Nonlinear dispersive equations: local and global

analysis

Terence Tao
Department of Mathematics, UCLA, Los Angeles, CA 90095
E-mail address: [email protected]
1991 Mathematics Subject Classification. Primary 35Q53, 35Q55, 35L15

The author is partly supported by a grant from the Packard foundation.

To Laura, for being so patient.
 Contents

Preface ix
Chapter 1. Ordinary differential equations 1
1.1. General theory 2
1.2. Gronwalls inequality 11
1.3. Bootstrap and continuity arguments 20
1.4. Noethers theorem 26
1.5. Monotonicity formulae 35
1.6. Linear and semilinear equations 40
1.7. Completely integrable systems 49
Chapter 2. Constant coefficient linear dispersive equations 55
2.1. The Fourier transform 62
2.2. Fundamental solution 69
2.3. Dispersion and Strichartz estimates 73
2.4. Conservation laws for the Schrodinger equation 82
2.5. The wave equation stress-energy tensor 89
2.6. X s,b spaces 97
Chapter 3. Semilinear dispersive equations 109
3.1. On scaling and other symmetries 114
3.2. What is a solution? 120
3.3. Local existence theory 129
3.4. Conservation laws and global existence 143
3.5. Decay estimates 153
3.6. Scattering theory 162
3.7. Stability theory 171
3.8. Illposedness results 180
3.9. Almost conservation laws 186

Chapter 4. The Korteweg de Vries equation 197

4.1. Existence theory 202
4.2. Correction terms 213
4.3. Symplectic non-squeezing 218
4.4. The Benjamin-Ono equation and gauge transformations 223
Chapter 5. Energy-critical semilinear dispersive equations 231
5.1. The energy-critical NLW 233
5.2. Bubbles of energy concentration 247
5.3. Local Morawetz and non-concentration of mass 257
5.4. Minimal-energy blowup solutions 262
vii
viii CONTENTS

5.5. Global Morawetz and non-concentration of mass 271

Chapter 6. Wave maps 277
6.1. Local theory 288
6.2. Orthonormal frames and gauge transformations 299
6.3. Wave map decay estimates 310
6.4. Heat flow 320
Chapter A. Appendix: tools from harmonic analysis 329
Chapter B. Appendix: construction of ground states 347
Bibliography 363
 Preface

Politics is for the present, but an equation is something for eternity.

(Albert Einstein)
This monograph is based on (and greatly expanded from) a lecture series given
at the NSF-CBMS regional conference on nonlinear and dispersive wave equations
at New Mexico State University, held in June 2005. Its objective is to present
some aspects of the global existence theory (and in particular, the regularity and
scattering theory) for various nonlinear dispersive and wave equations, such as the
Korteweg-de Vries (KdV), nonlinear Schrodinger, nonlinear wave, and wave maps
equations. The theory here is rich and vast and we cannot hope to present a
comprehensive survey of the field here; our aim is instead to present a sample of
results, and to give some idea of the motivation and general philosophy underlying
the problems and results in the field, rather than to focus on the technical details.
We intend this monograph to be an introduction to the field rather than an ad-
vanced text; while we do include some very recent results, and we imbue some more
classical results with a modern perspective, our main concern will be to develop
the fundamental tools, concepts, and intuitions in as simple and as self-contained
a matter as possible. This is also a pedagogical text rather than a reference; many
details of arguments are left to exercises or to citations, or are sketched informally.
Thus this text should be viewed as being complementary to the research literature
on these topics, rather than being a substitute for them.
The analysis of PDE is a beautiful subject, combining the rigour and technique
of modern analysis and geometry with the very concrete real-world intuition of
physics and other sciences. Unfortunately, in some presentations of the subject (at
least in pure mathematics), the former can obscure the latter, giving the impression
of a fearsomely technical and difficult field to work in. To try to combat this, this
book is devoted in equal parts to rigour and to intuition; the usual formalism of
definitions, propositions, theorems, and proofs appear here, but will be interspersed
and complemented with many informal discussions of the same material, centering
around vague Principles and figures, appeal to physical intuition and examples,
back-of-the-envelope computations, and even some whimsical quotations. Indeed,
the exposition and exercises here reflect my personal philosophy that to truly under-
stand a mathematical result one must view it from as many perspectives as possible
(including both rigorous arguments and informal heuristics), and must also be able
to translate easily from one perspective to another. The reader should thus be
aware of which statements in the text are rigorous, and which ones are heuristic,
but this should be clear from context in most cases.
To restrict the field of study, we shall focus primarily on defocusing equations,
in which soliton-type behaviour is prohibited. From the point of view of global
existence, this is a substantially easier case to study than the focusing problem, in
ix
x PREFACE

which one has the fascinating theory of solitons and multi-solitons, as well as various
mechanisms to enforce blow-up of solutions in finite or infinite time. However, we
shall see that there are still several analytical subtleties in the defocusing case,
especially when considering critical nonlinearities, or when trying to establish a
satisfactory scattering theory. We shall also work in very simple domains such
as Euclidean space Rn or tori Tn , thus avoiding consideration of boundary-value
problems, or curved space, though these are certainly very important extensions to
the theory. One further restriction we shall make is to focus attention on the initial
value problem when the initial data lies in a Sobolev space Hxs (Rd ), as opposed to
more localised choices of initial data (e.g. in weighted Sobolev spaces, or self-similar
initial data). This restriction, combined with the previous one, makes our choice of
problem translation-invariant in space, which leads naturally to the deployment of
the Fourier transform, which turns out to be a very powerful tool in our analysis.
Finally, we shall focus primarily on only four equations: the semilinear Schrodinger
equation, the semilinear wave equation, the Korteweg-de Vries equation, and the
wave maps equation. These four equations are of course only a very small sample of
the nonlinear dispersive equations studied in the literature, but they are reasonably
representative in that they showcase many of the techniques used for more general
equations in a comparatively simple setting.
Each chapter of the monograph is devoted to a different class of differential
equations; generally speaking, in each chapter we first study the algebraic struc-
ture of these equations (e.g. symmetries, conservation laws, and explicit solutions),
and then turn to the analytic theory (e.g. existence and uniqueness, and asymptotic
behaviour). The first chapter is devoted entirely to ordinary differential equations
(ODE). One can view partial differential equations (PDE) such as the nonlinear
dispersive and wave equations studied here, as infinite-dimensional analogues of
ODE; thus finite-dimensional ODE can serve as a simplified model for understand-
ing techniques and phenomena in PDE. In particular, basic PDE techniques such
as Picard and Duhamel iteration, energy methods, continuity or bootstrap argu-
ments, conservation laws, near-conservation laws, and monotonicity formulae all
have useful ODE analogues. Furthermore, the analogy between classical mechan-
ics and quantum mechanics provides a useful heuristic correspondence between
Schrodinger type equations, and classical ODE involving one or more particles, at
least in the high-frequency regime.
The second chapter is devoted to the theory of the basic linear dispersive mod-
els: the Airy equation, the free Schrodinger equation, and the free wave equation.
In particular, we show how the Fourier transform and conservation law methods,
can be used to establish existence of solutions, as well as basic estimates such as
the dispersive estimate, local smoothing estimates, Strichartz estimates, and X s,b
estimates.
In the third chapter we begin studying nonlinear dispersive equations in earnest,
beginning with two particularly simple semilinear models, namely the nonlinear
Schrodinger equation (NLS) and nonlinear wave equation (NLW). Using these
equations as examples, we illustrate the basic approaches towards defining and
constructing solutions, and establishing local and global properties, though we de-
fer the study of the more delicate energy-critical equations to a later chapter. (The
mass-critical nonlinear Schrodinger equation is also of great interest, but we will
not discuss it in detail here.)
 PREFACE xi

In the fourth chapter, we analyze the Korteweg de Vries equation (KdV), which
requires some more delicate analysis due to the presence of derivatives in the non-
linearity. To partly compensate for this, however, one now has the structures of
nonresonance and complete integrability; the interplay between the integrability on
one hand, and the Fourier-analytic structure (such as nonresonance) on the other,
is still only partly understood, however we are able to at least establish a quite
satisfactory local and global wellposedness theory, even at very low regularities,
by combining methods from both. We also discuss a less dispersive cousin of the
KdV equation, namely the Benjamin-Ono equation, which requires more nonlinear
techniques, such as gauge transforms, in order to obtain a satisfactory existence
and wellposedness theory.
In the fifth chapter we return to the semilinear equations (NLS and NLW),
and now establish large data global existence for these equations in the defocusing,
energy-critical case. This requires the full power of the local wellposedness and per-
turbation theory, together with Morawetz-type estimates to prevent various kinds
of energy concentration. The situation is especially delicate for the Schrodinger
equation, in which one must employ the induction on energy methods of Bourgain
in order to obtain enough structural control on a putative minimal energy blowup
solution to obtain a contradiction and thus ensure global existence.
In the final chapter, we turn to the wave maps equation (WM), which is some-
what more nonlinear than the preceding equations, but which on the other hand
enjoys a strongly geometric structure, which can in fact be used to renormalise
most of the nonlinearity. The small data theory here has recently been completed,
but the large data theory has just begun; it appears however that the geometric
renormalisation provided by the harmonic map heat flow, together with a Morawetz
estimate, can again establish global existence in the negatively curved case.
As a final disclaimer, this monograph is by no means intended to be a defini-
tive, exhaustive, or balanced survey of the field. Somewhat unavoidably, the text
focuses on those techniques and results which the author is most familiar with, in
particular the use of the iteration method in various function spaces to establish a
local and perturbative theory, combined with frequency analysis, conservation laws,
and monotonicity formulae to then obtain a global non-perturbative theory. There
are other approaches to this subject, such as via compactness methods, nonlinear
geometric optics, infinite-dimensional Hamiltonian dynamics, or the techniques of
complete integrability, which are also of major importance in the field (and can
sometimes be combined, to good effect, with the methods discussed here); however,
we will be unable to devote a full-length treatment of these methods in this text. It
should also be emphasised that the methods, heuristics, principles and philosophy
given here are tailored for the goal of analyzing the Cauchy problem for semilinear
dispersive PDE; they do not necessarily extend well to other PDE questions (e.g.
control theory or inverse problems), or to other classes of PDE (e.g. conservation
laws or to parabolic and elliptic equations), though there are certain many connec-
tions and analogies between results in dispersive equations and in other classes of
PDE.
I am indebted to my fellow members of the I-team (Jim Colliander, Markus
Keel, Gigliola Staffilani, Hideo Takaoka), to Sergiu Klainerman, and to Michael
Christ for many entertaining mathematical discussions, which have generated much
of the intuition that I have tried to place into this monograph. I am also very
xii PREFACE

thankful for Jim Ralston for using this text to teach a joint PDE course, and
providing me with careful corrections and other feedback on the material. I also
thank Soonsik Kwon and Shaunglin Shao for additional corrections. Last, but
not least, I am grateful to my wife Laura for her support, and for pointing out
the analogy between the analysis of nonlinear PDE and the electrical engineering
problem of controlling feedback, which has greatly influenced my perspective on
these problems (and has also inspired many of the diagrams in this text).

Terence Tao

Notation. As is common with any book attempting to survey a wide range

of results by different authors from different fields, the selection of a unified no-
tation becomes very painful, and some compromises are necessary. In this text I
have (perhaps unwisely) decided to make the notation as globally consistent across
chapters as possible, which means that any individual result presented here will
likely have a notation slightly different from the way it is usually presented in the
literature, and also that the notation is more finicky than a local notation would
be (often because of some ambiguity that needed to be clarified elsewhere in the
text). For the most part, changing from one convention to another is a matter of
permuting various numerical constants such as 2, , i, and 1; these constants are
usually quite harmless (except for the sign 1), but one should nevertheless take
care in transporting an identity or formula in this book to another context in which
the conventions are slightly different.
In this text, d will always denote the dimension of the ambient physical space,
which will either be a Euclidean space1 Rd or the torus Td := (R/2Z)d . (Chapter
1 deals with ODE, which can be considered to be the case d = 0.) All integrals on
these spaces will be with respect to Lebesgue measure dx. If x = (x1 , . . . , xd ) and
= (1 , . . . , d ) lie in Rd , we use x to denote the dot product x := x1 1 + . . . +
xd d , and |x| to denote the magnitude |x| := (x21 + . . . + x2d )1/2 . We also use hxi
to denote the inhomogeneous magnitude (or Japanese bracket ) hxi := (1 + |x|2 )1/2
of x, thus hxi is comparable to |x| for large x and comparable to 1 for small x.
In a similar spirit, if x = (x1 , . . . , xd ) Td and k = (k1 , . . . , kd ) Zd we define
k x := k1 x1 + . . . + kd xd T. In particular the quantity eikx is well-defined.
We say that I is a time interval if it is a connected subset of R which contains
at least two points (so we allow time intervals to be open or closed, bounded or
unbounded). If u : I Rd Cn is a (possibly vector-valued) function of spacetime,
we write t u for the time derivative u t , and x1 u, . . . , xd u for the spatial derivatives
u u
x1 , . . . , xd ; these derivatives will either be interpreted in the classical sense (when
u is smooth) or the distributional (weak) sense (when u is rough). We use x u :
I Rd Cnd to denote the spatial gradient x u = (x1 u, . . . , xd u). We can
iterate this gradient to define higher derivatives kx for k = 0, 1, . . .. Of course,

1We will be using two slightly different notions of spacetime, namely Minkowski space R1+d
and Galilean spacetime R Rd ; in the very last section we also need to use parabolic spacetime
R+ Rd . As vector spaces, they are of course equivalent to each other (and to the Euclidean
space Rd+1 ), but we will place different (pseudo)metric structures on them. Generally speaking,
wave equations will use Minkowski space, whereas nonrelativistic equations such as Schrodinger
equations will use Galilean spacetime, while heat equations use parabolic spacetime. For the most
part the reader will be able to safely ignore these subtle distinctions.
 PREFACE xiii

these definitions also apply to functions on Td , which can be identified with periodic
functions on Rd .
We use the Einstein convention for summing indices, with Latin indices ranging
Pd
from 1 to d, thus for instance xj xj u is short for j=1 xj xj u. When we come
to wave equations, we will also be working in a Minkowski space R1+d with a
Minkowski metric g ; in such cases, we will use Greek indices and sum from 0 to d
(with x0 = t being the time variable), and use the metric to raise and lower indices.
Thus for instance if we use the standard Minkowski metric dg 2 = dt2 + |dx|2 , then
0 u = t u but 0 u = t u. Rt Rs
In this monograph we always adopt the convention that s = t if t < s.
This convention will usually be applied only to integrals in the time variable.
We use the Lebesgue norms
Z
kf kLx (Rd C) := (
p |f (x)|p dx)1/p
Rd

for 1 p < for complex-valued measurable functions f : Rd C, with the

usual convention
kf kL d
x (R C)
:= ess sup |f (x)|.
xRd

In many cases we shall abbreviate Lpx (Rd C) as Lpx (Rd ), Lp (Rd ), or even Lp
when there is no chance of confusion. The subscript x, which denotes the dummy
variable, is of course redundant. However we will often retain it for clarity when
dealing with PDE, since in that context one often needs to distinguish between
Lebesgue norms in space x, time t, spatial frequency , or temporal frequency
. Also we will need it to clarify expressions such as kxf kLpx (Rd ) , in which the
expression in the norm depends explicitly on the variable of integration. We of
course identify any two functions if they agree almost everywhere. One can of
course replace the domain Rd by the torus Td or the lattice Zd , thus for instance
X
kf klpx (Zd C) := ( |f (k)|p )1/p .
kZd

One can replace C by any other Banach space X, thus for instance Lpx (Rd X)
is the space of all measurable functions u : Rd X with finite norm
Z
kukLpx(Rd X) := ( ku(x)kpX dx)1/p
Rd

with the usual modification for p = . In particular we can define the mixed
Lebesgue norms Lqt Lrx (I Rd C) = Lqt (I Lrx (Rd C)) for any time interval
I as the space of all functions u : I Rd C with norm
Z Z Z
q 1/q
kukLt Lrx (IRd C) := ( ku(t)kLr (Rd ) dt)
q =( ( |u(t, x)|r dx)q/r dt)1/q
x
I I Rd

with the usual modifications when q = or r = . One can also use this Banach
space notation to make sense of the Lp norms of tensors such as f , 2 f , etc.,
provided of course that such derivatives exist in the Lp sense.
In a similar spirit, if I is a time interval and k 0, we use Ctk (I X) to
denote the space of all k-times continuously differentiable functions u : I X with
xiv PREFACE

the norm
k
X
kukCtk (IX) := ktj ukL
t (IX)
.
j=0

We adopt the convention that kukCtk (IX) = if u is not k-times continuously

differentiable. One can of course also define spatial analogues Cxk (Rd X) of these
k
spaces, as well as spacetime versions Ct,x (I Rd X). We caution that if I is not
compact, then it is possible for a function to be k-times continuously differentiable
but have infinite Ctk norm; in such cases we say that u Ct,loc
k
(I X) rather than
k
u Ct (I X). More generally, a statement of the form u Xloc () on a domain
means that we can cover by open sets V such that the restriction u|V of u
to each of these sets V is in X(V ); under reasonable assumptions on X, this also
implies that u|K X(K) for any compact subset K of . As a rule of thumb, the
global spaces X() will be used for quantitative control (estimates), whereas the
local spaces Xloc () are used for qualitative control (regularity); indeed, the local
spaces Xloc are typically only Frechet spaces rather than Banach spaces. We will
need both types of control in this text, as one typically needs qualitative control to
ensure that the quantitative arguments are rigorous.
If (X, dX ) is a metric space and Y is a Banach space, we use C 0,1 (X Y ) to
denote the space of all Lipschitz continuous functions f : X Y , with norm

kf (x) f (x )kY
kf kC 0,1 (XY ) := sup .
x,x X:x6=x dX (x, x )

(One can also define the inhomogeneous Lipschitz norm kf kC 0,1 := kf kC 0,1 +kf kC 0 ,
but we will not need this here.) Thus for instance C 1 (Rd Rm ) is a subset
of C 0,1 (Rd Rm ), which is in turn a subset of Cloc 0
(Rd Rm ). The space
0,1
Cloc (X Y ) is thus the space of locally Lipschitz functions (i.e. every x X is
contained in a neighbourhood on which the function is Lipschitz).
In addition to the above function spaces, we shall also use Sobolev spaces H s ,
W , H s , W s,p , which are defined in Appendix A, and X s,b spaces, which are
s,p

defined in Section 2.6.

If V and W are finite-dimensional vector spaces, we use End(V W ) to denote
the space of linear transformations from V to W , and End(V ) = End(V V ) for
the ring of linear transformations from V to itself. This ring contains the identity
transformation id = idV .
If X and Y are two quantities (typically non-negative), we use X . Y or Y & X
to denote the statement that X CY for some absolute constant C > 0. We use
X = O(Y ) synonymously with |X| . Y . More generally, given some parameters
a1 , . . . , ak , we use X .a1 ,...,ak Y or Y &a1 ,...,ak X to denote the statement that
X Ca1 ,...,ak Y for some (typically large) constant Ca1 ,...,ak > 0 which can depend
on the parameters a1 , . . . , ak , and define X = Oa1 ,...,ak (Y ) similarly. Typical choices
of parameters include the dimension d, the regularity s, and the exponent p. We
will also say that X is controlled by a1 , . . . , ak if X = Oa1 ,...,ak (1). We use X
Y to denote the statement X . Y . X, and similarly X a1 ,...,ak Y denotes
X .a1 ,...,ak Y .a1 ,...,ak X. We will occasionally use the notation X a1 ,...,ak Y
or Y a1 ,...,ak X to denote the statement X ca1 ,...,ak Y for some suitably small
quantity ca1 ,...,ak > 0 depending on the parameters a1 , . . . , ak . This notation is
 PREFACE xv

somewhat imprecise (as one has to specify what suitably small means) and so we
shall usually only use it in informal discussions.
Recall that a function f : Rd C is said to be rapidly decreasing if we have
khxiN f (x)kL d <
x (R )

for all N 0. We then say that a function is Schwartz if it is smooth and all of its
derivatives x f are rapidly decreasing, where = (1 , . . . , d ) Zd+ ranges over
all multi-indices, and x is the differential operator
1 d
x := ( ) ...( ) .
x1 xd
In other words, f is Schwartz if and only if x f (x) = Of,,N (hxiN ) for all Zd+
and all N . We use Sx (Rd ) to denote the space of all Schwartz functions. As
is well known, this is a Frechet space, and thus has a dual Sx (Rd ) , the space
of tempered distributions. This space contains all locally integrable functions of
polynomial growth, and is also closed under differentiation, multiplication with
functions g of symbol type (i.e. g and all its derivatives are of polynomial growth)
and convolution with Schwartz functions; we will not present a detailed description
of the distributional calculus here.
 CHAPTER 1

Ordinary differential equations

Science is a differential equation. Religion is a boundary condition.

(Alan Turing, quoted in J.D. Barrow, Theories of everything)

This monograph is primarily concerned with the global Cauchy problem (or
initial value problem) for partial differential equations (PDE), but in order to as-
semble some intuition on the behaviour of such equations, and on the power and
limitations of the various techniques available to analyze these equations, we shall
first study these phenomena and methods in the much simpler context of ordinary
differential equations (ODE), in which many of the technicalities in the PDE anal-
ysis are not present. Conversely, the theory of ODEs, particularly Hamiltonian
ODEs, has a very rich and well-developed structure, the extension of which to non-
linear dispersive PDEs is still far from complete. For instance, phenomena from
Hamiltonian dynamics such as Kolmogorov-Arnold-Moser (KAM) invariant tori,
symplectic non-squeezing, Gibbs and other invariant measures, or Arnold diffusion
are well established in the ODE setting, but the rigorous theory of such phenomena
for PDEs is still its infancy.
One technical advantage of ODE, as compared with PDE, is that with ODE
one can often work entirely in the category of classical (i.e. smooth) solutions,
thus bypassing the need for the theory of distributions, weak limits, and so forth.
However, even with ODE it is possible to exhibit blowup in finite time, and in high-
dimensional ODE (which begin to approximate PDE in the infinite dimensional
limit) it is possible to have the solution stay bounded in one norm but become
extremely large in another norm. Indeed, the quantitative study of expressions
such as mass, energy, momentum, etc. is almost as rich in the ODE world as it is
in the PDE world, and thus the ODE model does serve to illuminate many of the
phenomena that we wish to study for PDE.
A common theme in both nonlinear ODE and nonlinear PDE is that of feedback
- the solution to the equation at any given time generates some forcing term, which
in turn feeds back into the system to influence the solution at later times, usually
in a nonlinear fashion. The tools we will develop here to maintain control of this
feedback effect - the Picard iteration method, Gronwalls inequality, the bootstrap
principle, conservation laws, monotonicity formulae, and Duhamels formula - will
form the fundamental tools we will need to analyze nonlinear PDE in later chapters.
Indeed, the need to deal with such feedback gives rise to a certain nonlinear way
of thinking, in which one continually tries to control the solution in terms of itself,
or derive properties of the solution from (slightly weaker versions of) themselves.
This way of thinking may initially seem rather unintuitive, even circular, in nature,
but it can be made rigorous, and is absolutely essential to proceed in this theory.
1
2 1. ORDINARY DIFFERENTIAL EQUATIONS

1.1. General theory

It is a capital mistake to theorise before one has data. Insensibly
one begins to twist facts to suit theories, instead of theories to suit
facts. (Sir Arthur Conan Doyle, A Study in Scarlet)
In this section we introduce the concept of an ordinary differential equation
and the associated Cauchy problem, but then quickly specialise to an important
subclass of such problems, namely the Cauchy problem (1.7) for autonomous first-
order quasilinear systems.
Throughout this chapter, D will denote a (real or complex) finite dimensional
vector space, which at times we will endow with some norm kkD ; the letter D stands
for data. An ordinary differential equation (ODE) is an equation which governs
certain functions u : I D mapping a (possibly infinite) time interval I R to
the vector space1 D. In this setup, the most general form of an ODE is that of a
fully nonlinear ODE

(1.1) G(u(t), t u(t), . . . , tk u(t), t) = 0

where k 1 is an integer, and G : Dk+1 I X is a given function taking values in

k
another finite-dimensional vector space X. We say that a function u Cloc (I D)
2
is a classical solution (or solution for short) of the ODE if (1.1) holds for all t I.
The integer k is called the order of the ODE, thus for instance if k = 2 then we
have a second-order ODE. One can think of u(t) as describing the state of some
physical system at a given time t; the dimension of D then measures the degrees
of freedom available. We shall refer to D as the state space, and sometimes refer
to the ODE as the equation(s) of motion, where the plural reflects the fact that X
may have more than one dimension. While we will occasionally consider the scalar
case, when D is just the real line R or the complex plane C, we will usually be
more interested in the case when the dimension of D is large. Indeed one can view
PDE as a limiting case of ODE as dim(D) .
In this monograph we will primarily consider those ODE which are time-
translation-invariant (or autonomous), in the sense that the function G does not
actually depend explicitly on the time parameter t, thus simplifying (1.1) to

(1.2) G(u(t), t u(t), . . . , tk u(t)) = 0

for some function G : Dk+1 X. One can in fact convert any ODE into a time-
translation-invariant ODE, by the trick of embedding the time variable itself into
the state space, thus replacing D with D R, X with X R, u with the function

1One could generalise the concept of ODE further, by allowing D to be a smooth manifold
instead of a vector space, or even a smooth bundle over the time interval I. This leads for instance
to the theory of jet bundles, which we will not pursue here. In practice, one can descend from this
more general setup back to the original framework of finite-dimensional vector spaces - locally
in time, at least - by choosing appropriate local coordinate charts, though often the choice of
such charts is somewhat artifical and makes the equations messier; see Chapter 6 for some related
issues.
2We will discuss non-classical solutions shortly. As it turns out, for finite-dimensional ODE
there is essentially no distinction between a classical and non-classical solution, but for PDE there
will be a need to distinguish between classical, strong, and weak solutions. See Section 3.2 for
further discussion.
 1.1. GENERAL THEORY 3

u(t) := (u(t), t), and G with the function3

G((u0 , s0 ), (u1 , s1 ), . . . , (uk , sk )) := (G(u0 , . . . , uk ), s1 1).
For instance, solutions to the non-autonomous ODE
t u(t) = F (t, u(t))
are equivalent to solutions to the system of autonomous ODE
t u(t) = F (s(t), u(t)); t s(t) 1 = 0
provided that we also impose a new initial condition s(0) = 0. This trick is not
always without cost; for instance, it will convert a non-autonomous linear equation
into an autonomous nonlinear equation.
By working with time translation invariant equations we obtain our first sym-
metry, namely the time translation symmetry
(1.3) u(t) 7 u(t t0 ).
More precisely, if u : I D solves the equation (1.2), and t0 R is any time
shift parameter, then the time-translated function ut0 : I + t0 D defined by
ut0 (t) := u(t t0 ), where I + t0 := {t + t0 : t I} is the time translation of I,
is also a solution to (1.2). This symmetry tells us, for instance, that the initial
value problem for this equation starting from time t = 0 will be identical (after
applying the symmetry (1.3)) to the initial value problem starting from any other
time t = t0 .
The equation (1.2) implicitly determines the value of the top-order derivative
tk u(t) in terms of the lower order derivatives u(t), t u(t), . . . , tk1 u(t). If the
hypotheses of the implicit function theorem4 are satisfied, then we can solve for
tk u(t) uniquely, and rewrite the ODE as an autonomous quasilinear ODE of order
k
(1.4) tk u(t) = F (u(t), t u(t), . . . , tk1 u(t)),
for some function F : Dk D. Of course, there are times when the implicit
function theorem is not available, for instance if the domain Y of G has a different
dimension than that of D. If Y has larger dimension than D then the equation is
often over-determined ; it has more equations of motion than degrees of freedom,
and one may require some additional hypotheses on the initial data before a solution
is guaranteed. If Y has smaller dimension than D then the equation is often under-
determined ; it has too few equations of motion, and one now expects to have a
multiplicity of solutions for any given initial datum. And even if D and Y have
the same dimension, it is possible for the ODE to sometimes be degenerate, in that
the Jacobian that one needs to invert for the implicit function theorem becomes
singular.
3Informally, what one has done is added a clock s to the system, which evolves at the fixed
rate of one time unit per time unit ( ds
dt
1 = 0), and then the remaining components of the system
are now driven by clock time rather than by the system time. The astute reader will note that this
new ODE not only contains all the solutions to the old ODE, but also contains some additional
solutions; however these new solutions are simply time translations of the solutions coming from
the original ODE.
4An alternate approach is to differentiate (1.2) in time using the chain rule, obtaining an
equation which is linear in tk+1 u(t), and provided that a certain matrix is invertible, one can
rewrite this in the form (1.4) but with k replaced by k + 1.
4 1. ORDINARY DIFFERENTIAL EQUATIONS

Degenerate ODE are rather difficult to study and will not be addressed here.
Both under-determined and over-determined equations cause difficulties for analy-
sis, which are resolved in different ways. An over-determined equation can often be
made determined by forgetting some of the constraints present in (1.2), for in-
stance by projecting Y down to a lower-dimensional space. In many cases, one can
then recover the forgotten constraints by using some additional hypothesis on the
initial datum, together with an additional argument (typically involving Gronwalls
inequality); see for instance Exercises 1.13, (6.4). Meanwhile, an under-determined
equation often enjoys a large group of gauge symmetries which help explain
the multiplicity of solutions to the equation; in such a case one can often fix a spe-
cial gauge, thus adding additional equations to the system, to make the equation
determined again; see for instance Section 6.2 below. In some cases, an ODE can
contain both over-determined and under-determined components, requiring one to
perform both of these types of tricks in order to recover a determined equation,
such as one of the form (1.4).
Suppose that u is a classical solution to the quasilinear ODE (1.4), and that
the nonlinearity F : Dk D is smooth. Then one can differentiate (1.4) in time,
use the chain rule, and then substitute in (1.4) again, obtain an equation of the
form
tk+1 u(t) = Fk+1 (u(t), t u(t), . . . , tk1 u(t))
for some smooth function Fk+1 : Dk D which can be written explicitly in terms
of G. More generally, by an easy induction we obtain equations of the form

(1.5) tk u(t) = Fk (u(t), t u(t), . . . , tk1 u(t))
for any k k, where Fk : Dk D is a smooth function which depends only on G
and k . Thus, if one specifies the initial data u(t0 ), . . . , tk1 u(t0 ) at some fixed time
t0 , then all higher derivatives of u at t0 are also completely specified. This shows in
particular that if u is k 1-times continuously differentiable and F is smooth, then
u is automatically smooth. If u is not only smooth but analytic, then from Taylor
expansion we see that u is now fixed uniquely. Of course, it is only reasonable to
expect u to be analytic if F is also analytic. In such a case, we can complement
the above uniqueness statement with a (local) existence result:
Theorem 1.1 (Cauchy-Kowalevski theorem). Let k 1. Suppose F : Dk D
is real analytic, let t0 R, and let u0 , . . . , uk1 D be arbitrary. Then there exists
an open time interval I containing t0 , and a unique real analytic solution u : I D
to (1.4), which obeys the initial value conditions
u(t0 ) = u0 ; t u(t0 ) = u1 , . . . , tk1 u(t0 ) = uk1 .
We defer the proof of this theorem to Exercise 1.1. This beautiful theorem
can be considered as a complete local existence theorem for the ODE (1.4), in
the case when G is real analytic; it says that the initial position u(t0 ), and the
first k 1 derivatives, t u(t0 ), . . . , tk1 u(t0 ), are precisely the right amount of
initial data 5 needed in order to have a wellposed initial value problem (we will
define wellposedness more precisely later). However, it turns out to have somewhat
5Conventions differ on when to use the singular datum and the plural data. In this text,
we shall use the singular datum for ODE and PDE that are first-order in time, and the plural
data for ODE and PDE that are higher order (or unspecified order) in time. Of course, in both
cases we use the plural when considering an ensemble or class of data.
 1.1. GENERAL THEORY 5

limited application when we move from ODE to PDE (though see Exercise 3.25).
We will thus rely instead primarily on a variant of the Cauchy-Kowalevski theorem,
namely the Picard existence theorem, which we shall discuss below.
Remark 1.2. The fact that the solution u is restricted to lie in a open interval
I, as opposed to the entire real line R, is necessary. A basic example is the initial
value problem
(1.6) ut = u2 ; u(0) = 1
where u takes values on the real line R. One can easily verify that the function
1
u(t) := 1t solves this ODE with the given initial datum as long as t < 1, and thus
is the unique real-analytic solution to this ODE in this region. But this solution
clearly blows up (i.e. ceases to be smooth) at t = 1, and so cannot be continued6
real analytically beyond this point.
There is a simple trick available to reduce a k th order ODE such as (1.4) to a
first order ODE, at the cost of multiplying the number of degrees of freedom by k,
or more precisely, replacing the state space D by the phase space Dk . Indeed, if
one defines the new function u : I Dk by
u(t) := (u(t), t u(t), . . . , tk1 u(t)),
then the equation (1.4) is equivalent to
t u(t) = F (u(t))
where F : Dk Dk is the function
F (u0 , . . . , uk1 ) = (u1 , . . . , uk1 , F (u0 , . . . , uk1 )).
Furthermore, u is continuously differentiable if and only if u is k times continuously
differentiable, and the k initial conditions
u(t0 ) = u0 ; t u(t0 ) = u1 ; ...; tk1 u(t0 ) = uk1
collapse to a single initial condition
u(t0 ) = u0 := (u0 , . . . , uk1 ).
Thus for the remainder of this chapter, we shall focus primarily on the initial value
problem (or Cauchy problem) of obtaining solutions u(t) to the first-order ODE7
(1.7) t u(t) = F (u(t)) for all t I; u(t0 ) = u0 .
where the interval I, the initial time t0 , the initial datum u0 D, and the nonlin-
earity F : D D are given. We will distinguish three types of solutions:

6One can of course consider a meromorphic continuation beyond t = 1, but this would require
complexifying time, which is a somewhat non-physical operation. Also, this complexification now
relies very heavily on the analyticity of the situation, and when one goes from ODE to PDE, it is
unlikely to work for non-analytic initial data. The question of whether one can continue a solution
in some weakened sense beyond a singularity is an interesting and difficult one, but we will not
pursue it in this text.
7One can interpret F as a vector field on the state space D, in which case the ODE is simply
integrating this vector field; see Figure 1. This viewpoint is particularly useful when considering
the Hamiltonian structure of the ODE as in Section 1.4, however it is not as effective a conceptual
framework when one passes to PDE.
6 1. ORDINARY DIFFERENTIAL EQUATIONS

F(u(t))

u(t)

u0

Figure 1. Depicting F as a vector field on D, the trajectory of

the solution u(t) to the first order ODE (1.7) thus follows the
arrows and integrates the vector field F . Contrast this classi-
cal solution interpretation of an ODE with the rather different
strong solution interpretation in Figure 2.

