∗ )f (v)(f (v 0 ) − f (v))dσdv∗ dv
2n
R ×S n−1 |v − v∗ |
Z
1 v − v∗
= B(|v − v∗ |, · σ)g(v∗ )(f 2 (v 0 ) − f 2 (v))dσdv∗ dv
2 R2n ×S n−1 |v − v∗ |
Z
1 v − v∗
− B(|v − v∗ |, · σ)g(v∗ )(f (v 0 ) − f (v))2 dσdv∗ dv.
2 R2n ×S n−1 |v − v∗ |
(8.51)
8.3. Properties of Boltzmann collision operator without cutoff 241
For the first term in (8.51), we have the following lemma.
Lemma 8.1. (Cancellation) For a.e. v∗ ∈ Rn ,

Z
B(|v − v∗ |, k · σ)(f 2 (v 0 ) − f 2 (v))dσdv = (f 2 ∗ A)(v∗ ), (8.52)
Rn ×S n−1
where
v − v∗
k= , (8.53)
|v − v∗ |
Z π

2 1 |z|
A(z) = S n−2
sinn−2 θ B( , cos θ) − B(|z|, cos θ) dθ.
0 cosn 2θ cos θ2
(8.54)
Proof. We do the calculation as if B were integrable and apply a limiting
procedure to conclude in the general case, that is, we choose a sequence
integrable function B̃j in S n−1 , such that limj→∞ B̃j = B. In fact the
right-hand side of (8.52) should be taken as a definition of the left-hand
side. Recall that
v + v∗ |v − v∗ | |v − v∗ |
v0 = + σ = v∗ + (k + |k|σ).
2 2 2
For each σ and with v∗ still fixed, we perform the change of variables v → v 0 .
This change of variables is well-defined on the set cos θ > 0, and it follows
either by a direct calculation or by using the cylindrical symmetry of this
transformation that its Jacobian determinant is
dv 0 1 1 1 (k 0 · σ)

= I + k ⊗ σ = n (1 + k · σ) = n−1 , (8.55)
dv 2 2 2 2
v 0 −v∗
where k 0 = |v 0 −v∗ | . Then
θ 1
k 0 · σ = cos
> √ .
2 2
Define the inverse transformation ψσ : v → ψσ (v 0 ) = v, from the figure
0
0
8.1, it follows that |v∗ − ψσ (v 0 )| = |vk−v ∗| 0
0 ·σ . Then changing the name v for
v,
|v − v∗ |
|v∗ − ψσ (v)| = . (8.56)
k·σ
Applying this change of variable to the first part in the left-hand side of
(8.52), then changing the name v 0 for v,
Z
B(|v − v∗ |, k · σ)g(v∗ )f 2 (v 0 )dσdv
Rn ×S n−1
Z
dv
= B(|v − v∗ |, 2(k 0 · σ)2 − 1)g(v∗ )f 2 (v 0 )| |dσdv 0
Rn ×S n−1 dv 0
Z
2n−1
= B(|ψσ (v 0 ) − v∗ |, 2(k 0 · σ)2 − 1)g(v∗ )f 2 (v 0 ) dσdv 0
k0 ·σ> √12 (k 0 · σ)2
Z
2n−1
= B(|ψσ (v) − v∗ |, 2(k · σ)2 − 1)g(v∗ )f 2 (v) dσdv. (8.57)
k·σ> √12 (k · σ)2
Using the fact that 2n−2 sinn−2 2θ cosn−2 2θ = sinn−2 θ and n-dimensional
spherical coordinates, we have
Z
2n−1
B(|ψσ (v) − v∗ |, 2(k · σ)2 − 1) dσ
k·σ> √12 (k · σ)2
Z
|v − v∗ | 2n−1
= B( , 2(k · σ)2 − 1) dσ
k·σ> √12 k·σ (k · σ)2
Z π4
|v − v∗ | 2n−1
= |S n−2
| sinn−2 θB( , cos 2θ) 2 dθ
0 cos θ cos θ
Z π2
θ |v − v ∗ | 2n−2
= |S n−2 | sinn−2 B( θ
, cos θ) 2 θ dθ
0 2 cos 2 cos 2
Z π2 n−2
sin θ |v − v∗ |
= |S n−2 | θ
B( , cos θ)dθ. (8.58)
0
n
cos 2 cos θ2
Define
A(v − v∗ )
Z π

2 1 |v − v∗ |
=S n−2
sinn−2 θ θ
B( θ
, cos θ) − B(|v − v∗ |, cos θ) dθ.
0 cosn 2 cos 2
(8.59)
This completes the proof.
Remark 8.1. If B(|v − v∗ |, k · σ) = |v − v∗ |γ b(k · σ), then

A(v − v∗ )
Z π

n−2
2 1 |v − v∗ | γ
=S sinn−2 θ ( ) b(cos θ) − |v − v∗ | γ
b(cos θ) dθ
0 cosn θ2 cos θ2
Z π
2 1
=S n−2
sinn−2 θ|v − v∗ |γ b(cos θ)( − 1)dθ. (8.60)
0 cosγ+n 2θ
Note that if γ > −n, then (8.60) is nonnegative; if γ = −n, it is Coulomb
potentials, they require a more careful analysis. Using Lemma 8.1, we find
that if 0 6 γ 6 2
Z
B(|v − v∗ |, k · σ)g(v∗ )(f 2 (v 0 ) − f 2 (v))dσdvdv∗
R2n ×S n−1 (8.61)
6 CkgkL12 kf kL21 ,
θ
which follows from the fact that 1 − cosn 2 6 n(1 − cos 2θ ) = 2n sin 4θ .
For the second term in (8.51), for simplicity, we only consider

Maxwellian molecules case, that is, cross section B(|v − v∗ |, k · σ) = b(k · σ).
Lemma 8.2. Assume g ∈ L1 (Rn ) and f ∈ L2 (Rn ), the following

Plancherel-type identity holds
Z
v − v∗
b( · σ)g(v∗ )(f (v 0 ) − f (v))2 dσdv∗ dv
R2n ×S n−1 |v − v∗ |
Z
ξ
= b( · σ) ĝ(0)|fˆ(ξ)|2 + ĝ(0)|fˆ(ξ + )|2 (8.62)
Rn ×S n−1 |ξ|

− ĝ(ξ − )fˆ(ξ + )fˆ(ξ) − ĝ(ξ − )fˆ(ξ + )fˆ(ξ) dξdσ.
Proof. We shall do the proof only in the case when b is integrable and
the result will follow by monotonicity. First, we have
(f (v 0 ) − f (v))2 = f 2 (v 0 ) − 2f (v 0 )f (v) + f 2 (v). (8.63)
0
We begin with the middle term 2f (v )f (v) in (8.63). By the pre-
postcollisional change of variables and Parseval’s identity and Bobylev’s
identity, we have
Z Z
b(k · σ)g(v∗ )f (v 0 )f (v)dσdv∗ dv = Q+ (g, f )f dv
1 + 1
= (Q (g, f ), f )L2 + (f, Q+ (g, f ))L2
2 2 (8.64)
1 1
= (F Q+ (g, f ), F f )L2 + (F f, F Q+ (g, f ))L2
2Z 2
1 ξ
= b( · σ) ĝ(ξ )f (ξ )f (ξ) + ĝ(ξ − )fˆ(ξ + )fˆ(ξ) dξdσ.
− ˆ + ˆ
2 |ξ|
R
For the third term f 2 (v) in (8.63), using the fact that S n−1 b(k · σ)dσ does
not depend on the unit vector k, we have
Z Z Z Z
b(k · σ)g(v∗ )f 2 (v)dσdv∗ dv = b(k · σ)dσ g(v∗ )dv∗ f 2 (v)dv
S n−1 Rn Rn
Z
ξ
= b( · σ)ĝ(0)|fˆ(ξ)|2 dξdσ. (8.65)
|ξ|
For the first term f 2 (v 0 ) in (8.63), we first make the change of variables
(v − v∗ , v∗ ) → (v1 , v∗ ), change the name v1 for v, and then use the change
of variables v → v 0 as in Lemma 8.1 to obtain
Z
v − v∗ v + v∗ |v − v∗ |
b( · σ)g(v∗ )f 2 ( + σ)dσdv∗ dv
|v − v∗ | 2 2
Z
v1 v1 + |v1 |σ 2
= b( · σ)g(v∗ )|τ−v∗ f ( )| dσdv∗ dv1
|v1 | 2
Z (8.66)
v v + |v|σ 2
= b( · σ)g(v∗ )|τ−v∗ f ( )| dσdv∗ dv
|v| 2
Z n−1
2
= b(ψ(v 0 , σ)g(v∗ ) v0 |τ−v∗ f (v 0 )|2 dσdv∗ dv 0 ,
( |v0 | · σ)2
where
v0
ψ(v 0 , σ) = 2( 0 · σ)2 − 1, τ−v∗ f = f (v∗ +).
|v |
R
Using the fact that S n−1 b(k · σ) does not depend on k and |F (τh f )| =
|F (f )|, and reversing the change of variables ξ → ξ + as in Lemma 8.1, we
have
Z Z
2n−1
(8.66) = g(v∗ ) b(ψ(ξ, σ) ξ |f (ξ)|2 dσdξ dv∗
( |ξ| · σ)2
Z (8.67)
ξ + 2
= ĝ(0) b( · σ)|f (ξ )| dσdξ.
|ξ|
Collecting (8.63), (8.64), (8.65) and (8.67), we can obtain (8.62).
Corollary 8.1. Assume g ∈ L1 (Rn ), f ∈ L2 (Rn ) and f > 0. Then

Z
v − v∗
b( · σ)g(v∗ )(f (v 0 ) − f (v))2 dσdv∗ dv
2n
R ×S n−1 |v − v∗ |
Z Z (8.68)
ˆ 2 ξ −
> |f (ξ)| b( · σ)(ĝ(0) − |ĝ(ξ )|)dσdξ.
Rn S n−1 |ξ|
Proof. Using Lemma 8.2 and the following inequality,
|fˆ(ξ + )|2 + |fˆ(ξ)|2 > |fˆ(ξ)|2 ,
we can obtain Corollary 8.1.
Lemma 8.3. Assume that b satisfies (8.21), Then there exists a positive
constant Cg depending on n, kgkL11 , kgkL log L and b, such that for |ξ| > 1
Z
ξ
b( · σ)(ĝ(0) − |ĝ(ξ − )|)dσ > Cg |ξ|ν . (8.69)
S n−1 |ξ|
Lemma 8.3 is a consequence of the two lemmas below, next we will use
Cg to denote a positive constant depending on n, kgkL11 , kgkL log L and b.
Lemma 8.4. There exists a positive constant C̃g depending on n, kgkL11

and kgkL log L , such that for all ξ ∈ R
ĝ(0) − ĝ(ξ) > C̃g (|ξ|2 ∧ 1). (8.70)
Proof. For some θ ∈ R, we have

ĝ(0) − ĝ(ξ)
Z
= g(v)(1 − cos(v · ξ + θ))dv
Rn
Z
v·ξ+θ
=2 g(v) sin2 ( )dv
Rn 2
Z
> 2 sin2 ε g(v)dv (8.71)
|v|6r,∀p∈Z,|v·ξ+θ−2pπ|>2ε
n Z o
kgkL11
> 2 sin2 ε kgkL1 − − g(v)dv
r |v|6r,∀p∈Z,|v·ξ+θ−2pπ|62ε
n Z o
kgk L 1
> 2 sin2 ε kgkL1 − 1
− sup g(v)dv .
r |A|6 4ε (2r)n−1 (1+ r|ξ| ) A
|ξ| π
If |ξ| > 1, Lemma 8.4 holds with the following constant

n Z o
kgkL11
C̃g = 2 sin2 ε kgkL1 − − sup g(v)dv , (8.72)
r |A|64ε(2r)n−1 + 2ε (2r)n A π
ε > 0 and r > 0 being chosen in such a way that this quantity is positive.
ε
If |ξ| 6 1, let δ = |ξ| in (8.71), Lemma 8.4 holds with the following
constant
Z
2 sin2 (δ|ξ|) n kgkL11 o
Cg = 2δ inf kgk L 1 − − sup g(v)dv ,
|ξ|61 δ 2 |ξ|2 r r
|A|64ε(2r)n−1 (1+ π ) A
(8.73)
where δ > 0 and r > 0 are chosen in such a way that this quantity is
positive.
Lemma 8.5. If b satisfies
K(ν)
sinn−2 θb(cos θ) ≈ as θ → 0, K(ν) > 0; (8.74)
θ1+ν
then for |ξ| > 1,
Z
ξ
b( · σ)(|ξ − |2 ∧ 1)dσ > K(ν)|ξ|ν . (8.75)
S n−1 |ξ|
Proof. Recall that

|ξ|2 ξ
|ξ − |2 = (1 − · σ).
2 |ξ|
Passing to n-dimensional spherical coordinates, we find for some θ0 > 0,
Z
ξ
b( · σ)(|ξ − |2 ∧ 1)dσ
S n−1 |ξ|
Z π2 |ξ|2
= S n−2 sinn−2 θb(cos θ) (1 − cos θ) ∧ 1 dθ (8.76)
0 2
Z θ0
K |ξ|2 θ2 dθ
> |S n−2 | ( ∧ 1) 1+ν .
2 0 2 θ
By the change of variables θ → |ξ|θ, the integral in (8.76) is also
Z θ0 2
θ dθ
|ξ|ν ( ∧ 1) 1+ν (8.77)
0 2 θ
so that when |ξ| > 1, Lemma 8.5 holds with
Z θ0 2
K θ dθ
K(v) = |S n−2 | ( ∧ 1) 1+ν .
2 0 2 θ
By Lemma 8.2, Corollary8.1 and Lemma 8.3, for the second term in
(8.51), we have
Z
v − v∗
b( · σ)g(v∗ )(f (v 0 ) − f (v))2 dσdv∗ dv > Cg kf k2H ν/2 .
R2n ×S n−1 |v − v∗ |
(8.78)
8.4 Regularity of solutions for spatially homogeneous case
In this section, we will give the application of the results of Section 3 in this
chapter; and prove the regularity of solutions for the spatially homogeneous
Boltzmann equation without angular cutoff