1
A classical solution of (1.7) is a function u Cloc (I D) which solves
(1.7) for all t I in the classical sense (i.e. using the classical notion of
derivative).
0
A strong solution of (1.7) is a function u Cloc (I D) which solves (1.7)
in the integral sense that
Z t
(1.8) u(t) = u0 + F (u(s)) ds
t0
8
holds for all t I;
A weak solution of (1.7) is a function u L (I D) which solves (1.8)
in the sense of distributions, thus for any test function C0 (I), one
has
Z Z Z Z t
u(t)(t) dt = u0 (t) + (t) F (u(s)) dsdt.
I I I t0
Later, when we turn our attention to PDE, these three notions of solution
shall become somewhat distinct; see Section 3.2. In the ODE case, however, we
fortunately have the following equivalence (under a very mild assumption on F ):
0
Lemma 1.3. Let F Cloc (D D). Then the notions of classical solution,
strong solution, and weak solution are equivalent.
Proof. It is clear that a classical solution is strong (by the fundamental the-
orem of calculus), and that a strong solution is weak. If u is a weak solution, then
8Recall that we are adopting the convention that R t = R s if t < s.
s t
 1.1. GENERAL THEORY 7

Forcing term
Nonlinearity F
F(u)

Integration
in time

Initial datum Solution

u u
0
Constant evolution

Figure 2. A schematic depiction of the relationship between the

initial datum u0 , the solution u(t), and the nonlinearity F (u). The
main issue is to control the feedback loop in which the solution
influences the nonlinearity, which in turn returns to influence the
solution.

it is bounded and measurable, hence F (u) is also bounded and measurable. Thus
Rt
the integral t0 F (u(s)) ds is Lipschitz continuous, and (since u solves (1.8) in the
sense of distributions) u(t) is also Lipschitz continuous, so it is a strong solution
(we allow ourselves the ability to modify u on a set of measure zero). Then F (u)
is continuous, and so the fundamental theorem of calculus and (1.8) again, u is in
1
fact in Cloc and is a classical solution. 
The three perspectives of classical, strong, and weak solutions are all important
in the theory of ODE and PDE. The classical solution concept, based on the differ-
ential equation (1.7), is particularly useful for obtaining conservation laws (Section
1.4) and monotonicity formulae (Section 1.5), and for understanding symmetries
of the equation. The strong solution concept, based on the integral equation (1.8),
is more useful for constructing solutions (in part because it requires less a priori
regularity on the solution), and establishing regularity and growth estimates on the
solution. It also leads to a very important perspective on the equation, viewing
the solution u(t) as being the combination of two influences, one coming from the
initial datum u0 and the other coming from the forcing term F (u); see Figure 2.
Finally, the concept of a weak solution arises naturally when constructing solutions
via compactness methods (e.g. by considering weak limits of classical solutions),
since continuity is not a priori preserved by weak limits.
To illustrate the strong solution concept, we can obtain the first fundamental
theorem concerning such Cauchy problems, namely the Picard existence theorem.
We begin with a simplified version of this theorem to illustrate the main point.
Theorem 1.4 (Picard existence theorem, simplified version). Let D be a finite-
dimensional normed vector space. Let F C 0,1 (D D) be a Lipschitz function on
D with Lipschitz constant kF kC 0,1 = M . Let 0 < T < 1/M . Then for any t0 R
and u0 D, there exists a strong (hence classical) solution u : I D to the Cauchy
problem (1.7), where I is the compact time interval I := [t0 T, t0 + T ].
8 1. ORDINARY DIFFERENTIAL EQUATIONS

Proof. Fix u0 D and t0 R, and let : C 0 (I D) C 0 (I D) be the

map Z t
(u)(t) := u0 + F (u(t )) dt .
t0
Observe from (1.8) that a strong solution is nothing more than a fixed point of the
map . It is easy to verify that is indeed a map from C 0 (I D) to C 0 (I D).
Using the Lipschitz hypothesis on F and the triangle inequality, we obtain
Z t Z t
k(u)(t) (v)(t)kD = k F (u(t )) F (v(t )) dt kD M ku(t ) v(t )kD dt
t0 t0

for all t I and u, v C 0 (I ), and thus

k(u) (v)kC 0 (ID) T M ku vkC 0 (ID) .
Since we have T M < 1, we see that will be a strict contraction on the complete
metric space C 0 (I D). Applying the contraction mapping theorem (Exercise 1.2)
we obtain a fixed point to , which gives the desired strong (and hence classical)
solution to the equation (1.7). 
Remark 1.5. An inspection of the proof of the contraction mapping theorem
reveals that the above argument in fact gives rise to an explicit iteration scheme
that will converge to the solution u. Indeed, one can start with the constant
solution u(0) (t) := u0 , and then define further iterates u(n) C 0 (I ) by
u(n) := (u(n1) ), or in other words
Z t
(n)
u (t) := u0 + F (u(n1) (t ) dt .
t0
These Picard iterates do not actually solve the equation (1.7) in any of the above
senses, but they do converge uniformly on I to the actual solution. See Figure 3.
Remark 1.6. The above argument is perhaps the simplest example of the
iteration method (also known as the contraction mapping principle method or the
inverse function theorem method ), constructing a nonlinear solution as the strong
limit of an iterative procedure. This type of method will be our primary means
of generating solutions which obey a satisfactory set of existence, uniqueness, and
regularity properties. Note that one needs to select a norm kkD in order to obtain
a quantitative estimate on the time of existence. For finite-dimensional ODE, the
exact choice of norm is not terribly important (as all norms are equivalent), but
selecting the norm in which to apply the contraction mapping theorem will become
decisive when studying PDE.
Because F is assumed to be globally Lipschitz (C 0,1 ), one can actually construct
a global solution to (1.7) in this case, by iterating the above theorem; see Exercise
0,1
1.10. However, in most applications F will only be locally Lipschitz (Cloc ), and
so we shall need a more general version of the above existence theorem. One
such version (which also gives some further information, namely some Lipschitz
continuity properties on the solution map) is as follows.
Theorem 1.7 (Picard existence theorem, full version). Let D be a finite-
dimensional normed vector space. Let t0 R, let be a non-empty subset of D,
and let N () := {u D : ku vkD < for some v } be the -neighbourhood
of for some > 0. Let F : D D be a function which is Lipschitz on the closed
 1.1. GENERAL THEORY 9

Forcing term
F(u) in C0
Nonlinearity F
Integration (dilates by M)
in time
(contracts by T)

Initial datum Solution

u in u in C
0
0
Constant evolution

Figure 3. The Picard iteration scheme. The map is basically

the loop from the solution u to itself. To obtain the fixed point,
start with the initial datum u0 as the first approximant to u, and
apply repeatedly to obtain further approximations to u. As long
as the net contraction factor T M is less than 1, the iteration
scheme will converge to an actual solution.

neighbourhood N () with some Lipschitz constant kF kC 0,1 (N ()) = M > 0, and

which is bounded by some A > 0 on this region. Let 0 < T < min(/A, 1/M ),
and let I be the interval I := [t0 T, t0 + T ]. Then for every u0 , there exists
a strong (hence classical) solution u : I N () to the Cauchy problem (1.7).
Furthermore, if we then define the solution maps St0 (t) : D for t I and
St0 : C 0 (I D) by setting St0 (t)(u0 ) := u(t) and St0 (u0 ) := u, then St0 (t)
and St0 are Lipschitz continuous maps, with Lipschitz constant at most 1T1 M .

Proof. Write := N () for short. For each u0 let u0 : C 0 (I

) C 0 (I ) be the map
Z t
u0 (u)(t) := u0 + F (u(s)) ds.
t0

As before, a strong solution to (1.7) is nothing more than a fixed point of the map
u0 . Since F is bounded by A on and T < /A, we see from the triangle
inequality that u0 will indeed map C 0 (I ) to C 0 (I ). Also, since F
has Lipschitz constant at most M on , we may argue as in the proof of Theorem
1.4 and conclude that u0 will thus be a strict contraction on the complete metric
space C 0 (I ) with contraction constant c := T M < 1, and hence will have a
fixed point u = u0 (u) C 0 (I ). This gives a strong (and hence classical)
solution to the equation (1.7).
Now let u0 and u0 be two initial data in , with corresponding solutions
St0 (u0 ) = u C 0 (I D), St0 (u0 ) = u C 0 (I D) constructed above. Observe
from construction that u0 (u) = u and u0 (u) = u0 (u) + u0 u0 = u + u0 u0 ,
thus
u u = u0 (u) u0 (u) + u0 u0 .
10 1. ORDINARY DIFFERENTIAL EQUATIONS

Taking norms and applying the contraction property and the triangle inequality,
we conclude
ku ukC 0 (ID) cku ukC 0 (ID) + ku0 u0 kD
and hence
1
ku ukC 0 (ID) ku0 u0 kD .
1c
This proves the desired Lipschitz property on St0 , and hence on each individual
St0 (t). 
Remark 1.8. The above theorem illustrates a basic point in nonlinear differ-
ential equations: in order to construct solutions, one does not need to control the
nonlinearity F (u) for all choices of state u, but only for those u that one expects to
encounter in the evolution of the solution. For instance, if the initial datum is small,
one presumably only needs to control F (u) for small u in order to obtain a local
existence result. This observation underlies many of the perturbative arguments
which we shall see in this text (see for instance Proposition 1.24 below).
Remark 1.9. In the next section we shall complement the Picard existence
theorem with a uniqueness theorem. The hypothesis that F is locally Lipschitz can
be weakened, but at the cost of losing the uniqueness; see Exercise 1.23.
Exercise 1.1. Begin the proof of the Cauchy-Kowalevski theorem by reducing
to the case k = 1, t0 = 0, and u0 = 0. Then, use induction to show that if the
higher derivatives tm u(0) are derived recursively as in (1.5), then we have some
bound of the form
ktm u(0)kD K m+1 m!
for all m 0 and some large K > 0 depending on F , where kkD is some arbitrary
norm on the finite-dimensional space D. Then, define u : I D for some sufficiently
small neighbourhood I of the origin by the power series
X
tm u(0) m
u(t) = t
m=0
m!
and show that t u(t) G(u(t)) is real analytic on I and vanishes at infinite order
at zero, and is thus zero on all of I.
Exercise 1.2. (Contraction mapping theorem) Let (X, d) be a complete non-
empty metric space, and let : X X be a strict contraction on X, thus there
exists a constant 0 < c < 1 such that d((u), (v)) cd(u, v) for all u, v X.
Show that has a unique fixed point, thus there is a unique u X such that
u = (u). Furthermore, if u0 is an arbitrary element of X and we construct the
sequence u1 , u2 , . . . X iteratively by un+1 := (un ), show that un will converge
to the fixed point u. Finally, we have the bound
1
(1.9) d(v, u) d(v, (v))
1c
for all v X.
Exercise 1.3. (Inverse function theorem) Let D be a finite-dimensional vector
1
space, and let Cloc (D D) be such that (x0 ) has full rank for some
x0 D. Using the contraction mapping theorem, show that there exists an open
neighbourhood U of x0 and an open neighbourhood V of (x0 ) such that is a
bijection from U to V , and that 1 is also Cloc
1
.
 1.2. GRONWALLS INEQUALITY 11

Exercise 1.4. Suppose we make the further assumption in the Picard existence
k
theorem that F Cloc (D D) for some k 1. Show that the maps St0 (t) and
k+1
S(t) are then also continuously k-times differentiable, and that u Cloc (I D).
Exercise 1.5. How does the Picard existence theorem generalise to higher or-
der quasilinear ODE? What if there is time dependence in the nonlinearity (i.e. the
ODE is non-autonomous)? The latter question can also be asked of the Cauchy-
Kowaleski theorem. (These questions can be answered quickly by using the reduc-
tion tricks mentioned in this section.)
Exercise 1.6. One could naively try to extend the local solution given by the
Picard existence theorem to a global solution by iteration, as follows: start with
the initial time t0 , and use the existence theorem to construct a solution all the
way up to some later time t1 . Then use u(t1 ) as a new initial datum and apply the
existence theorem again to move forward to a later time t2 , and so forth. What
goes wrong with this strategy, for instance when applied to the problem (1.6)?

1.2. Gronwalls inequality

It takes money to make money. (Proverbial)
As mentioned earlier, we will be most interested in the behaviour of ODE in
very high dimensions. However, in many cases one can compress the key features
of an equation to just a handful of dimensions, by isolating some important scalar
quantities arising from the solution u(t), for instance by inspecting some suitable
norm ku(t)kD of the solution, or looking at special quantities related to conservation
or pseudoconservation laws such as energy, centre-of-mass, or variance. In many
cases, these scalar quantities will not obey an exact differential equation themselves,
but instead obey a differential inequality, which places an upper limit on how quickly
these quantities can grow or decay. One is then faced with the task of solving
such inequalities in order to obtain good bounds on these quantities for extended
periods of time. For instance, if a certain quantity is zero or small at some time
t0 , and one has some upper bound on its growth rate, one would like to say that
it is still zero or small at later times. Besides the iteration method used already
in the Picard existence theorem, there are two very useful tools for achieving this.
One is Gronwalls inequality, which deals with linear growth bounds and is treated
here. The other is the continuity method, which can be used with nonlinear growth
bounds and is treated in Section 1.3.
We first give Gronwalls inequality in an integral form.
Theorem 1.10 (Gronwall inequality, integral form). Let u : [t0 , t1 ] R+ be
continuous and non-negative, and suppose that u obeys the integral inequality
Z t
(1.10) u(t) A + B(s)u(s) ds
t0

for all t [t0 , t1 ], where A 0 and B : [t0 , t1 ] R+ is continuous and nonnegative.

Then we have
Z t
(1.11) u(t) A exp( B(s) ds)
t0

for all t [t0 , t1 ].

12 1. ORDINARY DIFFERENTIAL EQUATIONS

Forcing term
Bu
Growth factor B
Integration
in time

Initial bound Solution

A u
Constant evolution

Figure 4. The linear feedback encountered in Theorem 1.10, that

causes exponential growth by an amount depending on the growth
factor B. Contrast this with Figure 2.

Remark 1.11. This estimate is absolutely sharp, since the function u(t) :=
Rt
A exp( t0 B(s) ds) obeys the hypothesis (1.10) with equality.
Proof. By a limiting argument it suffices to prove the claim when A > 0. By
(1.10) and the fundamental theorem of calculus, (1.10) implies
Z t Z t
d
(A + B(s)u(s) ds) B(t)(A + B(s)u(s) ds)
dt t0 t0

and hence by the chain rule

Z t
d
log(A + B(s)u(s) ds) B(t).
dt t0

Applying the fundamental theorem of calculus again, we conclude

Z t Z t
log(A + B(s)u(s) ds) log A + B(s) ds.
t0 t0

Exponentiating this and applying (1.10) again, the claim follows. 

There is also a differential form of Gronwalls inequality in which B is allowed

to be negative:
Theorem 1.12 (Gronwall inequality, differential form). Let u : [t0 , t1 ] R+
be absolutely continuous and non-negative, and suppose that u obeys the differential
inequality
t u(t) B(t)u(t)
for almost every t [t0 , t1 ], where B : [t0 , t1 ] R+ is continuous and nonnegative.
Then we have Z t
u(t) u(t0 ) exp( B(s) ds)
t0
for all t [t0 , t1 ].
 1.2. GRONWALLS INEQUALITY 13

Rt
Proof. Write v(t) := u(t) exp( t0 B(s) ds). Then v is absolutely continuous,
and an application of the chain rule shows that t v(t) 0. In particular v(t) v(t0 )
for all t [t0 , t1 ], and the claim follows. 

Remark 1.13. This inequality can be viewed as controlling the effect of linear
feedback; see Figure 4. As mentioned earlier, this inequality is sharp in the worst
case scenario when t u(t) equals B(t)u(t) for all t. This is the case of adversarial
feedback, when the forcing term B(t)u(t) is always acting to increase u(t) by
the maximum amount possible. Many other arguments in this text have a similar
worst-case analysis flavour. In many cases (in particular, supercritical defocusing
equations) it is suspected that the average-case behaviour of such solutions (i.e.
for generic choices of initial data) is significantly better than what the worst-case
analysis suggests, thanks to self-cancelling oscillations in the nonlinearity, but we
currently have very few tools which can separate the average case from the worst
case.
As a sample application of this theorem, we have
Theorem 1.14 (Picard uniqueness theorem). Let I be an interval. Suppose we
1
have two classical solutions u, v Cloc (I D) to the ODE
t u(t) = F (u(t))
0,1
for some F Cloc (D D). If u and v agree at one time t0 I, then they agree
for all times t I.
Remark 1.15. Of course, the same uniqueness claim follows for strong or weak
solutions, thanks to Lemma 1.3.
Proof. By a limiting argument (writing I as the union of compact intervals)
it suffices to prove the claim for compact I. We can use time translation invariance
to set t0 = 0. By splitting I into positive and negative components, and using the
change of variables t 7 t if necessary, we may take I = [0, T ] for some T > 0.
Here, the relevant scalar quantity to analyze is the distance ku(t) v(t)kD
between u and v, where kkD is some arbitrary norm on D. We then take the ODE
for u and v and subtract, to obtain
t (u(t) v(t)) = F (u(t)) F (v(t)) for all t [0, T ]
Applying the fundamental theorem of calculus, the hypothesis u(0) = v(0), and the
triangle inequality, we conclude the integral inequality
Z t
(1.12) ku(t) v(t)kD kF (u(s)) F (v(s))kD ds for all t [0, T ].
0

Since I is compact and u, v are continuous, we see that u(t) and v(t) range over
a compact subset of D. Since F is locally Lipschitz, we thus have a bound of the
form |F (u(s)) F (v(s))| M |u(s) v(s)| for some finite M > 0. Inserting this
into (1.12) and applying Gronwalls inequality (with A = 0), the claim follows. 

Remark 1.16. The requirement that F be Lipschitz is essential; for instance

the non-Lipschitz Cauchy problem
(1.13) t u(t) = p|u(t)|(p1)/p ; u(0) = 0
14 1. ORDINARY DIFFERENTIAL EQUATIONS

T+

u
u
0

Figure 5. The maximal Cauchy development of an ODE which

blows up both forwards and backwards in time. Note that in order
for the time of existence to be finite, the solution u(t) must go
to infinity in finite time; thus for instance oscillatory singularities
cannot occur (at least when the nonlinearity F is smooth).

for some p > 1 has the two distinct (classical) solutions u(t) := 0 and v(t) :=
max(0, t)p . Note that a modification of this example also shows that one cannot
expect any continuous or Lipschitz dependence on the initial data in such cases.
By combining the Picard existence theorem with the Picard uniqueness theo-
rem, we obtain
0,1
Theorem 1.17 (Picard existence and uniqueness theorem). Let F Cloc (D
D) be a locally Lipschitz function, let t0 R be a time, and let u0 D be an initial
datum. Then there exists a maximal interval of existence I = (T , T+ ) for some
T < t0 < T+ +, and a unique classical solution u : I D to the
Cauchy problem (1.7). Furthermore, if T+ is finite, we have ku(t)kD as
t T+ from below, and similarly if T is finite then we have ku(t)kD as
t T from above.
Remark 1.18. This theorem gives a blowup criterion for the Cauchy problem
(1.7): a solution exists globally if and only if the ku(t)kD norm does not go to
infinity9 in finite time; see Figure 5. (Clearly, if ku(t)kD goes to infinity in finite
time, u is not a global classical solution.) As we shall see later, similar blowup
criteria (for various norms D) can be established for certain PDE.

Proof. We define I to be the union of all the open intervals containing t0 for
which one has a classical solution to (1.7). By the existence theorem, I contains a
9We sometimes say that a solution blows up at infinity if the solution exists globally as
t , but that the norm ku(t)kD is unbounded; note that Theorem 1.17 says nothing about
whether a global solution will blow up at infinity or not, and indeed both scenarios are easily seen
to be possible.
 1.2. GRONWALLS INEQUALITY 15

neighbourhood of t0 and is clearly open and connected, and thus has the desired
form I = (T , T+ ) for some T < t0 < T+ +. By the uniqueness
theorem, we may glue all of these solutions together and obtain a classical solution
u : I D on (1.7). Now suppose for contradiction that T+ was finite, and that
there was some sequence of times tn approaching T+ from below for which ku(t)kD
stayed bounded. On this bounded set (or on any slight enlargement of this set) F
is Lipschitz. Thus we may apply the existence theorem and conclude that one can
extend the solution u to a short time beyond T+ ; gluing this solution to the existing
solution (again using the uniqueness theorem) we contradict the maximality of I.
This proves the claim for T+ , and the claim for T is proven similarly. 

The Picard theorem gives a very satisfactory local theory for the existence and
uniqueness of solutions to the ODE (1.7), assuming of course that F is locally
Lipschitz. The issue remains, however, as to whether the interval of existence
(T , T+ ) is finite or infinite. If one can somehow ensure that ku(t)kD does not
blow up to infinity at any finite time, then the above theorem assures us that the
interval of existence is all of R; as we shall see in the exercises, Gronwalls inequality
is one method in which one can assure the absence of blowup. Another common
way to ensure global existence is to obtain a suitably coercive conservation law
(e.g. energy conservation), which manages to contain the solution to a bounded set;
see Proposition 1.24 below, as well as Section 1.4 for a fuller discussion. A third
way is to obtain decay estimates, either via monotonicity formulae (see Section
1.5) or some sort of dispersion or dissipation effect. We shall return to all of these
themes throughout this monograph, in order to construct global solutions to various
equations.
Gronwalls inequality is causal in nature; in its hypothesis, the value of the
unknown function u(t) at times t is controlled by its value at previous times 0 <
s < t, averaged against a function B(t) which can be viewed as a measure of
the feedback present in the system; thus it is excessive feedback which leads to
exponential growth (see Figure 4). This is of course very compatible with ones
intuition regarding cause and effect, and our interpretation of t as a time variable.
However, in some cases, when t is not being interpreted as a time variable, one
can obtain integral inequalities which are acausal in that u(t) is controlled by an
integral of u(s) both for s < t and s > t. In many such cases, these inequalities
lead to no useful conclusion. However, if the feedback is sufficiently weak, and one
has some mild growth condition at infinity, one can still proceed as follows.

Theorem 1.19 (Acausal Gronwall inequality). Let 0 < < , 0 < <
and , > 0 be real numbers. Let u : R R+ be measurable and non-negative,
and suppose that u obeys the integral inequality
Z
(1.14) u(t) A(t) + min(e(st) , e(ts) )u(s) ds
R

for all t R, where A : R R+ is an arbitrary function. Suppose also that we

have the subexponential growth condition

sup e|t| u(t) < .

tR
16 1. ORDINARY DIFFERENTIAL EQUATIONS

Then if < min(, ) and is sufficiently small depending on , , , , , we

have

(1.15) u(t) 2 sup min(e (st) , e (ts) )A(s).
sR

for all t R.
Proof. We shall use an argument similar in spirit to that of the contraction
mapping theorem, though in this case there is no actual contraction to iterate as
we have an integral inequality rather than an integral equation. By raising and
(depending on , , ) if necessary we may assume < min( , ). We will
assume that there exists > 0 such that A(t) e|t| for all t R; the general
case can then be deduced by replacing A(t) by A(t) + e|t| and then letting 0,
noting that the growth of the e|t| factor will be compensated for by the decay of

the min(e (st) , e (ts) ) factor since < min( , ). Let B : R R+ denote
the function

B(t) := sup min(e (st) , e (ts) )A(s).
sR
|t|
Then we see that e A(t) B(t), that B is strictly positive, and furthermore
B obeys the continuity properties

(1.16) B(s) max(e (st) , e (ts)
)B(t)
for all t, s R.
Let M be the smallest real number such that u(t) M B(t) for all t R; our
objective is to show that M 2. Since B is bounded from below by e|t| , we see
from the subexponential growth condition that M exists and is finite. From (1.14)
we have Z
u(t) B(t) + min(e(st) , e(ts) )u(s) ds.
R
Bounding u(s) by M B(s) and applying (1.16), we conclude
Z

u(t) B(t) + M B(t) min(e( )(st) , e( )(ts) ) ds.
R

Since 0 < < and 0 < < , the integral is convergent and is independent of
t. Thus if is sufficiently small depending on , , , , we conclude that
1
u(t) B(t) + M B(t)
2
which by definition of M implies M 1 + 21 M . Since M is finite, we have M 2
as desired. 

The above inequality was phrased for a continuous parameter t, but it quickly
implies a discrete analogue:
Corollary 1.20 (Discrete acausal Gronwall inequality). Let 0 < < ,
0 < < , > 0, and 0 < < min(, ) be real numbers. Let (un )nZ be a
sequence of non-negative numbers such that
X
(1.17) un An + min(e(mn) , e(nm) )um
mZ
 1.2. GRONWALLS INEQUALITY 17

for all t R, where (An )nZ is an arbitrary non-negative sequence. Suppose also
that we have the subexponential growth condition
sup un e|n| < .
nZ

Then if is sufficiently small depending on , , , , , we have

(1.18) un 2 sup min(e (mn) , e (nm) )Am .
mZ

for all n Z.
This corollary can be proven by modifying the proof of the previous theorem,
or alternatively by considering the function u(t) := u[t] , where [t] is the nearest
integer to t; we leave the details to the reader. This corollary is particularly useful
for understanding the frequency distribution of solutions to nonlinear dispersive
equations, in situations when the data is small (so the nonlinear effects of energy
transfer between dyadic frequency ranges || 2n are weak). See for instance
[Tao5], [Tao6], [Tao7] for ideas closely related to this. One can also use these
types of estimates to establish small energy regularity for various elliptic problems
by working in frequency space (the smallness is needed to make the nonlinear effects
weak).
Exercise 1.7 (Comparison principle). Let I = [t0 , t1 ] be a compact interval,
and let u : I R, v : I R be two scalar absolutely continuous functions. Let
0,1
F Cloc (I R R), and suppose that u and v obey the differential inequalities
t u(t) F (t, u(t)); t v(t) F (t, v(t))
for all t I. Show that if u(t0 ) v(t0 ), then u(t) v(t) for all t [t0 , t1 ], and
similarly if u(t0 ) < v(t0 ), then u(t) < v(t) for all t [t0 , t1 ]. (Hint: for the first
claim, study the derivative of max(0, u(t) v(t))2 and use Gronwalls inequality.
For the second, perturb the first argument by an epsilon.)
Exercise 1.8. Give an example to show that Theorem 1.10 fails when B is
permitted to be negative. Briefly discuss how this is consistent with the fact that
Theorem 1.12 still holds for negative B.
Exercise 1.9 (Sturm comparison principle). Let I be a time interval, and let
2 0
u, v Cloc (I R) and a, f, g Cloc (I R) be such that
t2 u(t) + a(t)t u(t) + f (t)u(t) = t2 v(t) + a(t)t v(t) + g(t)v(t) = 0
for all t I. Suppose also that v oscillates faster than u, in the sense that g(t) f (t)
for all t I. Suppose also that u is not identically zero. Show that the zeroes of
v intersperse the zeros of u, in the sense that whenever t1 < t2 are times in I
such that u(t1 ) = u(t2 ) = 0, then v has at least one zero in the interval [t1 , t2 ].
(Hint: reduce to the case when t1 and t2 are consecutive zeroes of u, and argue by
contradiction. By replacing u or v with u or v if necessary one may assume that
u, v are nonnegative on [t1 , t2 ]. Obtain a first order equation for the Wronskian
ut v vt u.) This principle can be thought of as a substantial generalisation of the
observation that the zeroes of the sine and cosine functions intersperse each other.
0,1
Exercise 1.10. Let F Cloc (D D) have at most linear growth, thus
kF (u)kD . 1 + kukD for all u D. Show that for each u0 D and t0 R
there exists a unique classical global solution u : R D to the Cauchy problem
18 1. ORDINARY DIFFERENTIAL EQUATIONS

(1.7). Also, show that the solution maps St0 (t) : D D defined by St0 (u0 ) = u(t0 )
are locally Lipschitz, obey the time translation invariance St0 (t) = S0 (t t0 ), and
the group laws S0 (t)S0 (t ) = S0 (t + t ) and S0 (0) = id. (Hint: use Gronwalls in-
equality to obtain bounds on ku(t)kD in the maximal interval of existence (T , T+ )
given by Theorem 1.17.) This exercise can be viewed as the limiting case p = 1 of
Exercise 1.11 below.
Exercise 1.11. Let p > 1, let D be a finite-dimensional normed vector space,
0,1
and let F Cloc (D D) have at most pth -power growth, thus kF (u)kD . 1+kukpD
for all u D. Let t0 R and u0 D, and let u : (T , T+ ) D be the maximal
classical solution to the Cauchy problem (1.7) given by the Picard theorem. Show
that if T+ is finite, then we have the lower bound
ku(t)kD &p (T+ t)1/(p1)
as t approaches T+ from below, and similarly for T . Give an example to show
that this blowup rate is best possible.
0,1
Exercise 1.12 (Slightly superlinear equations). Suppose F Cloc (D D)
has at most x log x growth, thus
kF (u)kD . (1 + kukD ) log(2 + kukD )
for all u D. Do solutions to the Cauchy problem (1.7) exist classically for all time
(as in Exercise 1.10), or is it possible to blow up (as in Exercise 1.11)? In the latter
case, what is the best bound one can place on the growth of ku(t)kD in time; in
the former case, what is the best lower bound one can place on the blow-up rate?
Exercise 1.13 (Persistence of constraints). Let u : I D be a (classical)
solution to the ODE t u(t) = F (u(t)) for some time interval I and some F
0 1
Cloc (D D), and let H Cloc (D R) be such that hF (v), dH(v)i = G(v)H(v)
0
for some G Cloc (D R) and all v D; here we use
d
(1.19) hu, dH(v)i := H(v + u)|=0
d
to denote the directional derivative of H at v in the direction u. Show that if
H(u(t)) vanishes for one time t I, then it vanishes for all t I. Interpret this
geometrically, viewing F as a vector field and studying the level surfaces of H. Note
that it is necessary that the ratio G between hF, dHi and H be continuous; it is not
enough merely for hF, dHi to vanish whenever H does, as can be seen for instance
from the counterexample H(u) = u2 , F (u) = 2|u|1/2 , u(t) = t2 .
1
Exercise 1.14 (Compatibility of equations). Let F, G Cloc (D D) have
the property that
(1.20) hF (v), dG(v)i hG(v), dF (v)i = 0
for all v D. (The left-hand side has a natural interpretation as the Lie bracket
[F, G] of the differential operators F , G associated to the vector fields F and
G.) Show that for any u0 D, there exists a neighbourhood B R2 of the origin,
and a map u C 2 (B D) which satisfies the two equations
(1.21) s u(s, t) = F (u(s, t)); t u(s, t) = G(u(s, t))
for all (s, t) B, with initial datum u(0, 0) = u0 . Conversely, if u C 2 (B D)
solves (1.21) on B, show that (1.20) must hold for all v in the range of u. (Hint:
 1.2. GRONWALLS INEQUALITY 19

use the Picard existence theorem to construct u locally on the s-axis {t = 0}

by using the first equation of (1.21), and then for each fixed s, extend u in the
t direction using the second equation of R(1.21). Use Gronwalls inequality and
s
(1.20) to establish that u(s, t) u(0, t) 0 F (u(s , t)) ds = 0 for all (s, t) in a
neighbourhood of the origin.) This is a simple case of Frobeniuss theorem, regarding
when a collection of vector fields can be simultaneously integrated.
Exercise 1.15 (Integration on Lie groups). Let H be a finite-dimensional
vector space, and let G be a Lie group in End(H) (i.e. a group of linear transfor-
mations on H which is also a smooth manifold). Let g be the Lie algebra of G (i.e.
0,1
the tangent space of G at the identity). Let g0 G, and let X Cloc (R g)
1
be arbitrary. Show that there exists a unique function g Cloc (R G) such
that g(0) = g0 and t g(t) = X(t)g(t) for all t R. (Hint: first use Gronwalls
inequality and Picards theorem to construct a global solution g : R Mn (C)
to the equation t g(t) = X(t)g(t), and then use Gronwalls inequality again,
and local coordinate patches of G, to show that g stays on G.) Show that the
same claim holds if the matrix product X(t)g(t) is replaced by the Lie bracket
[g(t), X(t)] := g(t)X(t) X(t)g(t).
Exercise 1.16 (Levinsons theorem). Let L C 0 (R End(D)) be a time-
dependent linear transformation acting on a finite-dimensional Hilbert space D, and
let F C 0 (R D) be a time-dependent forcing term. Show that for every u0 D
0
there exists a global solution u Cloc (R D) to the ODE t u = L(t)u + F (t),
with the bound
Z t Z t
|u(t)| (|u0 |D + |F (s)|D ds) exp( k(L(t) + L (t))/2kop dt)
0 0
for all t 0. (Hint: control the evolution of |u(t)|2 = hu(t), u(t)iD .) Thus one can
obtain global control on a linear ODE with arbitrarily large coefficients, as long as
the largeness is almost completely contained in the skew-adjoint component of the
linear operator. In particular, if F and the self-adjoint component of L are both
absolutely integrable, conclude that u(t) is bounded uniformly in t.
Exercise 1.17. Give examples to show that Theorem 1.19 and Corollary 1.20
fail (even when A is identically zero) if or become too large, or if the hypothesis
that u has subexponential growth is dropped.
Exercise 1.18. Let , > 0, let d 1 be an integer, let 0 < d, and let
u : Rd R+ and A : Rd R+ be locally integrable functions such that one has
the pointwise inequality
Z
e|xy|
u(x) A(x) +
u(y) dy
Rd |x y|
for almost every x Rd . Suppose also that u is a tempered distribution in addition
to a locally integrable function. Show that if 0 < < and is sufficiently small
depending on , , , then we have the bound

u(x) 2ke |xy| A(y)kL d
y (R )

for almost every x Rd . (Hint: you will need to regularise u first, averaging on a
small ball, in order to convert the tempered distribution hypothesis into a pointwise
subexponential bound. Then argue as in Proposition 1.19. One can then take limits
at the end using the Lebesgue differentiation theorem.)
20 1. ORDINARY DIFFERENTIAL EQUATIONS

Exercise 1.19 (Singular ODE). Let F, G C 0,1 (D D) be Lipschitz maps

with F (0) = 0 and kF kC 0,1 (DD) < 1. Show that there exists a T > 0 for
which there exists a unique classical solution u : (0, T ] D to the singular
non-autonomous ODE t u(t) = 1t F (u(t)) + G(u(t)) with the boundary condi-
tion lim supt0 ku(t)kD /t < as t 0. (Hint: For uniqueness, use a Gronwall
inequality argument. For existence, construct iterates in the space of functions
{tv : v C 0 ([0, T ] D)}.) Show that u in fact extends to a C 1 function on [0, T ]
with u(0) = 0 and t u(0) = G(0). Also give an example to show that uniqueness
can break down when the Lipschitz constant of F exceeds 1. (You can take a very
simple example, for instance with F linear and G zero.)

1.3. Bootstrap and continuity arguments

If you have built your castles in the air, your work need not be lost;
that is where they should be. Now put the foundations under them.
(Henry David Thoreau, Walden)
The Picard existence theorem allows us to construct solutions to ODE such as
t u(t) = F (u(t)) on various time intervals. Once these solutions have been con-
structed, it is natural to then ask what kind of quantitative estimates and asymp-
totics these solutions satisfy, especially over long periods of time. If the equation is
fortunate enough to be solvable exactly (which can happen for instance if the equa-
tion is completely integrable), then one can read off the desired estimates from the
exact solution, in principle at least. However, in the majority of cases no explicit
solution is available10. Many times, the best one can do is to write the solution
u(t) in terms of itself, using the strong solution concept. For instance, if the initial
condition is u(t0 ) = u0 , then we have
Z t
(1.22) u(t) = u0 + F (u(s)) ds.
t0
This equation tells us that if we have some information on u (for instance, if we
control some norm kkY of u(s)), we can insert this information into the right-
hand side of the above integral equation (together with some knowledge of the
initial datum u0 and the nonlinearity F ), and conclude some further control of the
solution u (either in the same norm kkY , or in some new norm).
Thus we can use equations such as (1.22) to obtain control on u - but only if one
starts with some control on u in the first place. Thus it seems difficult to get started
when executing this strategy, since one often starts with only very little control on
u, other than continuity. Nevertheless there is a simple principle, of almost magical
power, that allows one to assume for free that u already obeys some quantitative
bound, in order to prove that u obeys another quantitative bound - as long as
the bound one ends up proving is slightly stronger than the bound one used as a
hypothesis (to avoid circularity). This principle - which is a continuous analogue
10Of course, the contraction mapping argument used in Theorem 1.7 does in principle give
a description of the solution, at least locally, as the limit of iterates of a certain integral map ,
and the Cauchy-Kowalevski theorem in principle gives a Taylor series expansion of the solution.
However in practice these expansions are rather complicated, and only useful for analyzing the
short-time behaviour and not long-time behaviour. Even if an explicit solution (e.g. involving
special functions) is available, it may be easier to read off the asymptotics and other features of
the equation from an analytic argument such as a bootstrap argument than from inspection of
the explicit solution.
 1.3. BOOTSTRAP AND CONTINUITY ARGUMENTS 21

Conclusion
C(t)
"nonlinear"
can extend t by
(closed) implication
an open set

("continuity")

Base case Hypothesis

H(t ) H(t)
0

Figure 6. A schematic depiction of the relationship between the

hypothesis H(t) and the conclusion C(t); compare this with Figure
2. The reasoning is noncircular because at each loop of the iteration
we extend the set of times for which the hypothesis and conclusion
are known to hold. The closure hypothesis prevents the iteration
from getting stuck indefinitely at some intermediate time.

of the principle of mathematical induction - is known as the bootstrap principle or

the continuity method 11. Abstractly, the principle works as follows.
Proposition 1.21 (Abstract bootstrap principle). Let I be a time interval,
and for each t I suppose we have two statements, a hypothesis H(t) and a
conclusion C(t). Suppose we can verify the following four assertions:
(a) (Hypothesis implies conclusion) If H(t) is true for some time t I, then
C(t) is also true for that time t.
(b) (Conclusion is stronger than hypothesis) If C(t) is true for some t I,
then H(t ) is true for all t I in a neighbourhood of t.
(c) (Conclusion is closed) If t1 , t2 , . . . is a sequence of times in I which con-
verges to another time t I, and C(tn ) is true for all tn , then C(t) is
true.
(d) (Base case) H(t) is true for at least one time t I.
Then C(t) is true for all t I.
Remark 1.22. When applying the principle, the properties H(t) and C(t) are
typically chosen so that properties (b), (c), (d) are relatively easy to verify, with
property (a) being the important one (and the nonlinear one, usually proven
by exploiting one or more nonlinear feedback loops in the equations under study).
The bootstrap principle shows that in order to prove a property C(t) obeying (c),
it would suffice to prove the seemingly easier assertion H(t) = C(t), as long as
H is weaker than C in the sense of (b) and is true for at least one time.

11The terminology bootstrap principle arises because a solution u obtains its regularity
from its own resources rather than from external assumptions - pulling itself up by its bootstraps,
as it were. The terminology continuity method is used because the continuity of the solution is
essential to making the method work.
22 1. ORDINARY DIFFERENTIAL EQUATIONS

Proof. Let be the set of times t I for which C(t) holds. Properties (d)
and (a) ensure that is non-empty. Properties (b) and (a) ensure that is open.
Property (c) ensures that is closed. Since the interval I is connected, we thus see
that = I, and the claim follows. 
More informally, one can phrase the bootstrap principle as follows:
Principle 1.23 (Informal bootstrap principle). If a quantity u can be bounded
in a nontrivial way in terms of itself, then under reasonable conditions, one can
conclude that u is bounded unconditionally.
We give a simple example of the bootstrap principle in action, establishing
global existence for a system in a locally stable potential well from small initial
data.
Proposition 1.24. Let D be a finite-dimensional Hilbert space, and let V
2
Cloc (D R) be such that such that V (0) = 0, V (0) = 0, and 2 V (0) is strictly
positive definite. Then for all u0 , u1 D sufficiently close to 0, there is a unique
2
classical global solution u Cloc (R D) to the Cauchy problem
(1.23) t2 u(t) = V (u(t)); u(0) = u0 ; t u(0) = u1 .
Furthermore, this solution stays bounded uniformly in t.
Remark 1.25. The point here is that the potential well V is known to be stable
near zero by hypothesis, but could be highly unstable away from zero; see Figure
7. Nevertheless, the bootstrap argument can be used to prevent the solution from
tunnelling from the stable region to the unstable region.

Proof. Applying the Picard theorem (converting the second-order ODE into
a first-order ODE in the usual manner) we see that there is a maximal interval
of existence I = (T , T+ ) containing 0, which supports a unique classical solution
2
u Cloc (I D) to the Cauchy problem (1.23). Also, if T+ is finite, then we have
limtT+ ku(t)kD + kt u(t)kD = , and similarly if T is finite.
For any time t I, let E(t) denote the energy
1
(1.24) E(t) := kt u(t)k2D + V (u(t)).
2
From (1.23) we see that
t E(t) = ht u(t), t2 u(t)i + ht u(t), V (u(t))i = 0
and thus we have the conservation law
1
E(t) = E(0) = ku1 k2D + V (u0 ).
2
If u0 , u1 are sufficiently close to 0, we can thus make E(t) = E(0) as small as
desired.
The problem is that we cannot quite conclude from the smallness of E that u
is itself small, because V could turn quite negative away from the origin. However,
such a scenario can only occur when u is large. Thus we need to assume that u is
small in order to prove that u is small. This may seem circular, but fortunately the
bootstrap principle allows one to justify this argument.
Let > 0 be a parameter to be chosen later, and let H(t) denote the statement
kt u(t)k2D + ku(t)k2D (2)2
 1.3. BOOTSTRAP AND CONTINUITY ARGUMENTS 23

V(u)

2 2

Figure 7. The potential well V in Proposition 1.24. As long as

kt u(t)k2D + ku(t)k2D is known to be bounded by (2)2 , the Hamil-
tonian becomes coercive and energy conservation will trap a parti-
cle in the region kt u(t)k2D + ku(t)k2D 2 provided that the initial
energy is sufficiently small. The bootstrap hypothesis can be re-
moved because the motion of the particle is continuous. Without
that bootstrap hypothesis, it is conceivable that a particle could
discontinuously tunnel through the potential well and escape,
without violating conservation of energy.

and let C(t) denote the statement

kt u(t)k2D + ku(t)k2D 2 .
Since u is continuously twice differentiable, and blows up at any finite endpoint of
I, we can easily verify properties (b) and (c) of the bootstrap principle, and if u0
and u1 are sufficiently close to 0 (depending on ) we can also verify (d) at time
t = 0. Now we verify (a), showing that the hypothesis H(t) can be bootstrapped
into the stronger conclusion C(t). If H(t) is true, then ku(t)kD = O(). We then
see from the hypotheses on V and Taylor expansion that
V (u(t)) cku(t)k2D + O(3 )
for some c > 0. Inserting this into (1.24), we conclude
1
kt u(t)k2D + cku(t)k2D E(0) + O(3 ).
2
This is enough to imply the conclusion C(t) as long as is sufficiently small, and
E(0) is also sufficiently small. This closes the bootstrap, and allows us to conclude
24 1. ORDINARY DIFFERENTIAL EQUATIONS

2A

A + F(u)

u forbidden
here
u
A 2A

Figure 8. A depiction of the situation in Exercise 1.21. Note the

impenetrable barrier in the middle of the u domain.

that C(t) is true for all t I. In particular, I must be infinite, since we know
that kt u(t)k2D + ku(t)k2D would blow up at any finite endpoint of I, and we are
done. 

One can think of the bootstrap argument here as placing an impenetrable

barrier
2 < kt u(t)k2D + ku(t)k2D (2)2
in phase space. Property (a) asserts that the system cannot venture into this
barrier. Properties (b), (c) ensure that this system cannot jump from one side
of the barrier to the other instantaneously. Property (d) ensures that the system
starts out on the good side of the barrier. We can then conclude that the system
stays in the good side for all time; see Figure 7. Note also the division of labour
in proving these properties. The properties (b), (c) are proven using the local
existence theory (i.e. Theorem 1.17). The property (d) comes from the hypotheses
on the initial datum. The property (a) requires some structural information on the
equation, in this case the existence of a conserved energy E(t) with enough locally
coercive properties to contain the system within the desired barrier. This pattern
of argument is very common in the analysis of nonlinear ODE and PDE, and we
shall see several more examples of this later in this monograph.
Exercise 1.20. Show by example that Proposition 1.21 fails if any one of its
four hypotheses are removed.

0
Exercise 1.21. Let I be a time interval, and u Cloc (I R+ ) be a non-
negative function obeying the inequality
(1.25) u(t) A + F (u(t))
 1.3. BOOTSTRAP AND CONTINUITY ARGUMENTS 25

for some A, > 0 and some function F : R+ R+ which is locally bounded.

Suppose also that u(t0 ) 2A for some t0 I. If is sufficiently small depending
on A and F , show that in fact u(t) 2A for all t I. Show that the conclusion
can fail if u is not continuous or is not small. Note however that no assumption
is made on the growth of F at infinity. Informally speaking, this means that if
one ever obtains an estimate of the form u A + F (u), then one can drop the
F (u) term (at the cost of increasing the main term A by a factor of 2) provided
that is suitably small, some initial condition is verified, and some continuity is
available. This is particularly useful for showing that a nonlinear solution obeys
almost the same estimates as a linear solution if the nonlinear effect is sufficiently
weak. Compare this with Principle 1.23.
0
Exercise 1.22. Let I be a time interval, and let u Cloc (I R+ ) obey the
inequality
u(t) A + F (u(t)) + Bu(t)
for some A, B, > 0 and 0 < < 1, and some locally bounded function F : R+
R+ . Suppose also that u(t0 ) A for some t0 I and A > 0. Show that if is
sufficiently small depending on A, A , B, , F , then we have u(t) . A+B 1/(1) for
all t I. Thus we can tolerate an additional u-dependent term on the right-hand
side of (1.25) as long as it grows slower than linearly in u.
Exercise 1.23 (Compactness solutions). Let t0 R be a time, let u0 D,
and let F : D D be a function which is continuous (and hence bounded) in a
neighbourhood of u0 . Show that there exists an open time interval I containing
t0 , and a classical solution u C 1 (I D) to the Cauchy problem (1.7). (Hint:
approximate F by a sequence of Lipschitz functions Fm and apply Theorem 1.17
to obtain solutions um to the problem t um = Fm (um ) on some maximal interval
(T,m , T+,m ). Use a bootstrap argument and Gronwalls inequality to show that
for some fixed open interval I (independent of m) containing t0 , the solutions um
will stay uniformly bounded and uniformly Lipschitz (hence equicontinuous) in this
interval, and that this interval is contained inside all of the (T,m , T+,m ). Then
apply the Arzela-Ascoli theorem to extract a uniformly convergent subsequence of
the u on I, and see what happens to the integral equations um (t) = um (t0 ) +
Rt m
F (u (s)) ds in the limit, using Lemma 1.3 if necessary.) This is a simple
t0 m m
example of a compactness method to construct solutions to equations such as (1.13),
for which uniqueness is not available.
Exercise 1.24 (Persistence of constraints, II). Let u : [t0 , t1 ] D be a classical
solution to the ODE t u(t) = F (u(t)) for some continuous F : D D, and let
1
H1 , . . . , Hn Cloc (D R) have the property that
hF (v), dHj (v)i 0
whenever 1 j n and v D is such that Hj (v) = 0 and Hi (v) 0 for all
1 i n. Show that if the statement
Hi (u(t)) 0 for all 1 i n
is true at time t = t0 , then it is true for all times t [t0 , t1 ]. Compare this result
with Exercise 1.13.
Exercise 1.25 (Forced blowup). Let k 1, and let u : [0, T ) R be a
classical solution to the equation tk u(t) = F (u(t)), where F : R R is continuous.
26 1. ORDINARY DIFFERENTIAL EQUATIONS

Suppose that u(0) > 0 and tj u(0) 0 for all 1 j < k, and suppose that one has
the lower bound such that F (v) & v p for all v u(0) and some p > 1. Conclude
the upper bound T .p,k u(0)(1p)/k on the time of existence. (Hint: first establish
that u(t) u(0) and tj u(t) 0 for all 1 j < k and 0 t < T , for instance by
using Exercise 1.24. Then bootstrap these bounds to obtain some estimate on the
doubling time of u, in other words to obtain an upper bound on the first time t for
which u(t) reaches 2u(0).) This shows that equations of the form tk u(t) = F (u(t))
can blow up if the initial datum is sufficiently large and positive.
Exercise 1.26. Use the continuity method to give another proof of Gronwalls
inequality (Theorem 1.10). (Hint: for technical reasons it may be easier to first
Rt
prove that u(t) (1 + )A exp( t0 B(s) ds) for each > 0, as continuity arguments
generally require an epsilon of room.) This alternate proof of Gronwalls inequal-
ity is more robust, as it can handle additional nonlinear terms on the right-hand
side provided that they are suitably small.

1.4. Noethers theorem

Now symmetry and consistency are convertible terms - thus Poetry
and Truth are one. (Edgar Allen Poe, Eureka: A Prose Poem)
A remarkable feature of many important differential equations, especially those
arising from mathematical physics, is that their dynamics, while complex, still con-
tinue to maintain a certain amount of unexpected structure. One of the most
important examples of such structures are conservation laws - certain scalar quan-
tities of the system that remain constant throughout the evolution of the system;
another important example are symmetries of the equation - that there often exists
a rich and explicit group of transformations which necessarily take one solution
of the equation to another. A remarkable result of Emmy Noether shows that
these two structures are in fact very closely related, provided that the differential
equation is Hamiltonian; as we shall see, many interesting nonlinear dispersive and
wave equations will be of this type. Noethers theorem is one of the fundamental
theorems of Hamiltonian mechanics, and has proven to be extremely fruitful in the
analysis of such PDE. Of course, the field of Hamiltonian mechanics offers many
more beautiful mathematical results than just Noethers theorem; it is of great in-
terest to see how much else of this theory (which is still largely confined to ODE)
can be extended to the PDE setting. See [Kuk3] for some further discussion.
Noethers theorem can be phrased symplectically, in the context of Hamiltonian
mechanics, or variationally, in the context of Lagrangian mechanics. We shall opt
to focus almost exclusively on the former; the variational perspective has certain
strengths for the nonlinear PDE we shall analyse (most notably in elucidating the
role of the stress-energy tensor, and of the distinguished role played by ground
states) but we will not pursue it in detail here (though see Exercises 1.44, 1.45,
2.60). We shall content ourselves with describing only a very simple special case of
this theorem; for a discussion of Noethers theorem in full generality, see [Arn].
Hamiltonian mechanics can be defined on any symplectic manifold, but for
simplicity we shall restrict our attention to symplectic vector spaces.
Definition 1.26. A symplectic vector space (D, ) is a finite-dimensional real
vector space D, equipped with a symplectic form : D D R, which is bilinear
and anti-symmetric, and also non-degenerate (so for each non-zero u D there
 1.4. NOETHERS THEOREM 27

1
exists a v D such that (u, v) 6= 0). Given any H Cloc (D R), we define the
0
symplectic gradient H Cloc (D D) to be the unique function such that
d
(1.26) hv, dH(u)i =
H(u + v)|=0 = ( H(u), v);
d
this definition is well-defined thanks to the non-degeneracy of and the finite
1
dimensionality of D. Given two functions H, E Cloc (D R), we define the
Poisson bracket {H, E} : D R by the formula
(1.27) {H, E}(u) := ( H(u), E(u)).
A Hamiltonian function on a phase space (D, ) is any function12 H Cloc
2
(D
R); to each such Hamiltonian, we associate the corresponding Hamiltonian flow
(1.28) t u(t) = H(u(t)).
Note that with this definition, Hamiltonian ODE (1.28) are automatically au-
tonomous (time-translation-invariant). However it is possible to consider time-
varying Hamiltonians also: see Exercise 1.42. Note that the Hamiltonian of an
equation is only determined up to a constant, since replacing H by H + C does not
affect the symplectic gradient of H.
Example 1.27. If
D := Rn Rn = {(q1 , . . . , qn , p1 , . . . , pn ) : q1 , . . . , qn , p1 , . . . , pn R}
for some n 1, and : D D R is the bilinear form
n
X
:= dqj dpj
j=1

or in other words
n
X
((q1 , . . . , qn , p1 , . . . , pn ), (q1 , . . . , qn , p1 , . . . , pn )) := pj qj pj qj
j=1
1
then (D, ) is symplectic, and for any H, E Cloc (D R) we have
H H H H
H = ( ,..., , ,..., )
p1 pn q1 qn
and
Xn
H E H E
{H, E} = .
j=1
p j qj qj pj
In particular, the Hamiltonian ODE associated to a Hamiltonian function H
2
Cloc (D R) is given by Hamiltons equations of motion
H H
(1.29) t qj (t) = (q(t), p(t)); t pj (t) = (q(t), p(t)),
pj qj
where we write
u(t) = (q(t), p(t)); q(t) = (q1 (t), . . . , qn (t)); p(t) = (p1 (t), . . . , pn (t)).

12One can weaken this hypothesis of continuous twice differentiability and still define a
Hamiltonian flow, but the theory becomes more delicate and we will not address it here.
28 1. ORDINARY DIFFERENTIAL EQUATIONS

Thus, for instance, if H takes the form

1
(1.30) H(q, p) = |p|2 + V (q)
2m
where |p|2 := p21 + . . . + p2n , m > 0 is a constant, and V Cloc
2
(Rn R), then
Hamiltons equations become Newtons laws of motion
1
t q(t) = p(t); t p(t) = V (q(t)).
m
Example 1.28. Let D = Cn be endowed with the symplectic form
n
X 1
(1.31) := Im(dzj dzj )
j=1
2

or in other words
n
X
((z1 , . . . , zn ), (z1 , . . . , zn )) := Im(zj zj ).
j=1

1
Then for any H, E Cloc (D R) we have
H H
H = (2i , . . . , 2i )
z1 zn
and
n
X H E
{H, E} = 4 Im( )
j=1
zj zj

where H H H
z := x i y and
H
z := H
x + i H
y . Thus for instance if H is the simple
harmonic oscillator
n
X 1
(1.32) H(z) := j |zj |2
j=1
2

for some 1 , . . . , n R, then the equations of motion are

t zj (t) = ij zj (t).
This is in fact the canonical form for any quadratic Hamiltonian; see Exercise
1.41.
2
Hamiltonian equations enjoy a number of good properties. Since H Cloc (D
0,1
R), the function H Cloc (D D) is locally Lipschitz, and so the Picard
existence theorem applies; in particular, for any bounded set D, there is a
positive time T > 0 for which we have Lipschitz flow maps S(t) = S0 (t) C 0,1 (
D) defined for |t| T . In the quadratic growth case when 2 H is bounded, then
0,1
H is globally Lipschitz, and one thus has global flow maps S(t) Cloc (D D)
defined for all times t R (by Exercise 1.10). These maps obey the group laws
S(t + t ) = S(t)S(t ), S(0) = id. Furthemore the S(t) are diffeomorphisms and
symplectomorphisms; see Exercise 1.4 and Exercise 1.32.
2
Let H, E Cloc (D R) be two Hamiltonians on a symplectic phase space
1
(D, ), and let u Cloc (I D) be a classical solution to the Hamiltonian ODE
 1.4. NOETHERS THEOREM 29

(1.28). From the chain rule and (1.26), (1.27), we thus have the Poisson equation 13
d
(1.33) E(u(t)) = {H, E}(u(t)).
dt
Let us say that a quantity E is conserved by (1.28) if E(u(t)) is constant for
any (classical) solution u : I D of (1.28). From (1.33) and the anti-symmetry
{H, E} = {E, H} of the Poisson bracket, we conclude

Theorem 1.29 (Noethers theorem). Let H and E be two Hamiltonians on a

symplectic phase space (D, ). Then the following are equivalent.
(a) {H, E} = 0.
(b) The quantity E is conserved by the Hamiltonian flow of H.
(c) The quantity H is conserved by the Hamiltonian flow of E.

If any of the above three properties hold, we say that H and E Poisson com-
mute. As stated, Noethers theorem is symmetric in H and E. However, this
theorem is often interpreted in a somewhat asymmetric way. Assume for sake of
argument that the flow maps SE (t) of E are globally defined (this is the case, for in-
stance, if E is quadratic growth). We view the flow maps SE (t) as a one-dimensional
group action on the phase space D. Noethers theorem then asserts that E is a con-
served quantity for the equation (1.28) if and only if H is symmetric (i.e. invariant)
with respect to the group actions SE (t); for a generalisation to higher-dimensional
group actions, see Exercise 1.34. Thus this theorem gives a very satisfactory link
between the symmetries of the Hamiltonian H to the conserved quantities of the
flow (1.28). The larger the group of symmetries, the more conserved quantities one
obtains14.
For instance, since H clearly Poisson commutes with itself, we see that H
itself is a conserved quantity, thus H(u(t0 )) = H(u0 ) for any classical solution
1
u Cloc (I D) to the Cauchy problem
(1.34) t u(t) = H(u(t)); u(t0 ) = u0 .
As another example, if (D, ) is the complex phase space given in Example 1.28,
and the Hamiltonian is invariant under phase rotations, thus
H(ei z1 , . . . , ei zn ) = H(z1 , . . . , zn ) for all z1 , . . . , zn C, R
P
then the total charge nj=1 |zj |2 is conserved by the flow. Indeed, the phase rotation
is (up to a factor of two) nothing more than the Hamiltonian flow associated to the
total charge.
Another class of important examples concerns the phase space (Rd Rd )N of
N particles in Rd , parameterised by N position variables q1 , . . . , qN Rd and N

13This equation is unrelated to the PDE u = , which is sometimes also referred to as

Poissons equation.
14Provided, of course, that the symmetries themselves come from Hamiltonian flows. Certain
symmetries, notably scaling symmetries, are thus difficult to place in this framework, as they
typically violate Louvilles theorem and thus cannot be Hamiltonian flows, though they do tend
to generate almost conserved quantities, such as that arising in the virial identity. Also, discrete
symmetries such as time reversal symmetry or permutation symmetry also are not in the range
of applicability for Noethers theorem.
30 1. ORDINARY DIFFERENTIAL EQUATIONS

Table 1. Some common symmetry groups and their associated

conservation laws (or approximate conservation laws, in the case of
the virial identity). Not all of these follow directly from Noethers
theorem as stated, and are best viewed instead using the La-
grangian approach to this theorem. In some cases, the interpre-
tation of the conserved quantity depends on the equation; for in-
stance spatial translation corresponds to momentum for wave and
Schrodinger equations, but corresponds instead to mass for KdV
type equations.

Symmetry Conserved quantity

time translation energy / Hamiltonian
spatial translation momentum / mass
spatial rotation angular momentum
Galilean transformation (renormalised) centre-of-mass
Lorentz transformation (renormalised) centre-of-energy
scaling (virial identity)
base space diffeomorphism stress-energy
phase rotation mass / probability / charge
gauge transform charge

momentum variables p1 , . . . , pN Rd , with the symplectic form

N
X N X
X d
:= dqj dpj = dqj,i dpj,i .
j=1 j=1 i=1

If a Hamiltonian H(q1 , . . . , qN , p1 , . . . , pN ) is invariant under spatial translations,

thus
H(q1 x, . . . , qN x, p1 , . . . , pN ) = H(q1 , . . . , qN , p1 , . . . , pN )
for all x, p1 , . . . , pN , q1 , . . . , qN Rd , then Noethers theorem implies that the total
P
momentum p = N j=1 pj is conserved by the flow. If the Hamiltonian takes the
form
N
X 1
(1.35) H(q1 , . . . , qN , p1 , . . . , pN ) = mi |pi |2 + V (q1 , . . . , qN )
j=1
2
2
for some (translation invariant) potential V Cloc ((Rd )N R), then the total
momentum takes the familiar form
N
X dqj
p= mj .
j=1
dt

Similarly, if the Hamiltonian is invariant under angular rotations U : Rd Rd ,

thus
H(U q1 , . . . , U qN , U p1 , . . . , U pN ) = H(q1 , . . . , qN , p1 , . . . , pN )
for all p1 , . . . , pN , q1 , . . . , qN Rd and U SO(d), then Noethers theorem (or more
precisely the generalisation in Exercise 1.34) implies that the angular momentum
PN V2 d
L := j=1 qj pj R is also preserved by the flow.
 1.4. NOETHERS THEOREM 31

2
i |z(t)| z(t)

z(t)

H = const

Figure 9. The complex scalar ODE z = i|z|2 z is a Hamiltonian

ODE with the conserved Hamiltonian H(z) := 41 |z|4 . This conser-
vation law coerces the solution z to stay inside a bounded domain,
and hence blowup does not occur. This is in contrast with the
similar-looking ODE z = +|z|2 z, which blows up in finite time
from any non-zero initial datum. Note also the rotation symmetry
of this equation, which by Noethers theorem implies conservation
of |z|2 .

Remark 1.30. Noethers theorem connects exact (Hamiltonian) symmetries

with exact conservation laws. There are a number of generalisations (both rig-
orous and informal) to this theorem. In particular, we expect approximate or
non-Hamiltonian symmetries to be related to approximate conservation laws. One
important instance of this heuristic involves conformal Killing vector fields, which
can be viewed as approximate symmetries of the underlying geometry; see Section
2.5.

Exercise 1.27. Let D be a real Hilbert space, and let J End(D) be a linear
map such that J 2 = id. Show that the bilinear form : D D R defined
by (u, v) := hu, Jvi is a symplectic form, and that H = JH (where is
the gradient with respect to the Hilbert space structure). This is the constant-
coefficient version of a more general fact, that a symplectic form can be combined
with an almost complex structure J to produce a Riemannian metric; this fact is
fundamental to the theory of symplectic topology, which is far beyond the scope of
this text (though see Section 4.3).
Exercise 1.28 (Linear Darboux theorem). Show that any symplectic space
(D, ) is equivalent, after a linear change of variables, to the standard symplectic
space in Example 1.27; in particular symplectic spaces are always even-dimensional.
32 1. ORDINARY DIFFERENTIAL EQUATIONS

(Hint: induct on the dimension of D. If the dimension is non-zero, use the non-
degeneracy of to locate two linearly independent vectors u, v D such that
(u, v) 6= 0. Then restrict to the symplectic complement {w D : (u, w) =
(v, w) = 0} and use the induction hypothesis.) Note that this change of variables
will usually not be unique. Conclude in particular that every symplectic phase
space has an even number of dimensions.
2
Exercise 1.29. Show that if H Cloc (D R) is a Hamiltonian which has a
non-degenerate local minimum at some u0 D (thus H(u0 ) = 0 and 2 H(u0 ) is
strictly positive definite), then one has global solutions to the associated Hamilton-
ian equation as soon as the initial datum u0 is sufficiently close to u0 . Note that
this generalises Proposition 1.24; indeed, one can proceed by a modification of the
proof of that proposition. Similarly, show that if H is a Hamiltonian which is glob-
ally coercive in the sense that limv |H(v)| = , then one has global solutions
to the associated Hamiltonian equation for arbitrary initial data.
Exercise 1.30. Show that if one applies the time reversal change of variable
t 7 t to a Hamiltonian equation, one obtains another Hamiltonian equation;
what is the new Hamiltonian?
Exercise 1.31. Let I be a time interval, and let (D, ), (D , ) be symplec-
1
tic vector spaces. Let u Cloc (I D) solve a Hamiltonian equation t u(t) =
H(u(t)) for some Hamiltonian H Cloc 2
(D R), and let u Cloc 1
(I D )
solve a Hamiltonian equation t u (t) = H (u (t)) for some Hamiltonian H

2
Cloc (D R). Show that the combined system (u, u ) Cloc
1
(I D D ) solves a
Hamiltonian equation on D D , with an appropriate symplectic form and

a Hamiltonian H H . This shows that a system of many non-interacting particles

is automatically Hamiltonian if each component particle evolves in a Hamiltonian
manner.
Exercise 1.32 (Preservation of symplectic form). Let (D, ) be a symplectic
2 2
space, let H Cloc (D R) be a Hamiltonian, and let u Cloc (R R R D)
be such that for each x, y R, the function t 7 u(t, x, y) solves the Hamiltonian
equation t u(t, x, y) = H(u(t, x, y)). Show that for each x, y R, the quantity
(x u(t, x, y), y u(t, x, y)) is conserved in time. Conclude in the quadratic growth
case (with 2 H bounded) that the solution maps S(t) are symplectomorphisms
(they preserve the symplectic form ). There is a local converse, namely that
any smooth one-parameter group of symplectomorphisms must locally arise from a
Hamiltonian, but we will not detail this here.
Exercise 1.33 (Liouvilles theorem). Let (D, ) be a symplectic space, and
let dm be a Haar measure on D. (One can define a canonical Haar measure,
namely Louville measure, by setting m := dim(D)/2 .) Let H Cloc
2
(D R) be a
Hamiltonian, and let be any open bounded set in D, thus we have a solution map
0
S(t) Cloc ( D) for any sufficiently small t. Show that S(t)() has the same
m-measure as . (Hint: use Exercise 1.32.) More generally, show that the previous
claim is true if we replace dm by the (non-normalised) Gibbs measure d :=
eH dm for any R. This constructs for us a small family of invariant measures
for the Hamiltonian flow; a major (and rather difficult) problem in the field is
to construct similar invariant measures for Hamiltonian PDE, and to investigate
to what extent these are the only invariant measures available. See for instance
[Kuk3], [Bou4].
 1.4. NOETHERS THEOREM 33

Exercise 1.34 (Moment maps). Let G be a finite-dimensional Lie group acting

(on the left) on a symplectic phase space (D, ), let g be the Lie algebra and let
g be the dual Lie algebra. We identify each Lie algebra element x g with a
vector field Xx on D in the obvious manner. Suppose we have a moment map
2
Cloc (D g ), in other words a map with the property that
Xx (u) = hx, (u)i for all u D, x g.
(For instance, if G is the additive real line R, then the group action is simply
the Hamiltonian flow maps S(t) associated to the Hamiltonian .) Show that if
2
H Cloc (D R) is a Hamiltonian which is G-invariant (thus H(gu) = H(u) for
all u D, g G), then is conserved by the Hamiltonian flow of H. Show that the
converse is also true if G is connected. Use this generalisation of Noethers theorem
to verify the claims concerning conservation of momentum and angular momentum
made above.
2
Exercise 1.35. If H1 , H2 , H3 Cloc (D R) are three Hamiltonians, verify
the Jacobi identity {H1 , {H2 , H3 }} + {H2 , {H3 , H1 }} + {H3, {H1 , H2 }} = 0 and the
Leibnitz rule
(1.36) {H1 , H2 H3 } = {H1 , H2 }H3 + H2 {H1 , H3 }
1
Exercise 1.36 (Poisson bracket vs. Lie bracket). If H Cloc (D R), let DH

denote the differential operator defined by DH E := {H, E} for E Cloc (D R).
d
Thus for instance, (1.33) asserts that dt E(u) = DH E(u) for all E Cloc (D R)
2
and all solutions u to (1.28). Show that for any Hamiltonians H1 , H2 Cloc (D
R), that [DH1 , DH2 ] = D{H1 ,H2 } , where [A, B] := AB BA denotes the Lie bracket
of A and B. This is the classical relationship between the Poisson and Lie brackets;
see Exercise 2.15 for some discussion of the quantum relationship between these
brackets also.
1
Exercise 1.37. A function E Cloc (D R) is said to be an integral of motion
of an ODE t u(t) = F (u(t)) if there is a function G : Cloc0
(D D ) assigning a
linear functional G(u) End(D R) to each u D, such that we have the identity
Z t1
E(u(t1 )) E(u(t0 )) = G(u)(t u(t) F (u(t))) dt
t0
1
for all time intervals [t0 , t1 ] and all functions u Cloc ([t0 , t1 ] D) (which may
or may not solve the ODE). Show that a Hamiltonian function E is an integral
of motion for a Hamiltonian ODE t u(t) = H(u(t)) if and only if E Poisson
commutes with H.
2
Exercise 1.38. Let H Cloc (D R) be a Hamiltonian. Show that the space
of all Hamiltonians E which Poisson commute with H form an algebra (thus the
space is a vector space and is also closed under pointwise multiplication), and is
2
also closed under all change of variable maps E 7 E for any Cloc (R R).
(In fact, these claims are valid for the space of integrals of motion for any first-order
ODE, not just the Hamiltonian ones.)
2
Exercise 1.39. Let H, E Cloc (D R) be two quadratic growth Hamil-
2 2
tonian functions (so H, E are bounded), and let SH (t) and SE (s) be the
associated flow maps for t, s R. Show that H, E Poisson commute if and only if
SH (t) and SE (s) commute for all t, s R. (Hint: use Exercise 1.14.)
34 1. ORDINARY DIFFERENTIAL EQUATIONS

2
Exercise 1.40. Let H Cloc (D R). Show that the flow maps SH (t) : D
D are linear for all times t if and only if H is a quadratic form, plus a constant.
Exercise 1.41 (Symplectic normal forms). Let (D, ) be a 2n-dimensional
symplectic vector space, and let H : D R+ be a positive definite quadratic form
on D. Show that there exists real numbers 1 . . . n > 0 and linear coordinate
functions z1 , . . . , zn End(D C) such that takes the form (1.31) and H
takes the form (1.32). (Hint: choose a real coordinate system on D (identifying it
with R2n ) so that H is just the standard Euclidean form H(x) = |x|2 . Then the
symplectic form is given by (x, y) = xJy for some anti-symmetric non-degenerate
real-valued 2n 2n matrix J. Analyze the eigenspaces and eigenvalues of J and use
this to construct the complex coordinates z1 , . . . , zd .) Conclude in particular that
the ellipsoid {z D : H(z) = 1} contains n periodic orbits for the Hamiltonian
flow with periods 2/1 , . . . , 2/n respectively. We refer to 1 , . . . , n as the
frequencies of the Hamiltonian H. One can devise analogues of this transformation
for more general Hamiltonians (which contain higher order terms in addition to a
quadratic component), leading to the theory of Birkhoff normal forms, which we
will not discuss here.
1
Exercise 1.42. Let (D, ) be a symplectic space, let H Cloc (R D R),
and consider the time-varying Hamiltonian equation
t u(t) = H(t, u(t)).
Show that it is possible to convert this time-varying Hamiltonian equation into a
time-independent equation on a symplectic vector space R2 D, by a trick similar
to that mentioned in Section 1.1. Does this equation preserve the Hamiltonian
H(t, u(t))? If not, is there some substitute quantity which is preserved?
Exercise 1.43. Let D = (Rd Rd )N be the phase space of N particles in d
dimensions. Suppose that a Hamiltonian equation is invariant under the Galilean
symmetry
(q1 , . . . , qN , p1 , . . . , pN ) 7 (q1 vt, . . . , qN vt, p1 m1 v, . . . , pN mN v)
for any v Rd and some fixed m1 , . . . , mN > 0, in the sense that whenever the
function
t 7 (q1 (t), . . . , qN (t), p1 (t), . . . , pN (t))
solves the Hamiltonian ODE, then so does the transformed function
t 7 (q1 vt, . . . , qN vt, p1 m1 v, . . . , pN mN v).
Conclude that the normalised centre of mass
N
X N
X
mj qj t pj
j=1 j=1

is an invariant of the flow. (Hint: convert t into another phase space variable as in
Exercise 1.42, so that Noethers theorem can be applied.)
Exercise 1.44 (Connection between Hamiltonian and Lagrangian mechanics,
I). Let (D, ) be the standard symplectic phase space in Example 1.27, and let
 1.5. MONOTONICITY FORMULAE 35

L Cloc (Rn Rn R); we use q1 , . . . , qn , q1 , . . . , qn to denote the variables of L.
Define the momentum coordinates p1 , . . . , pn by
L
(1.37) pj := (q, q)
qj
and assume that the coordinates (q1 , . . . , qn , q1 , . . . , qn ) and (q1 , . . . , qn , p1 , . . . , pn )

are diffeomorphic to each other. We then define the Hamiltonian H Cloc (D R)
by
H(q, p) := q p L(q, q)
where q is defined implicitly by (1.37). Show that if I is a bounded interval and
q C (I Rn ), then q is a formal critical point for the Lagrangian
Z
S(q) := L(q(t), t q(t)) dt
I
with endpoints held fixed, if and only if (q(t), p(t)) solves the Hamiltonian ODE
(1.29). (You may wish to first work with the n = 1 case and with the example
(1.30) to get some intuition.)
Exercise 1.45 (Connection between Hamiltonian and Lagrangian mechanics,
II). Let (D, ) be the standard symplectic phase space in Example 1.27, and let

H Cloc (D R) be a Hamiltonian phase function. Let I be a bounded time
interval. Show that if q, p C (I Rn ) obey the constraint
H
(1.38) t qj (t) = (q(t), p(t))
pj
(which can be viewed as an implicit definition of the momentum p(t) in terms of the
position q(t) and the velocity t q(t), at least if H is sufficiently non-degenerate),
then q and p obey the Hamiltonian ODE (1.29), if and only if q and p are formal
critical points of the Lagrangian
Z
S(q, p) := (t q(t)) p(t) H(q(t), p(t)) dt
I
subject to the constraint (1.38) and also fixing the values of q(t) and p(t) at the
endpoints. Explain why this connection is essentially the inverse of that in the
preceding exercise.