∂t f (t, v) = Q(f, f )(t, v), t > 0, v ∈ Rn ,
(8.79)
f (0, v) = f0 (v),
where cross section B
v − v∗ v − v∗
B(|v − v∗ |, · σ) = |v − v∗ |γ b( · σ),
|v − v∗ | |v − v∗ |
K(ν)
sinn−2 θb(cos θ) ≈ when θ → 0, K(ν) > 0.
θ1+ν
8.4. Regularity of solutions for spatially homogeneous case 247
From (8.33), it follows that the solutions f (t, v) of the Cauchy problem
(8.79) have the following conservation of mass, momentum and energy,
Z Z
f (t, v)dv = f0 (v)dv, (8.80)
Rn Rn
Z Z
f (t, v)vj dv = f0 (v)vj dv, j = 1, 2...n, (8.81)
Rn Rn
Z Z
|v|2 |v|2
f (t, v) dv = f0 (v) dv. (8.82)
Rn 2 Rn 2
For the existence of weak solutions for Cauchy problem (8.79), in 1998
Villani [230] showed that if the initial data have the finite mass, energy and
entropy,
Z
f0 (v)[1 + |v|2 + log(1 + f0 (v))]dv < +∞, (8.83)
Rn
then he constructed a weak solution f (t, v) of the Cauchy problem (8.79),
which satisfies the following:
f (t, v) > 0, f (t, v) ∈ C(R+ , S 0 ); t > 0, f (t, v) ∈ L12 ∩ LlogL, (8.84)
f (t, v) ∈ L1 ([0, T ]; L12+γ ), (8.85)
f (0, v) = f0 (v), (8.86)

Z Z
|v|2
f (t, v)ψ(v)dv = f0 (v)ψ(v)dv, ψ(v) = 1, v1 , ..., vn , , (8.87)
Rn Rn 2
Z Z
f (t, v)logf (t, v)dv 6 f0 (v)logf0 (v)dv, (8.88)
Rn Rn
Z Z Z t Z
f (t, v)ϕ(t, v)dv − f0 (v)ϕ(0, v)dv − dτ f (τ, v)∂τ ϕ(τ, v)dv
Rn Rn 0 Rn
Z t Z
= dτ Q(f, f )(τ, v)ϕ(τ, v)dv, ∀ϕ(t, v) ∈ C 1 (R+ ; C0∞ (Rn )), (8.89)
0 Rn
where the last integral in the right-hand side being defined by the following
formulae
Z
Q(f, f )(τ, v)ϕ(τ, v)dv
Rn
Z Z
1
= Bf (v∗ )f (v)(ϕ(v∗0 ) + ϕ(v 0 ) − ϕ(v∗ ) − ϕ(v))dvdv∗ dσ.
2 R2n S n−1
These make sense provided that f satisfies (8.84) and test function ϕ ∈
L∞ ([0, T ]; W 2,∞ ).
In this section, we assume that a weak solution to the Boltzmann equa-
tion (8.79) has already been constructed and that it satisfies the usual
entropic estimates, to say,
Z
f (t, v)[1 + |v|2 + log(1 + f (t, v))]dv < +∞, (8.90)
Rn
we are interested in regularity issues associated to such solutions. That is,
is this weak solution more regular than the initial datum, and if so, can we
have estimates on this regularity?
In this section, we will show that the weak solutions above constructed
f (t, v) of the Cauchy problem (8.79) with Maxwellian molecules case be-
longs to Hv+∞ . For hard potentials case, there exist some results only
in modified hard potentials cases now, in details, one can refer to [4; 60;
111]. For soft potentials cases, there seems no any results now.
In this section, we will introduce two methods of multiplier to consider
it. That is, we first choose a team of suitable multiplier, then mainly give
the sharp estimates of the commutators of the collision operator Q(f, f )
and pseudo-differential operators composed by multipliers.
First, we give the method of the Littlewood-Paley decomposition, in
details, one can refer to [3]. The Littlewood-Paley decomposition is defined
as in Section 3 in Chapter 1: Assume that supp ϕk ⊂ {ξ : 2k−1 6 |ξ| 6
P∞
2k+1 }, k ∈ N, supp ϕ0 ⊂ {ξ : |ξ| 6 2}, k=0 ϕk = 1; we can assume
ϕk is a radial function. Recall that for every multi-index α, there exists a
positive number Cα such that
2k|α| |Dα ϕk (ξ)| 6 Cα , k = 0, 1, 2...; ξ ∈ Rn . (8.91)
Then the Littlewood-Paley projection operator is defined by
d
4 ˆ n
k f (ξ) = ϕk (ξ)f (ξ), k = 0, 1, 2...; ξ ∈ R .
Lemma 8.6. Assume that the initial data f0 (v) satisfies (8.83), cross sec-
tion B(|v − v∗ |, k · σ) = b(k · σ) with b satisfying the following
K
sinn−2 θb(cos θ) ≈ 1+ν when θ → 0, K > 0. (8.92)
θ
Let f (t, v) be any weak non negative solution of the Cauchy problem (8.79),
satisfying (8.80)-(8.82) and (8.90). Then
k+1
X
(4k Q(f, f ), 4k f )L2 − (Q(f, 4k f ), 4k f )L2 6 Ckf0 kL1 ( k4k f kL2 ),
j=k−2
(8.93)
k+1
X
(4k f, 4k Q(f, f ))L2 − (4k f, Q(f, 4k f ))L2 6 Ckf0 kL1 ( k4k f kL2 ).
j=k−2
(8.94)
Remark 8.2. Let f (t, v) be a weak solution of the Boltzmann equation.

If we choose 4k f as test function (4k f ∈ L∞ ([0, T ]; W 2,∞ )), then by the
definition of weak solutions, the inner products in the left side of (8.93) and
(8.94) make sense. Unless f ∈ H ∞ (Rn ), f (t, v) cannot be chosen as a test
function in the definition of weak solutions. This is also one of the reasons
why we use the method of multiplier.
Proof. We only prove (8.93), the proof of (8.94) is similar with one of
(8.93). Applying Fourier transform on Boltzmann equation (8.79), using
(8.49) and (8.50) (Bobylev identity), we have
Z
\ ξ
∂t fˆ(ξ) = Q(f, f )(ξ) = b( · σ) fˆ(ξ − )fˆ(ξ + ) − fˆ(0)fˆ(ξ) dσ. (8.95)
S n−1 |ξ|
Multiplying (8.95) by ϕk (ξ), we have
d
∂t 4 \
k f (ξ) = 4k Q(f, f )(ξ)
Z
ξ
= b( · σ) fˆ(ξ − )fˆ(ξ + )ϕk (ξ) − fˆ(0)fˆ(ξ)ϕk (ξ) dσ.
S n−1 |ξ|
(8.96)
Then,
Z Z
d
∂t (4 d
k f , 4k f )L2 =
d
(∂t 4 d
k f )4k f dξ +
d
4 d
k f ∂t 4k f dξ
Rn Rn
= (4k\Q(f, f ), 4d d \
k f )L2 + (4k f , 4k Q(f, f ))L2
Z Z
ξ
= b( · σ) fˆ(ξ − )fˆ(ξ + )ϕ2k (ξ)fˆ(ξ) − fˆ(0)fˆ(ξ)ϕ2k (ξ)fˆ(ξ) dσdξ
n n−1 |ξ|
ZR ZS
ξ
+ b( · σ) fˆ(ξ − )fˆ(ξ + )ϕ2k (ξ)fˆ(ξ) − fˆ(0)fˆ(ξ)ϕ2k (ξ)fˆ(ξ) dσdξ.
Rn S n−1 |ξ|
(8.97)
Moreover, we have
Z
\ ξ
Q(f, 4k f )(ξ) = b( · σ) fˆ(ξ − )ϕk (ξ + )fˆ(ξ + ) − fˆ(0)ϕk (ξ)fˆ(ξ) dσ,
S n−1 |ξ|
(8.98)
\
(Q(f, d
4k f ), 4 d \
k f )L2 + (4k f , Q(f, 4k f ))L2
ZZ
ξ
= b( · σ) fˆ(ξ − )ϕk (ξ + )fˆ(ξ + )ϕk (ξ)fˆ(ξ) − fˆ(0)fˆ(ξ)ϕ2k (ξ)fˆ(ξ) dσdξ
|ξ|
ZZ
ξ
+ b( · σ) fˆ(ξ − )ϕk (ξ + )fˆ(ξ + )ϕk (ξ)fˆ(ξ) − fˆ(0)fˆ(ξ)ϕ2k (ξ)fˆ(ξ) dσdξ.
|ξ|
(8.99)
Thus, from (8.97) and (8.99) it follows that
(4k\ d
Q(f, f ), 4 \ d
k f )L2 − (Q(f, 4k f ), 4k f )L2
Z Z
ξ
= b( · σ)fˆ(ξ − )fˆ(ξ + )(ϕk (ξ) − ϕk (ξ + ))ϕk (ξ)fˆ(ξ)dσdξ.
R n S n−1 |ξ|
(8.100)
From the definitions of ξ + and ξ − , it follows that
ξ + |ξ|σ θ ξ π
ξ+ = , |ξ + | = |ξ| cos , · σ = cos θ, 0 6 θ 6 ,
2 2 |ξ| 2
|ξ|2 θ
6 |ξ + |2 6 |ξ|2 , |ξ|2 − |ξ + |2 = |ξ − |2 = |ξ|2 sin2 .
2 2
From the integral in (8.100), we note that 2k−1 6 |ξ| 6 2k+1 . From the
definition of ξ + , it follows that 2k−2 6 |ξ + | 6 2k+1 . Thus
\
fˆ(ξ + ) = 4 + \ + d + \ +
k−2 f (ξ ) + 4k−1 f (ξ ) + 4k f (ξ ) + 4k+1 f (ξ ). (8.101)
Plugging this expression of f (ξ + ) in (8.100), and by the fact that |fˆ(ξ − )| 6
kf kL1 6 kf0 kL1 , (8.100) is bounded by
Z Z
ξ d + k d
kf0 kL1 b( · σ)|4 k f (ξ )||Aξ ||4k f (ξ)|dσdξ
2 k−1 6|ξ|62k+1 S n−1 |ξ|
Z Z
ξ \ + k d
+ kf0 kL 1 b( · σ)|4 k−2 f (ξ )||Aξ ||4k f (ξ)|dσdξ
2k−1 6|ξ|62k+1 S n−1 |ξ|
Z Z
ξ \ + k d
+ kf0 kL1 b( · σ)|4 k−1 f (ξ )||Aξ ||4k f (ξ)|dσdξ
2k−1 6|ξ|62k+1 S n−1 |ξ|
Z Z
ξ \ + k d
+ kf0 kL1 b( · σ)|4 k+1 f (ξ )||Aξ ||4k f (ξ)|dσdξ,
2k−1 6|ξ|62k+1 S n−1 |ξ|
(8.102)
where Akξ = ϕk (ξ) − ϕk (ξ + ). By (8.91), we have
|Akξ | = |ϕk (ξ) − ϕk (ξ + )| 6 |ϕ̃k (|ξ|) − ϕ̃k (|ξ + |)|
1 + 1 |ξ|2 − |ξ + |2 θ (8.103)
. (|ξ| − |ξ |) . . sin2 .
2k 2k |ξ| + |ξ + | 2
We only estimate the first term in (8.102), the estimates of the other terms
in (8.102) can be obtained similarly. By (8.103) and Hölder inequality,
(8.102) is bounded by
Z Z
ξ d + k d
kf0 kL1 b( · σ)|4 k f (ξ )||Aξ ||4k f (ξ)|dσdξ
2k−1 6|ξ|62k+1 S n−1 |ξ|
Z Z
d ξ θ d +
. kf0 kL1 |4 k f (ξ)| b( · σ) sin2 |4 k f (ξ )|dσdξ
Rn S n−1 |ξ| 2
Z 1/2 nZ Z ξ θ d + 2 o1/2
. kf0 kL1 |4d 2
k f (ξ)| dξ b( · σ) sin2 |4 k f (ξ )|dσ dξ
|ξ| 2
. kf0 kL1 k4k f k2L2 . (8.104)
+
which follows from the change of variables ξ → ξ as in Lemma 8.1.
Theorem 8.1. Under the hypothesis of Lemma 8.6, Let f (t, v) be any weak
non negative solution of the Cauchy problem (8.79), satisfying (8.90). Then
for any s ∈ R+ and t > 0, one has f (t, v) ∈ H s (Rn ).
Proof. First, we have
∂t k4k f kL2 = (4k\ d

Q(f, f ), 4 d \
k f )L2 + (4k f , 4k Q(f, f ))L2 . (8.105)
From (8.51) it follows that

Q(f, 4k f ), 4k f 2
L
Z
v − v∗
= b( · σ)f (v∗ )4k f (v)(4k f (v 0 ) − 4k f (v))dσdv∗ dv
R2n ×S n−1 |v − v∗ |
Z
1 v − v∗
= b( · σ)f (v∗ )((4k f )2 (v 0 ) − (4k f )2 (v))dσdv∗ dv
2 R2n ×S n−1 |v − v∗ |
Z
1 v − v∗
− b( · σ)f (v∗ )(4k f (v 0 ) − 4k f (v))2 dσdv∗ dv
2 R2n ×S n−1 |v − v∗ |
= I1 − I2 . (8.106)
For the first term in (8.106), using Lemma 8.1 in this chapter, we get
Z
I1 = b(k · σ)((4k f )2 (v 0 ) − (4k f )2 (v))dσdvdv∗
2n
R ×S n−1
Z (8.107)
=A (4k f )2 (v∗ )dv∗ ,
Rn
where
Z π

2 1
A = S n−2 sinn−2 θ θ
− 1 b(cos θ)dθ. (8.108)
0 cosn 2
From 1 − cosn θ
2 6 n(1 − cos 2θ ) = 2n sin2 θ
4 it follows that
I1 . k4k f k2L2 . (8.109)
For the second term in (8.106), using Corollary 8.1 and Lemma 8.3 in this
chapter, we have
Z
1 v − v∗
I2 = b( · σ)f (v∗ )(4k f (v 0 ) − 4k f (v))2 dσdv∗ dv
2 R2n ×S n−1 |v − v∗ |
Z
> Cf0 (1 + |ξ|ν )ϕk (ξ)fˆ(ξ)dξ > Cf0 k4k f k2 ν2 .
H
Rn
(8.110)
Thus collecting (8.105), (8.106), (8.109), (8.110) and Lemma 8.6, we get
k+1
X
∂t k4k f kL2 + Cf0 k4k f k2H ν2 6 Ckf0 kL1 ( k4j f k2L2 ). (8.111)
j=k−2
Dividing(8.111) by 2kn , we have