1.5. Monotonicity formulae

If something cannot go on forever, it will stop. (Herbert Stein)
As we have already seen, conservation laws (such as conservation of the Hamil-
tonian) can be very useful for obtaining long-time existence and bounds for solutions
to ODE. A very useful variant of a conservation law is that of a monotonicity for-
mula - a quantity G(u(t), t) depending on the solution u(t), and perhaps on the time
t, which is always monotone increasing in time t, or perhaps monotone decreasing
in time t. These monotone quantities can be used to obtain long-time control of a
solution in several ways. For instance, if a quantity G(u(t), t) is large at some initial
time t0 and is monotone increasing, then clearly it will stay large for all later times
t > t0 ; conversely, if G(u(t), t) is bounded at time t0 , is monotone decreasing, and
is manifestly non-negative, then it will stay bounded for all later times t > t0 . If G
is monotone increasing, and is itself the time derivative of another quantity K(t),
then we also learn that K(t) is convex in time, which can be useful in a number of
36 1. ORDINARY DIFFERENTIAL EQUATIONS

ways. Finally, if one knows that G(u(t), t) is bounded uniformly in time (e.g. by
using conservation laws), and is monotone, then we conclude from the fundamental
theorem of calculus that the derivative t G(u(t), t) is absolutely integrable in time,
and thus decays to zero as t , at least in some averaged sense. This type of
long-time decay is especially useful for understanding the asymptotic behaviour of
the solution.
We will be most interested in monotonicity formulae in the setting of PDE.
However, we can present some simple ODE analogues of some of the more common
monotonicity formulae here, which may help motivate the otherwise miraculous-
seeming formulae which we will encounter in later chapters.
Unlike conservation laws, which can be systematically generated from symme-
tries via Noethers theorem, we do not have a fully automated way for producing
monotone or convex quantities other than trial and error (for instance by starting
with a conserved quantity such as energy or momentum and perturbing it some-
how to be monotone instead of conserved), although certain tactics (e.g. exploiting
conformal Killing fields, see Section 2.5) have proven to be generally quite fruit-
ful. Thus we shall content ourselves in this section by presenting some illustrative
examples of monotonicity formulae for ODE, each of which has an extension to a
PDE such as the nonlinear Schrodinger equation.
2
Example 1.31 (Virial identity). Let V Cloc (Rd R) be a twice contin-
uously differentiable potential, and consider a classical solution x : I Rd to
Newtons equations of motion

(1.39) t2 x(t) = V (x(t)).

Then we observe the virial identity

t2 (|x(t)|2 ) = 2t (x(t) t x(t))

= 2|t x(t)|2 + 2x(t) t2 x(t)
= 2|t x(t)|2 2x(t) V (x(t)).

This has a number of consequences. If V is radially decreasing, so that xV (x) 0

for all x Rd , then we thus conclude that |x(t)|2 is convex. If instead we have a
bound of the form
x V (x) CV (x)
for some C 2, then we can obtain the lower bound

t2 (|x(t)|2 ) 2CE

where E is the conserved energy

1
(1.40) E = E(t) = |t x(t)|2 + V (x(t)).
2
Thus |x(t)|2 is now strictly convex when the energy is positive. At a heuristic
level, we thus see that positive energy tends to repel solutions from the origin,
whereas negative energy tends to focus solutions towards the origin. For another
application, see Exercise 1.48. For the linear and nonlinear Schrodinger analogues
of these estimates, see (2.38), (3.72).
 1.5. MONOTONICITY FORMULAE 37

Example 1.32 (Morawetz identity). We continue the previous example. A

variant of the virial identity is the Morawetz identity
x(t)
t2 |x(t)| = t ( t x(t))
|x(t)|
|t x(t)|2 (x(t) t x(t))2 x(t)
= + 2 x(t)
|x(t)| |x(t)|3 |x(t)| t
|x(t) (t x(t))|2 x(t) V (x(t))
=
|x(t)| |x(t)|
x x
whenever x(t) 6= 0, where x (v) := v |x| h |x| , vi is the projection of a vector v to
the orthogonal complement of x. Now suppose that V is radially decreasing and
x(t)
non-negative, then the above identity shows that the quantity |x(t)| t x(t), which
measures the radial component of the velocity, is monotone increasing (and that
|x(t)| is convex). This is intuitively plausible; particles that move towards the origin
must eventually move away from the origin, but not vice versa, if the potential is
repulsive. On the other hand, we have
x(t)
| t x(t)| |t x(t)| 2E
|x(t)|
where the energy E is defined in (1.40). From the fundamental theorem of calculus,
we thus conclude that
Z Z
|x(t) (t x(t))|2 x(t) V (x(t))
(1.41) dt + dt 2 2E,
I |x(t)| I |x(t)|
provided that x does not pass through the origin in the time interval I. (This
latter hypothesis can be removed by limiting arguments; see Exercise 1.46.) If
I = R, this estimate is asserting in particular a certain decay for the angular
component x(t) (t x(t)) of the velocity; that particles following this law of motion
must eventually move in mostly radial directions. For the linear and nonlinear
Schrodinger analogues of these estimates, see (2.40), (3.37).

Example 1.33 (Local smoothing). Again continuing the previous example, one
can obtain a smoother analogue of the Morawetz inequality by replacing the non-
differentiable function |x| by the smoother function hxi := (1 + |x|2 )1/2 . One then
has15
x(t)
t2 hx(t)i = t ( t x(t))
hx(t)i
|t x(t)|2 (x(t) t x(t))2 x(t)
= + 2 x(t)
hx(t)i hx(t)i3 hx(t)i t
|t x(t)|2 |x(t)|2 |t x(t)|2 (x(t) t x(t))2 x(t) V (x(t))
= 3
+ .
hx(t)i hx(t)i3 hx(t)i
This time there is no need to exclude the case when x(t) = 0. In particular, if V is
radially decreasing and non-negative, we conclude that
x(t) |t x(t)|2
t ( t x(t))
hx(t)i hx(t)i3
15One can in fact deduce this new identity from the previous one by adding an extra dimension
to the state space Rd , and replacing x by (x, 1); we omit the details.
38 1. ORDINARY DIFFERENTIAL EQUATIONS

x(3)

x(2)

x(1)

Figure 10. A particle passing by the origin, encountering a repul-

sive force, will convert its ingoing momentum to outgoing momen-
tum. Since there is no way to convert outgoing momentum back
to ingoing momentum, we conclude that if the total energy (and
hence momentum) is bounded, then the particle cannot move past
the origin for extended periods of time. Note that this diagram is
slightly different from the one in Figure 1 because the equation is
second-order rather than first-order in time; the position controls
the acceleration rather than the velocity.

and hence by using the fundamental theorem of calculus we obtain the local smooth-
ing estimate
Z
|t x(t)|2
3
dt CE 1/2
I hx(t)i
for some absolute constant C > 0. This result is perhaps a little surprising, since
E 1/2 only seems to control the speed |t x(t)|, as opposed to the square of the speed
1
|t x(t)|2 . Intuitively, the reason for this is the localisation factor hx(t)i 3 , combined

with the integration in time. When the particle x(t) is travelling at very high
speeds, then |t x(t)|2 is much larger than |t x(t)|, but to compensate for this, the
1
particle only lives near the origin (where the localisation factor hx(t)i 3 is large)

for a brief time. In Section 2.4, we shall quantise this estimate to Schrodinger
equations; the ability to upgrade the speed to the square of the speed will become
a smoothing effect of half a derivative, which may help explain the terminology
local smoothing.

Example 1.34 (Interaction Morawetz). Consider an N -particle system of non-

interacting particles with masses m1 , . . . , mN > 0, with the classical solution x :
I (Rd )N given by Newton first law
mi t2 xi (t) = 0 for i = 1, . . . , N.
 1.5. MONOTONICITY FORMULAE 39

x (2) x i(2)
j

x i (1) x (1)
j

Figure 11. When two particles collide (i.e. pass through each
other), their mutual ingoing momentum is converted to mutual
outgoing momentum. As there is no mechanism to convert mutual
outgoing momentum back into mutual ingoing momentum, we thus
see that the total number of collisions (weighted by their mass and
relative velocity) is controlled by the total momentum.

It is easily verified that this system has a conserved energy

X1
E := mi |t xi (t)|2
i
2
P
and one trivially also has a conserved mass M := i mi . Let define a collision
to be a triplet (i, j, t) where t is a time and 1 i < j N are indices such that
xi (t) = xj (t). Let us make the assumption that only finitely many collisions occur
in the time interval I. If t is not one of the times where a collision occurs, we can
define the interaction momentum
X X xi (t) xj (t)
P (t) := mi mj t (xi (t) xj (t));
|xi (t) xj (t)|
1i<jN

roughly speaking, this measures how much the particles are receding from each
other. A computation shows that
X X |xi (t)xj (t) t (xi (t) xj (t))|2
t P (t) = mi mj 0
|xi (t) xj (t)|
1i<jN

when t is not a collision time. Each collision (i, j, t) causes a jump in P (t) by
2mi mj |t (xi (t) xj (t))|, thus P is monotone increasing. Using the crude bound
X X
|P (t)| mi mj |t xj (t)| M 2M E
i j

from Cauchy-Schwarz, we thus conclude the interaction Morawetz inequality

X
2mi mj |t (xi (t) xj (t))| 2M 2M E
(i,j,t)

where (i, j, t) runs over all collisions. There is a related (though not completely
analogous) inequality for the nonlinear Schrodinger equation; see (3.42).
40 1. ORDINARY DIFFERENTIAL EQUATIONS

Exercise 1.46. Let V : Rd R be twice continuously differentiable, radially

decreasing (so in particular V (0) = 0), and non-negative, and let u0 Rd . Show
that there is a unique global solution u : R Rd to (1.39) with initial datum
u(0) = u0 . Also, show that if x(t) is not identically zero, then x(t) can equal zero
for at most one time t0 R, and in such a case we can refine (1.41) to
Z Z
|x(t) (t x(t))|2 x(t) V (x(t))
dt + dt + 2|t x(t0 )| 2 2E.
R |x(t)| R |x(t)|
Exercise 1.47. With the same hypotheses as Exercise 1.46, show that for each
> 0 we have the estimate
Z
|t x(t)|2
1+
dt . E 1/2 .
R hx(t)i
This improves upon Example 1.33, which dealt with the case = 2. (Hint: use the
monotonicity formulae already established, as well as some new formulae obtained
by considering derivatives of expressions such as hx(t)i1 (x(t) t x(t)).) Show
that the estimate fails at the endpoint = 0, even when V = 0 (compare this with
(1.41)).
2
Exercise 1.48 (Virial identity). Suppose that x1 , . . . , xN Cloc (R Rd ) are
solutions to the system of ODE
X mi mj xj (t) xi (t)
mj t2 xj (t) = G 2
|xi (t) xj (t)| |xj (t) xi (t)|
i6=j

where the masses m1 , . . . , mN are positive, and G > 0 is an absolute constant; this
models the behaviour of N particles under Newtonian gravity. Assume that the xj
are all uniformly bounded in time and that |xi xj | never vanishes for any i 6= j.
Suppose also that the average kinetic and potential energies
Z T X N
1 1
K := lim mj |t xj (t)|2 dt;
T + 2T T 2
j=1
Z T X
1 Gmi mj
V := lim dt
T + 2T T |xi (t) xj (t)|
1i<jN

exist. Conclude the virial identity K = 21 T . (Hint: look at Example 1.31.) This
identity is of importance in astrophysics, as it allows one to infer the potential
energy (and hence the possible existence of dark matter) from measurements of the
kinetic energy (e.g. via red-shift measurements).

1.6. Linear and semilinear equations

Mathematics would certainly have not come into existence if one
had known from the beginning that there was in nature no exactly
straight line, no actual circle, no absolute magnitude. (Friedrich
Nietzsche, Human, All Too Human)
Let us now return to the general Cauchy problem (1.7). We shall assume that
we have a special solution, the vacuum solution u(t) 0; in other words, we assume
that F (0) = 0. If F is continuously differentiable, we can then perform a Taylor
expansion
F (u) = Lu + N (u)
 1.6. LINEAR AND SEMILINEAR EQUATIONS 41

where L End(D) is a linear operator, and N : D D vanishes faster than linearly

at 0, thus
N (kukD )
(1.42) lim = 0.
kukD 0 kukD
We shall refer to L as the linear component of F , and N as the genuinely nonlinear
component. Thus we can write our quasilinear equation as a semilinear equation 16
(1.43) t u(t) Lu(t) = N (u(t)).
If N = 0, we say that the equation is linear, otherwise it is nonlinear. In general,
linear equations are much better understood than nonlinear equations, as a vast
array of tools such as linear algebra, spectral theory, Fourier analysis, special func-
tions (explicit solutions), and the principle of superposition can be now invoked to
analyze the equation. A very profitable idea in solving equations such as (1.43) is
to treat the genuine nonlinearity N (u) as negligible, and thus to view the equation
(1.43) as a perturbation of the linear equation
(1.44) t u(t) Lu(t) = 0.
This perturbation strategy is reasonable if u is small (so that N (u), which vanishes
to at least second order, will be very small compared to Lu) or if one is only solving
the equation for very short times (so that the nonlinearity does not have much of
a cumulative influence on the solution). However, when considering large solutions
for long periods of time, the perturbation approach usually needs to be abandoned
in favour more global or non-perturbative techniques such as energy methods
or monotonicity formula methods, although if one establishes a sufficient amount
of decay on the solution in time, then one can often re-instate the perturbation
analysis in the asymptotic limit t , which can be used for instance to obtain a
scattering theory.
In accordance to this perturbation philosophy, let us study (1.43) by first con-
sidering the linear equation (1.44), say with initial datum u(0) = u0 D. (We know
from time translation invariance that the choice of initial time t0 is not particularly
relevant.) Then there is a unique global solution to (1.44), given by
X n
t n
u(t) = etL u0 = L u0 ;
n=0
n!
the finite dimensionality of D ensures that L is bounded, and hence this sum is
absolutely convergent. We refer to the linear operators etL End(D) as the linear
propagators; observe that we have the group law etL esL = e(t+s)L with e0L = id.
In particular, if u0 is an eigenvector of L, thus Lu0 = u0 for some C, then
the unique global solution to (1.44) is given by u(t) = et u0 .
It is thus clear that the eigenvalues of L will play an important role in the
evolution of the equation (1.44). If L has eigenvalues with negative real part, then
the equation is stable or dissipative in the corresponding eigenspaces, displaying
exponential decay in time as t + (but exponential growth as t ).
Conversely, if L has eigenvalues with positive real part, then the equation is unstable
16For ODE, there is little distinction between a quasilinear equation and a semilinear one.
For PDE, one usually requires in a semilinear equation that the genuinely nonlinear part N (u) of
the equation is lower order (i.e. involves fewer spatial derivatives) than the linear part Lu; some
authors require that N (u) contain no derivatives whatsoever.
42 1. ORDINARY DIFFERENTIAL EQUATIONS

Forcing term
N(u) Genuine nonlinearity
Linear evolution N
(inhomogeneous)
D

Initial datum Solution

u u
0 Linear evolution
(homogeneous)
u > ulin
0
Figure 12. The Duhamel formulation of a semilinear ODE, re-
lating the initial datum u0 , the solution u(t), and the nonlinearity
N (u). Again, compare with Figure 2.

or anti-dissipative in the corresponding eigenspaces, exhibiting exponential growth

as t +. We will be concerned primarily with the dispersive case, in between
the stable and unstable modes, in which the eigenvalues are all purely imaginary;
in particular, we will usually consider the case when L is skew-adjoint with respect
to a suitable Hilbert space structure on D. In such cases, we see from the spectral
theorem that there exists an orthogonal Fourier basis (e ) of D, with each e
being an eigenvector of L with some imaginary eigenvalue ih():
Le = ih()e .
The function h : R will be referred to as the dispersion relation of L. If we
then define the Fourier coefficients
(1.45) f() := hf, e i
for any f D and , then the solution to (1.44) can be given on the Fourier
side as
(1.46) d
u(t)() = eith() u0 ().
Thus each Fourier mode in (1.44) oscillates independently in time, with the time
oscillation frequency given by the dispersion relation 7 h(). The magnitude
d
|u(t)()| is conserved by the flow, so each Fourier coefficient simply moves in a
circle.
In order to perturb the linear equation (1.44) to the nonlinear equation (1.43),
we need the fundamental identity
Proposition 1.35 (Duhamels formula). Let I be a time interval, let t0 be a
time in I, and let L End(D), u C 1 (I D), f C 0 (I D). Then we have
(1.47) t u(t) Lu(t) = f (t) for all t I
 1.6. LINEAR AND SEMILINEAR EQUATIONS 43

if and only if
Z t
(1.48) u(t) = e(tt0 )L u(t0 ) + e(ts)L f (s) ds for all t I,
t0
Rt R t0
where we adopt the convention that t0
= t
if t < t0 .
Remark 1.36. The case L = 0 is just the fundamental theorem of calculus.
Indeed one can view Duhamels formula as the fundamental theorem of calculus,
twisted (i.e. conjugated) by the linear propagator etL ; this helps explain the simi-
larity between Figure 2 and Figure 12.
Proof. If we make the ansatz17 u(t) = etL v(t) for some v : I D in (1.47),
then (1.47) is equivalent to
t v(t) = etL f (t),
which by the fundamental theorem of calculus is equivalent to
Z t
v(t) = v(t0 ) + esL f (s) ds.
t0

The claim then follows by multiplying both sides by etL and using the group law. 

In view of this proposition, we see that if N : D D is continuous and u is

assumed to be continuous, then the Cauchy problem
t u Lu = N (u); u(0) = u0 ,
is equivalent to the integral equation
Z t
tL
(1.49) u(t) = e u0 + e(ts)L N (u(s)) ds.
0

This should be compared with the solution u(t) = etL u0 of the corresponding linear
problem; thus if we think of N as being small, (1.49) is a quantitative formulation
of the assertion that the nonlinear solution resembles the linear solution. One can
view (1.49) as the strong solution concept for u, adapted to the flow etL of the
linear operator L.
The equation (1.49) is a variant of (1.8), but is a little smarter in that it uses
the more accurate approximation etL u0 to the nonlinear solution u, as opposed to
the somewhat cruder approximation u0 . As a consequence, the error term in (1.49)
tends to be somewhat smaller than that in (1.8), as it involves just the genuinely
nonlinear component N of the nonlinearity. Just as (1.8) can be iterated using the
contraction mapping theorem to obtain the Picard existence theorem, the variant
(1.49) can also be iterated to obtain a variant of the Picard existence theorem,
which can exploit the special properties of the linear propagator etL to give better
bounds on the time of existence. To describe this iteration scheme, let us first work
abstractly, viewing (1.49) as instance of the more general equation
(1.50) u = ulin + DN (u)

17This is of course the technique of integrating factors, which is a special case of the method
of variation of parameters. The choice of ansatz u(t) = etL v(t) is inspired by the fact that one
solves the linear equation (1.44) if and only if v is constant.
44 1. ORDINARY DIFFERENTIAL EQUATIONS

Forcing term
Duhamel operator D N(u) in N Genuine nonlinearity N

Gain 1 / 2C0
Lose C 0

Initial datum Linear solution Solution

u in D u in B/2
0 lin u in B

Figure 13. The iteration scheme for Proposition 1.38. In prac-

tice, the object ulin arises as the linear evolution of some initial
datum u0 , as in Figure 12, though we do not use this in the state-
ment and proof of the Proposition.

where ulin (t) := etL u0 is the linear solution, and D is the Duhamel operator
Z t
DF (t) := e(ts)L F (s) ds.
0
A useful heuristic principle in trying to solve equations of this general abstract type
is
Principle 1.37 (Perturbation principle). If one is working on a time interval
[0, T ] such that DN (u) ulin , then u should evolve on [0, T ] as if it were linear
(in particular, the solution should exist and obey the same type of estimates that
ulin does). If one is working instead on a time interval where DN (u) ulin , one
should expect u to exhibit nonlinear behaviour (which could range from blowup or
excessive growth on one hand, to additional decay on the other, or something in
between such as nontrivial nonlinear oscillation).
This is of course a very vague principle, since terms such as , , or
nonlinear behaviour are not well defined. In practice, DN (u) will tend to be
small compared to ulin if the initial datum u0 is suitably small, or if the time t is
close to 0, so for small data or small times one expects linear-type behaviour. For
large data or large times, perturbation theory does not predict linear behaviour,
and one could now have nonlinear effects such as blowup18. To control solutions
in this regime one generally needs to augment the perturbation theory with other
tools such as conservation laws.
Let us now give a rigorous formulation of the first half of this principle, by using
the following variant of the contraction mapping theorem to construct solutions.
18Note however that it is possible for the nonlinear term to dominate the linear term but
still be able to construct and control solutions. This for instance occurs if there is an energy
cancellation that shows that the nonlinear term, while nominally stronger than the linear term,
is somehow almost orthogonal to the solution in the sense that it does not significantly increase
certain energies; we shall see several examples of this in the text. Thus the solutions will not
stay close to the linear solution but will still be bounded in various norms. In certain defocusing
dissipative settings it is even possible for the nonlinearity to always act to reduce the energy, thus
giving a better behaved solution than the linear equation.
 1.6. LINEAR AND SEMILINEAR EQUATIONS 45

Proposition 1.38 (Abstract iteration argument). Let N , S be two Banach

spaces. Suppose we are given a linear operator D : N S with the bound
(1.51) kDF kS C0 kF kN
for all F N and some C0 > 0, and suppose that we are given a nonlinear operator
N : S N with N (0) = 0, which obeys the Lipschitz bounds
1
(1.52) kN (u) N (v)kN ku vkS
2C0
for all u, v in the ball B := {u S : kukS }, for some > 0. (In other words,
kN kC 0,1 (B N ) 2C1 0 .) Then for all ulin B/2 there exists a unique solution
u B to the equation (1.50), with the map ulin 7 u Lipschitz with constant at
most 2. In particular we have
(1.53) kukS 2kulinkS .
This proposition is established by the arguments used to prove the contraction
mapping principle, and is left as an exercise. The idea of using this type of abstract
Duhamel iteration to tackle nonlinear PDE dates back to [Seg].
Remarks 1.39. Note that we have considerable freedom in selecting the spaces
S and N ; this freedom becomes very important when considering the low-regularity
local wellposedness theory of PDE. The Picard existence argument in Theorem 1.7
corresponds, roughly speaking, to the choice S = N = C 0 (I D), with (1.8)
taking the place of (1.49). There are a number of variations of this iteration scheme;
for instance, instead of measuring the solution u in a single norm S, one sometimes
is in a situation where u is measured both in a smooth norm S and a rough
norm S0 ; the solution may be large in the smooth norm but small in the rough norm.
In such cases it can still be possible to close an iteration argument using estimates
that combine both norms together; this becomes important in the large data theory
and in the persistence of regularity theory. While it is possible to build an abstract
framework for such schemes, the formulation becomes rather complicated, and so
when these types of situations arise (see for instance Proposition 3.11) we shall
simply perform the iteration by hand.
Remark 1.40. As with Remark 1.5, the proof of the above theorem provides
an explicit iteration scheme to construct the desired solution, starting with the
linear iterate u(0) := ulin and then constructing successive Duhamel iterates u(n) :=
ulin + DN (u(n1) ). This scheme often converges better than the one in Remark
1.5, though it is far from the most rapidly convergent scheme (and is usually not
used directly in numerical computations).
We illustrate the iteration method by establishing global existence for linearly
stable nonlinear equations from small data.
Proposition 1.41 (Linear stability implies nonlinear stability). Let D be a
finite-dimensional real Hilbert space, and let L End(D) be a linear operator which
is linearly stable in the sense that19 there exists > 0 such that hLu, ui kuk2D
19In the finite-dimensional case, linear stability is equivalent to the spectrum of L being
contained entirely in the interior of the left half-plane. In the infinite-dimensional case, the
relationship between stability and spectrum is more delicate, especially if L fails to be normal or
self-adjoint. Indeed, for PDE, nonlinear stability is often significantly more difficult to establish
than linear stability.
46 1. ORDINARY DIFFERENTIAL EQUATIONS

u(t)
N(u(t))

L u(t)

Figure 14. Proposition 1.41 from the vector field perspective of

Figure 1. If u is sufficiently small, then the dissipative effect of
the linear term Lu will dominate the effect of the nonlinearity
N (u), regardless of the orientation of N (u), causing u to decay
exponentially towards the origin.

2
for all u D. Let N Cloc (D D) be a function which vanishes to more than
first order at the origin in the sense of (1.42) (in fact, the C 2 hypothesis shows
that N will vanish to second order). If u0 D is sufficiently close to the origin,
there exists a unique classical solution u : [0, +) D to (1.43) with initial datum
u(0) = u0 , and furthermore there is an estimate of the form
(1.54) ku(t)kD 2et ku0 kD .
Proof. The uniqueness of u follows from the Picard uniqueness theorem, so
it suffices to establish existence, as well as the estimate (1.54). A simple Gronwall
argument (see Exercise 1.54) gives the dissipative estimate
(1.55) ketL u0 kD et ku0 kD
for all u0 D and t 0. Let us now define the spaces S to be the space of all
functions u C 0 ([0, +) D) whose norm
kukS := sup et ku(t)kD
t0

is finite, and let N be the space of all functions F C 0 ([0, +) D) then

kukN := sup et ku(t)kD
t0

is finite. Thus if we set ulin (t) := etL u0 then kulin kS ku0 kD . Next, observe from
(1.42) that N (0) = 0, and so by Taylor expansion and the hypothesis that N is
C 2 , we have N (u) = ON (kukD ) for kukD sufficiently small (the implied constant
depends of course on N ). From the fundamental theorem of calculus this implies
that
kN (u) N (v)kD N ku vkD (kukD + kvkD )
 1.6. LINEAR AND SEMILINEAR EQUATIONS 47

whenever kukD , kvkD are sufficiently small. In particular, if kukS , kvkS for
some sufficiently small , then we have
kN (u(t)) N (v(t))kD N et ku vkS et
and hence (if is sufficiently small)
1
kN (u) N (v)kN ku vkS .
2
whenever kukS , kvkS . Also, from the triangle inequality and (1.55) we have
1
kDF kS kF kN

where D is the Duhamel operator. From Proposition 1.38 we thus see that if
ku0 kD /2, then can thus construct a solution u to (1.47) with kukS 2kulinkS
2ku0kD , and the claim follows. 
2
Remark 1.42. If one merely assumes N to be Lipschitz rather than Cloc ,
then one can obtain some weaker decay estimates, for instance given any <

one can obtain an estimate ku(t)kD 2e t ku0 kD if ku0 kD is sufficiently small
depending on N and , . This can be achieved for instance by modifying the
Gronwall inequality argument in Exercise 1.54. We leave the details to the reader
as an exercise. Thus for dissipative equations one often has the luxury of conceding
an exponential factor in the time variable t, which is very helpful for keeping the
iteration under control. When we turn to nonlinear dispersive equations, no such
exponential factors are available for us to concede (although in some cases the
dispersion will give us polynomial type decay in t, which can be quite useful), and
we must proceed with a bit more care.
An important special case of the general equation (1.43) occurs when the gen-
uinely nonlinear component N is k-linear for some k 2, in the sense that
N (u) = Nk (u, u, . . . , u)
k
where Nk : X X is a function which is (real-)linear in each of the k variables.
In the k = 2 case we call N bilinear or quadratic, in the k = 3 case we call N
trilinear or cubic, and so forth. The condition k 2 is essentially forced upon us
by the condition (1.42). In these cases, the hypothesis (1.52) will hold for small
provided that N is bounded from S to N ; see Exercise 1.51.
When the nonlinearity is k-linear and the linear term L is skew-adjoint, one
can view the evolution (1.43) in terms of frequency interactions (using the Fourier
transform (1.45)). We illustrate this in the k = 2 case N (u) = N2 (u, u); to simplify
the exposition, we will also assume that N2 is not only real linear but is in fact
complex linear. The situation for antilinear nonlinearities and for higher orders
k > 2 requires some simple modifications to the discussion below which are left to
the reader. If we take Fourier transforms of (1.49) (using (1.46)), we obtain
Z t X ,
d
u(t)() = eith() u0 () + ei(ts)h() d 1 )u(s)(
c1 2 u(s)( d 2 ) ds
0 1 ,2

where c1 ,2 is the structure constant

c1 ,2 := hN2 (e1 , e2 ), e i.
48 1. ORDINARY DIFFERENTIAL EQUATIONS

Typically, the structure constants will usually be zero; given any mode , only a
few pairs of modes 1 , 2 can interact to excite that mode; a typical constraint
in order for c1 ,2 to be non-zero is of the form = 1 + 2 (where one places
some group structure on the space of frequencies). Making the renormalisation
d
u(t)() := eith() a (t), which is suggested to us by the variation of parameters
method, we obtain an integral system of equations for the functions (a (t)) :
Z t
(1.56) a (t) = u0 () + eis(h(1 )+h(2 )h()) c1 ,2 a1 (s)a2 (s) ds.
0
Thus each a (t) is initially set equal to u0 (), but as time evolves, the -modes
a (t) is influenced by the bilinear interactions of the pairs of modes a1 (t), a2 (t)
that can excite the -mode. The resonance function h(1 ) + h(2 ) h() plays a
key role in the analysis. If this quantity is large in magnitude, then the integral in
(1.56) will be highly oscillatory, and thus likely to be rather small; in this case, we
say that the interaction between the modes 1 , 2 , is non-resonant. The dominant
contribution to (1.56) typically consists instead of the resonant interactions, in
which the resonance function is zero or small. In order to obtain an iterative
scheme for solving this equation (using for instance Proposition 1.38), especially at
low regularities, one often has to spend some effort to control the resonant portions
of interaction, either by showing that the resonant interactions are fairly rare, or
by extracting some null structure from the structure coefficients c1 ,2 , which
causes them to vanish whenever the resonance function vanishes. We will see some
examples of this in later sections.
Exercise 1.49. Prove Proposition 1.38. (Hint: review the proof of Theorem
1.4, Theorem 1.7, and Exercise 1.2.)
Exercise 1.50 (Stability). Let the notation and hypotheses be as in Proposi-
tion 1.38. Suppose that ulin B/2 , and we have an approximate solution u B
to the equation (1.50), in the sense that u = ulin + DN (u) + e for some e S. Let
u B be the actual solution to (1.50) given by the above Proposition. Show that
ku ukS 2kekS . Note that this generalises the Lipschitz claim in Proposition
1.38.
Exercise 1.51. Let N , S be Banach spaces as in Proposition 1.38, and suppose
that one is given a k-linear nonlinearity N (u) = Nk (u, . . . , u), which maps S to N
with the k-linear estimate
kN (u1 , . . . , uk )kN C1 ku1 kS . . . kuk kS
for all u1 , . . . , uk S and some constant C1 > 0. Show that the hypothesis (1.52)
holds for u, v B with := 2kC10 C1 .
Exercise 1.52 (Second order Duhamel). Let L End(D). Show that the
solution to the homogeneous linear second-order ODE
tt u Lu = 0
with initial datum u(0) = u0 , t u(0) = u1 is given by u(t) = U0 (t)u0 + U1 (t)u1 for
some operators U0 : R D D, U1 : R D D. Show that the unique classical
2
solution u Cloc (D R) to the inhomogeneous linear second-order ODE
tt u Lu = f
 1.7. COMPLETELY INTEGRABLE SYSTEMS 49

0
with initial datum u(t0 ) = u0 , t u(t0 ) = u1 , where f Cloc (D D) and t0 R,
is given by the Duhamel formula
Z t
u(t) = U0 (t t0 )u0 + U1 (t t0 )u1 + U1 (t s)f (s) ds.
t0

(Hint: convert the second-order equation to a first order one, then use Proposition
1.35.)
Exercise 1.53 (Duhamel vs. resolvents). Let L, V, L0 End(D) be such that
L = L0 + V . Use Duhamels formula to show that
Z t Z t
etL = etL0 + e(ts)L0 V esL ds = etL0 + e(ts)L V esL0 ds.
0 0

If is a scalar such that the resolvent operators R() := (L )1 and R0 () :=

(L0 )1 exist, establish the resolvent identity
R() = R0 () R0 ()V R() = R0 () R()V R0 ()
and discuss the relationship between the above identities using the Fourier duality
between t and .
Exercise 1.54. Let L be as in Proposition 1.41, and let u solve the equation
(1.44). Use Gronwalls inequality to establish the bound ku(t)kD et ku(0)kD
for all t 0. (Hint: establish a monotonicity formula for ku(t)k2D , and then use the
version of Gronwalls inequality mentioned in Exercise 1.7.)
Exercise 1.55 (Stable manifold). Let D be a finite-dimensional real Hilbert
space, and let L End(D) be weakly linearly stable in the sense that hLu, ui 0
0,1
for all u D. Let N Cloc (D D) have the property that hN (u), ui 0
for all u D. Show that for any u0 D, there exists a unique classical solution
u : [0, +) D to (1.43) with initial datum u(0) = u0 , which is uniformly bounded
and obeys the estimate
Z
|hLu(t), u(t)i| dt ku0 k2D .
0

Conclude in particular that if V is the subspace V := {u D : hLu, ui = 0}, that

dist(u(t), V ) 0 as t +.

1.7. Completely integrable systems

Tyger! Tyger! burning bright
In the forests of the night,
What immortal hand or eye
Could frame thy fearful symmetry?
(William Blake, The Tyger)
We have already encountered Hamiltonian ODE in Section 1.4, which enjoy
at least one conserved integral of motion, namely the Hamiltonian H itself. This
constrains the Hamiltonian flow to a codimension one subspace of the symplectic
phase space D. Additional conserved integrals of motion can constrain the flow
further. It turns out that the largest number of independent conserved integrals that
one can have in a Hamiltonian system is half the dimension of the phase space (see
50 1. ORDINARY DIFFERENTIAL EQUATIONS

Exercise 1.56). When this occurs, we say that the system is completely integrable 20;
the phase space splits completely into the conserved quantities E1 , . . . , EN (also
called action variables), together with the dynamic variables (also called angle
variables) induced by the N flows corresponding to E1 , . . . , EN .
Example 1.43 (Simple harmonic oscillator). Let D = Cn be the phase space in
Example 1.28, and let H be the Hamiltonian (1.32). Then there are n independent
conserved quantities
E1 := |z1 |2 ; . . . ; En := |zn |2
and n angle variables 1 , . . . , n T, defined for most points in phase space by
polar coordinates
z1 = |z1 |ei1 ; . . . ; zn = |zn |ein .
Then the Hamiltonian flow in these action-angle coordinates becomes linear:
j Ej (t) = 0; j j (t) = j .
Also, observe that the Hamiltonian H is just a linear combination of the basic
conserved quantities E1 , . . . , En , which is of course consistent with the fact that H
is itself conserved. More generally, any linear system (1.44) in which L is skew-
adjoint will lead to a completely integrable system.
There are many ways to determine if a system is completely integrable. We
shall discuss only one, the method of Lax pairs.
Definition 1.44. Consider an ODE
(1.57) t u(t) = F (u(t)),
0,1
where F Cloc (D D) and D is a finite-dimensional phase space. Let H be a
finite-dimensional complex Hilbert space, and let End(H) be the space of linear
maps from H to itself (for instance, if H = Cn , then End(H) is essentially the ring
of n n complex matrices). A Lax pair for the ODE (1.57) is any pair L, P
1
Cloc (D End(H)) of functions such that we have the identity
(1.58) t L(u(t)) = [L(u(t)), P (u(t))]
for all classical solutions u : I D to the ODE (1.57), or equivalently if
(F (u) )L(u) = [L(u), P (u)] for all u D.
Here [A, B] := AB BA denotes the usual Lie bracket of the matrices A and B.
Remark 1.45. Geometrically, (1.58) asserts that the matrix L(u(t)) evolves
via infinitesimal rotations that are orthogonal to L(u(t)). In many cases, P
will take values in the Lie algebra g of some Lie group G in End(H), and L will
take values either in the Lie algebra g or the Lie group G; note that the equation
(1.58) is consistent with this assumption, since Lie algebras are always closed under
the Lie bracket (see also Exercise 1.15).