∂t k4k f kL2 k4k f k2L2
kn
+ (Cf0 2kν − Ckf0 kL1 )
2 2kn (8.112)
k4 2 2
k4k+1 f k2L2
k−2 f kL2 k4k−1 f kL2
6 Ckf0 kL1 + + .
2(k−2)n 2(k−1)n 2(k+1)n
k4k f k2
Let Uk (t) = 2kn
L2
, Ck = Cf0 2kν − Ckf0 kL1 and β = Ckf0 kL1 . Then
(8.112) can be rewritten by:
∂t Uk (t) + Ck Uk (t) 6 β(Uk−2 (t) + Uk−1 (t) + Uk+1 (t)). (8.113)

From Bernstein’s inequality (Polyya Plancherel Nikoolski inequality), it
follows that
k4k f kL2 6 C(ϕ)2nk k4k f kL1 6 C(ϕ)2nk kf kL1 6 C(ϕ)2nk kf0 kL1 .
For k > 0, t > 0,
Uk (t) 6 M, with M = C(ϕ)kf0 kL1 . (8.114)

By the definition of Ck , we show that Ck is nondecreasing, Ck > 0 with
large enough k (without loss of generality, we can assume that Ck > 0 with
k > k0 > 3). From (8.113) and (8.114), it follows that Uk (t) satisfy the
conditions of the following Lemma 8.7. Then for all integer p > 1, there
exist constants Ap and Dp such that
1
Uk (t) 6 M Ap e−Ck−2(p−1)t + M Dp ,t > 0 (8.115)
(Ck−2(p−1) )p
for all k > 2(p − 1) + k0 + 2. Thus, for fixed s > 0, t > 0 and an integer
p > 1 which we shall choose just below; using the definition of Sobolev
spaces, we have
∞
X
kf k2H s ∼ 22ks k4k f k2L2
k=0
2(p−1)+k0 +1 ∞
X X
= 22ks k4k f k2L2 + 22ks k4k f k2L2 (8.116)
k=0 k=2(p−1)+k0 +2
2(p−1)+k0 +1 ∞
X X
= 22ks k4k f k2L2 + 2k(2s+n) Uk .
k=0 k=2(p−1)+k0 +2
In order to prove f ∈ H s , it remains to show that the last series appearing
above is a convergent one. From (8.115) it follows that
∞
X
2k(2s+n) Uk
k=2(p−1)+k0 +2
∞
X
6 2k(2s+n) M Ap e−Ck−2(p−1)t (8.117)
k=2(p−1)+k0 +2
∞
X 1
+ 2k(2s+n) M Dp .
(Ck−2(p−1) )p
k=2(p−1)+k0 +2
Using the definition of Ck , we have
Ck−2(p−1) ∼ C(f0 , p)2kνp , if k is large enough. (8.118)
Choosing p such that νp > 2s + n + 1 for instance, yields that the two series
on the right hand side of the last inequality are convergent.
Lemma 8.7. (Iteration) Let β, M be two non negative numbers. Given
positive integer k0 , assume that {Ck }k>k0 is a sequence of positive numbers,
and is non-decreasing in k, and satisfies: there exists a positive α such that
Ck+1 − Ck > α for all k > k0 . {Uk }k>k0 = {Uk (t)}k>k0 is another sequence
with t ∈ R+ satisfying:
∂t Uk (t) + Ck Uk (t) 6 β(Uk−2 (t) + Uk−1 (t) + Uk+1 (t)),
0 6 Uk (t) 6 M, ∀k > k0 + 2, ∀t > 0.
Then for any integer p > 1, there exist constants Ap , Dp such that, for all
k > 2(p − 1) + k0 + 2, one has
1
Uk (t) 6 M Ap e−Ck−2(p−1)t + M Dp , t > 0.
(Ck−2(p−1) )p
Proof. It is done by iteration. First start from the fact that Uk 6 M ,

and using this in the right hand side of the differential inequality, leads to
the above conclusion, for p = 1. Then, start from the above first iteration,
and replace again in the rhs of the above inequality, and just repeat the
process.
Next, we will introduce another method of multiplier, in details, one
can refer to [171]. For convenience, we assume n = 3. Let f (t, v) be a weak
solution of the Cauchy problem (8.79). For any fixed T0 > 0, it follows
that f (t, v) ∈ L1 (R3 ) ⊂ H −2 (R3 ), t ∈ [0, T0 ]. For t ∈ [0, T0 ], N > 0 and
0 < δ < 1, define multiplier:
N t−4 N T0 + 4
Mδ (t, ξ) = (1 + |ξ|2 ) 2 (1 + δ|ξ|2 )−N0 , N0 =
2
and the corresponding pseudo-differential operator Mδ (t, Dv ) is defined by
Mδ (t, Dv ) = F −1 Mδ (t, ξ)F .
Then for any δ ∈ (0, 1),
Mδ (t, Dv )2 f ∈ L∞ ([0, T0 ]; W 2,∞ (R3 )), Mδ (t, Dv )f ∈ C([0, T0 ]; L2 (R3 )).
Lemma 8.8. Under the hypothesis of Lemma 8.6, Let f (t, v) be any weak
non negative solution of the Cauchy problem (8.79), satisfying (8.90). Then
(Q(f, f ), Mδ2 f )L2 − (Q(f, Mδ f ), Mδ f )L2 6 Cf kf0 kL1 kMδ2 f kL2 , (8.119)
where the constant Cf is independent of 0 < δ < 1.
Proof. Similarly with the proof of Lemma 8.6, we have

(Q(f, f ), Mδ2 f )L2 − (Q(f, Mδ f ), Mδ f )L2
\ [ 2 \ [
= (Q(f, f ), M δ f )L2 − (Q(f, Mδ f ), Mδ f )L2
Z Z
ξ
= b( · σ)fˆ(ξ − )fˆ(ξ + )(Mδ (t, ξ) − Mδ (t, ξ + ))Mδ (t, ξ)fˆ(ξ)dσdξ.
n
R S n−1 |ξ|
(8.120)
It suffices to prove
N T0 +4 θ
Mδ (t, ξ) − Mδ (t, ξ + ) 6 N0 2 2 Mδ (t, ξ + ) sin2 . (8.121)
2
Define
N t−4
M̃δ (t, s) = (1 + s) 2 (1 + δs)−N0 , s = |ξ|2 , s+ = |ξ + |2 ,
so that
Mδ (t, ξ) = M̃δ (t, |ξ|2 ).
From the mean value theorem, it follows that there exists s+ < s̃ < s such
that
∂ M̃δ
M̃δ (t, s) − M̃δ (t, s+ ) = (t, s̃)(s − s+ ).
∂s
By the definition of M̃δ ,
∂ M̃δ Nt − 4 δN0
(t, s) = − M̃δ (t, s).
∂s 2(1 + s) 1 + δs
By the following
s δs M̃ (t, s̃) θ
δ N T0 +4
, 6 1; 6 2 2 , s − s+ = s sin2 ,
1 + s 1 + δs +
M̃δ (t, s ) 2
we have
N T0 +4 θ
M̃δ (t, s) − M̃δ (t, s+ ) 6 N0 2 2 M̃δ (t, s+ ) sin2 .
2
Then
Z Z
ξ
· σ)fˆ(ξ − )fˆ(ξ + )(Mδ (t, ξ) − Mδ (t, ξ + ))Mδ (t, ξ)fˆ(ξ)dσdξ
b(
Rn S n−1 |ξ|
Z Z
ξ θ
6 b( · σ) sin2 |fˆ(ξ − )||fˆ(ξ + )|Mδ (t, ξ + )Mδ (t, ξ)|fˆ(ξ)|dσdξ
R n S n−1 |ξ| 2
6 Ckf0 kL1 kMδ f k2L2 .
Thus using Lemma 8.8, we give another proof of Theorem 8.1. We
note that any weak solution f has the following properties: Mδ2 f ∈
L∞ ([0, T0 ]; W 2,∞ (R3 )) and Mδ f ∈ C([0, T0 ]; L2 (R3 )). From the definition
of weak solution, we choose ψ(t, v) = Mδ2 f (t, v) as test function. For any
t ∈ (0, T0 ),
Z Z
f (t, v)Mδ2 f (t, v)dv − f0 (v)Mδ2 f (0, v)dv
R3 R3
Z t Z
− dτ f (τ, v)∂τ (Mδ2 f (τ, v))dv (8.122)
0 R3
Z t Z
= dτ Q(f, f )(τ, v)Mδ2 f (τ, v)dv.
0 R3
Moreover, we have
Z t Z
dτ f (τ, v)∂τ (Mδ2 f (τ, v))dv
0 R3
Z t Z
Mδ2 f (τ + h, v) − Mδ2 f (τ, v)
= lim dτ (f (τ, v) + f (τ + h, v)) dv
h→0 0 R3 2h
Z t Z
= lim dτ
h→0 0 R3
n M f (τ + h, v)
δ
Mδ (τ + h)f (τ, v) + Mδ (τ + h)f (τ + h, v)
2h
M f (τ, v) o
δ
− Mδ (τ )f (τ, v) + Mδ (τ )f (τ + h, v) dv
2h
Z t Z
(Mδ f )2 (τ + h, v) − (Mδ f )2 (τ, v)
= lim dτ dv
h→0 0 R3 2h
Z t Z
1 n
+ lim dτ Mδ (τ + h)(f (τ, v))Mδ (τ + h)(f (τ + h, v))
h→0 0 R3 2h
o
− Mδ (τ )(f (τ, v))Mδ (τ )(f (τ + h, v)) dv
= J1 + J2 . (8.123)
For J1 , we have Z Z
t
(Mδ f )2 (τ + h, v) − (Mδ f )2 (τ, v)
J1 = lim dτ dv
h→0 0 R3 2h
Z Z h oZ
1 n t+h
= lim dτ − dτ (Mδ f )2 (τ, v)dv
h→0 2h 0 0 R3
Z Z
1 1
= (Mδ f )2 (t, v)dv − (Mδ f )2 (0, v)dv. (8.124)
2 R3 2 R3
For J2 , we have
Z t Z
1 n
J2 = lim dτ Mδ (τ + h)(f (τ, v))Mδ (τ + h, v)(f (τ + h, v))
h→0 0 R3 2h
o
− Mδ (τ )(f (τ, v))Mδ (τ )(f (τ + h, v)) dv
Z t Z
1 n
= lim dτ f (τ, v)Mδ2 (τ + h, v)(f (τ + h, v))
h→0 0 R3 2h
o
− f (τ, v)Mδ2 (τ )(f (τ + h, v)) dv
Z t Z n o
1
= lim dτ f (τ, v) Mδ2 (τ + h) − Mδ2 (τ ) (f (τ + h, v)) dv
h→0 0 R3 2h
Z t Z
1
= dτ f (τ, v)(∂τ Mδ2 (τ ))(f (τ, v))dv. (8.125)
2 0 R 3
By Z(8.123), (8.124) and (8.125), Z (8.122) can be rewritten by

1 2 1
f (t, v)Mδ f (t, v)dv − f0 (v)Mδ2 f (0, v)dv
2 R3 2 R3
Z t Z Z t Z
= dτ f (τ, v)(∂τ Mδ2 )(f (τ, v))dv + dτ (Q(f, f ), Mδ2 f )L2 (τ, v)dv.
0 R3 0 R3
(8.126)
Similarly with the first proof of Theorem 8.1, or by the content of Section
3 in this chapter, we have
n o
kMδ f k2H ν2 6 Cf (−Q(f, Mδ f ), Mδ f )L2 + kMδ f k2L2 . (8.127)
From Lemma 8.8, it follows that
n o
ef (−Q(f, f ), M 2 f )L2 + kMδ f k2 2 .
kMδ f k2H ν2 6 C (8.128)
δ L
Since
∂t Mδ (t, ξ) = N Mδ (t, ξ) loghξi,
we obtain
Z t Z Z t
1
2
dτ f (τ, v)(∂τ Mδ (τ ))(f (τ, v))dv 6 2N k(log Λ) 2 Mδ f (τ )k2L2 dτ,
0 R3 0
(8.129)
−1 −1
where log Λ = F loghξiF and Λ = F hξiF . This together with (8.126)
and (8.128), imply
Z t
1 ν
kMδ f (t)k2L2 + kΛ 2 Mδ f (τ )k2L2 dτ
2Cf 0
Z t Z t
1
2 2
6 kMδ f (0)kL2 + 2N k(log Λ) Mδ f (τ )kL2 dτ +
2 kMδ f (τ )k2L2 dτ.
0 0
(8.130)
Using loghξi 6 hξi and Gagliardo-Nirenberg interpolation inequality, for
any ε > 0, we have
1 Z t ν
2
kMδ f (t)kL2 + −ε kΛ 2 Mδ f (τ )k2L2 dτ
2Cf 0
Z t (8.131)
2 2
6 kMδ f (0)kL2 + Cε,N kMδ f (τ )kL2 dτ.
0
1
By choosing ε = 4Cf > 0, there exists a constant Cf,N depending only on
Cf , N , T0 and being independent of δ ∈ (0, 1), such that for any t ∈ (0, T0 ),
Z t
kMδ f (t)k2L2 6 kMδ f (0)k2L2 + Cf,N kMδ f (τ )k2L2 dτ. (8.132)
0
Then Gronwall inequality yields
kMδ f (t)k2L2 6 eCf,N,t kMδ (0)f0 )k2L2 . (8.133)
Since
kMδ f (t)kL2 = k(1 − δ∆)−N0 f (t)kH N t−4 ,
and
kMδ (0)f0 kL2 (R3 ) = k(1 − δ∆)−N0 f0 kH −4 (R3 ) 6 kf0 kH −4 (R3 ) 6 Ckf0 kL1 (R3 ) ,
we obtain
k(1 − δ∆)−N0 f (t)kH N t−4 6 C̃eCf,N,t kf0 kL1 (R3 ) ,
where the constant C̃ is independent of δ ∈ (0, 1). Finally, for any given
t > 0 since N can be arbitrarily large, by letting δ → 0, we have f (t) ∈
H +∞ (R3 ).
Remark 8.3. Recently, the existence of the classical global solutions of the
Boltzmann equation without angular cut-off was obtained by Gressman
and Strain [83]. There are some recent works by Chen, Li and Xu [34;
36] which have been devoted to the study of the regularity for the Landau
equation, which can be regarded as a limit of the Boltzmann equation when
the collisions become grazing. Mouhot and Villani [172] for the first time
establish the Landau damping in a nonlinear context.
Appendix A
Notations
We list some notations used in this book, most of them are familiar for the
PDE readers.
(1) R = the set of real, C = the set of complex number, N = the

set of natural number, Z = the set of integers, Z+ = N ∪ {0},
R+ = [0, ∞).
(2) a ∨ b = max(a, b); a ∧ b = min(a, b).
(3) p0 denotes the duality number of p, i.e., 1/p+1/p0 = 1, ∀ p ∈ [1, ∞].
(4) C > 1 and 0 < c < 1 express universal constants that can be
different at different places.
(5) A . B means A 6 CB; A ∼ B means A . B and B . A.
(6) Lp := Lp (Rn ) is the Lebesgue space,
Z 1/p
kf kp := kf kLp(Rn ) = |f (x)|p dx
Rn
(7) Let f = (f1 , ..., fn ) be a vector function,

kf kp = (kf1 k2p + ... + kfn k2p )1/2 .
(8) For x = (x1 , ..., xn ) ∈ Rn , we write |x| = (|x1 |2 + ... + |xn |2 )1/2 ,
and sometimes we also denote |x| = |x1 | + ... + |xn |.
(9) Let A ⊂ Rn , we denote by |A| the measure of A.
(10) Let T and S be operators, we use the notation T ∼ S to express
that T can be roughly regarded as S (when S is easily understood).
259
Appendix B
Definition of scattering operator
We give an exact definition of the scattering operator by the taking NLS