20This definition unfortunately does not rigorously extend to the infinite dimensional phase
spaces one encounters in PDE. Indeed, we do not yet have a fully satisfactory definition of what
it means for a PDE to be completely integrable, though we can certainly identify certain very
suggestive symptoms of complete integrability of a PDE, such as the presence of infinitely many
conserved quantities, a Lax pair formulation, or the existence of explicit multisoliton solutions.
 1.7. COMPLETELY INTEGRABLE SYSTEMS 51

A trivial example of a Lax pair is when L : D Mn (C) is constant, and P

is chosen to commute with L; we shall be more interested in non-trivial examples
when L, and more precisely the spectrum (eigenvalues) of L, admit some genuine
variation across the phase space D. A simple example is provided by the one-
dimensional harmonic oscillator
t u = iu
in the phase space D = C, with H = C2 and Lax pair
   
i(|u|2 ) iu2 i 0
(1.59) L(u) := ; P (u) :=
iu2 i( |u|2 ) 0 i
where the spectral parameter is an arbitrary complex number. Here, L and P take
values in the Lie algebra su2 (C) of SU2 (C), the group of 2 2 unitary matrices.
The higher-dimensional harmonic oscillator in Example 1.28 can also be given a
Lax pair by taking direct sums of the above example; we omit the details.
Now we show how Lax pairs lead to conserved quantities.
Proposition 1.46. Suppose that an ODE (1.57) is endowed with a Lax pair
L : D End(H), P : D End(H). Then for any non-negative integer k, the
moment tr(Lk ) is preserved by the flow (1.57), as is the spectrum (L) := { C :
L not invertible}.
Proof. We begin with the moments tr(Lk ). Let u : I D solve (1.57). From
the Leibnitz rule and the first trace identity
(1.60) tr(AB) = tr(BA)
we have
t tr(L(u(t))k ) = ktr(L(u(t))k1 t L(u(t))) = ktr(L(u(t))k1 [L(u(t)), P (u(t))]).
But from the second trace identity
(1.61) tr(A[B, C]) = tr(B[C, A]) = tr(C[A, B])
(which follows easily from the first trace identity), and the obvious fact that Lk1
commutes with L, we obtain t tr(L(u(t))k ) = 0 as desired.
One can conclude conservation of the spectrum (L) from that of the moments
by using the characteristic polynomial of L. For a more direct approach (which
does not rely as much on the finite dimensionality of L), see Exercise 1.57. 
The quantities tr(L), tr(L2 ), . . . may seem like an infinite number of conserved
quantities, but they are of course not all independent. For instance in the example
(1.59), all the quantities tr(Lk ) are functions of a single conserved quantity |z|2 .
This makes the number of conserved quantities equal to the half the (real) dimension
of the phase space C, and so this equation is completely integrable.
One special case of solutions to a completely integrable system arises when the
spectrum (L) of the Lax operator is unexpectedly simple, for instance if L is a
rank one operator. This often leads to very algebraically structured solutions such
as solitary waves (solitons). For instance, in Example 1.43, the case when L is
rank one corresponds to that of a single excited mode, when only one of the zj is
non-zero, which can be viewed as a rather trivial instance of a solitary wave. The
more general task of reconstructing the solution given the spectral information on
L (and certain supplemental scattering data associated to the initial datum u0 )
is known as inverse scattering and is a very rich subject involving some beautiful
52 1. ORDINARY DIFFERENTIAL EQUATIONS

analysis, algebra, and geometry. It is well outside the scope of this monograph; we
refer the reader to [HSW] for an introduction.
We now give some non-trivial examples of completely integrable systems. The
first is the periodic Toda lattice

(1.62) t an = an (bn+1 bn ); t bn = 2(a2n a2n1 )

where n ranges over a cyclic group Z/N Z, and an : R R, bn : R R are real-

valued functions of time; this can be viewed as a discrete version of the periodic
Korteweg-de Vries (KdV) equation. To place this lattice in Lax pair form, we let H
be an N -dimensional real Hilbert space with orthonormal basis {en : n Z/N Z},
and for any given state u = ((an , bn ))nZ/N Z we define L = L(u) : H H and
P = P (u) : H H on basis vectors by

Len := an en+1 + bn en + an1 en1

P en := an en+1 an1 en1 .

One can easily verify the Lax pair equation (1.58) by testing it on basis vectors.
Note that L is self-adjoint and P is skew-adjoint, which is of course consistent with
(1.58). The Toda lattice enjoys N independent conserved quantities arising from
L, including the trace
X
tr(L) = bn
nZ/N Z

and the second moment

X
tr(L2 ) = b2n + 2a2n ;
nZ/N Z

one may verify by hand that these quantities are indeed preserved by (1.62). The
equation (1.62) is not a Hamiltonian flow using the standard symplectic form on
the state space, but can be transformed into a Hamiltonian flow (with Hamiltonian
2tr(L2 )) after a change of variables, see Exercise 1.59. One can create higher order
Toda flows by using higher moments of L as Hamiltonians, but we will not pursue
this here.
Another example of a completely integrable system is the periodic Ablowitz-
Ladik system

(1.63) t Fn = i(1 |Fn |2 )(Fn1 + Fn+1 ),

where n ranges over a cyclic group Z/N Z, and Fn : R C are complex-valued

functions of time with |Fn | < 1 for all n (this property is preserved by the flow).
This is a discrete analogue of the cubic defocusing periodic nonlinear Schrodinger
equation. To define a Lax pair (L, P ) for this equation, we take H to be a complex
Hilbert space spanned by 2N orthonormal basis vectors {vn , wn : n Z/N Z}. The
Lax operator L = L(F ) : H H is then defined on basis elements by
p
Lvn := 1 |Fn |2 vn+1 + Fn wn
p
Lwn+1 := Fn vn+1 + 1 |Fn |2 wn+1 ;
 1.7. COMPLETELY INTEGRABLE SYSTEMS 53

note that L is in fact a unitary operator (a discrete analogue of a Dirac operator),

with adjoint L = L1 given by
p
L vn+1 := 1 |Fn |2 vn Fn wn+1
p
L wn := Fn vn + 1 |Fn |2 wn+1 ;
The P operator is a little trickier to define. We first define the reflection operator
J : H H as
Jvn := vn ; Jwn := wn
and then the diagonal operator D = J8 ([L, J]2 + [L , J]2 ) by
Fn1 Fn + Fn1 Fn Fn1 Fn + Fn1 Fn
Dvn := vn ; Dwn := wn
2 2
and then define P by
LJL + L JL
P := i( D).
2
The verification of (1.58) is rather tedious but straightforward. Note that P is
skew-adjoint, which is consistent with (1.58) and the unitarity of L.
A completely integrable system contains some quite low-dimensional invariant
sets; in many cases (as with the harmonic oscillator), these invariant sets take
the form of torii. A very interesting question concerns the stability of such in-
variant surfaces; if one makes a perturbation to the Hamiltonian (destroying the
complete integrability), does the invariant surface similarly perturb? The answer
can be surprisingly subtle, involving the theory of Kolmogorov-Arnold-Moser torii,
Nekhoroshev stability, and Arnold diffusion, among other things. We will not at-
tempt to describe this theory here, but refer the reader to [Kuk3] for a discussion
of these topics in the context of Hamiltonian PDE.
Exercise 1.56 (Lagrangian submanifolds). Call a linear subspace V of a sym-
plectic phase space (D, ) null if (v, v ) = 0 for all v, v V . Show that if V is
null, then the dimension of V cannot exceed half the dimension of D. (Hint: look at
the symplectic complement V := {u D : (v, u) = 0 for all v V }.) Conclude
that if E1 , . . . , Ek are functions which Poisson commute with a given Hamilton-
ian H, then for each u D the gradients E1 (u), . . . , Ek (u) span a space of
dimension at most half the dimension of D.
Exercise 1.57 (Conservation of spectrum). Let the notation and hypotheses
be as in Proposition 1.46. Suppose that for some time t0 I and some L
we have (L), thus there exists a non-zero eigenvector 0 H such that
L(u(t0 ))0 0 = 0. Now let : I H solve the Cauchy problem
t (t) = P (u(t))(t); (t0 ) = 0 .
Show that such a solution exists, and furthermore we have
L(u(t))(t) (t) = 0
for all t I. (Hint: use Exercise 1.13). Conclude that the spectrum (L) is an
invariant of the flow.
Exercise 1.58 (Lax pairs vs. Hamiltonian mechanics). uppose that a sym-
1
plectic phase space (D, ) is endowed with maps L Cloc (D End(H)) and
1
R Cloc (D End(End(H))). Suppose we also have the R-matrix identity
{tr(AL), tr(BL)} = tr(BR([L, A]) AR([L, B]))
54 1. ORDINARY DIFFERENTIAL EQUATIONS

for all A, B Mn (C), where {, } denotes the Poisson bracket. Conclude the Poisson
commutation relations
{tr(AL), tr(Lk )} = ktr(A[L, Rt (Lk1 )])
and
{tr(Lm ), tr(Lk )} = 0
for all m, k 0 and A End(H), where Rt : D End(End(H)) is the transpose
of R, thus tr(AR(B)) = tr(BR(A)) for all A, B End(H). (Hint: take advantage
of the trace identities (1.60), (1.61) and the Leibnitz rule (1.36)). Conclude that
the Hamiltonian flows given by the Poisson-commuting Hamiltonians tr(Lk ) each
have a Lax pair (L, Pk ) with Pk := kRt (Lk1 ).
Exercise 1.59 (Hamiltonian formulation of Toda). Let D = RN RN be the
phase space in Exercise 1.27, where we shall abuse notation and write the phase
space variables as pn , qn where n ranges over the cyclic group Z/N Z. Consider the
Hamiltonian X 1
H(q, p) = p2 + V (qn+1 qn )
2 n
nZ/N Z
where V : R R is the Toda potential V (x) := ex + x 1. Show that the
associated Hamiltonian flow is equivalent to the Toda equations (1.62) after making
the Flaschka change of variables
1 1
an := e(qn+1 qn )/2 ; bn := pn .
2 2
Furthermore, show that H = 2tr(L2 ).
Exercise 1.60. Suppose we are given initial data Fn (0) for n Z/N Z with
|Fn (0)| < 1 for all n Z/N Z. Show that there is a unique global classical solution
to (1.63) with this initial data, and that we have |Fn (t)| < 1 for all n Z/N Z and
t R.
 CHAPTER 2

Constant coefficient linear dispersive equations

God runs electromagnetics by wave theory on Monday, Wednes-

day, and Friday, and the Devil runs them by quantum theory on
Tuesday, Thursday, and Saturday. (Sir William Bragg)
Having concluded our discussion of ODE, we begin the analysis of dispersive1
PDE. In this chapter, we shall begin with the study of constant-coefficient linear
dispersive PDE, which are the simplest example of a dispersive equation. Further-
more, much of the theory of nonlinear PDE, especially for short times or small
data, is obtained by perturbation of the linear theory; thus it is essential to have a
satisfactory theory of the linear equation before proceeding to the nonlinear one.
To simplify the discussion2, our partial differential equations shall always take
as their spatial domain either a Euclidean space Rd , or the standard torus Td =
(R/2Z)d ; functions on the latter domain can of course be viewed as periodic
functions on the former domain, and so we shall give our definitions for Rd only,
as the generalisation to Td will be clear. Also, we shall begin by focusing on PDE
which are first-order in time. A constant-coefficient linear dispersive PDE then
takes the form
(2.1) t u(t, x) = Lu(t, x); u(0, x) = u0 (x)
3 d
where the field u : R R V takes values in a finite-dimensional Hilbert space
V , and L is a skew-adjoint constant coefficient differential operator in space, thus
taking the form X
Lu(x) := c x u(x),
||k
where k 1 is an integer (the order of the differential operator), = (1 , . . . , d )
Zn+ ranges over all multi-indices with || := 1 + . . . + d less than or equal to k,
1Informally, dispersion will refer to the fact that different frequencies in this equation
will tend to propagate at different velocities, thus dispersing the solution over time. This is in
contrast to transport equations such as (2.2), which move all frequencies with the same velocity
(and is thus a degenerate case of a dispersive equation), or dissipative equations such as the heat
equation t u = u, in which frequencies do not propagate but instead simply attenuate to zero.
The wave equation (2.9) is partly dispersive - the frequency of a wave determines the direction of
propagation, but not the speed; see Principle 2.1.
2
The study of linear dispersive equations in the presence of potentials, obstacles or other
boundary conditions, or on curved manifolds or in variable coefficient situations, is of great impor-
tance in PDE, with applications to spectral theory, geometry, and even number theory; however,
we will not attempt to investigate these topics here.
3We shall say that the field is real if V is a real vector space, and complex if V is a complex
vector space. We say that the field is scalar if V is one-dimensional, vector if V is viewed as a
vector space, tensor if V is viewed as a tensor space, etc. For instance, a field taking values in Cd
would be a complex vector field. We will not use the term field in the algebraic sense in this
text.
55
56 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

x is the partial derivative

1 d
x := ( ) ...( ) ,
x1 xd
and c End(V ) are coefficients that do not depend on x. This operator is clas-
sically only defined on k-times continuously differentiable functions, but we may
extend it to distributions or functions in other function spaces in the usual manner;
thus we can talk about both classical and weak (distributional) solutions to (2.1).
We can also write L = ih(D), where D is the frequency operator
1 1 1
D := = ( x1 , . . . , xd )
i i i
and h : Rd End(V ) is the polynomial
X
h(1 , . . . , d ) = i||1 c 11 . . . dd .
||k

We assume that L is skew-adjoint, thus

Z Z
hLu(x), v(x)iV dx = hu(x), Lv(x)iV dx

for all test functions u, v; this is equivalent to requiring that coefficients of the
polynomial h be self-adjoint, so in the scalar case we require h to be real-valued.
Note that we do not restrict the time variable to an interval I; this is because the
solutions we shall construct to (2.1) will automatically exist globally in time. We
refer to the polynomial h as the dispersion relation of the equation (2.1).
A somewhat degenerate example of an equation of the form (2.1) is the phase
rotation equation
t u(t, x) = iu(t, x); u(0, x) = u0 (x)
where u is a complex field and R; this has the explicit solution u(t, x) =
eit u0 (x), and the dispersion relation is h() = . Another degenerate example is
the transport equation
(2.2) t u(t, x) = v x u(t, x); u(0, x) = u0 (x)
d
for some constant vector v R ; this has the explicit solution u(t, x) = u0 (x vt),
and the dispersion relation is h() = v . More interesting examples (many of
which arise from physics) can be constructed if one either raises the order of L, or
makes u vector-valued instead of scalar. Examples of the former include the free
Schrodinger equation
~
(2.3) it u + u = 0,
2m
Pd 2
where u : R Rd V is a complex field and = j=1 x 2 is the Laplacian
j
and Plancks constant ~ > 0 and the mass m > 0 are fixed scalars, as well as the
one-dimensional Airy equation
(2.4) t u + xxx u = 0
where u : R R R is a real scalar field. The dispersion relations here are
~
h() = 2m ||2 and h() = 3 respectively. Examples of the latter include vacuum
Maxwells equations
(2.5) t E = c2 x B; t B = x E; x E = x B = 0
 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS 57

in three dimensions d = 3, where E, B : R1+3 R3 are real vector4 fields and the
speed of light c > 0 is constant; the constraints that E and B be divergence-free
are not of the dynamical form (2.1), but nevertheless they end up being compatible
with the flow (Exercise 2.16). The Maxwell equations are a special case of the
abelian Yang-Mills equations
(2.6) F = 0; F + F + F = 0
V2 1+d
where F : R1+d R is an real anti-symmetric two-form field, and R1+d =
5
R R is endowed with the standard Minkowski metric g , defined using the
d

spacetime interval dg 2 = c2 dt2 + dx21 + dx22 + dx23 (with the convention x0 = t),
and which is used to raise and lower indices in the usual manner.
Another example from physics is the Dirac equation
mc
(2.7) i u = u,
~
where 0 , . . . , 3 End(V ) are the gamma matrices, acting on a four-dimensional
complex vector space V , known as spinor space, via the commutation relations
(2.8) + = 2g idV
where g is the Minkowski metric, the mass m 0 is non-negative, and u :
R1+3 V is a spinor field; see Exercise 2.1 for one construction of spinor space.
It is also of interest to consider dispersive equations which are second-order in
time. We will not give a systematic description of such equations here, but instead
only mention the two most important examples, namely the wave equation
(2.9) u = 0; u(0, x) = u0 (x); t u(0, x) = u1 (x)
1+d
where u : R V is a field, and  is the dAlembertian operator
1
 = = 2 t2 + ,
c
and the slightly more general Klein-Gordon equation
m 2 c2
(2.10) u = u; u(0, x) = u0 (x); t u(0, x) = u1 (x)
~2
where the mass m 0 is fixed.
Equations which involve c are referred to as relativistic, while equations in-
volving ~ are quantum. Of course, one can select units of space and time so that
c = ~ = 1, and one can also normalise m = 1 without much difficulty; these con-
stants need to be retained however if one wants to analyze the non-relativistic limit
c of a relativistic equation, the classical limit ~ 0 of a quantum equation,
or the massless limit m 0 of a massive equation.

4A more geometrically covariant statement would be that E and B combine to form a rank
two tensor field F , as in the example of the Yang-Mills equations below. Indeed, F should
be interpreted as the curvature of a connection; cf. Section 6.2.
5We shall use R1+d to denote Minkowski space with the Minkowski metric, and R Rd to
denote a spacetime without any Minkowski metric. Relativistic equations such as the Maxwell,
Yang-Mills, Dirac, Klein-Gordon, and wave equations live on Minkowski space, and thus interact
well with symmetries of this space such as the Lorentz transformations, whereas non-relativistic
equations such as the Airy and Schrodinger equation have no relation to the Minkowski metric
or with any related structures such as Lorentz transformations, the light cone, or the raising and
lowering conventions associated with the metric.
58 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

The constant-coefficient dispersive equations have a number of symmetries. All

are invariant under time translation u(t, x) 7 u(t t0 , x) and spatial translation
u(t, x) 7 u(t, x x0 ). Several also enjoy a scaling symmetry (Exercise 2.9). There
is also usually a time reversal symmetry, though the precise nature of the symmetry
varies. For instance, for the Schrodinger equation one takes u(t, x) 7 u(t, x),
for the Airy equation one takes u(t, x) 7 u(t, x), and for the wave and Klein-
Gordon equations one takes u(t, x) 7 u(t, x). For tensor equations such as Dirac,
Maxwell, and Yang-Mills, one has to apply the time reversal to the tensor space as
well as to the spacetime domain. The equations in Rd typically enjoy a rotation
and reflection invariance, which for scalar equations is simply u(t, x) 7 u(t, U x) for
all orthogonal matrices U O(d); in particular, this implies (once one has a reason-
able uniqueness theory) that radially symmetric data leads to radially symmetric
solutions. Again, for tensor equations one has to rotate the tensor as well as the
domain. The Schrodinger equation also has a very useful Galilean invariance (Exer-
cise 2.5), while the relativistic equations have a similarly useful Lorentz invariance
(Exercise 2.6). Finally, the Schrodinger equation also enjoys the pseudo-conformal
symmetry (Exercise 2.28), while the wave equation similarly enjoys the conformal
symmetry (Exercise 2.14).
The above equations are all connected to each other in various ways; some of
them are given in the exercises.
Exercise 2.1 (Spinor space). Let V = C4 be endowed with the sesquilinear
form
{(z1 , z2 , z3 , z4 ), (w1 , w2 , w3 , w4 )} := z1 w1 + z2 w2 z3 w3 z4 w4
and let , . . . , 3 End(V ) be given as
0

1
0 (z1 , z2 , z3 , z4 ) = (z1 , z2 , z3 , z4 )
c
1 (z1 , z2 , z3 , z4 ) = (z4 , z3 , z2 , z1 )
2 (z1 , z2 , z3 , z4 ) = (iz4 , iz3 , iz2 , iz1 )
3 (z1 , z2 , z3 , z4 ) = (z3 , z4 , z1 , z2 ).
Show that we have the commutation relations (2.8). Furthermore, for all u, v V ,
we have the symmetry { u, v} = {u, v}, and the 4-vector {u, u} is positive
time-like in the sense that
{u, 0 u} 0 and {u, u}{u, u} {u, u}2 0.
Exercise 2.2 (Maxwell vs. Yang-Mills; Dirac vs. Klein-Gordon). Show that
2
any Ct,x,loc solution to Maxwells equation (2.5) or the abelian Yang-Mills equa-
tion (2.6), also solves the wave equation (2.9). Conversely, if A = (A )=0,...,d
2
Ct,x,loc (R1+d R1+d ) solves the wave equation, show that the curvature F :=
A A solves the abelian Yang-Mills equation. Also explain why Maxwells
equation is a special case of the abelian Yang-Mills equation. In a similar spirit,
show that any continuously twice differentiable solution to Diracs equation (2.7)
2
also solves the Klein-Gordon equation (2.10), and conversely if Ct,x,loc (R1+3

V ) solves the Klein-Gordon equation then u := + m solves the Dirac equa-
tion.

Exercise 2.3 (Airy vs. Schrodinger). Let u Ct,x,loc (R R C) solve the
2
Schrodinger equation it u + x u = 0, with all derivatives uniformly bounded. Let
 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS 59

N > 1 be a large number, and let v : R R R be the real scalar field

 
iN x+iN 3 t x + 3N 2 t
v(t, x) := Re e u(t, ) .
3N
Show that v is an approximate solution to the Airy equation (2.4), in the sense that
t v + x3 v = Ou (N 3/2 ).
This suggests that solutions to the Airy equation can behave, in certain circum-
stances, like suitably rescaled and modulated solutions to the Schrodinger equation.
See [BC], [CCT], [Schn] for some nonlinear developments of this idea.
Exercise 2.4 (Group velocity of waves). Let h : Rd R be a polynomial with

real coefficients, and let L := iP (D). Show that if Ct,x,loc (Rd C) has all

derivatives bounded, and > 0, then the complex scalar field u Ct,x,loc (RRd
C) defined by
u(t, x) := eix0 +ith(0 ) ((x + h(0 )t))
is an approximate solution to (2.1) in the sense that
t u = Lu + O (2 ).
This suggests that (sufficiently broad) solutions to (2.1) which oscillate in space
with frequency 0 , should travel at group velocity h(0 ), and oscillate in time
with frequency h(0 ); see also Principle 2.1. In the case of the Schrodinger equation
(2.3), conclude (on a heuristic level) de Broglies law mv = ~, where v denotes
the group velocity. (Note the phase velocity in this case will be twice the group
velocity, 2v. More generally, an inspection of the phase x 0 + th(0 ) shows that
the phase velocity is |00 | h( 0)
|0 | .)

2
Exercise 2.5 (Galilean invariance). Let u Ct,x,loc (R Rd V ) be a com-
plex field, let v Rd , and let u Ct,x,loc
2
(R Rd C) be the field
2
u(t, x) := eimxv/~ eimt|v| /2~
u(t, x vt).
show that u solves the Schrodinger equation (2.3) if and only if u does.
2
Exercise 2.6 (Lorentz invariance). Let u Ct,x,loc (R1+d V ) be a field, let
v R be such that |v| < c < , and let uv Ct,x,loc (R1+d C) be the field
d 2

t v x/c2 xv vt
uv (t, x) := u( p , x xv + p )
2
1 |v| /c 2 1 |v|2 /c2
v v
where xv := (x |v| ) |v| is the projection of x onto the line parallel to v (with the
convention that xv = 0 when v = 0). Thus for instance, if c = 1, v = v1 e1 for some
1 < v1 < 1 and x := (x2 , . . . , xd ), then
t v1 x1 x1 v1 t
uv (t, x1 , x) := u( p ,p , x).
1 v12 1 v12
Show that uv solves the wave equation (2.9) if and only if u does, and similarly
for the Klein-Gordon equation (2.10). (Hint: show that the Minkowski metric is
2
preserved by the Lorentz transformation (t, x) 7 ( tvx/c
2 2
, x xv + xv vt
2 2
).)
1|v| /c 1|v| /c
What is the analogous symmetry for the Dirac, Maxwell, and Yang-Mills equations?
60 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Exercise 2.7 (Schrodinger vs. Klein-Gordon). Let u Ct,x,loc (R1+d V ) be
a complex field solving the Klein-Gordon equation (2.10). Show that if one applies
2
the change of variables u = eitmc /~ v then one obtains
~ ~
it v + v = 2 v.
2m 2mc2 t
This suggests that the Klein-Gordon equation can converge to the Schrodinger
equation in the non-relativistic limit c , though one has to be extremely
careful with this heuristic due to the double time derivative on the right-hand side.
(A more robust approximation is given in the next exercise.)
Exercise 2.8 (Schrodinger vs. Dirac). Let u : R R3 V be a spinor field
solving the Schrodinger equation
~
ic 0 t u u = 0
2m
with all derivatives uniformly bounded. Let v : R R3 V be the spinor field
2 0 ~ imc2 0 t/~ j
v := eimc t/~ u e xj u
2imc
where the j index is summed over 1, 2, 3. Show that v is an approximate solution to
the Dirac equation (2.7) (and hence the Klein-Gordon equation) in the sense that
mc 1
i v = v + O~,m,u ( 2 )
~ c
Thus in the non-relativistic limit c , certain solutions of the Dirac and
Klein-Gordon equations resemble suitably rescaled and modulated solutions of the
Schrodinger equation. See [MNO], [MNO2] for some nonlinear developments of
this idea. By using this correspondence between Schrodinger and Klein-Gordon,
one can also establish in a certain sense that the Lorentz invariance degenerates to
the Galilean invariance in the non-relativistic limit c ; we omit the details.
Exercise 2.9 (Scaling symmetry). Show that if P : Rd C is a homoge-
neous polynomial of degree k, and L = P (), then the equation (2.1) is invariant
under the scaling u(t, x) 7 u( tk , x ) for any > 0. Thus for instance, with the
Schrodinger equation the time variable has twice the dimension of the space vari-
able, whereas for the Airy equation the time variable has three times the dimension
of the space variable. For relativistic equations such as the wave equation, space
and time have the same dimension.
2
Exercise 2.10 (Wave vs. Klein-Gordon). Let u Ct,x,loc (R Rd V ) be a
2
complex field, and define v Ct,x,loc (R Rd+1 C) by
v(t, x1 , . . . , xn , xd+1 ) = eimcxd+1 /~ u(t, x1 , . . . , xd ).
Show that v solves the d + 1-dimensional wave equation (2.9) if and only if u solves
the d-dimensional Klein-Gordon equation (2.10). This allows one to use the method
of descent (analyzing a lower-dimensional PDE by a higher-dimensional PDE) to
obtain information about the Klein-Gordon equation from information about the
wave equation. As the name implies, the method of descent is largely one-way;
it identifies general solutions to lower-dimensional PDE with special solutions to
higher-dimensional PDE, but does not yield much information on general solutions
to the higher-dimensional PDE.
 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS 61

2
Exercise 2.11 (Wave vs. Schrodinger). Let u Ct,x,loc (R Rd V ) be a
2
complex field, and define v Ct,x,loc (R Rd+1 V ) by
t xd+1
v(t, x1 , . . . , xd , xd+1 ) = ei(t+xd+1 ) u( , x1 , . . . , xd ).
2
Set ~ = c = m = 1. Show that v solves the d + 1-dimensional wave equation (2.9) if
and only if u solves the d-dimensional Schrodinger equation (2.3). See also Exercise
3.2.
2
Exercise 2.12 (Wave vs. wave). Suppose that the field u Ct,x,loc (R
d
(R \{0}) V ) solves the d-dimensional wave equation u = 0, thus by abuse of
notation we can write u(t, x) = u(t, |x|) and consider u = u(t, r) now as a function
2
from R (0, ) C. Conclude the radial field v Ct,x,loc (R (Rd+2\{0}) V )
1
defined (again abusing notation) by v(t, r) := r r u(t, r) solves the d+2-dimensional
wave equation. Thus in the radial case at least, it is possible to construct solutions
to the d + 2-dimensional equation out of the d-dimensional equation.

Exercise 2.13 (1 + 1 waves). Show that if the field u Ct,x,loc (R R V )
2 2
solves the one-dimensional wave equation t u x u = 0 with initial data u(0, x) =
u0 (x), t u(0, x) = u1 (x), then
Z
u0 (x + t) + u0 (x t) 1 x+t
u(t, x) := + u1 (y) dy
2 2 xt
for all t, x R. This is a rare example of a PDE which can be solved entirely by
elementary means such as the fundamental theorem of calculus.

Exercise 2.14 (Conformal invariance). Let + R1+d be the forward light

cone {(, ) R1+d : > c||}, and let u Ct,x,loc
2
(+ V ) be a field. Let
2
u Ct,x,loc (+ V ) be the conformal inversion of u, defined by
t x
u(t, x) := (t2 |x|2 )(d1)/2 u( , ).
t2 |x|2 t2 |x|2
Establish the identity
d1 t x
u(t, x) = (t2 |x|2 ) 2 1 u( , ).
t2 |x|2 t2 |x|2
In particular u solves the wave equation (2.9) with c = 1 if and only if u does. (One
can use in fact hyperbolic polar coordinates to recast the wave equation as a wave
equation on hyperbolic space H d , in such a way that conformal inversion amounts
simply to time reversal. Another approach is to observe that Kelvin inversion
1
(t, x) 7 t2 |x| 2 (t, x) is a conformal transformation of Minkowski space.)

Exercise 2.15 (Quantum vs. classical dynamics). Let ~p = (p1 , . . . , pd ) denote

the momentum operator p~ := i~; show that when applied to the approximate
solutions to the Schrodinger equation in Exercise 2.4 that we have de Broglies
p|2 p21 +...+p2d
law ~p mv = ~. Let H := |~ 2m = 2m denote the Hamiltonian (this is
2
the analogue of the classical Hamiltonian H = |~
p| 1 2
2m = 2 m|v| in the absence of
d
a potential). If u Ct Sx (R R ) is a classical solution to the Schrodinger
62 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

equation (2.3) and E : Sx (Rd ) Sx (Rd ) is any (time-independent) continuous

linear operator, establish the Heisenberg equation
d i
hEu(t), u(t)iL2x (Rd ) = h [H, E]u(t), u(t)iL2x (Rd )
dt ~
R
where [H, E] := HE EH is the Lie bracket and hf, giL2x (Rd ) := Rd f (x)g(x).
Compare this with (1.33) and with Exercise 1.36. The precise connection between
the Lie bracket and the Poisson bracket involves the subjects of semiclassical analy-
sis, microlocal analysis, and geometric quantization, which we will not discuss here.
Exercise 2.16 (Consistency of Maxwells equations). Suppose that E, B
Ct Sx (R1+3 R3 ) solve the dynamic component t E = c2 x B; t B =
x E of Maxwells equations (2.5). Suppose also that the static components
of the equation hold at time t = 0, thus x E(0, x) = x B(0, x) = 0. Show that
Maxwells equations
R in fact hold for all time. (Hint: compute the time derivative
of the quantity R3 | E(t, x)|2 + | B(t, x)|2 dx using integration by parts.)