(7.54) as an example.
Definition B.1. Let X be a Banach space, S(t) = eit∆ be the evolution

semigroup associated to (7.54), and u(t) be the global solution of (7.54)
with the initial datum u0 ∈ X. If the limit
u+ = lim S(−t)u(t)
t→∞
exists in X, we say that u+ is the asymptotic state of u0 at +∞. Also, if
the limit
u− = lim S(−t)u(t)
t→−∞
exists in X, we say that u− is the asymptotic state of u0 at −∞. In other
words, u(t) behaves as t → ±∞ like the solutions S(t)u± of the linear
Schrödinger equation (or the free Schrödinger equation, i.e. iut + ∆u = 0).
−1
The inverse operators Ω± = U± of the operators U± : u0 7→ u± are called
the forward/backward wave operators. Note that the uniqueness aspect of
the H 1 -wellposedness theory ensures that the wave operators are injective.
If they are also surjective, in other word, if every H 1 -wellposed solution
is global and scatters in H 1 as t → +∞, we say that we also have the
asymptotic completeness. If the forward wave operator and the backward
wave operator exist simultaneously, the mapping S = Ω−1 + ◦ Ω− : u− 7→ u+
is called the scattering operator.
Remark B.1. The scattering theory involves two essential factors: the ex-
istence of the forward/backward wave operators and the asymptotic com-
pleteness. Generally speaking, the existence of wave operators is relatively
easy to establish as long as the power p is suitable, and especially if a small-
ness condition is assumed, both in focusing and defocusing cases. However,
261
262 Definition of scattering operator
the asymptotic completeness is a bit harder and requires some decay esti-
mates even in the defocusing case.
Remark B.2. To reduce the number of cases slightly, we shall only consider
scattering from t = 0 to t = +∞ or vice versa. One can certainly consider
scattering back and forth between t = 0 to t = −∞, or between t = −∞
and t = +∞, but the theory is more or less the same in each of these cases.
In general, for the well-posedness and the scattering theory of nonlinear

Schrödinger equations, we usually use the integral version of the equation,
that is, the Duhamel formula
Z t
u(t) = S(t − t0 )u(t0 ) − i S(t − τ )(|u(τ )|p−1 u(τ ))dτ. (B.1)
t0
For the scattering theory of the equation (7.54) in H 1 , we need to show
the global solution of (7.54) with the initial datum u(0) = u0 ∈ H 1 scatters
to a solution S(t)u+ of the associated linear equation as t → +∞ in H 1 ,
that is,
ku(t) − S(t)u+ kH 1 → 0, as t → +∞,
or equivalently,
kS(−t)u(t) − u+ kH 1 → 0, as t → +∞.
In other words, we require that the function S(−t)u(t) converges in H 1 as
t → +∞. From the Duhamel formula (B.1), we know
Z t
S(−t)u(t) = u0 − i S(−τ )(|u(τ )|p−1 u(τ ))dτ.
0
Thus, u scatters in H 1 as t → +∞ if and only if the improper integral
Z ∞
S(−τ )(|u(τ )|p−1 u(τ ))dτ (B.2)
0
is conditionally convergent in H 1 , in which case the asymptotic state u+ is
given by the formula
Z ∞
u+ = u0 − i S(−τ )(|u(τ )|p−1 u(τ ))dτ. (B.3)
0
We can regard the asymptotic state u+ as a nonlinear perturbation of the
initial state u0 . Comparing (B.1) with (B.3), we eliminate u0 to obtain the
identity
Z ∞
u(t) = S(t)u+ + i S(t − τ )(|u(τ )|p−1 u(τ ))dτ, (B.4)
t
which can be viewed as the limiting case t0 = +∞ of (B.1).
Appendix C
Some fundamental results
C.1 Gagliardo-Nirenberg inequality in Sobolev

spaces
The Gagliardo-Nirenberg inequality is a fundamental tool in the study of

partial differential equations, some special cases of which were discovered
by Gagliardo [75], Ladyzhenskaya [153] and Nirenberg [181]. The general
version can be stated as follows.
Theorem C.1. Let 1 6 p, p0 , p1 6 ∞, `, m ∈ N∪{0}, ` < m, `/m 6 θ 6 1,

1 ` 1−θ 1 m
= + +θ − . (C.1)
p n p0 p1 n
If m − ` − n/p0 is an integer, we further assume that `/m 6 θ < 1. Then
for any u ∈ C0∞ (Rn ),
X X
kDα ukLp (Rn ) . kuk1−θ
p n
L (R )
0
kDα ukθLp1 (Rn ) . (C.2)
|α|=` |α|=m
The proof of the Gagliardo-Nirenberg inequality is based on the global-

derivative analysis in Lp spaces, which is rather complicated, see [74; 87],
for instance.
C.2 Convexity Hölder inequality in sequence spaces `sp
Analogous to the convexity Hölder inequality in Triebel-Lizorkin spaces, we

have the following convexity Hölder inequality in sequence spaces `sp .
Lemma C.1. Let 0 < q 6 ∞, −∞ < s1 , s0 < ∞ with s0 6= s1 , 0 < θ < 1,
s = (1 − θ)s0 + θs1 . We have
k2sj aj k`q . k2s0 j aj k1−θ
`∞ k2
s1 j
aj kθ`∞ . (C.3)
263
264 Some fundamental results
Proof. We can assume that {aj } = {aj }j>0 , aj > 0 and s1 < s0 . Put
Ci = supj>0 2si j aj , i = 0, 1. It suffices to consider the case C1 > 0. We
have C1 6 C0 . Take j0 such that

C0 /2s0 j , j > j0 ,
min(C0 /2s0 j , C1 /2s1 j ) =
C1 /2s1 j , j 6 j0 .
It is easy to see that C0 ∼ C1 2(s0 −s1 )j0 . So,
k2s0 j aj k1−θ
`∞ k2
s1 j
aj kθ`∞ ∼ C1 2(s0 −s1 )j0 (1−θ) . (C.4)
On the other hand,
aj 6 C0 /2s0 j , j > j0 ; aj 6 C1 /2s1 j , 0 6 j 6 j0 .
A simple calculation yields
k2sj aj k`q . C1 2(s0 −s1 )j0 (1−θ) , (C.5)
C.3 Inclusion between homogeneous Triebel-Lizorkin

spaces
The inclusion among homogeneous Triebel-Lizorkin spaces is very useful

but Triebel [224] only claimed that it is probably true. Here we give the
details of the proof, which is analogous to that of Theorem 1.2.
Theorem C.2. Let 1 6 p1 < p2 < ∞, 1 6 r, q 6 ∞ and −∞ < s2 < s1 <

∞ satisfy s1 − n/p1 = s2 − n/p2 . Then we have
Ḟps11,q ⊂ Ḟps22,r . (C.6)
Proof. By `r ⊂ `q for q > r, it suffices to show that

Ḟps11,∞ ⊂ Ḟps22,1 . (C.7)
We can assume that kf kḞps1,∞ = 1. Recalling the equivalent norm on Lp ,
1
we have
Z ∞ ( )
X
p2
kf kḞ s2 ∼ tp2 −1 x : 2ks2 |(4k f )(x)| > t dt, (C.8)
p2 ,1 0
k∈Z
where |{· · · }| denotes the measure of the set {· · · }. It is easy to see that
∞
X
2ks2 |4k f | . 2K(s2 −s1 ) sup 2ks1 |4k f |. (C.9)
k
k=K+1
C.4. Riesz-Thorin interpolation theorem 265
By Corollary 1.1,
k4k f k∞ . 2kn/p1 k4k f kp1 . 2k(n/p1 −s1 ) kf kFps1,∞ . (C.10)

1
Hence, for any K ∈ Z,

K
X K
X
2ks2 |4k f | . 2k(s2 −s1 +n/p1 ) . 2Kn/p2 . (C.11)
k=−∞ k=−∞
Choosing K ∈ Z such that C2Kn/p2 ∼ t/2, we have 2K ∼ tp2 /n . If

P ks2
k∈Z 2 |(4k f )(x)| > t, then it follows from (C.9) and (C.11) that
∞
X
C2K(s2 −s1 ) sup 2ks1 |4k f | > 2ks2 |4k f | > t/2. (C.12)
k∈Z k=K+1
Collecting (C.8) and (C.12), we have

Z ∞

kf kpḞ2s2 . tp2 −1 x : sup 2ks1 |(4k f )(x)| > ctp2 /p1 dt
p2 ,1 0 k
Z ∞
p1 −1

. τ x : sup 2 ks1
|(4 f )(x)| > τ dτ
k
k
0
. 1, (C.13)
C.4 Riesz-Thorin interpolation theorem
Theorem C.3. Let 1 6 pi , qi 6 ∞, p0 6= p1 , q0 6= q1 satisfy
T : Lpi → Lqi , i = 0, 1.
Assume that θ ∈ (0, 1) satisifies
1 1−θ θ 1 1−θ θ
= + , = + .
p p0 p1 q q0 q1
Then we have T : Lp → Lq and
kT kLp→Lq 6 kT k1−θ θ
Lp0 →Lq0 kT kLp1 →Lq1 .
C.5 Hardy-Littlewood-Sobolev inequality
Up to now, the singular integration is still a difficult problem in the theory of

harmonic analysis. Hardy-Littlewood-Sobolev inequality is a fundamental
tool in this subject, cf. [202]. Let 0 < α < n,
Z
f (y)
Iα f (x) = dy.
R n |x − y|n−α
Proposition C.1. Let 1 < p, q < ∞ and 0 < α < n satisfy 1/p = 1/q +
α/n. Then we have
kIα f kq . kf kp . (C.14)
C.6 Van der Corput lemma
Lemma C.2. Let ϕ ∈ C0∞ (R). Assume that P ∈ C 2 (R) satisfies that for
any ξ ∈ supp ϕ, |P (k) (ξ)| > 1. Moreover, we assume that the following
alternative condition holds
(i) k > 2;
(ii) k = 1 and P 0 (x) is a monotone function.
Then we have Z

eiλP (ξ) ϕ(ξ)dξ . λ−1/k (kϕk∞ + kϕ0 k1 ).

C.7 Littlewood-Paley square function theorem
The Littlewood-Paley square function theorem is one of the most impor-

tant results in the early stage of harmonic analysis. It seems that Triebel-
Lizorkin spaces are also inspired by this theorem.
Proposition C.2. Let 1 < p < ∞, s ∈ R. Then
kukḢ s ∼ kukḞ s , kukHps ∼ kukFp,2
s . (C.15)
p p,2
In particular,
kukLp ∼ kukḞ 0 , kukLp ∼ kukFp,2
0 . (C.16)
p,2
The proof of the Littlewood-Paley square function theorem can be found

in [202].
C.8. Complex interpolation in modulation spaces 267
C.8 Complex interpolation in modulation spaces
We have the following
Theorem C.4. Let 0 < p, q, pi , qi 6 ∞, s, si ∈ R with i = 0, 1 and

1 1−θ θ 1 1−θ θ
s = (1 − θ)s0 + θs1 , = + , = + . (C.17)
p p0 p1 q q0 q1
Then we have
(Mps00,q0 , Mps11,q1 )θ = Mp,q
s
.
Recall that the result of Theorem C.4 is quite similar to Besov spaces,
indeed, if (C.17) holds, then we have
(Bps00 ,q0 , Bps11 ,q1 )θ = Bp,q
s
. (C.18)
See [245].
C.9 Christ-Kiselev lemma
The Christ-Kiselev lemma (cf. [46]) is very useful for the study of nonlinear
dispersive equations. There are a series of generalizations to the Christ-
Kiselev lemma in recent years, see Molinet-Ribaud [166] and Smith-Sogge
[201]. The following result is due to Wang-Han-Huang [243]. Denote
Z ∞ Z t
0 0 0
T f (t) = K(t, t )f (t )dt , Tre f (t) = K(t, t0 )f (t0 )dt0 . (C.19)
−∞ 0
If T : Y1 → X1 implies that Tre : Y1 → X1 , then T : Y1 → X1 is said to

be a well restricted operator.
Proposition C.3. Let T be as in (C.19). We have the following conclu-

sions.
(1) If ∧3i=1 pi > (∨3i=1 qi ) ∨ (q1 q3 /q2 ), then T : Lqx11 Lqx22 Lqt 3 (R3 ) →
Lpx11 Lpx22 Lpt 3 (R3 ) is a well restricted operator.
(2) If q1 < ∧3i=1 pi , then T : Lqt 1 Lqx21 Lqx32 (R3 ) → Lpx11 Lpx22 Lpt 3 (R3 ) is a
well restricted operator.
(3) If p1 > (∨3i=1 qi ) ∨ (q1 q3 /q2 ), then T : Lqx11 Lqx22 Lqt 3 (R3 ) →
Lpt 1 Lpx21 Lpx32 (R3 ) is a well restricted operator.
(4) If ∧3i=1 pi > (∨3i=1 qi ) ∨ (q1 q3 /q2 ), then T : Lqx11 Lqx22 Lqt 3 (R3 ) →
Lpx22 Lpx11 Lpt 3 (R3 ) is a well restricted operator.
In higher spatial dimensions, we have similar results.
Proposition C.4. Let T be as in (C.19). We have the following results.

(1) If min(p1 , p2 , p3 ) > max(q1 , q2 , q3 , q1 q3 /q2 ), then
T : Lqx11 Lqx22 ,...,xn Lqt 3 (Rn+1 ) → Lpx11 Lpx22 ,...,xn Lpt 3 (Rn+1 )
is a well restriction operator.
(2) If p0 > (∨3i=1 qi ) ∨ (q1 q3 /q2 ), then
T : Lqx11 Lqx22 ,...,xn Lqt 3 (Rn+1 ) → Lpt 0 Lpx11 ...Lpxnn (Rn+1 )
(3) If q0 < min (p1 , p2 , p3 ), then
T : Lqt 0 Lqx11 ...Lqxnn (Rn+1 ) → Lpx11 Lpx22 ,...,xn Lpt 3 (Rn+1 )
(4) If min(p1 , p2 , p3 ) > max(q1 , q2 , q3 , q1 q3 /q2 ), then
T : Lqx12 Lqx21 ,x3 ,...,xn Lqt 3 (Rn+1 ) → Lpx11 Lpx22 ,...,xn Lpt 3 (Rn+1 )
Bibliography
[1] M. J. Ablowitz, R. Haberman, Nonlinear evolution equations in two and three

dimensions, Phys. Rev. Lett., 35 (1975), 1185–1188.
[2] R. Alexandre, L. Desvillettes, C. Villani and B. Wennberg, Entropy dissipation and
long range interactions, Arch. Rat. Mech. Anal., 152 (2000), 327–355.
[3] R. Alexandre and M. El Safadi, Littlewood–Paley theory and regularity issues in
Boltzmann homogeneous equations. I. Non-cutoff case and Maxwellian molecules,
Math. Models Methods Appl. Sci., 15 (2005), 907–920.
[4] R. Alexandre and M. El Safadi, Littlewood–Paley theory and regularity issues
in Boltzmann homogeneous equations. II. Non cutoff case and non Maxwellian
molecules, Discrete Contin. Dyn. Syst., 24 (2009), 1–11.
[5] R. Aulaskari, J. Xiao, R. H. Zhao, On subspaces and subsets of BM OA and U BC,
Analysis, 15 (1995), 101–121.
[6] H. Bahouri and J. Shatah, Decay estimates for the critical semilinear wave equa-
tions, Ann. Inst. H. Poincaré, Anal. Non Linéaire, 15 (1997), 783–789.
[7] M. Beals, Self-spreading and strength of singularities for solutions to semilinear
wave equation, Ann. of Math., 118 (1983), 187–214.
[8] I. Bejenaru, S. Herr, J. Holmer, D. Tataru, On the 2D Zakharov system with L2 -
Schrödinger data, Nonlinearity, 22 (2009), 1063–1089.
[9] I. Bejenaru and T. Tao, Sharp well-posedness and ill-posedness results for a
quadratic non-linear Schrödinger equation, J. Funct. Anal., 233 (2006), 228–259.
[10] T. B. Benjamin, Internal waves of permanent form in fluids of great depth, J. Fluid
Mech., 29 (1967), 559–592.
[11] A. Bényi, K.A. Okoudjou, Local well-posedness of nonlinear dispersive equations on
modulation spaces, Bull. London Math. Soc., 41 (2009), 549–558.
[12] A. Bényi, K. Gröchenig, K.A. Okoudjou and L.G. Rogers, Unimodular Fourier
multiplier for modulation spaces, J. Funct. Anal., 246 (2007), 366–384.
[13] J. Bergh and J. Löfström, Interpolation Spaces, Springer–Verlag, 1976.
[14] J. Bourgain, Fourier transform restriction phenomena for certain lattice subsets
and applications to nonlinear evolution equations I, Schrödinger equations, Geom.
Funct. Anal., 3 (1993), 107–156.
[15] J. Bourgain, Fourier transform restriction phenomena for certain lattice subsets and
applications to nonlinear evolution equations II, The KdV-equation, Geom. Funct.
Anal., 3 (1993), 209–262.
[16] J. Bourgain, On the Cauchy problem for the Kadomtsev–Petviashvili equation,
Geom. Funct. Anal., 3 (1993), 315–341.
[17] J. Bourgain, Refinements of Strichartz’ inequality and applications to 2D-NLS with
critical nonlinearity, Int. Math. Res. Not., (1998), no. 5, 253–283.
269
270 Bibliography
[18] J. Bourgain, Global solutions of nonlinear Schrödinger equation, Amer. Math. Soc.,
1999.
[19] J. Bourgain, Global well posedness of defocusing critical nonlinear Schrödinger
equation in the radial case, J. Amer. Math. Soc., 12 (1999), 145–171.
[20] J. Bourgain, J. Colliander, On well-posedness of the Zakharov system, Int. Math.
Res. Not., 11 (1996), 515–546.
[21] P. Brenner, On space-time means and everywhere defined scattering operators for
nonlinear Klein–Gordon equations, Math. Z., 186 (1984), 383–391.
[22] P. Brenner, On scattering and everywhere-defined scattering operators of nonlonear
Klein-Gordon equations, J. Differential Equations, 56 (1985), 310–344.
[23] H. Brezis and P. Mironescu, Gagliardo–Nirenberg, composition and products in
fractional Sobolev spaces, J. Evol. Equ., 1 (2001), 387–404.
[24] L. Carleson, Some analytical problems related to statistical mechanics, Euclidean
Harmonic Analysis, Lecture Notes in Math., Vol. 779, Springer, Berlin, 1980, 5–45.
[25] T. Cazenave, F. B. Weissler, Some remarks on the nonlinear Schrödinger equation
in the critical case, Lecture Notes in Math., 1394 (1989), 18–29.
[26] T. Cazenave and F.B. Weissler, The Cauchy problem for the critical nonlinear
Schrödinger equation in H s , Nonlinear Anal. TMA, 14 (1990), 807–836.
[27] T. Cazenave, Semilinear Schrödinger equations, Courant Lecture Notes in Mathe-
matics, Vol. 10, 2003.
[28] C. Cercignani, The Boltzmann Equation and Its Applications, Springer–Verlag,
1988.
[29] J. Y. Chemin, Théormes d’unicité pour le système de Navier–Stokes tridimensionnel,
Journal d’Anal. Math. 77 (1999), 27–50.
[30] J. Y. Chemin, I. Gallagher, On the global wellposedness of the 3D Navier–Stokes
equations with large initial data. Ann. Sci. École Norm. Sup. 39 (2006), 679–698.
[31] J. Y. Chemin, I. Gallagher, Wellposedness and stability results for the Navier-Stokes
equations in R3 , Ann. Inst. H. Poincaré, Anal. Non Linéaire, 26 (2009), 599–624.
[32] J. Y. Chemin, I. Gallagher, Large, global solutions to the Navier–Stokes equations,
slowly varying in one direction, Tran. Amer. Math. Soc., 362 (2010), 2859–2873.
[33] J. Y. Chemin, P. Zhang, On the global wellposedness to the 3-D incompressible
anisotropic Navier–Stokes equations. Comm. Math. Phys., 272 (2007), 529–566.
[34] H. Chen, W. X. Li, C. J. Xu, Propagation of Gevrey regularity for solutions of
Landau equations, Kinet. Relat. Models, 1 (2008), 355–368.
[35] H. Chen, W. X. Li, C. J. Xu, Gevrey hypoellipticity for linear and non-linear Fokker–
Planck equations, J. Differential Equations, 246 (2009), 320–339.
[36] H. Chen, W. X. Li, C. J. Xu, Analytic smoothness effect of solutions for spatially
homogeneous Landau equation, J. Differential Equations, 248 (2010), 77–94.
[37] H. Chen, W. X. Li, C. J. Xu, Gevrey hypoellipticity for a class of kinetic equations,
Commun. Part. Diff. Equations, 36 (2011), 693–728.
[38] J. C. Chen, D. S. Fan and L. Sun, Asymptotic estimates For unimodular Fourier
multipliers on modulation spaces, Preprint.
[39] Cheskidov A., Shvydkoy R, The regularity of weak solutions of the 3D Navier-Stokes
−1
equations in B∞,∞ , Arch. Ration. Mech. Anal., 195 (2010), 159–169.
[40] J. Colliander, Nonlinear Schrödinger equations, Lecture notes, 2004.
[41] J. E. Colliander, J. M. Delort, C. E. Kenig and G. Staffilani, Bilinear estimates and
applications to 2D NLS, Trans. Amer. Math. Soc., 353 (2001), 3307–3325.
[42] J. Colliander, M. Keel, G. Staffilani, H. Takaoka and T. Tao, Sharp global well-
posedness for KdV and modified KdV on R and T, J. Amer. Math. Soc., 16 (2003),
705–749.
Bibliography 271
[43] J. Colliander, M. Keel, G. Staffilani, H. Takaoka and T. Tao, Viriel, Morawetz, and
interaction Morawetz inequalities, Lecture notes, 2004.
[44] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, and T. Tao, Global well-posedness
and scattering for the energy-critical nonlinear Schrödinger equation in R3 , Ann. of
Math., 167 (2008), 767–865.
[45] J. Colliander, M. Keel, G. Staffilani, H. Takaoka and T. Tao, Resonant de-
compositions and the I-method for cubic nonlinear Schrödinger equation on R2 ,
arXiv:0704.2730.
[46] M. Christ, A. Kiselev, Maximal functions associated to filtrations, J. Funct. Anal.,
179 (2001), 406–425.
[47] M. Christ, J. Colliander and T. Tao, Asymptotics, frequency modulation and low
regularity ill-posedness for canonical defocusing equations, Amer. J. Math., 125
(2003), 1235–1293.
[48] P. A. Clarkson and J. A. Tuszyriski, Exact solutions of the multidimensional deriva-
tive nonlinear Schrödinger equation for many-body systems near criticality, J. Phys.
A: Math. Gen. 23 (1990), 4269–4288.
[49] E. Cordero, F. Nicola, Some new Strichartz estimates for the Schrödinger equation.
J. Differential Equations, 245 (2008), 1945–1974.
[50] E. Cordero, F. Nicola, Remarks on Fourier multipliers and applications to the wave
equation, J. Math. Anal. Appl., 353 (2009), 583–591.
[51] E. Cordero, F. Nicola, Sharpness of some properties of Wiener amalgam and mod-
ulation spaces. Bull. Aust. Math. Soc., 80 (2009), 105–116.
[52] S. B. Cui, Pointwise estimates for a class of oscillatory integrals and related Lp –Lq
estimates, J. Fourier Anal. Appl., 11 (2005), 441–457.
[53] S. B. Cui, Pointwise estimates for oscillatory integrals and related Lp -Lq estimates,
II Multidimensional case, J. Fourier Anal. Appl., 12 (2006), 605–627.
[54] J. M. Delort, Existence globale et comportement asymptotique pour léquation de
Klein–Gordon quasi-linéaire à données petites en dimension 1, Ann. Sci. École
Norm. Sup., 34 (2001), 1–61.
[55] J. M. Delort, D. Y. Fang, R. Y. Xue, Global existence of small solutions for quadratic
quasilinear Klein-Gordon systems in two space dimensions. J. Funct. Anal., 211
(2004), 288–323.
[56] C. Deng, J. H. Zhao, S. B. Cui, Well-posedness of a dissipative nonlinear elec-
trohydrodynamic system in modulation spaces, Nonlinear Anal., TMA, 73 (2010),
2088–2100.
[57] L. Desvillettes, About the regularization properties of the non cut-off Kac equation,
Comm. Math. Phys., 168 (1995), 417–440.
[58] L. Desvillettes, Regularization properties of the 2-dimensional non radially symmet-
ric non cutoff spatially homogeneous Boltzmann equation for Maxwellian molecules,
Trans. Theory Stat. Phys., 26 (1997), 341–357.
[59] L. Desvillettes, About the use of the Fourier transform for the Boltzmann equation,
Summer School on Methods and Models of Kinetic Theory (MMKT 2002). Riv.
Mat. Univ. Parma, 7 (2003), 1–99.
[60] L. Desvillettes and B. Wennberg, Smoothness of the solution of the spatially homo-
geneous Boltzmann equation without cutoff, Comm. Partial Differential Equations,
29 (2004), 133–155.
[61] W. Y. Ding, Y. D. Wang, Schrödinger flow of maps into symplectic manifolds,
Science in China Ser. A, 41 (1998), 746–755.
[62] W. Y. Ding and Y. D. Wang, Local Schrödinger flow into Kähler manifolds, Sci.
China Ser. A, 44 (2001), 1446–1464.
272 Bibliography
[63] J. M. Dixon and J. A. Tuszynski, Coherent structures in strongly interacting many-