2.1. The Fourier transform

His life oscillates, as everyones does, not merely between two poles,
such as the body and the spirit, the saint and the sinner, but
between thousands, between innumerable poles. (Herman Hesse,
Steppenwolf)
The spatial Fourier transform, and the closely related spacetime Fourier trans-
form, is an exceptionally well-suited6 tool to analyze constant coefficient linear
dispersive equations such as (2.1). This is ultimately due to the invariance of these
equations under translations in either space or time. A brief summary of the proper-
ties of the Fourier transform that we shall need (as well as other notation appearing
here, such as the Sobolev spaces Hxs (Rd )) can be found in Appendix A.
One hint that the Fourier transform will be useful for solving equations such as
(2.1) comes from the simple observation that given any frequency 0 Rd and any
P : Rd R, the plane wave eix0 +ith(0 ) solves the equation (2.1) with L = ih(D)
(see Exercise 2.4). From the principle of superposition for linear equations, we thus
see that we can construct solutions to (2.1) as superpositions of plane waves.
In order to obtain an adequate wellposedness theory for dispersive equations,
it is often necessary to restrict attention to solutions which not only have some
smoothness, but also some decay. To get some idea of the problems involved,
consider the complex scalar field u : (, 0) Rd C defined by
1 2
u(t, x) := d/2 eim|x| /2~t .
|t|
This field can be verified to be a smooth solution to the Schrodinger equation (2.3)
for all negative times, but becomes singular at time t = 0. The problem is that
while this solution is smooth, it does not decay at all as x . For an even
worse example of bad behaviour of the Schrodinger equation - namely breakdown
of uniqueness even for smooth solutions - see Exercise 2.24.
6In some ways, it is too well suited; there are a number of results in this field which are
so easily proven using the Fourier transform, that non-Fourier-based alternative proofs have not
been adequately explored, and as such one encounters difficulty extending those results to variable-
coefficient, curved space, or nonlinear settings in which the Fourier transform is less useful.
 2.1. THE FOURIER TRANSFORM 63

To avoid these issues, we begin by restricting attention to the Schwartz space

Sx (Rd ). To simplify the discussion, let us now only consider scalar equations, so
that the dispersion relation h : Rd R is now real-valued; the vector-valued case
introduces a number of interesting new technical issues which we will not discuss
1
here. If u Ct,loc Sx (R Rd) is a classical solution to (2.1), then by taking Fourier
transforms of (2.1) we conclude
d
t u(t)() d
= ih()u(t)()
which has the unique solution

(2.11) d
u(t)() = eith() u
c0 ().
Note that as h() is real and cu0 is Schwartz, the function eith() u
c0 () is then also
Schwartz for any t R. We may thus apply the Fourier inversion formula and
obtain the solution
Z
(2.12) u(t, x) = eith()+ix u
c0 () d;
Rd

because of this, we shall let etL = eith(D) denote the linear propagator
Z
etL u0 (x) := eith()+ix u
c0 () d.
Rd

This propagator is defined initially for Schwartz functions, but can be extended
by standard density arguments to other spaces. For instance, Plancherels theorem
allows one to extend etL to be defined on the Lebesgue space L2x (Rd ), or more
generally to the inhomogeneous Sobolev space7 Hxs (Rd ) for any s R, as well
as the homogeneous Sobolev spaces Hxs (Rd ) (see Appendix A). It is clear that
etL is a unitary operator on these spaces, and in particular on L2x (Rd ) (which by
Plancherels identity is equivalent to Hx0 (Rd ) = Hx0 (Rd ), except for an inessential
factor of (2)d/2 ):
ketL f kHxs (Rd ) = kf kHxs (Rd ) ; ketL f kH s (Rd ) = kf kH s (Rd ) ketL f kL2x (Rd ) = kf kL2x (Rd ) .
x x

One can of course also extend these propagator to tempered distributions by duality.
Propagators are examples of Fourier multipliers and as such, they commute
with all other Fourier multipliers, including constant coefficient differential oper-
ators, translations, and other propagators. In particular they commute with the
fractional differentiation and integration operators his for any s R.
The Fourier transform can also be defined8 on the torus Td . If f Cx (Td
C) is smooth, the Fourier transform f : Zd C is defined by
Z
1
f(k) := f (x)eikx dx.
(2)d Td

7In the notation of the next chapter, the function u(t) = etL u is the unique strong H s (Rd )
0 x
solution to the Cauchy problem (2.1). As a rule of thumb, as long as one restricts attention to
strong solutions in a space such as Hxs , the linear evolution is completely non-pathological.
8Of course, the Fourier transform can in fact be defined on any reasonable abelian group,
and even (with some modifications) on most non-abelian groups; but we will not pursue these
issues here. Also, completely integrable PDE often come with a scattering transform which can
be viewed as a nonlinear version of the Fourier transform, but again we will not discuss this here.
64 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

One can show that f(k) is a rapidly decreasing function of k, and the inversion
formula is given by X
f (x) = f(k)eikx .
kZd
Much of the preceding discussion extends to the periodic setting, with some minor
changes; we leave the details to the reader. One advantage of the periodic setting is
that the individual Fourier modes eikx are now themselves square-integrable (and
more generally lie in all the Sobolev spaces Hxs (Td ), thus for instance keikx kHxs (Td )
is equal to hkis ). This makes it easier to talk about the evolution of individual
Fourier modes, as compared to the non-periodic case in which the Fourier modes
lie in a continuum.
The spatial Fourier transform f (x) 7 f() brings into view the oscillation of a
function in space. In the analysis of dispersive PDE, it is also important to analyze
the oscillation in time, which leads to the introduction of the spacetime Fourier
transform. If u : R Rd C is a complex scalar field, we can define its spacetime
Fourier transform u : R Rd C formally as
Z Z
u(, ) := u(t, x)ei(t +x) dtdx.
R Rd
To begin with, this definition is only sensible for sufficiently nice functions such as
those which are Schwartz in both space and time, but one can then extend it to
much more general functions and to tempered distributions by density arguments
or duality. Of course, in such case one does not always expect u to be well-behaved,
for instance it could be a measure or even a tempered distribution. Formally at
least, we have the inversion formula
Z Z
1
u(t, x) = ei(t +x) u(, ) d d.
(2)d+1 R Rd
The advantage of performing this transform is that it not only diagonalises the
linear operator L in (2.1), but also diagonalises the time derivative t . Indeed, if
u is any tempered distributional solution to (2.1) (which will in particular include
classical solutions which grow at most polynomially in space and time) with L =
P () = ih(/i) as before, then on taking the spacetime Fourier transform we
obtain
i u(, ) = ih()u(, )
and thus
( h())u(, ) = 0.
The theory of distributions then shows that u(, ) is supported in the characteristic
hypersurface {(, ) : = h()} of the spacetime frequency space R Rd , and in
fact takes the form
u(, ) = ( h())a()
for some spatial tempered distribution a, where is the Dirac delta. In the case of
the Schwartz solution (2.12), we have a = u0 , thus
u(, ) = ( h())u0 ().
For comparison, if we consider u0 as a function of spacetime via the trivial extension
u0 (t, x) := u0 (x), then we have
u0 (, ) = ( )u0 ().
 2.1. THE FOURIER TRANSFORM 65

Thus in the spacetime frequency space, one can think of the linear solution u
to (2.1) as the time-independent field u0 , twisted by the transformation (, ) 7
( + h(), ).
In applications to nonlinear PDE, it is often not feasible to use the spacetime
Fourier transform to the solution u directly, because often u is only defined on a
spacetime slab I Rd instead of the entire spacetime R1+d . This necessitates some
sort of artificial extension9 of the solution u from I to R in order to take advantage
of the features of the spacetime Fourier transform. Nevertheless, the spacetime
Fourier transform (and in particular the Sobolev spaces X s,b = Xs,b =h() adapted
to the characteristic hypersurface) has been proven to be a very useful tool in both
the study of linear and nonlinear dispersive equations; see Section 2.6.
Another approach to analyzing these PDE proceeds by taking the Fourier trans-
form in the time variable only, keeping the spatial variable untouched. This ap-
proach is well suited for settings in which the operator L has variable-coefficients, or
has a domain which is curved or has a boundary, and leads to spectral theory, which
analyzes the behaviour of propagators such as etL in terms of resolvents (L z)1 ,
as well as the closely related spectral measure of L. This perspective has proven
to be immensely useful for the linear autonomous setting, but has had less success
when dealing with non-autonomous or nonlinear systems, and we will not pursue
it here.
We summarise some of the above discussion in Table 1, as well as in the fol-
lowing principle.
Principle 2.1 (Propagation of waves). Suppose that a solution u solves a scalar
dispersive equation (2.1) with initial datum u0 . If u0 has spatial frequency roughly
0 (in other words, the Fourier transform u0 is concentrated near 0 ), then u(t)
will have spatial frequency roughly 0 for all times, and u will also oscillate in time
with frequency roughly h(0 ). In physical space, u will travel with velocity roughly
h(0 ). These heuristics are only valid to accuracies consistent with the spatial
and frequency uncertainty of u; the wave is initially coherent, but as time progresses,
the frequency uncertainty (and hence velocity uncertainty) overwhelms the spatial
uncertainty, leading to dispersion. (In particular, the uncertainty principle itself is
a mechanism for dispersion over time.)
Thus for instance, fields of frequency 0 will propagate at velocities ~/v under
the Schrodinger evolution, 3 2 for the Airy evolution, and c/|| for the wave
evolution, subject to limitations given by the uncertainty principle. See Exercise
2.4 for a partial verification of this principle. For the wave equation, thi principle
suggests that waves propagate at the speed of light c; we shall expand upon this
finite speed of propagation property in Section 2.5. One can also use techniques from
oscillatory integrals, in particular the method of stationary phase, to make rigorous
formulations of the above principle, but we will prefer to leave it as an informal
heuristic. The situation for systems (as opposed to scalar equations) can be more
complicated; one usually has to decompose h() (which is now an operator) into
eigenspaces, and thus decompose the wave of frequency into polarised components,
each of which can propagate in a different direction. We will not discuss this here.

9One could also work with various truncated variants of the Fourier transform, such as
cosine bases or the Laplace transform. However, the advantages of such tailored bases are minor,
66 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Table 1. Some different perspectives to analyzing an evolution

equation, and the coordinates used in that analysis. This list is not
intended to be exhaustive, and there is of course overlap between
the various perspectives.

Approach Sample tools

Spatial variable (x) Elliptic theory, gauge fixing, coercivity estimates
Causality (t) Gronwall, bootstraps, Duhamel, time subdivision
Physical space (x, t) Integration by parts, substitutions, fund. solution
Spacetime geom. (x ) Vector fields, null surfaces, conformal maps
Frequency space (, t) Littlewood-Paley theory, multipliers, paraproducts
Spectral theory (x, ) Spectral measure, resolvents, eigenfunctions
Spacetime freq. (, ) X s,b spaces, dispersion relation, null structure
Phase space (x, , t) Bicharacteristics, pseudodifferential operators, FIOs
Geom. optics (x, , t, ) Eikonal and Hamilton-Jacobi equations, WKB
Hamiltonian (D, t) Noethers theorem, nonsqueezing, normal forms
Lagrangian (u) Stress-energy tensor, symmetries, variational methods

Exercise 2.17 (Translation operators). Show that for any x0 Rd , the prop-
agator exp(x0 ) is the operation of translation by x0 , thus exp(x0 )f (x) =
f (x x0 ). Compare this with Taylors formula in the case that f is real analytic.
Exercise 2.18 (Wave propagators). Using the spatial Fourier transform, show
2
that if u Ct,loc Sx (R Rd ) is a field obeying the wave equation (2.9) with c = 1,
then
d sin(t||)
u(t)() = cos(t||)u0 () + u1 ()
||
for all t R and Rd ; one can also write this as

sin(t )
u(t) = cos(t )u0 + u1

or using the spacetime Fourier transform as
1 sgn( )
u(, ) = (| | ||)( u0 () + u1 ())
2 2i||
(note that some care is required to ensure that the product of ( ||) and sgn( )
actually makes sense). Thus the characteristic hypersurface for the wave equation is
the light cone {(, ) : | | = ||}. As these formulae make sense for any distributions
u0 , u1 , we shall refer to the function or distribution u(t) generated by these formulae
as the solution to the wave equation with this specified data. Show that if (u0 , u1 )
Hxs (Rd )Hxs1 (Rd ) for some s R, then we have u Ct0 Hxs (RRd )Ct1 Hxs1 (R
Rd ), and in fact we have the bounds
ku(t)kHxs1 (Rd ) + kt u(t)kHxs1 (Rd ) .d,s ku0 kHxs (Rd ) + ku1 kHxs1 (Rd ) )
and
ku(t)kHxs (Rd ) .d,s hti(ku0 kHxs (Rd ) + ku1 kHxs1 (Rd ) )

and are usually outweighed by the loss of algebraic structure and established theory incurred by
abandoning the Fourier transform.
 2.1. THE FOURIER TRANSFORM 67

for all times t R. Thus the solution u stays bounded to top order, but lower
order norms of u can grow linearly in time.
2
Exercise 2.19 (Klein-Gordon propagators). If u Ct,loc Sx (R Rd ) is a clas-
sical solution to the Klein-Gordon equation (2.10) with c = m = 1, then
d sin(thi)
u(t)() = cos(thi)u0 () + u1 ()
hi
for all t R and Rd ; one can also write this as

sin(t 1 )
u(t) = cos(t 1 )u0 + u1
1
or
1 sgn( )
u(, ) = (| | hi)( u0 () + u1 ()).
2 2ihi
Thus the characteristic hypersurface here is the two-sheeted hyperboloid {(, ) :
| |2 ||2 = 1}. Again we extend this formula to distributions to define the notion
of a distributional solution to the Klein-Gordon equation. Show that if (u0 , u1 )
Hxs (Rd ) Hxs1 (Rd ), then we have u Ct0 (R H s (Rd )) Ct1 (R Hxs1 (Rd )),
and in fact we have the bounds
ku(t)kHxs (Rd ) + kt u(t)kHxs1 (Rd ) .d,s ku0 kHxs (Rd ) + ku1 kHxs1 (Rd )
for all times t R. Thus the Klein-Gordon equation has slightly better regularity
behaviour in (inhomogeneous) Sobolev spaces than the wave equation.
Exercise 2.20 (Geometry of characteristic surfaces). Interpret Exercises 2.3-
2.11 using either the spatial Fourier transform or the spacetime Fourier transform,
viewing each of these exercises as an assertion that one dispersion relation can
be approximated by or transformed into another. Several of the exercises should
correspond to a specific geometric approximation or transformation (e.g. approxi-
mating a cubic by a quadratic or a polynomial by a tangent, or a hyperboloid by
two paraboloids; or by relating a cone to its conic sections). The exercises involving
vector equations such as Diracs equation are a little trickier to interpret this way,
as one also needs to analyze the eigenspaces of the dispersion relation h().
Exercise 2.21 (Duhamel formula). Let I be a time interval. Suppose that
u Ct1 Sx (I Rd ) and F Ct0 Sx (I Rd ) solve the equation t u = Lu + F , where
L = ih(D) is skew-adjoint. Establish the Duhamel formula
Z t
(2.13) u(t) = e(tt0 )L u(t0 ) + e(ts)L F (s) ds
t0

for all t0 , t I (compare with (1.48)).

Exercise 2.22 (Wave Duhamel formula). Let I be a time interval. Suppose
that u Ct2 Sx (I Rd ) and F Ct0 Sx (I Rd ) are fields such that u = F .
Establish the Duhamel formula
(2.14)
Z t
sin((t t0 ) ) sin((t s) )
u(t) = cos((tt0 ) )u(t0 )+ t u(t0 ) F (s) ds
t0
forall t0 , t I (compare with Exercise 1.52).
68 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Exercise 2.23 (Invariant energy). Show that if u Ct,loc Sx (R1+d ) is a clas-
sical solution to the wave equation with c = 1, then the invariant energy C(t) :=
ku(t)k2 1/2 d +kt u(t)k2 1/2 d is independent of the choice of time t. (Hint: use
Hx (R ) Hx (R )
the spatial Fourier transform, and either the explicit solution for u or differentiate
in time. One can also use the method sketched below.) Furthermore, it is also
invariant under the Lorentz transformation in Exercise 2.6. (This is rather tricky.
One way is to show that u can be expressed in the form
XZ d
u(t, x) = f ()eixit||
Rd ||
for some locally integrable functions f+ (), f (), and that for any such represen-
tation we have Z
1 X d
C(t) = |f ()|2 .
(2)d Rd ||
Then apply the Lorentz transform to u and see what that does to f . The measure
d 2 2
|| on the light cone || = 0 can also be interpreted more invariantly as the
Dirac measure 2( 2 ||2 ).)

Exercise 2.24 (Illposedness of Schrodinger in Cx,loc ). Give an example of a

smooth solution u Ct,x,loc (RR C) to the Schrodinger equation it u+x2 u = 0
which vanishes on the upper half-plane t 0 but is not identically zero. (Hint: Find
a function f (z) which is analytic10 on the first quadrant {Re(z), Im(z) 0}, which
decays fast enough on the boundary of this quadrant (which is a contour consist-
R 2
ing of the positive real axis and upper imaginary axis) so that f (z)eixz+itz dz
converges nicely for any xR R and t R, is equal to zero for t 0 by Cauchys
2
theorem, but is such that f (z)eitz dz is not identically zero. For the latter fact,
you can use one of the uniqueness theorems for the Fourier transform.) This shows
that one no longer has a satisfactory uniqueness theory once we allow our solutions
to grow arbitrarily large (so that the theory of tempered distributions no longer
applies); thus infinite speed of propagation, combined with extremely large reserves
of energy at infinity, can conspire to destroy uniqueness even in the smooth cat-

egory. Show that such an example does not exist if we replace Ct,x,loc by Ct,x (so
that all derivatives of u are uniformly bounded).
Exercise 2.25. Let u Ct0 Hx1 (R1+d )Ct1 L2x (R1+d ) be an energy class solution
to the wave equation with c = 1. Show that for any bounded time interval I we
have the bound
Z
k u(t, x) dtkHx2 (Rd ) .d ku(0)kHx1 (Rd ) + kt u(0)kL2x(Rd ) .
I
(This can be done either by direct integration by parts in physical space, or by the
spatial Fourier transform.) Thus integrating a solution in time leads to a gain of
regularity in space. This phenomenon is a consequence of the oscillation of u in
time, and fails if one places absolute values inside the time integral on the left-hand
side; see however the Strichartz estimates discussed in Section 2.3, which are also
10The author thanks Jared Wunsch, John Garnett, and Dimitri Shlyakhtenko for discussions
which led to this example. A key point is that the functions involved grow so fast that they are not
tempered distributions (despite being smooth), and thus beyond the reach of the distributional
Fourier transform.
 2.2. FUNDAMENTAL SOLUTION 69

a kind of smoothing effect formed by averaging in time. (Thanks to Markus Keel

for this problem.)

2.2. Fundamental solution

The wide wings flap but once to lift him up. A single ripple starts
from where he stood. (Theodore Roethke, The Heron)
In the previous section, we have given the solution u(t) to the linear disper-
sive equation (2.1) as a spatial Fourier multiplier applied to the initial datum u0 ,
see (2.11). For simplicity let us take u to be a complex scalar field and u0 to
be Schwartz. Since multiplication on the Fourier transform side intertwines with
convolution on the spatial side, we thus have
Z
u(t, x) = u0 Kt (x) = u0 (x y)Kt (y) dy
Rd
where the fundamental solution (or Riemann function) Kt is the (distributional)
inverse Fourier transform of the multiplier eith() :
Z
1
(2.15) Kt (x) := ei(x+th()) d.
(2)d Rd
One can also think of Kt as the propagator etL applied to the Dirac delta function
.
The integral in (2.15) is not absolutely convergent, and thus does not make
sense classically. However in most cases one can solve this problem by a limiting
procedure, for instance writing
Z
1 2
(2.16) Kt (x) = lim ei(x+th()) e|| d.
0 (2)d Rd

The integrals on the right-hand side are absolutely convergent, and the limit often
also exists in the sense of (tempered) distributions. For instance, in the case of the
Schrodinger equation (2.3), with h() = 2m~
||2 , we have
1 2
(2.17) Kt (x) = eit~/2m = eim|x| /(2~t)
(2i~t/m)d/2
for all t 6= 0, where one takes the standard branch of the complex square root with
branch cut on the negative real axis (see Exercise 2.26). Thus we have the formula
Z
1 2
(2.18) u(t, x) = d/2
eim|xy| /(2~t) u0 (y) dy
(2i~t/m) Rd
for t 6= 0 and all Schwartz solutions to the free Schrodinger equation (2.3) with ini-
tial datum u(0, x) = u0 (x). This simple formula is even more remarkable when one
observes that it is not obvious11 at all that u(t, x) actually does converge to u0 (x)
in the limit t 0. It also has the consequence that the Schrodinger evolution is
instantaneously smoothing for localised data; if u0 is so localised as to be absolutely
integrable, but is not smooth, then from (2.18) shows that at all other times t 6= 0,
the solution u(t) is smooth (but not localised). Thus the Schrodinger evolution can
instantaneously trade localisation for regularity (or vice versa, since the equation is
time reversible). This effect is related to the local smoothing phenomenon, which we
11Indeed, the problem of establishing pointwise convergence back to u when u is only
0 0
assumed to lie in a Sobolev space Hxs (Rd ) is still a partially unsolved problem, and is considered
to be quite difficult. See for instance [Sjo], [Veg].
70 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

discuss later in this chapter. It can also be explained using the heuristic of Heisen-
bergs law mv = ~ (from Principle 2.1); the high frequencies of u travel very fast
and radiate quickly away from the origin where they are initially localised, leaving
only the low frequencies, which are always smooth, to remain near the origin.
Remark 2.2. The instantaneous smoothing effect is also closely related to
the infinite speed of propagation for the Schrodinger equation; a solution which is
compactly supported at time t = 0 will instantly cease to be compactly supported
at any later time, again thanks to (2.18). Indeed, one can heuristically argue that
any equation which is both time reversible and enjoys finite speed of propagation
(as well as some sort of uniqueness for the evolution), such as the wave and Klein-
Gordon equations, cannot enjoy any sort of fixed-time smoothing effect (though
other nontrivial fixed-time estimates may well be available).
Next we consider the fundamental solution Kt (x) for the Airy equation (2.4).
Here we have h() = 3 , and thus we have
Z
1 3
Kt (x) = ei(x+t ) d.
2 R
A simple rescaling argument then shows
Kt (x) = t1/3 K1 (x/t1/3 )
where we adopt the convention that (t)1/3 = (t1/3 ). The function K1 is essen-
tially the Airy function, and is known to be bounded (see Exercise 2.30). Thus we
see that Kt = O(t1/3 ). This should be compared with the decay of O(t1/2 ) which
arises from the one-dimensional Schrodinger equation. The difference in decay is
explained by the different dispersion relations of the two equations (h() = 3 for
Airy, h() = 21 2 for Schrodinger). From Exercise 2.4 or Principle 2.1, the relation-
ship between group velocity and frequency for the Airy equation is v = 3 2 , as
opposed to v = for Schrodinger. Thus high frequencies move even faster in Airy
than in Schrodinger (leading to more smoothing), but low frequencies move more
slowly12 (leading to less decay).
Now we turn to the wave equation with c = 1, for which the situation is more
complicated. First of all, as the equation is second-order in time, there are two
fundamental solutions of importance, namely
Z
1
Kt0 (x) := cos(t )(x) = cos(2t||)eix d
(2)d Rd
and Z
sin(t ) 1 sin(2t||) ix
Kt1 (x) := (x) = e d
(2)d Rd ||
(see Exercise 2.18, 2.22). In principle, this is not a difficulty, as one can easily
verify (e.g. using the Fourier transform) that the two solutions are related by the
formula Kt0 = t Kt1 . However, in this case the fundamental solutions are only
distributions rather than smooth functions, because one now only has finite speed
of propagation (see Section 2.5) and hence no fixed-time smoothing effects exist.
Nevertheless, the above formulae do give some insight as to the character of these
12The reader may wonder at this point whether the vague heuristics from Principle 2.1 can be
placed on a more rigorous footing. This is possible, but requires the tools of microlocal analysis,
and in particular the principle of stationary phase and the wave packet decomposition. We will
not attempt a detailed description of these tools here, but see for instance [Stei2].
 2.2. FUNDAMENTAL SOLUTION 71

distributions. Indeed one easily sees that Kt0 and Kt1 are radially symmetric, real,
and verify the scaling rules
1 0 x sgn(t) 1 x
Kt0 (x) = K ( ); Kt1 (x) = K ( ).
|t|d 1 t |t|n1 1 t
In the case when d is odd, one can describe Kt0 and Kt1 explicitly (see Exercise 2.31),
and then one can then use the method of descent (as in Exercise 2.33) to obtain
a formula for even dimensions d. It will however be more important for us not to
have explicit formulae for these fundamental solutions, but instead to obtain good
estimates on the solutions and on various smoothed out versions of these solutions.
A typical estimate is as follows. Let Sx (Rd ) be a Schwartz function, and for
any > 0 let := n (x); thus for large this resembles an approximation to
the identity. Then we have the pointwise estimates
(2.19) |Kt0 (x)| .,d d hti(d1)/2
and
(2.20) |Kt1 (x)| .,d d1 hti(d1)/2
for all t R and x Rd . These estimates can be proven via the Fourier transform
and the principle of stationary phase; we shall indicate an alternate approach using
commuting vector fields in Exercise 2.65. Roughly speaking, these estimates assert
that Kt0 and Kt1 decay like t(d1)/2 , but only after integrating Kt0 and Kt1 d+1 2
times and d12 times respectively.
The situation for the Klein-Gordon equation (2.10) is even more complicated;
formulae for the explicit solution are in principle obtainable from the method of
descent (see Exercise 2.33) but they are not particularly simple to work with. It
is convenient to split the frequency domain into the nonrelativistic region, when
~|| mc, and the relativistic region, when ~|| mc. The basic rule of thumb
here is that the Klein-Gordon equation behaves like the Schrodinger equation in the
non-relativistic region and like the wave equation in the relativistic region. For some
more precise formulations of this heuristic, see [MN] and the references therein.
One can also construct fundamental solutions for these equations on the torus
Td = Rd /2Zd . In the case of the wave and Klein-Gordon equations, which have
finite speed of propagation, there is little difference between the two domains (for
short times at least). However, the fundamental solution for dispersive equations
such as the Schrodinger equation become significantly more complicated to con-
trol on torii, with some very delicate number theoretic issues arising; as such, the
concept of a fundamental solution has only limited importance in these settings.
Exercise 2.26 (Gaussian integrals). Use contour integration to establish the
identity r
Z
2 2 /4
ex ex dx = e

whenever , are complex numbers with Re() > 0, where one takes the standard
branch of the square root. Use this and (2.16) to establish (2.17). (You may wish
~
to first reduce to the case when d = 1 and 2m = 1. With a rescaling trick one can
also assume t = 1.)
Exercise 2.27 (Schrodinger fundamental solution). Show that up to multipli-
cation by a constant, the fundamental solution (2.17) for the Schrodinger equation
72 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

is the only tempered distribution which is invariant under spatial rotations, Galilean
transforms, time reversal symmetry, and the scaling u(t, x) 7 n u( t2 , x ). This
can be used to give an alternate derivation of the fundamental solution (except for
the constant).
1
Exercise 2.28 (Pseudoconformal transformation). Let u Ct,loc Sx (R Rd )
be a complex scalar field, and define the field v : R Rd C by
1 1 x 2
v(t, x) := d/2
u( , )ei|x| /2t ,
(it) t t
with the convention
1
v(0, x) = u0 (x).
(2)d/2
Establish the identity
1 1 1 1 1 x 2
(it v + v)(t, x) = 2 (it u + u)( , )ei|x| /2t
2 t (it)d/2 2 t t
for t 6= 0. In particular, show that if u solves the Schrodinger equation (2.3)
with ~ = m = 1, then v also solves the Schrodinger equation (even at time t = 0).
Thus the pseudoconformal transformation manages the remarkable feat of swapping
the initial datum u0 (x) with its Fourier transform u0 (x) (up to trivial factors).
Also verify that the psedoconformal transformation u 7 v is its own inverse. For
an additional challenge, see if you can link this transformation to the conformal
transformation in Exercise 2.14 using Exercise 2.11.
Exercise 2.29 (Van der Corput lemma). Let I R be a compact interval.
If Cx2 (I R) is either convex or concave, R and |x (x)| for all x I
and some > 0, establish the estimate | I ei(x) dx| 2 . (Hint: write ei(x) =
1 i(x)
i (x) x e and integrate by parts.) From this and induction, conclude that if
k
k 2 andR i(x) C (I R)
is such that |xk (x)| for all x I and some > 0,
1/k
then | e dx| .k . Obtain a similar estimate for integrals of the form
R i(x)I
R
e (x) dx when has bounded variation.
Exercise 2.30 (Airy fundamental solution). Let Kt (x) be the fundamental
solution of the Airy function. Establish the bounds K1 (x) = ON (hxiN ) for any
N 0 and x > 0, and K1 (x) = O(hxi1/4 ) for any x 0. (Hint: when x 1, use
repeated integration by parts. When x is bounded, use van der Corputs lemma.
When x 1, split the integral up into pieces and apply van der Corputs lemma
or integration by parts to each.) Explain why this disparity in decay is consistent
with Principle 2.1.
Exercise 2.31 (Wave fundamental solution). Let d 3 be odd and c = 1,
)
and consider the fundamental solutions Kt0 = cos(t ) and Kt1 = sin(t


for the wave equation. Show that these distributions are given by
Z Z
0 1 (d3)/2 d2
Kt (x)(x) dx = cd t ( t ) (t (t)d);
3 t d1
ZR Z S
1
Kt1 (x)(x) dx = cd ( t )(d3)/2 (td2 (t)d)
R 3 t S d1
for all test functions , where d is surface measure on the sphere S d1 , and cd
is a constant depending only on d. (Hint: since Kt0 and Kt1 are radial, it suffices
 2.3. DISPERSION AND STRICHARTZ ESTIMATES 73

to establish this when is also radial. Now apply Exercise 2.12 and Exercise 2.13.
Alternatively, one can compute the Fourier transforms of the distributions listed
above and verify that they agree with those of Kt0 and Kx1 . Yet another approach is
to show that the expressions given above solve the wave equation with the specified
initial data.) What happens in the case d = 1? (Use Exercise 2.13.)
Exercise 2.32 (Sharp Huygens principle). Let d 3 be odd, and let u
2
Ct,loc Sx (R Rd ) be a classical solution to the wave equation (2.9), such that the
initial data u0 , u1 is supported on a closed set Rd . Show that for any time t,
u(t) is supported on the set t := {x Rd : |x y| = ct for some y }.
Exercise 2.33 (Klein-Gordon fundamental solution). Let d 1 and c = ~ = 1,
let Kt0 and Kt1 be the fundamental solutions for the Klein-Gordon equation in d
dimensions, and let Kt0 and Kt1 be the fundamental solutions for the wave equation
in d + 1 dimensions. Establish the distributional identities
Z
Ktj (x1 , . . . , xn ) = Ktj (x1 , . . . , xn , xd+1 )eimxd+1 dxd+1
R
for j = 0, 1. (Hint: use Exercise 2.10.) In principle, this gives an explicit description
of the fundamental solution of the Klein-Gordon equation, although it is somewhat
unwieldy to work with in practice.

2.3. Dispersion and Strichartz estimates

Like as the waves make towards the pebbled shore, so do our min-
utes hasten to their end. (William Shakespeare, Sonnet 60)
In order to be able to perturb linear dispersive equations such as (2.1) to
nonlinear dispersive equations, it is crucial that we have some efficient ways to
control the size of solutions to the linear problem in terms of the size of the initial
datum (or of the forcing term, if one exists). Of course, to do this, one has to
quantify the notion of size by specifying a suitable function space norm. It turns
out for semilinear equations such as NLS and NLW, the mixed Lebesgue norms
Lqt Lrx (I Rd ), and more generally the mixed Sobolev norms Lqt Wxs,r (I Rd ), are
particularly useful13.
To make the discussion more concrete, let us consider the Schrodinger equation
(2.3) in Rd with ~ = m = 1, so the propagator operator is simply14 eit/2 . We first
ask what fixed-time estimates are available: if we fix a time t 6= 0, and we know
the initial datum u0 lies in some Sobolev space, what do we learn about the size of
the solution u(t) = eit/2 u0 ? To avoid technicalities let us take the solution to be
Schwartz; the general case can usually be handled by limiting arguments.
Since eit/2 is unitary, we obtain the L2x conservation law
(2.21) keit/2 u0 kL2x (Rd ) = ku0 kL2x (Rd )
and then (since eit/2 commutes with other Fourier multipliers)
keit/2 u0 kHxs (Rd ) = ku0 kHxs (Rd ) ;
13There are also Besov refinements of the Strichartz spaces which convey slightly more precise
information on the distribution of the solution among low, medium, and high frequencies, as well
as Lorentz refinements that extend a little more control over the distribution of large and small
values in physical space, but this is a more technical topic which we will skip lightly over here.
14The factor of 1/2 in the exponent is not particularly important, and all the estimates in
this section hold if it is omitted; we retain it for compatibility with other sections of the book.
74 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

one can also obtain this from Plancherels identity. From the fundamental solution
(2.18) and the triangle inequality we also have the dispersive inequality
d/2
(2.22) keit/2 u0 kL d .d t
x (R )
ku0 kL1x (Rd ) .
This shows that if the initial datum u0 has suitable integrability in space, then
the evolution will have a power-type decay in time; the L2x mass of the solution is
conserved, but is dispersed over an increasingly larger region as time progresses (see
Exercise 2.34). We can interpolate this (using the Marcinkeiwicz real interpolation
theorem, see e.g. [Sad]; one can also modify the arguments from Exercise A.5)
with (2.21) to obtain the further estimate
1 1
(2.23) keit/2 u0 kLp (Rd ) .d td( p 2 ) ku0 kLpx (Rd )
x

for all 1 p 2, where p is the dual exponent of p, defined by the formula

1 1 p q
p + p = 1. These are the complete range of Lx to Lx fixed-time estimates available
(see Exercise 2.35). In particular, the Schrodinger flow does not preserve any Lpx
norm other than the L2x norm. We can insert fractional differentiation operators as
before and conclude
1 1
keit/2 u0 kW s,p (Rd ) .d td( p 2 ) ku0 kWxs,p (Rd )
x

for all s R. By using Sobolev embedding, one can trade some of the regularity
on the left-hand side for integrability, however one cannot hope for any sort of
smoothing effect that achieves more regularity on the left-hand side than on the
left (see Exercise 2.36). This is in contrast with the smoothing effects of dissipative
propagators such as the heat kernels et .
These decay estimates are useful in the long-time asymptotic theory of nonlin-
ear Schrodinger equations, especially when the dimension d is large and the initial
datum u0 has good integrability properties. However in many situations, the initial
data is only assumed to lie in an L2x Sobolev space such as Hxs (Rd ). Fortunately,
by combining the above dispersive estimates with some duality arguments, one can
obtain an extremely useful set of estimates, known as Strichartz estimates, which
can handle this type of data:
Theorem 2.3 (Strichartz estimates for Schrodinger). [GV], [Yaj], [KTao] Fix
d 1 and ~ = m = 1, and call a pair (q, r) of exponents admissible if 2 q, r ,
2 d d
q + r = 2 and (q, r, d) 6= (2, , 2). Then for any admissible exponents (q, r) and
(q, r) we have the homogeneous Strichartz estimate
(2.24) keit/2 u0 kLqt Lrx (RRd ) .d,q,r ku0 kL2x (Rd )
the dual homogeneous Strichartz estimate
Z
(2.25) k eis/2 F (s) dskL2x (Rd ) .d,q,r kF kLq Lr (RRd )
t x
R
and the inhomogeneous (or retarded) Strichartz estimate
Z

(2.26) k ei(tt )/2 F (t ) dskLqt Lrx (Rd ) .d,q,r,q,r kF kLq Lr (RRd )
t x
t <t

The non-endpoint version of this theorem (when q, q 6= 2) had been established

in [GV], [Yaj], and of course the original work of Strichartz [Stri] (which in turn
had precursors in [Seg3], [Tomas]). The more delicate endpoint cases are treated
in [KTao]. The estimates are known to fail in a number of ways at the endpoint
 2.3. DISPERSION AND STRICHARTZ ESTIMATES 75

(q, r, d) = (2, , 2), see [Mon], although the homogeneous estimate can be salvaged
if one assumes spherical symmetry [Stef], [Tao2], [MNNO]. The exponents in
the homogeneous estimates are best possible (Exercise 2.42), but some additional
estimates are available in the inhomogeneous case [Kat8], [Fos].
Because the Schrodinger evolution commutes with Fourier multipliers such as
||s or his , it is easy to convert the above statements into ones at regularities
Hxs (Rd ) or Hxs (Rd ). In particular, if u : I Rd C is the solution to an inhomo-
geneous Schrodinger equation
1
it u + u = F ; u(0) = u0 Hxs (Rd ),
2
given by Duhamels formula (2.13) on some time interval I containing 0, then by
applying his to both sides and using the above theorem, we obtain the estimates

kukLqt Wxs,r (IRd ) .d,q,q,r,s ku0 kHxs (Rd ) + kF kLq W s,r (IRd )
t x

for any admissible (q, r) and (q, r), though one has to take some care with the
endpoint r = because of issues with defining Sobolev spaces there. Similarly if
we replace the Sobolev spaces Hxs , Wxs,r , Wxs,r by their homogeneous counterparts
Hxs , Wxs,r , Wxs,r . One can also use Sobolev embedding and (if I is bounded) Holders
inequality in time to increase the family of exponents for which Strichartz estimates
are available; see Figure 1 for some examples.
We shall give a proof of this non-endpoint cases of this theorem shortly, but we
first give an abstract lemma, the Christ-Kiselev lemma [CKis] which is very useful
in establishing retarded Strichartz estimates. A proof of the lemma as formulated
here can be found in [SSog] or [Tao2].

Lemma 2.4 (Christ-Kiselev lemma). Let X, Y be Banach spaces, let I be a time

interval, and let K C 0 (I I B(X Y )) be a kernel taking values in the space
of bounded operators from X to Y . Suppose that 1 p < q is such that
Z
k K(t, s)f (s) dskLqt (IY ) Akf kLpt (IX)
I

for all f Lpt (I X) and some A > 0. Then one also has
Z
k K(t, s)f (s) dskLqt (IY ) .p,q Akf kLpt(IX) .
sI:s<t

The principle that motivates this lemma is that if an operator is known to

be bounded from one space to another, then any reasonable localisation of that
operator (in this case, to the causal region s < t of time interactions) should also
be bounded. The hypothesis that p < q is unfortunately necessary; see Exercise
2.40.
We can now prove the non-endpoint cases of the Strichartz estimate.