body systems: II, Classical solutions and quantum fluctuations, J. Phys. A: Math.
Gen., 22 (1989), 4895–4920.
[64] B. Dodson, Global well-posedness and scattering for the defocusing, L2 -critical,
nonlinear Schrödinger equation when d > 3, arXiv:0912.2467.
nonlinear Schrödinger equation when d = 2, arXiv:1006.1375.
nonlinear Schrödinger equation when d = 1, arXiv:1010.0040.
[67] H. Dong, D. Du, On the local smoothness of solutions of the Navier–Stokes equa-
tions, J. Math. Fluid Mech., 9 (2007), 139–152.
[68] H. Dong, D. Du, The Navier–Stokes equations in the critical Lebesgue space, Comm.
Math. Phys., 292 (2009), 811–827.
[69] X. T. Duong, L. X. Yan, New function spaces of BMO type, the John-Nirenberg
inequality, interpolation, and applications, Comm. Pure Appl. Math. 58 (2005),
1375–1420.
[70] L. Escauriaza, G. A. Seregin, and V. Šverák, L3,∞ -solutions of Navier–Stokes equa-
tions and backward uniqueness, Uspekhi Mat. Nauk, 58 (2003), 3–44.
[71] D. Y. Fang, H. Pecher, S. J. Zhong, Low regularity global well-posedness for the
two-dimensional Zakharov system, Analysis (Munich), 29 (2009), 265–281.
[72] D. Y. Fang, J. Xie and T. Cazenave, Scattering for the focusing energy-subcritical
NLS, Science in China, Ser A, Math., in press.
[73] H. G. Feichtinger, Modulation spaces on locally compact Abelian group, Technical
Report, University of Vienna, 1983. Published in: “Proc. Internat. Conf. on
Wavelet and Applications”, 99–140. New Delhi Allied Publishers, India, 2003.
[74] A. Friedmann, Partial Differential Equations, Holt, Rinehart and Winston, New
York, 1969.
[75] E. Gagliardo, Proprieta di alcune classi di funzioni in pia variabili, Richerche Mat.,
7 (1958), 102–137; 9 (1959), 24–51.
[76] V. Georgiev, P. Popivanov, Global solution to the two-dimensional Klein–Gordon
equation, Comm. Partial Differential Equations 16 (1991), 941–995.
[77] Y. Giga, Solutions for semilinear parabolic equations in Lp and regularity of weak
solutions of the Navier–Stokes system, J. Differential Equations,61 (1986), 186–212.
[78] J. Ginibre and G. Velo, On a class of nonlinear Schrödinger equations II, Scattering
theory J. Funct. Anal., 32 (1979), 33–71.
[79] J. Ginibre and G. Velo, Time decay of finite energy solutions of the nonlinear Klein-
Gordon and Schrödinger equations, Ann. Inst. H. Poincaré, Phys. Theor., 43
(1985), 399–442.
[80] J. Ginibre and G. Velo, Generalized Strichartz inequalities for the wave equation,
J. Funct. Anal., 133 (1995), 50–68.
[81] J. Ginibre, Y. Tsutsumi and G. Velo, On the Cauchy problem for the Zakharov
System, J. Funct. Anal., 151 (1997), 384–436.
[82] L. Grafakos, Classical and modern Fourier analysis, Pearson/Prentice Hall, 2004.
[83] P. T. Gressman, R. M. Strain, Global classical solutions of the Boltzmann equation
without angular cut-off, arXiv:1011.5441.
[84] M. Grillakis, On nonlinear Schrödinger equations, Comm. Partial Differential Equa-
tions, 25 (2000), 1827–1844.
[85] P. Gröbner, Banachräume Glatter Funktionen und Zerlegungsmethoden, Doctoral
thesis, University of Vienna, 1992.
[86] K. Gröchenig, Foundations of Time–Frequency Analysis, Birkhäuser, Boston, MA,
Bibliography 273
2001.
[87] B. L. Guo, Viscosity Elimination Method and the Viscosity of Difference Scheme,
Chinese Sci. Publisher, 2004.
[88] B. L. Guo, S. J. Ding, Landau–Lifshitz equations, Frontiers of Research with the
Chinese Academy of Sciences, 1. World Scientific, 2008.
[89] B. L. Guo and B. X. Wang, The Cauchy problem for the Davey–Stewartson systems,
Comm. Pure Appl. Math., 52 (1999), 1477–1490.
[90] Z. H. Guo, Global Well-posedness of Korteweg-de Vries equation in H −3/4 (R), J.
Math. Pures Appl., 91 (2009), 583–597.
[91] Z. H. Guo, Local well-posedness and a priori bounds for the modified Benjamin–Ono
equation without using a gauge transformation, arXiv:0807.3764.
[92] Z. H. Guo, Local Well-posedness for dispersion generalized Benjamin–Ono equations
in Sobolev spaces, arXiv:0812.1825.
[93] Z. H. Guo, The Cauchy problem for a class of derivative nonlinear dispersive equa-
tions, Doctoral thesis, Peking Univ., (2009).
[94] Z. H. Guo, L. Z. Peng, B. X. Wang, Decay estimates for a class of wave equations,
J. Funct. Anal., 254 (2008), 1642–1660.
[95] Z. H. Guo, L. Z. Peng and B. X. Wang, On the local regularity of the KP-I equation
in anisotropic Sobolev space, J. Math. Pure Appl., 94 (2010), 414–432.
[96] Z. H. Guo and B. X. Wang, Global well posedness and inviscid limit for the
Korteweg–de Vries–Burgers equation, J. Differential Equations, 246 (2009), 3864–
3901.
[97] M. Hadac, Well-posedness for the Kadomtsev–Petviashvili-II equation and general-
izations, Trans. Amer. Math. Soc., 360 (2008), 6555–6572.
[98] M. Hadac, S. Herr and H. Koch, Well-posedness and scattering for the KP-II equa-
tion in a critical space, Ann. Inst. H. Poincaré, Anal. Non Linéaire, 26 (2009),
917–941.
[99] H. Hajaiej, L. Molinet, T. Ozawa, B. X. Wang, Necessary and sufficient conditions
for the fractional Gagliardo–Nirenberg inequalities and applications to the Navier–
Stokes and generalized boson equations, arXiv:1004.4287.
[100] C. C. Hao, L. Hsiao and B. X. Wang, Wellposedness for the fourth order nonlinear
Schrödinger equations, J. Math. Anal. Appl., 320 (2006), 246–265.
[101] C. C. Hao, L. Hsiao and B. X. Wang, Wellposedness of Cauchy problem for the
fourth order nonlinear Schrödinger equations in multi-dimensional spaces, J. Math.
Anal. Appl., 328 (2007), 58–83.
[102] N. Hayashi, The initial value problem for the derivative nonlinear Schrödinger equa-
tion in the energy space. Nonlinear Anal., TMA, 20 (1993), 823–833.
[103] N. Hayashi, E.I. Kaikina, P.I. Naumkin, Damped wave equation in the subcritical
case, J. Differential Equations, 207 (2004), 161–194.
[104] N. Hayashi, E.I. Kaikina, P.I. Naumkin, Damped wave equation with a critical
nonlinearity, Trans. Amer. Math. Soc., 358 (2006), 1117–1163.
[105] N. Hayashi and T. Ozawa, Finite energy solutions of nonlinear Schrödinger equa-
tions of derivative type, SIAM J. Math. Anal., 25 (1994), 1488–1503.
[106] N. Hayashi, T. Ozawa, Modified wave operators for the derivative nonlinear
Schrödinger equation. Math. Ann., 298 (1994), 557–576.
[107] N. Hayashi, T. Ozawa, Remarks on nonlinear Schrödinger equations in one space
dimension. Differential Integral Equations 7 (1994), 453–461.
[108] S. Herr, A. Ionescu, C. Kenig and H. Koch, A para-differential renormalization
technique for nonlinear dispersive equations, arXiv:0907.4649.
[109] Z. H. Huo and Y. L. Jia, The Cauchy problem for the fourth-order nonlinear
274 Bibliography
Schrödinger equation related to the vortex filament, J. Differential Equations, 214