Partial proof of Theorem 2.3. We shall only prove this theorem in the
non-endpoint cases when q, q 6= 2. In these cases we can argue using the T T
method (as was carried out for the closely related restriction problem in harmonic
analysis in [Tomas]) as follows. Let (q, r) be admissible. Applying Minkowskis in-
equality, (2.23) and the Hardy-Littlewood-Sobolev theorem of fractional integration
76 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

1/q

D
D

C1

B B
1
A1 A
1/r
Figure 1. The Strichartz game board for Schrodinger equations
10 30/13
in Hx1 (R3 ). The exponent pairs are A = L 2
t Lx , B = Lt Lx ,
6 4
D = Lt Lx , A1 = Lt Lx , B1 = Lt,x , C1 = Lt Lx , A = L1t L2x ,
2 6 10
6/5
D = L2t Lx . If the initial datum u0 lies in Hx1 (R3 ), and one
derivative of the forcing term F lies in a space the closed interval
between A and D , then one derivative of the solution u lies in
every space in the closed interval between A and D. Endpoint
Sobolev embedding can then move left, place the solution it-
self in any space between A1 and C1 (though the endpoint C1
requires a Besov modification to the Strichartz spaces due to fail-
ure of endpoint Sobolev embedding; see [CKSTT11]). If I is
bounded, Holder in time can also move up, lowering the r index
to gain a power of |I|. If one is working with Hx1 instead of Hx1
(or is restricting to high frequencies), then non-endpoint Sobolev
embedding is also available, allowing one to enter the pentagonal
region between AD and A1 C1 . If one restricts to low frequencies,
then Bernsteins inequality (A.5) can move further to the left than
A1 C1 .

(see (A.10)), we conclude that

Z Z
i(ts)/2
k e F (s) dskLt Lrx (RRd ) k
q kei(ts)/2 F (s)kLrx (Rd ) dskLqt (R)
R R
1
.d,r kkF kLrx (Rd ) 1 1 kL (R)
q
|t|d( 2 r ) t
.d,q,r kF kLq Lr (RRd )
t x

2
whenever 2 < r and 2 < q are such that + dr = d2 , and for any Schwartz
q
function F in spacetime. Applying Holders inequality, we conclude that
Z Z
| hei(ts)/2 F (s), F (t)i dsdt| .d,q,r kF k2 q r d
,
Lt Lx (RR )
R R
 2.3. DISPERSION AND STRICHARTZ ESTIMATES 77

R
where hF, Gi = Rd F (x)G(x) dx is the usual inner product. But the left-hand side
R is/2
factorises as k R e F (s) dsk2L2 (Rd ) , and thus we obtain the dual homogeneous
x
Strichartz estimate
Z
k eis/2 F (s) dskL2x (Rd ) .d,q,r kF kLq Lr (RRd )
t x
R

which is (2.25). By duality we then obtain (2.24). Composing those two estimates,
we conclude
Z
k eis/2 F (s) dskLqt Lrx (Rd ) .d,q,r,q,r kF kLq Lr (RRd )
t x
R

and (2.26) then follows from the Christ-Kiselev lemma. 

Strichartz estimates can be viewed in two ways. Locally in time, they describe
a type of smoothing effect, but reflected in a gain of integrability rather than
regularity (if the datum is in L2x , the solution u(t) is in Lrx with r > 2 for most of
the time), and only if one averages in time. (For fixed time, no gain in integrability
is possible; see Exercise 2.35.) Looking globally in time, they describe a decay effect,
that the Lrx norm of a solution u(t) must decay to zero as t , at least in some
Lqt -averaged sense. Both effects of the Strichartz estimate reflect the dispersive
nature of the Schrodinger equation (i.e. that different frequencies propagate in
different directions); it is easy to verify that no such estimates are available for the
dispersionless transport equation (2.2), except with the trivial pair of exponents
(q, r) = (, 2).
Remark 2.5. Readers who are familiar with the uncertainty principle (Prin-
ciple A.1) can interpret the Strichartz estimates for the homogeneous Schrodinger
equation as follows. Consider a solution u to the homogeneous Schrodinger equa-
tion with L2x norm O(1), and with frequency N (i.e. the Fourier transform is
supported in the region || N ), for each time t. The uncertainty principle shows
that at any given time t, the most that the solution u(t) can concentrate in physical
space is in a ball of radius 1/N ; the L2x normalisation then shows the solution can
be as large as N d/2 on this ball. However, the Strichartz estimates show (roughly
speaking) that such a strong concentration effect can only persist for a set of times
of measure 1/N 2 ; outside of this set, the solution must disperse in physical space
(compare with Proposition A.4). Note that this is also very consistent with Prin-
ciple 2.1, since concentration in a ball of radius 1/N would induce the frequency
uncertainty of N , hence a velocity uncertainty of N , which should disperse
the ball after time 1/N 2 .
Similar Strichartz estimates can be established for any linear evolution equation
which enjoys a dispersive estimate, such as the Airy equation. The wave equation
also obeys a dispersive inequality, see (2.19), (2.20), but the presence of the regular-
izing factor means that one requires some more harmonic analysis (in particular,
some Littlewood-Paley theory and the theory of complex analytic interpolation) in
order to utilise this estimate properly. Nevertheless, it is still possible to establish
suitable Strichartz estimates for the wave equation. Indeed, we have
Theorem 2.6 (Strichartz estimates for wave equation). Let I be a time interval,
and let u : I Rd C be a Schwartz solution to the wave equation u = F with
78 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

1/q A A
1

E
E

F1

A1 A
1/r
Figure 2. The Strichartz game board for wave equations in
Hx1 (R3 ). The exponent pairs are A = L 2 2
t Lx , E = Lt Lx , A1 =
6 6/5
Lt Lx , F1 = Lt Lx , A = Lt Lx , E = Lt Lx , A1 = L1t Lx . If
4 12 1 2 2 1
1 3 2 3
the initial data (u0 , u1 ) lies in Hx (R ) Lx (R ), and a suitable
derivative of the forcing term F on a space between A and E
(excluding E ), then a certain derivative of u lies in every space
between A and E (excluding E), and so by Sobolev u itself lies in
every space between A1 and E (excluding E). At the endpoint A ,
no derivatives on F are required. Also by Sobolev embedding, it
would have sufficed to place F in any space between A1 and E
(excluding E). Other moves, similar to those discussed in Figure
1, are available.

c = 1, and with initial data u(t0 ) = u0 , t u(t0 ) = u1 for some t0 I. Then we

have the estimates
kukLqt Lrx (IRd ) + kukCt0 Hxs (IRd ) + kt ukC 0 Hxs1 (IRd )
t
 
.q,r,s,n ku0 kH s (Rd ) + ku1 kHxs1 (Rd ) + kF kLq Lr (IRd )
x t x

whenever s 0, 2 q, q and 2 r, r < obey the scaling condition

1 d d 1 d
(2.27) + = s= + 2
q r 2 q r
and the wave admissibility conditions
1 d1 1 d1 d1
+ , + .
q 2r q 2r 4
For a proof of this theorem, see [Kat8], [GV2], [Kap], [LSog], [Sog], [SStru3],
[KTao]. Again, the exponents are sharp for the homogeneous version of these
estimates (i.e. with F = 0; see Exercise 2.43) but not the inhomogeneous; see [Har],
[Obe], [Fos] for some inhomogeneous estimates not covered by the above theorem.
The endpoint case r = contains a number of subtleties, see [FW] for some
 2.3. DISPERSION AND STRICHARTZ ESTIMATES 79

discussion. As with the Schrodinger equation, the Strichartz estimates for the wave
equation encode both local smoothing properties of the equation (the estimates
obtained from Strichartz lose fewer derivatives than the fixed-time estimates one
would get from Sobolev embedding), as well as global decay estimates, though
when the wave equation compared against the Schrodinger equation in the same
dimension, the estimates are somewhat inferior; this is due to the weaker dispersion
in the wave equation (different frequencies move in different directions, but not in
different speeds), the finite speed of propagation, and the concentration of solutions
along light cones. This estimate only controls the homogeneous Sobolev norm, but
the lower order term in the inhomogeneous Sobolev norm can often be obtained by
an integration in time argument.
An important special case of the wave equation Strichartz estimates is the
energy estimate
kukC 0 Hxs1 (IRd ) + kt ukC 0 Hxs1 (IRd )
t t
(2.28)
. kx u0 kHxs1 (Rd ) + ku1 kHxs1 (Rd ) + kF kL1 Hxs1 (IRd )
t

which can also be proven by other means, for instance via the Fourier transform
(Exercise 2.45), or by using the stress energy tensor (see Exercise 2.59). This
estimate has the useful feature of gaining a full degree of regularity; the forcing
term F is only assumed to have s 1 degrees of regularity, but the final solution u
has s degrees of regularity. One also has the useful variant
kukCt0 Hxs (IRd ) + kt ukC 0 Hxs1 (IRd )
t
(2.29)
.s h|I|i(ku0 kHxs (Rd ) + ku1 kHxs1 (Rd ) + kF kL1 Hxs1 (IRd ) ).
t

The other Strichartz estimates gain fewer than one full degree of regularity, but
compensates for this by improving the time and space integrability. One specific
Strichartz estimate of interest (and one which was discovered in the original paper
[Stri]) is the conformal Strichartz estimate for the wave equation
kukL2(d+1)/(d1) (RRd ) .d ku(0)kH 1/2 (Rd ) + kt u(0)kH 1/2 (Rd )
t,x x x
(2.30)
+ kukL2(d+1)/(d+3)(RRd ) ,
t,x

valid for d 2.
Strichartz estimates also exist for the Klein-Gordon equation but are more
difficult to state. See [MSW], [Nak4], [MN]; for short times one can just rely
on the wave equation Strichartz estimates (treating the lower order mass term as a
perturbation), while for long times it is easiest to proceed by treating the relativistic
and non-relativistic regimes separately. Some Strichartz estimates are also known
in periodic settings, but primarily of L4t L4x or L6t L6x type, and are proven using the
spacetime Fourier transform, as one cannot exploit dispersion in these settings (as
one can already see by considering plane wave solutions). See Section 2.6. More
generally, Strichartz estimates are closely related to restriction estimates for the
Fourier transform; see [Tao11] for a survey of the restriction phenomenon, which
has been used to obtain a number of refinements (such as bilinear refinements) to
the Strichartz inequalities.
Exercise 2.34 (Asymptotic Lpx behaviour of Schrodinger). Let u0 Sx (Rd )
be a non-zero Schwartz function whose Fourier transform u0 () is supported in the
ball || 1. Show that we have the estimate |eit u0 (x)| .N,u0 htiN hxiN for all
80 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

times t, all N > 0, and all x such that |x| 5|t|. (Hint: use rotational symmetry to
make x a multiple of the unit vector e1 , thus x = (x1 , 0, . . . , 0). Then use the Fourier
representation (2.12) of eit u0 (x), followed by repeated integration by parts in the
x1 direction.) In the region |x| < 5|t|, obtain the estimate |eit u0 (x)| .u0 htid/2 .
1 1
Conclude that keit u0 kLpx (Rd ) d,u0 ,p htid( p 2 ) for all 1 p . (Hint: first
establish this for p = 2, obtain the upper bound for all p, and then use Holders
inequality to obtain the lower bound.)
Exercise 2.35 (Necessary conditions for fixed-time Schrodinger). Suppose 1
p, q and R are such that the fixed-time estimate
keit/2 u0 kLqx (Rd ) Ct ku0 kLpx (Rd )
for all u0 Sx (Rd ) and t 6= 0, and some C independent of t and u0 . Using
scaling arguments, conclude that = d2 ( 1q p1 ). Using (2.34) (and time translation
invariance), conclude that q p and q = p (and thus 1 p 2). In particular the
Schrodinger operators are not bounded on any Lpx (Rd ) space except when p = 2.
Indeed, a good rule of thumb is that dispersive evolution operators only preserve
L2x-based spaces and no others (with the notable exception of the one-dimensional
wave evolution operators, which are not truly dispersive).
Exercise 2.36 (Schrodinger Strichartz cannot gain regularity). Using Galilean
invariance (Exercise 2.5), show that no estimate of the form
kei/2 u0 kWxs2 ,q (Rd ) Cku0 kWxs1 ,p (Rd )
or
keit/2 u0 kLqt Wxs2 ,r ([0,1]Rd ) Cku0 kWxs1 ,p (Rd )
can hold with C independent of u0 , unless s2 s1 .
Exercise 2.37 (Decay of finite energy Schrodinger solutions). Show that the
admissible space L 2
t Lx which appears in Theorem 2.3 can be replaced by the
slightly smaller space Ct0 L2x . Similarly, if u0 Hx1 (R3 ) and u : R R3 C
is the solution to the Schrodinger equation, show that limt ku(t)kLpx(R3 ) for
2 < p 6 and that
lim khxi u(t)kL2x + khxi u(t)kL2x = 0
t

for any > 0 (see [Tao8] for an application of these estimates to nonlinear
Schrodinger equations).
Exercise 2.38 (Pseudoconformal invariance of Strichartz). (Pieter Blue, pri-
vate communication) Show that if q, r are Schrodinger-admissible exponents, then
the space Lqt Lrx (R Rd ) is preserved under the pseudoconformal transformation
(Exercise 2.28). Conclude that Theorem 2.3 is invariant under the pseudocon-
formal transformation. (It is also invariant under space and time translations,
homogeneity, scaling, phase rotation, time reflection, spatial rotation and Galilean
transformations.)
Exercise 2.39 (Conformal invariance of Strichartz). Show that the conformal
Strichartz estimate (2.30) is invariant under space and time translations, homogene-
ity, scaling, phase rotation, time reflection, spatial rotation, Lorentz transformation
(Exercise 2.6), and conformal inversion (Exercise 2.14).
 2.3. DISPERSION AND STRICHARTZ ESTIMATES 81

Exercise 2.40. Show that Lemma 2.4 fails at the endpoint p = q, even when
X and Y are scalar. (Hint: take p = q = 2 and consider truncated versions of the
R (s)
Hilbert transform Hf (t) = p.v. fts ds.)
Exercise 2.41. Using Exercise 2.30, establish some Strichartz estimates for
the Airy equation. (As it turns out, when it comes to controlling equations of
Korteweg-de Vries type, these estimates are not as strong as some other estimates
for the Airy equation such as local smoothing and X s,b estimates, which we shall
discuss later in this chapter.)
Exercise 2.42. Using the scale invariance from Exercise 2.9, show that the
condition q2 + dr = d2 is necessary in order for (2.24). Next, by superimposing
multiple time-translated copies of a single solution together (thus replacing u(t, x)
PN
by j=1 u(t tj , x) for some widely separated times t1 , . . . , tN ) show that the
condition q 2 is also necessary. (The double endpoint (q, r, d) = (2, , 2) is
harder to rule out, and requires a counterexample constructed using Brownian
motion; see [Mon].) Note that similar arguments will establish the necessity of
(2.27), as well as the requirements q, q 2 in Theorem 2.6.
Exercise 2.43 (Knapp example). Consider the solution u : R1+d C to the
wave equation with c = 1 given by
Z
u(t, x) := eix eit|| d
11 2;|2 |,...,|n |

where 0 < < 1 is a small number. Show that u(t, x) is comparable in magnitude
to d1 whenever |t| 1/2 , |x1 + t| 1 and |x2 |, . . . , |xn | 1/. Use this to
conclude that the condition q1 + d1 d1
2r 4 in Theorem 2.6 cannot be improved.
One can also obtain essentially the same example by starting with a bump function
initial datum, and applying a Lorentz transform (see Exercise 2.6) with velocity
v := (1 2 )e1 ; the strange rectangular region of spacetime obtained above is then
explainable using the familiar phenomena of time dilation and length contraction
in special relativity.
Exercise 2.44 (Stein example). Let 1/2 < 1, and let g L2x (R3 ) be the
function
1B(2e3 ,2)\B(e3 ,1) (x)
g(x) := ,
|x|2 hlog |x|i
where e3 is the third unit vector of R3 and B(x, r) denotes the ball of radius
r

centred at x. Show that kgkL2x(R3 ) = O (1), but that the solution u(t) := sin(t


g
to the homogeneous wave equation with initial position 0 and initial velocity g obeys
u(te3 , t) = for all 1 t 2. (Hint: use Exercise 2.31.) In particular, Theorem
2.6 fails at the endpoint (q, r) = (2, ), a fact first observed (with a more indirect
proof) in [KM].
Exercise 2.45. Prove the energy estimate (2.28) using (2.14), the Fourier
transform, and Minkowskis inequality. Note that the implied constant in the .
notation is absolute (not depending on s or d). Then deduce (2.29).
Exercise 2.46 (Besov version of Strichartz estimates). For each dyadic num-
ber N , let PN be the Littlewood-Paley multiplier at frequency N , as defined in
82 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Appendix A. By means of complex interpolation, establish the inequalities

X
kukLqt Lrx (IRd ) .q,r ( kPN uk2Lq Lr (IRd ) )1/2
t x
N

whenever 2 q, r (so in particular whenever q, r are admissible exponents for

Strichartz estimates), as well as the dual estimate
X
( kPN F k2 q r d
)1/2 .q,r kF kLq Lr (IRd )
Lt Lx (IR ) t x
N

for the same range of q, r. (Note that to apply interpolation properly for the first
inequality, you will need to write u as an appropriate linear operator applied to
the functions uN = PN u.) By exploiting the fact that PN commutes with the
Schrodinger operator it + , establish the estimate
X
( kPN eit/2 u0 k2Lqt Lrx (RRd ) )1/2 .d,q,r ku0 kL2x (Rd )
N

for all Schrodinger-admissible q, r. Similarly establish analogues of (2.25) and

(2.26).

2.4. Conservation laws for the Schrodinger equation

Knowledge about life is one thing; effective occupation of a place
in life, with its dynamic currents passing through your being, is
another. (William James, The Varieties of Religious Experience)
In Hamiltonian ODE, one often encounters conserved quantities E(t) of the flow
such as energy, momentum, mass, and so forth. One has similar conserved quan-
tities in Hamiltonian PDE, but they come with additional structure. Namely, the
typical conserved quantity E(t) that one encounters is not just a scalar
R quantity, but
is also an integral of some density e0 (t, x), thus for instance E(t) = Rd e0 (t, x) dx.
The conservation of E(t) can then be manifested in a more local form by the point-
wise conservation law 15
(2.31) t e0 (t, x) + xj ej (t, x) = 0
16
for some vector-valued quantity ej , which is thus the current associated to the
density e0 . The conservation of E then follows (formally, at least) by integrating
the continuity equation in space and using the divergence theorem (assuming some
suitable spatial decay on the current, of course). Thus in PDE, conservation laws
come in both integral and differential forms. The differential form of the conserva-
tion law is significantly more flexible and powerful than the integral formulation, as
it can be localised to any given region of spacetime by integrating against a suitable
cutoff function (or contracting against a suitable vector field). Of course, when one
does so, one may no longer get a perfectly conserved quantity, but one often obtains
a quantity which is almost conserved (the derivative in time is small or somehow
lower order) or perhaps monotone. Thus from a single conservation law one can
generate a variety of useful estimates, which can serve to constrain the direction of

15Roman indices such as j and k will be summed over the spatial indices 1, . . . , d in the usual
manner; Greek indices such as and will be summed over 0, . . . , d.
16In some cases E and e are themselves vector-valued instead of scalar, in which case e
0 j
will be tensor-valued.
 2.4. CONSERVATION LAWS FOR THE SCHRODINGER EQUATION 83

propagation of a solution, or at least of various components of that solution (e.g.

the high frequency components).
These conservation laws are particularly useful for controlling the long-time
dispersive behaviour, or short-time smoothing behaviour, of nonlinear PDE, but
to illustrate the concepts let us just work first with the linear analogues of these
PDE. To make the discussion concrete, we focus on the Schrodinger equation (2.3)
with ~ = m = 1, thus
1
(2.32) it u + u = 0;
2

to avoid technicalities we consider only smooth solutions u Ct,loc Sx (R Rd ). In
practice one can usually extend the estimates here to rougher solutions by limiting
arguments or duality.
Before we discuss the differential form of the conservation laws, let us first
recall the integral form of the conservation laws and their connection (via Noethers
theorem) to various symmetries (cf. Table 1 from Chapter 1); we will justify their
conservation later when considering the differential form of these laws, and also
more informally in Exercise 2.47 below. The symmetry of phase rotation, u(t, x) 7
ei u(t, x), leads to the scalar conserved quantity
Z
M (t) := |u(t, x)|2 dx,
Rd

which is variously referred to as the total probability, charge, or mass in the lit-
erature. The symmetry of space translation, u(t, x) 7 u(t, x x0 ), leads to the
vector-valued conserved quantity p~(t), defined in coordinates as
Z
pj (t) := Im(u(t, x)xj u(t, x)) dx,
Rd

which is the total momentum. The symmetry of time translation, u(t, x) 7 u(t
t0 , x), leads to the conserved quantity
Z
1
E(t) := |u(t, x)|2 dx,
2 Rd
known as the total energy or Hamiltonian. The symmetry of Galilean invariance
(Exercise 2.5) leads to the conserved quantity
Z
x|u(t, x)|2 dx t~
p(t),
Rd

which is the normalised centre-of-mass. The pseudo-conformal symmetry (Exercise

2.28) does not directly lead to a conserved quantity, as it is a discrete symmetry
rather than a continuous one. However, it can be used to conjugate some of the pre-
ceding symmetries and generate new ones. The pseudo-conformal symmetry leaves
the mass invariant, as that symmetry commutes with phase rotation; it also swaps
momentum conservation with the conservation of normalised centre-of-mass, as the
pseudo-conformal symmetry intertwines spatial transation and Galilean invariance.
The energy conjugates to the pseudo-conformal energy
1
(2.33) k(x + it)u(t)k2L2x (Rd )
2
84 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

which is also conserved. In particular, we see from the triangle inequality and
conservation of energy that

(2.34) k|x|u(t)kL2x (Rd ) k|x|u(0)kL2x (Rd ) + tku(0)kL2x(Rd ) ,

which is an estimate establishing a kind of approximate finite speed of propagation

result for finite energy solutions. See also Exercise 2.49.
Now we analyze the local versions of these conserved quantities. It is convenient
to introduce the pseudo-stress-energy tensor T , defined for , = 0, 1, . . . , n and
on the spacetime R Rd by

T00 = |u|2
T0j = Tj0 = Im(uxj u)
1
Tjk = Re(xj uxk u) jk (|u|2 )
4

for all j, k = 1, . . . , d, where jk is the Kronecker delta (thus it equals 1 when j = k

and zero otherwise). A direct computation using (2.32) then verifies (for smooth
fields u, at least) the local conservation laws

(2.35) t T00 + xj T0j = 0; t Tj0 + xk Tjk = 0.

The quantity T00 is known as the mass density, the quantity T0j = Tj0 is known
as both the mass current and the momentum density, and the quantity Tjk is the
momentum current or stress tensor. For smooth, rapidly decreasing solutions u,
this verifies the conservation of mass and momentum asserted earlier, since
Z Z
M (t) = T00 (t, x) dx; pj (t) = T0j (t, x) dx.
Rd Rd

Conservation of energy also has a local formulation, though not one related to the
pseudo-stress-energy tensor17 T (Exercise 2.48).
By multiplying first equation in (2.35) by a smooth function a(x) of at most
polynomial growth, integrating in space, and then integrating by parts, we obtain
the identity
Z Z
2
(2.36) t a(x)|u(t, x)| dx = xj a(x)Im(u(t, x)xj u(t, x)) dx.
Rd Rd

This identity can be used to control the local rate of change of mass by quantities
such as the energy (see Exercises 2.50, 2.51).

17See however Exercise 3.30.

 2.4. CONSERVATION LAWS FOR THE SCHRODINGER EQUATION 85

One can also obtain a useful identity by differentiating (2.36) in time again and
using (2.35) to obtain the useful identity
Z Z
t2 a(x)|u(t, x)|2 dx = t xj a(x)T0j (t, x) dx
Rd Rd
Z
= xj a(x)xk Tjk (t, x) dx
d
Z R
(2.37) = (xj xk a(x))Tjk (t, x) dx
d
ZR
= (xj xk a(x))Re(xj uxk u) dx
Rd
Z
1
|u(t, x)|2 2 a(x) dx.
4 Rd
This identity has some useful consequences for various values of a. For instance,
letting a(x) = 1 we recover mass conservation, and letting a(x) = xj we recover
the conservation of normalised centre of mass. Setting a(x) = |x|2 , we obtain the
virial identity
Z Z
2 2 2
(2.38) t |x| |u(t, x)| dx = 2 |u(t, x)|2 dx = 4E(t);
Rd Rd
R
thus the energy controls the convexity of the quantity Rd |x|2 |u(t, x)|2 dx; compare
with the classical counterpart in Example 1.31. (Actually, this identity can easily
be seen to be equivalent to the pseudo-conformal conservation law after using mass
and momentum conservation.) Setting18 a(x) = |x|, we obtain
Z Z
x
t Im(u(t, x) u(t, x)) dx = (xj xk a(x))Re(xj uxk u) dx
Rd |x| Rd
(2.39) Z
1
|u(t, x)|2 2 a(x) dx.
4 Rd
Now observe that
/ u|2 /|x|,
(xj xk a(x))Re(xj uxk u) = |
x
where |/ u|2 := |u|2 | |x| u|2 is the angular component of the gradient. If we
specialise to three dimensions, we also have 2 a = 4. If we integrate in time,
we thus obtain the Morawetz identity
(2.40)
Z T Z Z T
/ u(t, x)|2
|
dxdt + 4 |u(t, 0)|2 dt
T R3 |x| T
Z Z
x x
= Im(u(T, x) u(T, x)) dx Im(u(T, x) u(T, x)) dx
R3 |x| R3 |x|
for any time T > 0. Using Lemma A.10, and observing (using the Fourier trans-
form) that ku(T )kH 1/2(R3 ) = ku(0)kH 1/2 (R3 ) , we conclude the Morawetz estimate
x x
Z Z Z
/ u(t, x)|2
|
(2.41) dxdt + |u(t, 0)|2 dt . ku(0)k2H 1/2 (R3 ) .
R R3 |x| R x

18Strictly speaking, (2.37) does not directly apply here because a is not smooth, but this can
be fixed by a regularisation argument; see Exercise 2.57.
86 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

This estimate can be thought of as a variant of a Strichartz estimate, obtaining

some smoothing and decay near the spatial origin x = 0; it should be compared
with Example 1.32. It has a significant advantage over the Strichartz estimate, in
that it extends with only minor modification to the case of defocusing nonlinear
Schrodinger equations; see Section 3.5, in which we also discuss an interaction
version of the above estimate. One can also extend this three-dimensional estimate
to higher dimensions. For lower dimensions there are some additional subtleties;
see [Nak2].
2 d
Exercise 2.47. Let us formally consider R Lx (R C) as a symplectic phase
space with symplectic form (u, v) = 2 Rd Im(u(x)v(x)) dx. Show that the
Schrodinger equation (2.3) with ~ = m = 1 is then the formalR Hamiltonian flow
associated to the (densely defined) Hamiltonian H(u) := 12 Rd |u|2 dx. Also use
this flow to formally connect the symmetries and conserved quantities mentioned
in this section via Noethers theorem (ignoring irrelevant constants such as factors
of 2). See [Kuk3] for a more rigorous treatment of this infinite-dimensional Hamil-
tonian perspective, and [SSul] for a Lagrangian perspective of the material in this
section.
Exercise 2.48. Obtain a local conservation law (2.31) for the energy density
e0 = 12 |u|2 for the Schrodinger equation.

Exercise 2.49 (Local smoothing from localised data). Let u Ct,loc Sx (R
R ) be a smooth solution to the Schrodinger equation (2.32) with ~ = m = 1. By
d

using mass conservation and the pseudo-conformal conservation law, establish the
bound
hRi
ku(t)kL2x (BR ) .d khxiu(0)kL2x (Rd )
|t|
for all t 6= 0 and R > 0, where BR := {x Rd : |x| R} is the spatial ball of radius
R. This shows that localisation of initial data leads to a local gain of regularity
(by a full derivative, in this case) at later times, together with some local decay in
time.

Exercise 2.50 (Local near-conservation of mass). Let u Ct,loc Sx (R Rd )
be a smooth solution to the Schrodinger equation (2.32) with ~ = m = 1, and with
energy E = 12 ku(0)k2L2 (Rd ) . Show that for any R > 0 and t 6= 0 we have
x
Z Z
E 1/2 |t|
( |u(t, x)|2 dx)1/2 ( |u(0, x)|2 dx)1/2 + Od ( ).
|x|R |x|2R R
(Hint: apply (2.36) with a(x) = 2 (x/R), where is a bump function supported on
the ball of radius 2 which equals 1 on Rthe ball of radius 1, and obtain a differential
inequality for the quantity M (t) := ( Rd 2 (x/R)|u(t, x)|2 dx)1/2 .) This estimate
(and the generalisation in Exercise 2.51 below) is especially useful for the energy-
1/2
critical nonlinear Schrodinger equation, as the error term of CE R |t| depends only
on the energy of u and not on other quantities such as the mass; see [Gri5], [Bou7],
[Bou9], [Tao9] for some applications of this estimate.
Exercise 2.51 (Local near-conservation of mass, II). [Bou7], [Bou9] Let the
notation and assumptions be as in Exercise 2.50. Prove the more general inequality
Z Z
E 1/2 |t|
( |u(t, x)|2 dx)1/2 ( |u(0, x)|2 dx)1/2 + Od ( )
|x|R |x|2n R Rn1/2
 2.4. CONSERVATION LAWS FOR THE SCHRODINGER EQUATION 87

for any integer n 1; this improves the error term at the cost of enlarging the ball
in the main term. (Hint: We may use time reversal symmetry to take t > 0. Use
R t Rof energy and the pigeonhole principle to locate an integer 1 j n
conservation
such that 0 2j R|x|2j+1 R |u(t, x)|2 dxdt . Et/n. Then apply the argument used
to establish Exercise 2.50, but with R replaced by 2j R.) This example shows that
one can sometimes exploit the pigeonhole principle to ameliorate boundary effects
caused by cutoff functions; for another instance of this idea, see [CKSTT11]. In
this particular case it turns out that one can use Exercise 2.50 (with R replaced by
2n1 R) to replace the n1/2 denominator by the much stronger 2n term, however
there are other situations in which the pigeonhole principle is the only way to obtain
a satisfactory estimate.

Exercise 2.52 (Weighted Sobolev spaces). For any integer k 0, define the
weighted Sobolev space Hxk,k (Rd ) be the closure of the Schwartz functions under
the norm
k
X
kukHxk,k (Rd ) := khxij ukHxkj (Rd ) .
j=0

Establish the estimate

keit/2 f kHxk,k (Rd ) .k,d htik kf kHxk,k (Rd )

for all Schwartz f , either using the Fourier transform or by using mass conser-
vation laws (and higher order variants of these laws, exploiting the fact that the
Schrodinger equation commutes with derivatives). (Hint: if you are really stuck,
try working out the k = 1 case first.)

Exercise 2.53 (Local smoothing for Schrodinger). [Sjo],[Veg],[CS] Let u

Ct,loc Sx (R R3 ) be a smooth solution to the Schrodinger equation (2.32) with
~ = m = 1. Establish the homogeneous local smoothing estimate
Z Z
hxi1 |u(t, x)|2 + hxi3 |u(t, x)|2 dxdt . ku(0, x)k2H 1/2 (R3 )
x
R R3

for all > 0. (Hint: One can take to be small. Then adapt the Morawetz
argument in the text with a(x) := hxi hxi1 .) This shows a local gain of half a
derivative for the homogeneous Schrodinger equation. Give an informal explanation
as to why this estimate is consistent with the Heisenberg law v = that arises from
Principle 2.1; compare it also with Example 1.33.

Exercise 2.54 (Local smoothing for Schrodinger, II). Let u C 3

t,loc Sx (RR )
be a smooth solution to the inhomogeneous Schrodinger equation

it u + u = F.

Establish the dual local smoothing estimate

Z Z
sup ku(t)kH 1/2 (R3 ) . ku(0)kH 1/2 (R3 ) + hxi1+ |F (t, x)|2 dxdt
x x
tR R R3
88 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

for any > 0. (Hint: use Exercise 2.53 and Duhamels formula.) Then establish
the retarded local smoothing estimate
Z Z
hxi1 |u(t, x)|2 + hxi3 |u(t, x)|2 dxdt
R R 3
Z Z
. ku(0, x)k2H 1/2 (R3 ) + hxi1+ |F (t, x)|2 dxdt.
x
R R3

(Hint: use the same argument19 as in Exercise 2.53. An additional interaction term
between u and F will appear, and can be controlled using Cauchy-Schwarz.) This
shows a local gain of a full derivative for the in homogeneous Schrodinger equation.

Exercise 2.55 (Local smoothing for Airy equation). [Kat2], [KF] Show that

smooth solutions u Ct,loc Sx (R R R) to the Airy equation t u + x3 u = 0
obey the conservation law

(2.42) t (u2 ) = x3 (u2 ) + 3x (u2x )

where ux := x u, and use this to conclude the local smoothing estimate

Z TZ Z
2 T + R3
(2.43) ux dxdt . u(0, x)2 dx
0 |x|R R2 R

for all T > 0 and R > 0. (Hint: first integrate (2.42) against the constant function
1 to establish L2x conservation, and then integrate instead against a suitable cutoff
function which equals 1 for x > 2R, zero for x < 2R, and increases steadily for
R < x < R.)

Exercise 2.56 (Sharp local smoothing for Airy equation). [KPV2] With the
notation as in the preceding exercise, prove the sharper estimate
Z Z
ux (t, x0 )2 dt . u(0, x)2 dx
R R
3
for any x0 R, so that the factor T +R
R2 in (2.43) can be replaced with R. (Hint: use
translation invariance to set x0 = 0, and use the Fourier representation of ux (t, 0),
followed by Plancherels theorem. Unlike (2.43), it seems difficult to establish this
estimate purely using conservation law techniques, thus suggesting some limitations
to that method.) Give an informal explanation as to why (2.43) is consistent
with the dispersion relation v = 3 2 that arises from Principle 2.1. What is the
analogous estimate for the one-dimensional Schrodinger equation?

Exercise 2.57. Justify the derivation of (2.39) from (2.37) (for Ct,loc S solu-
p x
tions u to the Schrodinger equation) by applying (2.37) with a(x) := 2 + |x|2
and then taking limits as 0. These types of regularisation arguments are quite
common in the theory of linear and nonlinear PDE, and allow one to extend the
validity of many formal computations well beyond the regularities that would be
needed to justify them classically.