(2005), 1–35.
[110] Z. H. Huo and Y. L. Jia, A refined well-posedness for the fourth-order nonlinear
Schrödinger equation related to the vortex filament, Comm. Partial Differential
Equations, 32 (2007), 1493–1510.
[111] Z. H. Huo, Y. Morimoto, S. Ukai and T. Yang, Regularity of solutions for spatially
homogeneous Boltzmann equation without angular cutoff, Kinet. Relat. Models,
1 (2008), 453–489.
[112] R. Ikehata, K. Nishihara, H. J. Zhao, Global asymptotics of solutions to the Cauchy
problem for the damped wave equation with absorption, J. Differential Equations,
226 (2006), 1–29.
[113] A. D. Ionescu and C. E. Kenig, Global well-posedness of the Benjamin–Ono equation
in low-regularity spaces, J. Amer. Math. Soc., 20 (2007), 753–798.
[114] A. Ionescu and C. E. Kenig, Low-regularity Schrödinger maps, II: Global well posed-
ness in dimensions d > 3, Commun. Math. Phys., 271 (2007), 523–559.
[115] A. D. Ionescu, C. E. Kenig and D. Tataru, Global well-posedness of the KP-I initial-
value problem in the energy space, Invent. Math., 173 (2008), 265–304.
[116] T. Iwabuchi, Navier-Stokes equations and nonlinear heat equations in modulation
spaces with negative derivative indices. J. Differential Equations, 248 (2010), 1972–
2002.
[117] F. John, Plane waves and spherical means, Applied to partial differential equations,
Springer, 1981.
[118] L.V. Kapitanskii, Weak and yet weak solutions of semilinear wave equations, Comm.
Partial Differential Equations, 19 (1994), 1629–1676.
[119] T. Kato, Strong Lp solutions of the Navier–Stokes equations in Rm with applications
to weak solutions, Math. Z., 187 (1984), 471–480.
[120] T. Kato, On nonlinear Schrödinger equations, Ann.Inst.H.Poincaré, Phys. Theor.
46 (1987), 113–129.
[121] M. Keel and T. Tao, End point Strichartz estimates, Amer. J. Math., 120 (1998),
955–980.
[122] C. E. Kenig, On the local and global well-posedness theory for the KP-I equation,
Ann. Inst. H. Poincaré, Anal. Non Linéaire, 21 (2004), 827–838.
[123] C. E. Kenig. G. S. Koch, An alternative approach to regularity for the Navier–
Stokes equations in critical spaces, Preprint.
[124] C.E. Kenig, and F. Merle, Global well-posedness, scattering and blow-up for the
energy-critical, focusing, non-linear Schrödinger equation in the radial case. Invent.
Math., 166 (2006), 645–675.
[125] C. E. Kenig, G. Ponce, C. Rolvent, L. Vega, The genreal quasilinear untrahyperbolic
Schrodinger equation, Adv. Math., 206 (2006), 402–433.
[126] C. E. Kenig, G. Ponce and L. Vega, Oscillatory integrals and regularity of dispersive
equations, Indiana Univ. Math. J., 40 (1991), 33–69.
[127] C. E. Kenig, G. Ponce and L. Vega, Well-posedness of the initial value problem for
the Korteweg-de Vries equation, J. Amer. Math. Soc., 4 (1991), 323–347.
[128] C. E. Kenig, G. Ponce and L. Vega, The Cauchy problem for the Korteweg-de Vries
equation in Sobolev spaces of negative indices, Duke Math. J., 71 (1993), 1–21.
[129] C.E. Kenig, G. Ponce and L. Vega, Well-posedness and scattering results for gen-
eralized KdV equation via contraction principles, Comm. Pure Appl. Math., 46
(1993), 527–620.
[130] C. E. Kenig, G. Ponce, L. Vega, Small solutions to nonlinear Schrödinger equation,
Ann. Inst. Henri Poincaré, Sect C, 10 (1993), 255–288.
Bibliography 275
[131] C. E. Kenig, G. Ponce and L. Vega, Quadratic forms for the 1-D semilinear
Schrödinger equation, Trans. Amer. Math. Soc., 346 (1996), 3323–3353.
[132] C. E. Kenig, G. Ponce and L. Vega, A bilinear estimate with applications to the
KdV equation, J. Amer. Math. Soc., 9 (1996), 573–603.
[133] C. E. Kenig, G. Ponce and L. Vega, Smoothing effects and local existence theory
for the generalized nonlinear Schrödinger equations, Invent. Math., 134 (1998),
489–545.
[134] C. E. Kenig, G. Ponce and L. Vega, On the ill-posedness of some canonical dispersive
equations, Duke Math. J., 106 (2001), 617–633.
[135] C. E. Kenig, G. Ponce, L. Vega, The Cauchy problem for quasi-linear Schrödinger
equations, Invent. Math., 158 (2004), 343–388.
[136] C. E. Kenig and H. Takaoka, Global wellposedness of the modified Benjamin–Ono
equation with initial data in H 1/2 , Int. Math. Res. Not., 2006, Art. ID 95702,
1–44.
[137] R. Killip, D. Li, M. Visan and X. Y. Zhang, Characterization of minimal-mass
blowup solutions to the focusing mass-critical NLS, SIAM J. Math. Anal., 41
(2009), 219–236.
[138] R. Killip, T. Tao, and M. Visan, The cubic nonlinear Schrödinger equation in two
dimensions with radial data, J. Euro. Math. Soc., 11 (2009), 1203–1258.
[139] R. Killip, M. Visan, The focusing energy-critical nonlinear Schrödinger equation in
dimensions five and higher, Amer. J. Math. 132 (2010), no. 2, 361–424..
[140] R. Killip, M. Visan, and X. Y. Zhang, The mass-critical nonlinear Schrödinger
equation with radial data in dimensions three and higher, Anal. of PDE, 1 (2008),
229–266.
[141] N. Kishimoto, Well-posedness of the Cauchy problem for the Korteweg-de Vries
equation at the critical regularity, Preprint.
[142] N. Kishimoto, Low-regularity Local Well-posedness for Quadratic Nonlinear
Schrödinger Equations, Master Thesis, Kyoto University, 2008.
[143] S. Klainerman, Long-time behavior of solutions to nonlinear evolution equations,
Arch. Rational Mech. Anal., 78 (1982), 73–98.
[144] S. Klainerman, Global existence of small amplitude solutions to nonlinear Klein–
Gordon equations in four space-time dimensions, Comm. Pure Appl. Math., 38
(1985), 631–641.
[145] S. Klainerman and M. Machedon, Space-time estimates for null forms and the local
existence theorem, Comm. Pure Appl. Math., 46 (1993), no. 9, 1221–1268.
[146] S. Klainerman and M. Machedon, Smoothing estimates for null forms and applica-
tions, Duke Math. J., 81 (1995), 99–133.
[147] S. Klainerman, G. Ponce, Global small amplitude solutions to nonlinear evolution
equations, Commun. Pure Appl. Math., 36 (1983), 133–141.
[148] M. Kobayashi, Modulation spaces M p,q for 0 < p, q 6 ∞, J. of Funct. Spaces and
Appl., 4 (2006), no.3, 329–341.
[149] M. Kobayashi, Dual of modulation spaces, J. of Funct. Spaces and Appl., 5 (2007),
1–8.
[150] M. Kobayashi, Y. Sawano, Molecular decomposition of the modulation spaces, Os-
aka J. Math., 47 (2010), 1029–1053.
[151] H. Koch and D. Tataru, Well-posedness for the Navier–Stokes equations, Adv.
Math., 157 (2001), 22–35.
[152] B. G. Konopelchenko, B. T. Matkarimov, On the inverse scattering transform of
the Ishimori equations, Phys. Lett., 135 (1989), 183–189.
276 Bibliography
[153] O. Ladyzhenskaya, The solvability “in the large” of the boundary value problem for
the Navier-Stokes equation in the case of two space variables, Comm. Pure Appl.
Math., 12 (1959), 427–433.
[154] O. Ladyzhenskaya, The Mathematical Theory of Viscous Incompressible Flow, Sec-
ond English edition, Mathematics and its Applications, Vol. 2, Gordon and Breach,
Science Publishers, New York–London–Paris, 1969.
[155] P. G. Lemarié-Rieusset, Recent developments in the Navier–Stokes problem, A CRC
Press Company, 2002.
[156] S.P. Levandosky, Decay estimates for fourth order wave equations, J. Differential
Equations, 143 (1998), 360–413.
[157] F. Linares and G. Ponce, On the Davey–Stewartson systems, Ann. Inst. H.
Poincaré, Anal. NonLinéaire, 10 (1993), 523–548.
[158] H. Lindblad and C.D. Sogge, On existence and scattering with minimal regularity
for semilinear wave equations, J. Funct. Anal., 130 (1995), 357–426.
[159] J. L. Lions, Some Methods for Solving Nonlinear Boundary Value Problems, Trans-
lated by B. L. Guo and L. R. Wang, Sun Yat-sen University Press, 1992.
[160] P. L. Lions, Regularity and compactness for Boltzmann collision operator without
cut-off. C. R. Acad. Sci. Paris Series I, 326 (1998), 37–41.
[161] W. Littman, Fourier transforms of surface-carried measures and differentiability of
surface averages, Bull. of Amer. Math. Soc., 69 (1963), 766–770.
[162] S. Machihara and T. Ozawa, Interpolation inequalities in Besov spaces, Proc. Amer.
Math. Soc., 131 (2002), 1553–1556.
−1
[163] R. May, Rôle de l’space de Besov B∞,∞ dans le contrôle de l’explosion évetuelle
en temps fini des solutions réguli‘eres des équations de Navier–Stokes, C. R. Acad.
Sci. Paris, 323 (2003), 731–734.
[164] A. Miyachi, F. Nicola, S. Riveti, A. Taracco and N. Tomita, Estimates for unimodu-
lar Fourier multipliers on modulation spaces, Proc. Amer. Math. Soc., 137 (2009),
3869–3883.
[165] L. Molinet and F. Ribaud, On the low regularity of the Korteweg–de Vries–Burgers
equation, Int. Math. Res. Not., 37 (2002), 1979–2005.
[166] L. Molinet, F. Ribaud, Well-posedness results for the generalized Benjamin–Ono
equation with small initial data, J. Math. Pures Appl., 83 (2004), 277–311.
[167] L. Molinet, J.C. Saut and N. Tzvetkov, Ill-posedness issues for the Benjamin–Ono
and related equations, SIAM J. Math. Anal., 33 (2001), 982–988.
[168] L. Molinet, J. C. Saut and N. Tzvetkov, Global well-posedness for the KP-I equation,
Math. Ann., 324 (2002) 255–275, 328 (2004), 707–710.
[169] L. Molinet, J. C. Saut and N. Tzvetkov, Well-posedness and ill-posedness results
for the Kadomtsev–Petviashvili-I equation, Duke Math. J., 115 (2002), 353–384.
[170] C. S. Morawetz and W. A. Strauss, Decay and scattering of solutions of a nonlinear
relativistic wave equation, Comm. Pure Appl. Math.,25 (1972), 1–31.
[171] Y. Morimoto, S. Ukai, C. J. Xu and T. Yang, Regularity of the solution to the
spatially homogeneous Boltzmann equation without angular cutoff, Discrete Contin.
Dyn. Syst., 24 (2009), 187–212.
[172] C. Mouhot and C. Villani, On the Landau damping, arXiv:0904.2760.
[173] M. Nakamura and T. Ozawa, Nonlinear Schrödinger equations in the Sobolev space
of critical order, J. Funct. Anal., 155 (1998), 364–380
[174] M. Nakamura and T. Ozawa, The Cauchy problem for nonlinear Klein–Gordon
equations in the Sobolev spaces, Publ. Res. Inst. Math. Sci., 37 (2001), 255–293.
[175] K. Nakanishi, Energy scattering for Hartree equations, Math. Res. Lett., 6 (1999),
107–118.
Bibliography 277
[176] K. Nakanishi, Energy scattering for nonlinear Klein–Gordon and Schrödinger equa-
tions in spatial dimensions 1 and 2, J. Funct. Anal., 169 (1999), 201–225.
[177] K. Nakanishi, Scattering theory for nonlinear Klein–Gordon equation with Sobolev
critical power, Int. Math. Res. Notices, 1999, 31–60.
[178] K. Nakanishi, Unique global existence and asymptotic behavior of solutions for wave
equations with non-coercive nonlinearity, Comm. Partial Differential Equations, 24
(1999), 185–221.
[179] K. Nakanishi, Remarks on the energy scattering for nonlinear Klein–Gordon and
Schrödinger equations, Tohoku Math. J., 53 (2001), 285–303.
[180] K. Nakanishi, H. Takaoka and Y. Tsutsumi, Counterexamples to bilinear estimates
related to the KdV equation and the nonlinear Schrödinger equation, Methods of
Appl. Anal., 8 (2001) 569–578.
[181] L. Nirenberg, On elliptic partial differential equations, Ann. Sc. Norm. Sup. Pisa,
Ser. III, 13, (1959), 115–162.
[182] H. Ono, Algebraic solitary waves in stratified fluids, J. Phys. Soc. Japan, 39 (1975)
1082–1091.
[183] F. Oru, Rôle des oscillations dans quelques probl‘emes d’nalyse non-linéaire, Doc-
torat de Ecole Normale Supérieure de Cachan, 1998.
[184] T. Ozawa, On the nonlinear Schrödinger equations of derivative type, Indiana Univ.
Math. J., 45 (1996), 137–163.
[185] T. Ozawa, K. Tsutaya, Y. Tsutsumi, Global existence and asymptotic behavior of
solutions for the Klein–Gordon equations with quadratic nonlinearity in two space
dimensions, Math. Z., 222 (1996), 341–362.
[186] T. Ozawa and J. Zhai, Global existence of small classical solutions to nonlinear
Schrödinger equations, Ann. I. H. Poincaré, Anal. Non Linéaire, 25 (2008), 303–
311.
[187] M. Paicu, Équation anisotrope de Navier–Stokes dans des espaces critiques, Rev.
Mat. Iberoamericana, 21 (2005), 179–235.
[188] Y. P. Pao, Boltzmann collision operator with inverse power intermolecular poten-
tial,, I, II. Commun. Pure Appl. Math., 27 (1974), 407–428, 559–581.
[189] H. Pecher, Low energy scattering for nonlinear Klein–Gordon equations, J. Funct.
Anal., 63 (1985), 101–122.
[190] H. Pecher, Lp -Abschätzungen und klassische Lösungen für nichtlineare Wellengle-
ichungen. I, Math. Z., 150 (1976), 159–183.
[191] H. Pecher, Nonlinear small data scattering for the wave and Klein–Gordon equa-
tions, Math.Z., 185 (1984), 261–270.
[192] J. Peetre, Applications de la théorie des espaces d’interpolation dans l’analyse har-
monique, Ricerche Mat., 15 (1966), 1–34.
[193] J. Rauch, Partial Differential Equations, Graduate Texts in Math., Springer–Verlag,
1991.
[194] J. Rauch and M. Reed, Nonlinear microlocal analysis of semilinear hyperbolic sys-
tems in one spatial dimension, Duke Math. J., 49 (1982), 397–475.
[195] M. Ruzhansky, J. Smith, Dispersive and Strichartz estimates for hyperbolic equa-
tions with constant coefficients, MSJ Memoirs, 22, Mathematical Society of Japan,
Tokyo, 2010.
[196] E. Ryckman, M. Visan, Global well-posedness and scattering for the defocusing
energy-critical nonlinear Schrödinger equation in R1+4 , Amer. J. Math., 129 (2007),
1–60.
[197] I.E. Segal, Dispersion for nonlinear relativistic equation II, Ann. Sc. Ec. Norm
Sci., 4 (1968), 459–497.
278 Bibliography
[198] I.E. Segal, Space-time decay for solutions of wave equations, Adv. Math., 22 (1976),
304–311.
[199] J. Shatah, Global existence of small classical solutions to nonlinear evolution equa-
tions, J. Differential Equations, 46 (1982), 409–423.
[200] J. Shatah, Normal forms and quadratic nonlinear Klein–Gordon equations, Comm.
Pure Appl. Math., 38 (1985), 685–696.
[201] H. F. Smith, C. D. Sogge, Global Strichartz estimates for nontrapping perturbations
of Laplacian, Comm. Partial Differential Equations, 25 (2000), 2171–2183.
[202] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Prince-
ton Univ. Press, Princeton, New Jersey, 1970.
[203] E. M. Stein, Harmonic Analysis, Princeton Univ. Press, Princeton, New Jersey,
1992.
[204] W. Strauss, Nonlinear scattering theory at low energy, J. Funct. Anal., 42 (1981),
110–133 and 43 (1981), 281–293.
[205] W. Strauss, Nonlinear Wave Equations, Lecture Notes, Vol. 73, Amer. Math.
Soc., Providence, RI. 1989.
[206] R. S. Strichartz, Restriction of Fourier transform to quadratic surfaces and decay
of solutions of wave equations, Duke Math. J., 44 (1977), 705–714.
0
[207] M. Sugimoto, Lp –Lp -estimates for hyperbolic equations, Math. Res. Lett., 2
(1995), 171–178.
[208] M. Sugimoto, Estimates for hyperbolic equations with non-convex characteristics,
Math. Z., 222 (1996), 521–531.
[209] M. Sugimoto, M. Ruzhansky, A smoothing property of Schrödinger equations and
a global existence result for derivative nonlinear equations, Advances in analysis,
315–320, World Sci. Publ., Hackensack, NJ, 2005.
[210] M. Sugimoto amd N. Tomita, The dilation property of modulation spaces and their
inclusion relation with Besov spaces, J. Funct. Anal., 248 (2007), 79–106.
[211] C. Sulem and P.L. Sulem, The Nonlinear Schrödinger Equation: Self-Focusing and
Wave Collaps, Appl. Math. Sci., 139, Springer–Verlag, 1999.
[212] H. Takaoka, Well-posedness for the Kadomtsev–Petviashvili-II equation, Adv. Dif-
ferential Equations, 5 (2000), 1421–1443.
[213] H. Takaoka, Well-posedness for the one-dimensional nonlinear Schrödinger equation
with the derivative nonlinearity, Adv. Differential Equations, 4 (1999), 561–580.
[214] H. Takaoka and N. Tzvetkov, On the local regularity of the Kadomtsev–Petviashvili-
II equation, Int. Math. Res. Not., 2001 (2001) 77–114.
[215] T. Tao, Spherically averaged endpoint Strichartz estimates for the two-dimensional
Schrödinger equation, Comm. Partial Differential Equations, 25 (2000), 1471–1485.
[216] T. Tao, Multilinear weighted convolution of L2 -functions, and applications to non-
linear dispersive equations, Amer. J. Math., 123 (2001), 839–908.
[217] T. Tao, Global well-posedness and scattering for the higher-dimensional energy-
critical non-linear Schrödinger equation for radial data, New York J. Math., 11
(2005), 57–80.
[218] T. Tao, Nonlinear dispersive equations, volume 106 of CBMS Regional Conference
Series in Mathematics, Published for the Conference Board of the Mathematical
Sciences, Washington, DC, 2006.
[219] T. Tao, Scattering for the quantic generalized Korteweg-de Vries equation, J. Dif-
[220] T. Tao, M. Visan and X. Y. Zhang, The nonlinear Schrödinger equation with com-
bined power-type nonlinearities, Comm. Partial Differential Equations, 32 (2007),
1281–1343.
Bibliography 279
[221] D. Tataru, Local and global results for wave maps I, Comm. Partial Differential
Equations, 23 (1998), 1781–1793.
[222] J. Toft, Continuity properties for modulation spaces, with applications to pseudo-
differential calculus, I, J. Funct. Anal., 207 (2004), 399–429.
[223] P. Tomas, A restriction theorem for the Fourier transform, Bull. Amer. Math. Soc.,
81 (1975), 477–478.
[224] H. Triebel, Theory of Function Spaces, Birkhäuser–Verlag, 1983.
[225] H. Triebel, Modulation spaces on the Euclidean n-spaces, Z. Anal. Anwendungen,
2 (1983), 443–457.
[226] Y. Tsutsumi, L2 -solutions for nonlinear Schrödinger equations and nonlinear
groups. Funkcial. Ekvac., 30 (1987), 115–125.
[227] J. A. Tuszynski and J. M. Dixon, Coherent structures in strongly interacting many-
body systems: I, Derivation of dynamics, J. Phys. A: Math. Gen., 22 (1989),
4877–4894.
[228] S. Ukai, Local solutions in Gevrey classes to the nonlinear Boltzmann equation
without cutoff, Japan J. Appl. Math., 1 (1984), 141–156.
[229] C. Villani, A Review of Mathematical Topics in Collisional Kinetic Theory, Hand-
book of Fluid Mechanics, eds. S. Friedlander and D. Serre (North–Holland, 2002).
[230] C. Villani, On a new class of weak solutions to the spatially homogeneous Boltzmann
and Landau equations, Arch. Rational Mech. Anal., 143 (1998), no. 3, 273–307.
[231] C. Villani, Regularity estimates via entropy dissipation for the spatially homoge-
neous Boltmann equation, Rev. Mat. Iberoamericana, 15 (1999), 335–352.
[232] M. Visan, The defocusing energy-critical nonlinear Schrödinger equation in higher
dimensions, Duke Math. J., 138 (2007), 281–374.
[233] H. Wadade, Remarks on the Gagliardo–Nirenberg type inequality in the Besov and
the Triebel-Lizorkin spaces in the limiting case, J. Fourier Anal. Appl., 15 (2009),
857–870.
[234] B. X. Wang, The Cauchy Problem for the Nonlinear Schrödinger, Klein–Gordon
Equations and Their Coupled Equations, Doctoral Thesis, Inst. Appl. Phys. and
Comput. Math., Dec. 1993, 1–224.
[235] B. X. Wang, Bessel (Riesz) potentials on Banach function spaces and their applica-
tions I, Theory, Acta Math. Sinica (N.S.), 14 (1998), 327–340.
[236] B. X. Wang, On existence and scattering for critical and subcritical nonlinear Klein–
Gordon equations in H s , Nonlinear Analysis, TMA, 31 (1998), 173–187.
[237] B. X. Wang , On scattering of solutions for the critical and subcritical nonlinear
Klein-Gordon equations, Discrete Contin. Dyn. Syst., 5 (1999), 753–763.
[238] B. X. Wang, The limit behavior of solutions for the complex Ginzburg–Landau
equation, Commun. Pure Appl. Math., 55 (2002), no.4, 481–508.
[239] B. X. Wang, The smoothness of scattering operators for the sinh–Gordon and non-
linear Schrödinger equations, Acta Math. Sinica (Engl. Ser.), 18 (2002), 549–564.
[240] B. X. Wang, Exponential Besov spaces and their applications to certain evolution
equations with dissipation, Commun. Pure and Appl. Anal., 3 (2004), 883–919.
[241] B. X. Wang, Nonlinear scattering theory for a class of wave equations in H s , J.
Math. Anal. Appl., 296 (2004), 74–96.
[242] B. X. Wang, Concentration Phenomenon for the L2 Critical and Super Critical
Nonlinear Schrödinger Equation in Energy Spaces, Commun. Contemp. Math., 8
(2006), 309–330.
[243] B. X. Wang, L. J. Han, C. Y. Huang, Global well-Posedness and scattering for
the derivative nonlinear Schrödinger equation with small rough data, Ann. I. H.
Poincaré, Anal. Non Linéaire, 26 (2009), 2253–2281.
280 Bibliography
[244] B. X. Wang, C. C. Hao and H. Hudzik, Energy scattering for the nonlinear
Schrödinger equations with exponential growth in lower spatial dimensions, J. Dif-
[245] B. X. Wang and C. Y. Huang, Frequency-uniform decomposition method for the
generalized BO, KdV and NLS equations, J. Differential Equations, 239 (2007),
213–250.
[246] B. X. Wang and H. Hudzik, The global Cauchy problem for the NLS and NLKG
with small rough data, J. Differential Equations, 231 (2007), 36–73.
[247] B. X. Wang, L. F. Zhao, B. L. Guo, Isometric decomposition operators, function
λ and their applications to nonlinear evolution equations, J. Funct. Anal.,
spaces Ep,q
233 (2006), 1–39.
[248] W. K. Wang, T. Yang, Point-wise estimates and Lp convergence rates to diffusion
waves for p-system with damping, J. Differential Equations 187 (2003), 310–336.
[249] W. K. Wang, W. J. Wang, The pointwise estimates of solutions for semilinear
dissipative wave equation in multi-dimensions. J. Math. Anal. Appl., 366 (2010),
226–241.
[250] F. B. Weissler, Existence and nonexistence of global solutions for a semilinear heat
equation, Israel Math. J. 39 (1981), 29–40.
[251] F. B. Weissler, The Navier–Stokes initial value problem in Lp , Arch. Rational Mech.
Anal. 74 (1980), 219–230.
[252] N. Wiener, Tauberian theorems, Ann. of Math., 33 (1932), 1–100.
[253] J. Xiao, Homothetic variant of fractional Sobolev space with application to Navier–
Stokes system, Dyn. Partial Differ. Equ. 4 (2007), 227–245.
[254] D. C. Yang, W. Yuan, A new class of function spaces connecting Triebel–Lizorkin
spaces and Q spaces, J. Funct. Anal., 255 (2008), 2760–2809.
[255] D. C. Yang, W. Yuan, New Besov-type spaces and Triebel–Lizorkin-type spaces
including Q spaces, Math. Z., 265 (2010), 451–480.
[256] K. Yosida, Functional Analysis, Springer–Verlag, 1981.
[257] W. Yuan, W. Sickel, D. C. Yang Morrey and Campanato Meet Besov, Lizorkin and
Triebel, Lect. Note in Math., 2005 (2010), 1-281.
[258] V. E. Zakharov, E. A. Kuznetson, Multi-scale expansions in the theory of systems
integrable by inverse scattering method, Physica D, 18 (1986), 455–463.
[259] V. E. Zakharov, E. I. Schulman, Degenerated dispersion laws, motion invariant and
kinetic equations, Physica D, 1 (1980), 185–250.
[260] Y. L. Zhou, B. L. Guo, Weak solution of system of ferromagnetic chain with several
variables. Sci. Sinica Ser. A, 30 (1987), 1251–1266.
[261] Y. L. Zhou, B. L. Guo and S. Tan, Existence and uniqueness of smooth solution for
system of ferro-magnetic chain. Sci. China Ser. A 34 (1991), 257–266.
Index
s
Bp,q ,9 Boltzmann equation
C (Rn ), 2
∞
Global (absolute) Maxwellian, 238
s
E2,1 , 45 Bobylev’s identities , 238
s
Fp,q , 9 Boltzmann collision operator, 233
Lp (`q ), 281 Boltzmann equation, 233
Lpx1i Lp(x2j )j6=i Lpt 2 , 189 Cross section, 233
s
Mp,q , 160 Cutoff cross section, 234
s
Xp,q , 10 Hard potentials, 235
k , 160 Homogeneous Boltzmann equation,
S, 2 233
S 0, 2 Local Maxwellian, 238
s
Ḃp,q , 20 Maxwellian potentials, 235
s
Ḟp,q , 20 Non cutoff cross section, 234
s Soft potentials, 235
Ẋp,q , 20
S˙ , 20
S˙ 0 , 20 Christ-Kiselev lemma, 267
`q (Lp ), 9 Complex interpolation in modulation
`s,q γ p spaces, 267
(L (R, L )), 179
φ ∗ ψ, 3 Convexity Hölder’s inequality, 21
φ̃, 3 Critical power in Ḣ s for NLS, 83
4k , 9 Critical space Ḣ s for NLS, 83
e k , 192
Critical space for NS equation, 35
{σk }k∈Zn , 160
p ∧ q, p ∨ q, 11 Dilation operator, 3
Dispersion, 51
Action functional, 206 Dispersion relation, 92
Affine transform, 7 Dyadic decomposition operator, 9
Almost orthogonality of k , 198
Asymptotic completeness, 261 Embedding theorem
Asymptotic state, 261 Embedding between Besov and
modulation spaces, 164
Bernstein’s inequality, 14 Embedding in homogeneous
281
282 Index
Triebel-Lizorkin spaces, 264 Isomorphism, 2

Embeddings on Besov spaces, 14
Embeddings on homogeneous Klein-Gordon semi-group
Besov space, 21 Strichartz estimates in modulation
Embeddings on homogeneous spaces, 183
Triebel space, 21 truncated decay in modulation
Embeddings on Triebel spaces, 14 spaces, 178, 179
Fourier (inverse) transform F (F −1 ), Lagrangian, 206

2 Legendre transformation, 207
Fourier multiplier, 5 Linear PDE
Fourier multiplier space Mp , 5 Klein-Gordon equation, 62, 67
Free transport, 229 Linear heat equation, 35
Frequency-uniform decomposition Schrödinger equation, 51
operator, 160 Wave equation, 66
Function space Littlewood-Paley decomposition
BM O−1 , 49 operator, 9
Anisotropic Lebesgue space, 189 Littlewood-Paley square function
Besov space Bp,qs
,9 theorem, 22, 266
s
Homogeneous Besov spaceḂp,q , 20
Minkowski’s inequality, 11
Homogeneous Triebel-Lizorkin
s Morawetz action, 215
space Ḟp,q , 20
s Morawetz inequality, 216
Modulation space Mp,q , 160
Multi-index α, 1
Schwartz spaceS := S (Rn ), 2
Multiplier theorem
Space of tempered distributions
Bernstein multiplier theorem, 7
S 0 = S 0 (Rn ), 2
s Mihlin multiplier theorem, 8
Triebel space Fp,q ,9
Nöther’s theorem, 205
Gagliardo-Nirenberg inequality Nöther’s theorem, 209
Fractional derivative cases, 25 Nakanishi-Morawetz inequality, 217,
Sobolev space cases, 263 220
Galilean operator, 214 NLKG, 88
Galilean transformation, 213 NLS, 75
Gauge transformation, 146 NLW, 89
Gevrey class, 44 Nonlinear PDE
Derivative nonlinear Klein-Gordon
Hamilton’s equation, 208 equation, 203
Hamilton’s principle, 206 Hamilton–Jacobi equation, 41
Hamiltonian, 207 Navier-Stokes equation, 34
Hardy inequality, 219 Nonlinear dispersive equation, 75
Hardy-Littlewood-Sobolev inequality, Nonlinear Klein–Gordon equation,
266 88
Heat semi-group Nonlinear Schrödinger equation, 75
Lr → Lp estimate, 36 Nonlinear wave equation, 89
Time-space mixed estimate, 36 Semi-linear parabolic equation, 41
Index 283
Zakharov system, 204 Schrödinger semi-group, 54, 67, 73,

NS, 34 176
Schrödinger semi-group with
Paley-Wiener-Schwartz theorem, 4 higher order, 54, 67
Pseudo-conformal conservation law, Wave semi-group, 54, 66, 73
214 Short time Fourier transform, 159
Pseudo-conformal transformation, Strichartz inequality, 51, 63, 68
213
Tempered distribution
Riesz-Thorin interpolation theorem, Composition of distribution, 4
265 Dilation of distribution, 4
Generalized derivative of
Scattering operator, 222, 261 distribution, 4
Schrödinger semi-group Translation of distribution, 4
Local smoothing effect, 99 Translation operator, 3
Maximal function estimate, 193
Maximal function estimates, 99 Van der Corput lemma, 266
Smooth effect, 191 Virial identity, 216
Strichartz estimate in modulation Virial potential, 215
spaces, 182 Vlasov equation, 229
truncated decay in modulation
spaces, 177 Wave operators, 261
Uniform boundedness in Wiener decomposition, 157
modulation spaces, 176
Semi-group
Heat semi-group, 35, 36
Klein-Gordon semi-group, 62, 67,
74, 178

14634.harmonic Analysis Method For Nonlinear Evolution Equations, I by Baoxiang Wang

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

14634.harmonic Analysis Method For Nonlinear Evolution Equations, I by Baoxiang Wang

Enviado por

Direitos autorais:

Formatos disponíveis

p =0

8209.9789814360739- tp.indd 1 7/1/11 3:37 PM

8209.9789814360739- tp.indd 2 7/1/11 3:37 PM

British Library Cataloguing-in-Publication Data

HARMONIC ANALYSIS METHOD FOR NONLINEAR EVOLUTION

ZhangJi - Harmonic analysis Method.pmd 1 6/15/2011, 1:23 PM

The sole purpose of science is the glory of human spirit. —— C. G.

On the other hand, the theory of harmonic analysis is also promoted by

Beijing, May 1, 2011

3. Strichartz estimates for linear dispersive equations 51

4. Local and global wellposedness for nonlinear dispersive equations 75

5. The low regularity theory for the nonlinear dispersive

6. Frequency-uniform decomposition techniques 157

6.5 Derivative nonlinear Schrödinger equations . . . . . . . . 187

7. Conservations, Morawetz’ estimates of nonlinear

8. Boltzmann equation without angular cutoff 227

Appendix A Notations 259

Appendix B Definition of scattering operator 261

Appendix C Some fundamental results 263

In this chapter we begin by considering the Fourier transform on the

1.1 Schwartz space, tempered distribution, Fourier trans-

Let α = (α1 , ..., αn ) be a multi-index, denote

We denote by C ∞ (Rn ) the set of all infinitely differentiable functions de-

It is easy to see that pk (·) is a norm on S and so, it is a semi-norm on

where B is a bounded set in S . Let φ, ψ ∈ S , the Fourier (inverse)

Proposition 1.1. Let φ ∈ S . Then we have

Proposition 1.2. Let φ, ψ ∈ S . Then

1.1. Schwartz space, tempered distribution, Fourier transform 3

We define the convolution φ ∗ ψ on S in the following way

Proposition 1.3. Let φ, ψ ∈ S . Then

We denote by σλ and τh the dilation and the translation operators:

Proposition 1.4. Let φ ∈ S , λ ∈ R \ {0}, h ∈ Rn . Then

the other proofs are easily obtained.

Proposition 1.5. F : S 0 → S 0 is a continuously linear bijection map,

Now we can define the convolution of f ∈ S 0 and φ ∈ S . Let us recall

where φ̃ = φ(−·). Following this idea, we define

first investigate the corresponding operation for Schwartz functions, which

Proposition 1.6. Let f ∈ S 0 and φ ∈ S . Then f ∗ φ is an infinitely

Proposition 1.7 (Paley-Wiener-Schwartz theorem).

Proposition 1.7 is a fundamental result in the framework of the Fourier

1.2 Fourier multiplier on Lp

We denote by Lp := Lp (Rn ) the Lebesgue space, 1 6 p 6 ∞, the norm on

1.2. Fourier multiplier on Lp 5

Definition 1.1. Let ρ ∈ S 0 . If there exists C > 0 such that2

then ρ is said to be a multiplier on Lp . The set of all multipliers on Lp is

kρkMp = sup{kF −1 ρF f kp : f ∈ S , kf kp = 1}. (1.6)

If ρ ∈ Mp , 1 6 p < ∞, since S is dense in Lp , one can extend F −1 ρF

Proposition 1.8. Let 1 6 p 6 q 6 2. Then we have

(1) Mp = Mp0 , k · kMp = k · kMp0 ;

Proof. First, we prove (1). Let ρ ∈ Mp . For any f, g ∈ S , kf kp =

|(F −1 ρF f, g)| 6 kF −1 ρF f kp kgkp0 6 kρkMp . (1.7)

Secondly, we show the result of (2). Due to 1 6 p 6 q 6 2, we see that

Proposition 1.9. Mp is a Banach algebra.

Proof. Obviously, k · kMp is a norm. By (3) in Proposition 1.8, one sees

1.2. Fourier multiplier on Lp 7

In order to emphasize that Mp depends on Rn , we write Mp = Mp (Rn ).

Proposition 1.10. Let a : Rn → Rm (n > m) be a surjective affine

Proof. It suffices to consider the case that a : Rn → Rm is a linear

Proposition 1.11 (Bernstein multiplier theorem). Let L > n/2 be

p ∈ Rn , where A is a linear transform of Rn .

Proof. By Proposition 1.8, it suffices to consider the case p = 1. Obvi-

Proposition 1.12 (Mihlin multiplier theorem). Let L > n/2 be an

Proof. See [13].