19We thank Jared Wunsch for pointing out this simple argument.
 2.5. THE WAVE EQUATION STRESS-ENERGY TENSOR 89

2.5. The wave equation stress-energy tensor

A man must drive his energy, not be driven by it. (William Fred-
erick Book)
Having considered conservation laws for the Schrodinger equation, we turn to
the wave equation
(2.44) u = F,

where we take smooth scalar fields u Ct,loc Sx (R1+d C) for simplicity. In
particular u has sufficient regularity and decay at infinity to justify all integration
by parts computations.
While it is possible to view the wave equation as a Hamiltonian system (see
Exercise 2.58), the geometry of Minkowski space suggests that one can also alter-
nately view the wave equation (and the associated conservation laws) in a more
Lorentz-invariant way; thus our perspective here will be slightly different from
that in the preceding section. Indeed, one can view the wave equation in a La-
grangian manner
R instead, viewing u as a formal critical point of the Lagrangian
functional 12 R1+d u u dg. This functional is formally invariant under diffeo-
morphic changes of variable; by considering the variation of the functional along
an infinitesimal such change of variable (see Exercise 2.60) one can then construct
a stress-energy tensor T for this equation, which in this case is
1
(2.45) T := Re( u u) g Re( u u).
2
In coordinates with the normalisation c = 1, we have
1 1
T00 = T00 = |t u|2 + |u|2
2 2
T0j = T0j = Re(t uxj u)
jk
Tjk = Tjk = Re(xj uxk u) (|u|2 |t u|2 ).
2
This tensor is real and symmetric. The quantity T00 is known as the energy den-
sity, the quantity T0j is the energy current or momentum density, and Tjk is the
momentum current or stress tensor. The similarity with the Schrodinger pseudo-
stress-energy tensor is not accidental; see Exercise 3.30.
The tensor T is clearly symmetric, and a quick computation using (2.44)
yields the divergence equation
(2.46) T = Re(( u)F ).
Henceforth we consider the homogeneous equation F 0, so that T is divergence
free. In coordinates, we thus have
t T00 + xj T0j = 0; t T0j + xk Tjk = 0
(compare with (2.35)). This already yields conservation of the total energy
Z
E[u(t)] = E(t) := T00 (t, x) dx
Rd
and the total momentum
Z
j
p (t) := T0j (t, x) dx,
Rd
90 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

assuming sufficient spatial decay of course on the solution u and its derivatives.
It turns out that these conservation laws can be nicely localised in spacetime
by exploiting the positivity property
(2.47) T v v 0
whenever v is a time-like or light-like vector (so v v 0). Indeed from (2.45) we
have
1
T v v = (v u)2 (v v )( u)( u)
2
which is clearly non-negative in the light-like case v v = 0, and in the timelike
case one can check positivity by dividing the gradient u into the component
parallel to v and the component Minkowski-orthogonal to v (on which the metric
g is spacelike, hence non-negative); alternatively one can use Lorentz invariance to
reduce to the case20 v = t . More generally we have T v w 0 whenever v, w
are time-like or light-like, and both future-oriented (see Exercise 3.41). To exploit
this positivity, we can use Stokes theorem to obtain the identity
Z Z Z
(2.48) T X n dS = T X n dS + (T X )dg
1 0
for an arbitrary smooth vector field X , where is an open region in spacetime
bounded below by a spacelike hypersurface 0 and above by a spacelike hypersur-
face 1 , n is the positive timelike unit normal and dS is the induced measure from
the metric g (which is positive on the spacelike surfaces 0 , 1 ); if is unbounded
then of course we need to assume suitable decay hypotheses on u or X. For in-
stance, if 0 = {(t, x) : t = 0}, 1 = {(t, x) : t = t1 } and X = t for some arbitrary
time t1 R we obtain the conservation of energy E(t1 ) = E(0), while if we instead
take X = xj we obtain conservation of momentum pj (t1 ) = pj (0). Now suppose
that T > t1 > 0, and we take 0 to be the disk {(t, x) : t = 0, |x| T } and 1 to
be the truncated cone21
1 := {(t, x) : 0 < t < t1 , |x| = T t} {(t, x) : t = t1 , |x| T t1 }.
Setting X = t , we conclude the energy flux identity
Z Z
(2.49) T00 (t1 , x) dx + FluxT [0, t1 ] = T00 (0, x) dx
|x|T t1 |x|T

where FluxT [t0 , t1 ] is defined for 0 t0 < t1 < T by

Z
FluxT [t0 , t1 ] := T X n dS.
t0 <t<t1 ,|x|=T t

Intuitively, this identity asserts that the energy at the top of the truncated cone,
plus the energy flux escaping the sides of the cone, is equal to the original energy
at the base of the cone. From (2.47) we see that FluxT [0, t1 ] is non-negative, and
so we have the localised energy monotonicity formula
Z Z
T00 (t1 , x) dx T00 (0, x) dx
|x|T t1 |x|T

20We of course equate a vector field X with the associated first-order differential operator
X in the usual manner.
21Strictly speaking, is not quite spacelike, which causes dS to degenerate to zero and n
1
to elongate to infinity. But the area form n dS remains well defined in the limit; we omit the
standard details.
 2.5. THE WAVE EQUATION STRESS-ENERGY TENSOR 91

In particular, we have obtained the finite speed of propagation property: if u solves

the wave equation, and vanishes on the ball |x| T at time t = 0, then it vanishes
on the smaller ball |x| T t1 for all times 0 < t1 < T ; this reflects the well-known
fact that solutions to wave equations cannot propagate faster than the speed of light
(which has been normalised to c = 1 here). Also, from the energy flux identity and
energy conservation we have
(2.50) FluxT [0, t1 ] E(0);
thus if the solution has finite energy, then FluxT [0, t1 ] is monotone increasing in
t1 and is also bounded. It therefore converges22 to some limit as t1 T ; since
FluxT [t0 , t1 ] = FluxT [0, t1 ] FluxT [0, t0 ], we conclude in particular the flux decay
property
lim FluxT [t0 , t1 ] = 0.
t0 ,t1 T

This shows that near the tip (T , 0) of the cone, an asymptotically vanishing amount
of energy will escape the sides of the cone. There is however still the possibility of
energy concentration, in which the energy stays inside the cone and concentrates to
the tip as t T ; we shall discuss this possibility further in Section 5.1.
To exploit (2.48) and (2.47) further, it is of interest to understand the diver-
gence23 of vector fields of the form T X . Indeed we see from (2.46) and the
symmetry of T that the vector field T X has divergence
1
(2.51) (T X ) = T
2
where is the deformation tensor
(2.52) := X + X = LX g ,
where LX denotes the Lie derivative. In particular, if the vector field X is a
Killing vector field (i.e. the diffeomorphism induced by X preserves the metric),
then LX g = 0 and hence T X is divergence-free. In particular, we obtain
conservation of the quantity
Z
T0 X (t, x) dx
Rd
For instance, the vector fields t and xj , which correspond to time translation and
spatial translation respectively, yield the conservation of energy and momentum
respectively. If instead we take the rotation vector field xj xk xk xj for some
fixed 1 j < k n, which is clearly Killing, we obtain the conservation of angular
momentum
Z Z
xj T0k (t, x) xk T0j (t, x) dx = Re t u(t, x)(xk xj xj xk )u(t, x) dx.
Rd Rd

22This innocuous statement from basic real analysis - that every monotone bounded sequence
converges - is surprisingly useful in the analysis of PDE, as it allows one to convert a monotonicity
formula (say for a sequence an ) into a decay estimate (for the Cauchy differences an am ). The
drawback is that the decay is qualitative only; there is no uniform control on the decay of the
an am , although one can see that these differences cannot be too large for too many disjoint
intervals [n, m].
23We thank Markus Keel and Sergiu Klainerman for sharing some unpublished notes on this
topic, which were very helpful in preparing the material in this section.
92 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Taking instead the Lorentz vector field xj t + txj , which is also Killing, we obtain
conservation of normalised centre-of-mass
Z Z
xj T00 (t, x) tT0j (t, x) dx = xj T00 (t, x) dx tpj .
Rd Rd
Thus Lorentz invariance plays the role in the wave equation that Galilean invariance
does for the Schrodinger equation.
Unfortunately, there are no further Killing vector fields for Minkowski space
(other than taking linear combinations of the ones listed above); this is the hyper-
bolic analogue of the well-known fact that the only orientation-preserving isometries
of Euclidean space (i.e. the rigid motions) are the translations and rotations, and
combinations thereof; see Exercise 2.62. However, there is a slightly larger class of
conformal Killing vector fields, where the deformation tensor does not vanish,
but is instead a scalar multiple of g , thus = g for some scalar function
. Inserting this and (2.45) into (2.51), and observing that g g = (d + 1), we
conclude that for a conformal Killing field we have
d1
(T X ) = Re( u u).
4
Using the equation (2.44), we can rewrite this further as
d1
(T X ) = (|u|2 ),
8
which rearranges as
d1 2
(2.53) P = |u| 
8
where P is the vector field
d1
P := T X + ( (|u|2 ) ( )|u|2 ).
8
This is quite close to being a conservation law; note that P contains terms which
are quadratic in the first derivatives of u, but the divergence of P only contains
terms of zeroth order in u.
To give some examples of (2.53), let us first consider the Morawetz vector field
X associated to (t2 +|x|2 )t +2txj xj (i.e. X0 = (t2 +|x|2 ) and Xj = 2txj ); this is
the pullback of the time translation vector field t under the conformal inversion
in Exercise 2.14. This vector field is easily verified to be conformal Killing with
= 4t. Since (4t) = 0, we thus see that the vector field
d 1 0 2
P := (t2 + |x|2 )T0 + 2txj Tj + (d 1)tReu u g |u|
2
is divergence free. In particular we see that the conformal energy
Z
Q[u[t], t] := (t2 + |x|2 )T00 (t, x) 2txj T0j (t, x) + (d 1)tReut u(t, x)
R d
(2.54)
d1
|u(t.x)|2 dx
2
is preserved in time, and in particular is equal to
Z
d1 2
Q[u[0], 0] = |x|2 T00 (t, x) |u| (0, x) dx.
R d 2
 2.5. THE WAVE EQUATION STRESS-ENERGY TENSOR 93

This conservation law is the analogue of the conservation of pseudo-conformal en-

ergy (2.33) for the Schrodinger equation (cf. the last part of Exercise 2.28). The
quantity Q[u[t], t] is not obviously non-negative, but it can eventually be rearranged
x x
using the null frame vector fields L := t + |x| x , L := t |x| x as
Z
1
|(t + |x|)Lu + (d 1)u|2 + |(t |x|)Lu + (d 1)u|2 + 2(t2 + |x|2 )|/ u|2 dx;
4 Rd
see for instance [Kla3] (for a related computation, see Exercise A.16). One thus
obtains the decay laws
k(t + |x|)Lu(t) + (d 1)u(t)kL2x (Rd ) , k(t |x|)Lu(t) + (d 1)u(t)kL2x (Rd ) ,
k(t + |x|)|
/ u|kL2x (Rd ) . k|x|x,t u(0)kL2x (Rd )
(compare with (2.34)). On one hand, these estimates are somewhat weak compared
with the decay of t(d1)/2 ordinarily expected for the wave equation; the decay
obtained here is more of the order of 1/t (although near the light cone one does
not obtain any decay in the L direction). On the other hand, the argument here
is extraordinarily robust and in particular extends to very rough manifolds; see
[Kla3] for some further discussion.
Another application of (2.53) arises by taking the scaling vector field x ,
which is conformal Killing with = 2. This shows that the vector field
d1
(2.55) T x + Reu u
2
is divergence free, and thus
Z
d1
tT00 xj T0j Reut u dx
Rd 2
is a conserved quantity. This particular conservation law is not of much direct use,
as the top order terms tT00 xj T0j do not have a definite sign24. However, if one
localises this law to the cone {|x| t}, then one recovers some positivity (at least
to top order), and one can obtain a useful estimate, especially in the context of
establishing non-concentration for the semilinear wave equation. See Section 5.1.
Another use of Killing vector fields X (introduced by Klainerman [Kla]) lies
in the fact that they commute with the dLambertian , in the sense that
(X u) = X (u)
for all smooth u. This can be seen by direct computation (using Exercise 2.62)
or by noting that the dLambertian is determined entirely by the metric g, and
is hence preserved by the infinitesimal diffeomorphism associated to X . We can
iterate this and obtain
(K1 . . . Kk u) = K1 . . . Kk u
whenever K1 , . . . , Kk are one of the Killing first order differential operators t , xj ,
xj xk xk xj , or txj + xj t . In particular, if u is a smooth solution to the wave
equation (2.44), then so does K1 . . . Kk u. In particular we have energy conservation
E[K1 . . . Kk u](t) = E[K1 . . . Kk u](0).

24A general principle is that if the top order terms in an expression are ill-behaved, then no
amount of structure arising from the lower-order terms will rescue this. Thus one should always
attend to the top order terms first.
94 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

given sufficient decay of u and its derivatives, of course.

Let us apply this to a smooth solution with initial data u(0), t u(0) supported
in the ball {|x| 1}; by finite speed of propagation we then see that for future
times t > 0, the solution is supported in the cone {|x| 1 + t}. Then we see that
(2.56) E[K1 . . . Kk u](t) .k,u 1
for all k 0 and all Killing vector fields K1 , . . . , Kk . This can be used to establish
some decay of the solution u. Let us fix a time t 1 and look at the region
t/2 < |x| 1 + t using polar coordinates x = r for some t/2 < r 1 + t and
S d1 . We can then conclude from (2.56) (using the Killing vector fields xj
and xj xj xk xj ) that
Z Z
|l rm x,t u(t, r)|2 rd1 drd .l,m,u 1
S d1 t/2<r1+t

for all l, m 0. Note that we may replace rd1 by td1 since r is copmarable to t.
If in particular we define the spherical energies
X Z
f (r) := ( |l x,t u(t, r)|2 d)1/2
0ld Sd1

then a simple application of Cauchy-Schwarz and Minkowskis inequality yields

Z
|rm f (r)|2 dr .u t1d
t/2<r1+t

for m = 0, 1. Using the Poincare inequality

Z Z
|f (x)|2 . |f (y)|2 dy + |f (y)|2 dy
I I
whenever x I and I is an interval of length at least 1, we see that
|f (r)|2 .u t1d
for all t/2 < r 1 + t. Applying the Sobolev embedding theorem on the sphere (or
by using Poincare type inequalities) we then obtain the pointwise bound
|t,x u(t, x)| .u t(d1)/2
for all t 1 and t/2 < r 1 + t. If we combine this with both the finite speed of
propagation and some Lorentz transforms (see Exercise 2.6) to cover the interior
region r t/2, we conclude the more global estimate
(d1)/2
(2.57) kt,x u(t)kL d .u hti
x (R )
.
This can be used for instance to establish the dispersive bounds in (2.19), (2.20)
(Exercise 2.65). In fact, more precise estimates than (2.57) are available, which
establish more concentration properties near the light cone; see [Sog] for a more
detailed treatment of this method (which is especially useful for establishing global
solutions of semilinear and quasilinear wave equations from small, rapidly decaying
initial data).
Exercise 2.58 (Hamiltonian formulation of wave equation). Let us formally
1/2 1/2
consider Hx (Rd R) Hx (Rd R R) as a symplectic phase space with
symplectic form ((u0 , u1 ), (v0 , v1 )) = Rd u0 v1 u1 v0 . Show that u is a formal
 2.5. THE WAVE EQUATION STRESS-ENERGY TENSOR 95

solution to the wave equation (2.32) if and only if the curve t 7 (u(t), t u(t)) follows
the formal Hamiltonian flow associated to the (densely defined) Hamiltonian
Z
1
H(u0 , u1 ) := |u0 |2 + |u1 |2 dx.
2 Rd

Exercise 2.59 (Energy estimate). Let u Ct,loc Sx (R1+d C) be a Schwartz
solution to the inhomogeneous wave R equation u = F , and let T be as above.
By analyzing the energies E(t) := Rd T00 (t, x) dx, establish the energy identity
Z Z
(2.58) t T00 (t, x) dx = t u(t, x)F (t, x) dx
Rd Rd
and conclude the energy estimate (2.28) with s = 1. Then use the commutativity
of the wave equation with Fourier multipliers to establish this estimate for gen-
eral values of s. This is an example of the energy method ; see also the proof of
Proposition 3.3 for another instance of this technique.
Exercise 2.60 (Lagrangian derivation of stress-energy tensor). In this exer-
cise we shall work formally, ignoring issues of differentiability or integrability, and
will assume familiarity with (pseudo-)Riemannian geometry. Given any Lorentzian
metric g on R1+d , and any scalar field u : R1+d R, define the Lagrangian
functional Z
S(u, g) := L(u, g) dg
R1+d
p
where dg = det(g)dxdt is the usual measure induced by g, and L(u, g) is a local
quantity which is invariant under diffeomorphic changes of variable. Let X be an
arbitrary vector field on R1+d , and let gs be the deformation of the metric g for
s R along the vector field X, thus
d
(gs ) |s=0 = g ; (gs ) |s=0 = LX g = ,
ds
where LX is the Lie derivative, and = X + X is the deformation tensor
(here denotes the Levi-Civita covariant derivative, see Exercise 6.5). Similarly,
let us be the deformation of u along X, thus
d
us |s=0 = u; us |s=0 = LX u = X u.
ds
d
As L is invariant under diffeomorphic changes of variable, we have ds S(us , gs ) = 0.
Use this fact to conclude that if for fixed g, u is a critical point of the Lagrangian
S(u, g), then we have the integral conservation law
Z
T dg = 0,
R1+d
where the stress-energy tensor T is defined by
L 1
T := g L.
g 2
Conclude that T is divergence-free. In the special case L(u, g) := g u u,
with g equal to the Minkowski metric, show that this definition of the stress-energy
tensor co-incides with (2.45). See [SStru3] for further discussion of the Lagrangian
approach to wave equations.
96 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Exercise 2.61. Obtain a pointwise angular momentum conservation law for

the Schrodinger equation.
Exercise 2.62 (Classification of Killing fields). Show that if X is a Killing
vector field for Minkowski space R1+d , then X = 0. (Hint: consider various
permutations of the identity = 0.) Conclude that the only Killing vector
fields are the linear combinations of the ones given in the text.
Exercise 2.63 (Uniqueness of classical solutions). Show that for any smooth

initial data u0 Cx,loc (Rd ), u1 Cx,loc (Rd ) there exists a unique smooth solution
1+d
u Ct,x,loc (R ) to the wave equation (2.44) with initial data u(0, x) = u(x),
t u(0, x) = u1 (x). Note that we are not assuming any boundedness or decay at
infinity. (Hint: use finite speed of propagation, and the existence and uniqueness
theory in the Schwartz class.) This should be contrasted with the breakdown of

uniqueness in the smooth category Cx,loc in the case of infinite speed of propagation,
see Exercise 2.24.
Exercise 2.64 (Morawetz inequality for wave equation). Let u Ct Sx (R1+3 )
be a real-valued smooth solution to the wave equation (2.32) with finite energy
E(t) = E < . By contracting the stress-energy tensor against the radial vector
x
field X associated |x| x , obtain the identity
/ u|2
| 1
(T X ) = (|u|2 )
|x| 2|x|
and then conclude the Morawetz inequality
Z Z
/ u(t, x)|2
|
dxdt + |u(t, 0)|2 dt . E.
R1+3 |x| R
(Hint: multiply the previous identity by a smooth cutoff in time to a large interval
[T, T ], then integrate by parts, and use the energy bound to control all error
terms. You will have to deal with the singularity at x = 0 in some fashion (e.g. by
approximating |x| by (2 + |x|2 )1/2 , or by removing a small neighbourhood of the
origin. You may need Hardys inequality, Lemma A.2.) Compare this with (2.41).
Exercise 2.65. Use (2.57) to prove (2.19), (2.20). (Hint: first use (2.57) and
a scaling argument to establish (2.19), (2.20) when is itself the derivative of
a compactly supported bump function, then use translation invariance to replace
compactly supported bump function with Schwartz function. Finally, use some
form of dyadic decomposition (e.g. Littlewood-Paley decomposition) to handle the
general case.)
Exercise 2.66. Obtain a conserved stress-energy tensor for the Klein-Gordon
equation, which collapses to the one given for the wave equation above when c = 1
and m = 0, and collapses instead to the pseudo-stress-energy tensor given for the
Schrodinger equation in the limit c and ~ = m = 1, using the connection in
Exercise 2.7.
Exercise 2.67. Obtain conserved stress-energy tensors for the Maxwell and
abelian
R Yang-Mills equations with
R c = 1, in such a way that the conserved energies
are 21 R3 |E|2 +|B|2 dx, and 12 Rd |F0i |2 +|Fij |2 respectively. For the Dirac equation
with m = 0, show that the rank three stress-energy tensor
1
T := { u, u} g { u, u}
2
 2.6. X s,b SPACES 97

is divergence-free in all three variables. Is there an analogue of this tensor in the

massive case m 6= 0?

Exercise 2.68 (Equipartition of energy). Suppose that u is a Ct,loc S solution
R x
to the Klein-Gordon equation u = m u, thus the energy E := Rd 21 |u|2 +
2

1 2 m2 2
2 |t u| + 2R |u| dx is conserved in time. By considering the time derivative of the
expression Rd u(t, x)t u(t, x) dx, establish the estimate
Z Z
|t u|2 |u|2 m2 |u|2 dxdt = O(E/m)
I Rd
for arbitrary time intervals I. Thus in a time-averaged
R sense, the energy in E will
1 2
be equally split
R between the kinetic component 2 R d |t u| dx and the potential
1 2 2 2
component 2 R2 |x u| + m |u| .
Exercise 2.69 (Carleman inequality). Let u Cx (Rd R) be a compactly
supported solution to the Poisson equation u = F , and define the stress-energy
tensor
1 1 1
T := u u g u u = u u g (u2 ) + g uF
2 4 2
where g now denotes the Euclidean metric on Rd instead of a Minkowski metric.
Establish the divergence identity
T = F u
(cf. (2.46)). Now contract this against the vector field e2txj xj for some j = 1, . . . , d
and t 6= 0 and integrate by parts to obtain the identity
Z Z
2t|xj (etxj u)|2 = etxj F xj (etxj u)
Rd Rd
and conclude the Carleman-type inequality
1
kxj (etxj u)kL2x (Rd ) ketxj F kL2x (Rd ) .
2|t|
Conclude the following unique continuation property: if u is a scalar or vector field
on Rd which is smooth and compactly supported with u = O(|u|), then u vanishes
identically. (Hint: if u is compactly supported, then ketxj ukL2x can be controlled by
a bounded multiple of kxj (etxj u)kL2x . Now let t .) This shows, for instance,
that Schrodinger operators +V with smooth bounded potentials V cannot have
any compactly supported eigenfunctions (bound states); we shall also use it to show
a rigidity property of harmonic maps in Exercise 6.40. For a different proof, see
Exercise B.6.

2.6. X s,b spaces

I dreamed a thousand new paths... I woke and walked my old one.
(Chinese proverb)
Let us now return to a general scalar constant-coefficient dispersive linear equa-
tion
t u = Lu
where L = ih(/i) for some real-valued polynomial h; again, the vector-valued
case is more complicated and will not be treated here. As discussed in Section
98 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

2.1, the spacetime Fourier transform u of solutions to this equation will be sup-
ported on the hypersurface {(, ) : = h()}. If one then localises the solution in
time (for instance by multiplying u by some smooth cutoff function (t)), then the
uncertainty principle (or the intertwining of multiplication and convolution by the
Fourier transform) then suggests that the Fourier transform fu will be concentrated
in the region {(, ) : = h() + O(1)}.
Now consider a nonlinear perturbation of the above equation, such as
t u = Lu + N (u).
At first glance one may expect the presence of the nonlinearity to distort the Fourier
support of the solution substantially, so that u or fu now has a substantial portion
which lies far away from the characteristic hypersurface = h(). Certainly one has
a significant distortion if one does not localise in time first (indeed, the nonlinear
solution need not even exist globally in time). However, if one applies a suitably
short time cutoff , then it turns out that for many types of nonlinearities u,
and for surprisingly rough classes Hxs (Rd ) of initial data, the localised Fourier
transform f u still concentrates near the characteristic hypersurface. The reason
for this is a dispersive smoothing effect for the operator t L away from the
hypersurface = h(), which can be viewed as the analogue of the more familiar
elliptic regularity phenomenon for elliptic equations (if Lu = f and L is elliptic,
then u is smoother than f ).
There are a number of ways to capture this dispersive smoothing effect, but one
particularly convenient way is via the X s,b -spaces (also known as Fourier restriction
spaces, Bourgain spaces, or dispersive Sobolev spaces). The full name of these
spaces25 is Xs,b d d
=h() (R R ), thus these spaces take R R as their domain and
are adapted to a single characteristic hypersurface = h(). Roughly speaking,
these spaces are to dispersive equations as Sobolev spaces are to elliptic equations.
In a standard Sobolev space Hxs (Rd ), one can differentiate the function using the
elliptic derivative hi s times and still remain square-integrable; for the space
X s,b (R Rd ), one can differentiate s times using the elliptic derivative hi and
b times using the dispersive derivative t L, and still remain square-integrable.
The precise definition is as follows.
Definition 2.7 (X s,b spaces). Let h : Rd R be a continuous function, and
let s, b R. The space Xs,b d
=h() (R R ), abbreviated X
s,b
(R Rd ) or simply X s,b
is then defined to be the closure of the Schwartz functions St,x (R Rd ) under the
norm
kukX s,b (RRd ) := khis h h()ib u(, )kL2 L2 (RRd ) .
=h()

These spaces in their modern form were introduced by Bourgain [Bou], al-
though they appear in the context of one-dimensional wave equations in earlier
work of Beals [Bea] and Rauch-Reed [RR], and implicitly in the work of Klainer-
man and Machedon [KM]. A good survey of these spaces and their applications
can be found in [Gin]. Multilinear estimates for these spaces were systematically
studied in [Tao4].

25The terminology H s, = H s, (R1+d ) is also occasionally used in the literature, as these

=h()
spaces resemble product Sobolev spaces.
 2.6. X s,b SPACES 99

1111111111
0000000000
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111

0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
Figure 3. A solution to a linear dispersive equation (such as
the Airy equation t u + xxx u = 0) will have its spacetime
Fourier transform concentrated perfectly on the characteristic sur-
face = h(). Solutions to nonlinear perturbations of that disper-
sive equation (such as the KdV equation t u + xxx u = 6ux u)
will typically, after localisation in time, have spacetime Fourier
transform supported near the characteristic surface; thus the non-
linearity does not significantly alter the spacetime Fourier path
of the solution, at least for short times. The X s,b spaces are an
efficient tool to capture this clustering near the characteristic sur-
face.

In the case b = 0, the choice of dispersion relation = h() is irrelevant,

and the X s,b space is simply the space L2t Hxs , as can be seen by an application
of Plancherels theorem in time. In the case h = 0, the X s,b space becomes the
product space Htb Hxs , and for general h it is a conjugate of this space (Exercise
2.70). The spatial domain Rd can be replaced with other abelian groups such as
the torus Td with minimal modification (just as Sobolev spaces and similarly be
defined for the torus), indeed we have
kukX s,b (RTd ) := khkis h h()ib u(, k)kL2 l2k (RZd )
=h(k)

where u is the spatially periodic, temporally non-periodic Fourier transform

Z Z
1
u(, k) := u(t, x)ei(t +k) dxdt.
(2)d R Td
Most of the results stated here for the non-periodic setting will extend without any
difficulty to the periodic setting; we leave the verification of these details to the
reader.
The spaces Xs,b tL
=h() are well adapted to the solutions u(t) = e u(0) of the
linear dispersive equation t u = Lu, where L := ih(D) = ih(/i), as the following
lemma shows:
100 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Lemma 2.8 (Free solutions lie in X s,b ). Let f Hxs (Rd ) for some s R, and
let L = ih(/i) for some polynomial h : Rd R. Then for any Schwartz time
cutoff Sx (R), we have
k(t)etL f kX s,b (RRd ) .,b kf kHxs (Rd ) .
=h()

Proof. A computation shows that the spacetime Fourier transform of (t)etL f

at (, ) is simply ( h())f(). Since is rapidly decreasing, the claim follows.

We now discuss the basic properties of the X s,b spaces. We first observe the
easily verified fact that the X s,b spaces are Banach spaces. We have the trivial
nesting

Xs=h()
,b
Xs,b
=h()
whenever s s and b b. From Parsevals identity and Cauchy-Schwarz we have
the duality relationship
s,b
(Xs,b
=h() ) = X =h() .
Also, these spaces interpolate nicely in the s and b indices, as can be seen using the
Stein complex interpolation theorem (see e.g. [Stei2]). These two facts can save
some effort when proving certain estimates regarding the X s,b spaces, particularly
the multilinear estimates.
Now we study the invariance and stability properties of these spaces. The X s,b
spaces are invariant under translations in space and time, but they are usually not
invariant under frequency modulations (e.g. multiplication by a spatial phase eix
or a temporal phase eit ). The behaviour under complex conjugation is given by
the identity
kukX s,b = kukX s,b
=h() =h()

and thus one has conjugation invariance when h is odd, but not necessarily other-
wise.
When b > 1/2, one can view the X s,b spaces as being very close to free solutions
(i.e. solutions to the equation t u = Lu). This is formalised in the following lemma:
Lemma 2.9. Let L = iP (/i) for some polynomial P : Rd R, let s R,
and let Y be a Banach space of functions on R Rd with the property that
keit0 etL f kY . kf kHxs (Rd )
for all f Hxs (Rd ) and 0 R. Then we have the embedding
kukY .b kukX s,b (RRd ) .
=h()

Conversely, free solutions will lie in X s,b once suitably truncated in time; see
Lemma 2.11.
Proof. By Fourier inversion we have
Z Z
1
u(t, x) = u(, )eit eix dd.
(2)d+1 R Rd
If we write = h() + 0 , and set
Z
1
f0 (x) := u(h() + 0 , )eix d
(2)d Rd
 2.6. X s,b SPACES 101

we have Z
tL 1
e f0 (x) = u(h() + 0 , )eith() eix d
(2)d Rd
and thus have the representation
Z
1
u(t) = eit0 etL f d0 .
2 R
Taking Y norms and using Minkowskis inequality and the hypothesis on Y , we
obtain Z
kukY . kf0 kHxs (Rd ) d0 ,
R
and hence by Cauchy-Schwarz and the hypothesis b > 1/2
Z
kukY .b ( h0 i2b
R
kf0 k2Hxs (Rd ) d0 )1/2 .
Using Plancherels theorem, the right-hand side rearranges to equal Cb kukX s,b , and
the claim follows. 
Observe that the same argument applies when P is merely a continuous function
rather than a polynomial, though in this case L will be a Fourier multiplier rather
than a differential operator. Applying this to Y = Ct0 Hxs , we obtain the immediate
corollary
Corollary 2.10. Let b > 1/2, s R, and h : Rd R be continuous. Then
for any u Xs,b d
=h() (R R ) we have

kukCt0Hxs (RRd ) .b kukX s,b (RRd ) .

=h()

Furthermore, the X s,b spaces enjoy the same Sobolev embeddings that free
solutions to the equation ut = Lu do. For instance, by combining Lemma 2.9 with
(2.24) (and observing that the spaces Lqt Lrx are invariant under multiplication by
phases such as eit0 ), one concludes that
kukLqt Lrx (RRd ) .q,r,b kukX 0,b (RRd )
=||2

for all Schrodinger-admissible (q, r).

It turns out that the X s,b spaces are only well suited to analyzing nonlinear
dispersive equations when one localises in time. Fortunately, these spaces are easy
to localise:
Lemma 2.11 (X s,b is stable wrt time localisation). Let St (R) be a Schwartz
function in time. Then we have
k(t)ukX s,b (RRd ) .,b kukX s,b (RRd )
=h() =h()

for any s, b R, any h : Rd R, and any field u St,x (R Rd ). Furthermore,

if 1/2 < b b < 1/2, then for any 0 < T < 1 and > 0 we have

k(t/T )ukX s,b (RRd )
.,b,b T bb kukX s,b (RRd ) .
=h() =h()

The second estimate in this lemma is useful in the large data theory, as it allows
one to keep certain X s,b norms of a solution small by localizing to a sufficiently
small time interval.
102 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Proof. Let us first understand how the X s,b spaces behave with respect to
temporal frequency modulation u(t, x) 7 eit0 u(t, x). From the crude estimate
h 0 h()ib .b h0 i|b| h h()ib
and elementary Fourier analysis, we conclude that
keit0 ukX s,b (RRd ) .b h0 i|b| kukX s,b (RRd ) .
=h() =h()
R
If we now use Fourier expansion in time to write (t) = R (0 )eit0 d0 and use
Minkowskis inequality, we conclude
Z
k(t)ukX s,b (RRd ) .b ( |(0 )|h0 i|b| d0 )kukX s,b (RRd ) .
=h() =h()
R

Since is Schwartz, is rapidly decreasing, and the first claim follows.

Now we prove the second claim. By conjugating by his we may take s = 0.
By composition it suffices to treat the cases 0 b b or b b 0; by duality
we may then take 0 b b. By interpolation with the trivial case b = b we may
take b = 0, thus we are now reduced to establishing
k(t/T )ukL2t L2x (RRd ) .,b T b kukX 0,b (RRd )
=h()

for 0 < b < 1/2. By partitioning frequency space we can divide into two cases, one
where u is supported on the region h h()i 1/T , and one where h h()i 1/T .
In the former case we will have
kukX 0,0 (RRd ) T b kukX 0,b (RRd )
=h() =h()

and the claim then follows from the boundedness of . In the latter case, we use a
variant of Corollary 2.10, noting for any time t that
d
ku(t)kL2x(Rd ) . ku(t)()k L2 (Rd )
Z
.k |u(, )| d kL2 (Rd )
h h()i1/T
Z
.b T b1/2 k( h h()i2b |u(, )|2 d )1/2 kL2 (Rd )

= T b1/2 kukX s,b (RRd )

=h()

thanks to Plancherel, the triangle inequality, Cauchy-Schwarz, and the localisation

of u. Integrating this against (t/T ), the claim follows. 

The X s,b spaces react well to Fourier multipliers, in much the same way that
ordinary Sobolev spaces H s do. If Dk is a Fourier multiplier of order k, in the sense
that
d
D k f () = m()f ()

for all Schwartz functions f Sx (Rd ) and some measurable multiplier m : Rd C

obeying the growth condition |m()| . hik , then D can be extended to spacetime
functions by acting on each time separately, thus (Du)(t) = D(u(t)), or in terms of
the spacetime Fourier transform
g
D k u(, ) = m()u(, )
 2.6. X s,b SPACES 103

and then one easily verifies that D maps Xs,b sk,b

=h() continuously to X =h() for any
s, b R and any h:
kDk ukX sk,b . kukX s,b .
=h() =h()

This is analogous to the well-known estimate kDk ukH sk . kukH s for Sobolev
spaces. In the case when k is a non-negative integer, we have the converse
(2.59) kukX s,b . kukX sk,b + kkx ukX sk,b ,
=h() =h() =h()

which is proven by repeating the above arguments. Similarly, if h : Rd R is a

polynomial and L := ih(/i), then we have

(^
t L)u(, ) = i( h())u(, )

and hence
k(t L)ukX s,b1 . kukX s,b .
=h() =h()

It is natural to ask whether there is a converse inequality, in the spirit of (2.59).

This is indeed the case:
Proposition 2.12 (X s,b energy estimate). Let h : Rd R be a polynomial,

let L := ih(/i), and let u Ct,loc Sx (R Rd ) be a smooth solution to the equation
ut = Lu + F . Then for any s R and b > 1/2, and any compactly supported
smooth time cutoff (t), we have
k(t)ukX s,b (RRd ) .,b ku(0)kHxs (Rd ) + kF kX s,b1 (RRd ) .
=h() =h()

Proof. To abbreviate the notation we shall write Xs,b d

=h() (R R ) simply as
X s,b . Let [R, R] be a time interval containing the support of . By truncating F
smoothly in time, using a compactly supported cutoff that equals 1 on [R, R] we
may assume (using Lemma 2.11) that F is supported on [2R, 2R] and is Schwartz
in spacetime. Also, by applying his to both u and F if necessary, we may take
s = 0.
Let us first suppose that u vanishes at time 2R, thus u(2R) = 0. By
Duhamels formula (2.13) (with t0 = 2R) we thus have
Z t Z
(t)u(t) = (t) e(ts)L F (s) ds = (t) (t s)1[0,+) (t s)e(ts)L F (s) ds
R

where is a smooth compactly supported function which equals 1 on [3R, 3R].

By Lemma 2.11, it would thus suffice to show that
Z
k (t s)1[0,+) (t s)e(ts)L F (s) dskX 0,b .,b kF kX 0,b1 .
R
R
A routine computation shows that the spacetime Fourier transform of R (t
s)1[0,+) (t s)e(ts)L F (s) ds at (, ) is equal to
Z
( (t)1[0,+) (t)eit( h()) dt)F (, ).
R

The expression inside the parentheses can be shown (via integration by parts) to
be at most O (h i1 ). The claim then follows.
104 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

Now we handle the general case. We split u(t) = (u(t) e(t+2R)L u(2R)) +
tL 2RL
e e u(2R). For the first term, the preceding argument applies, and we have

k(t)(u(t) e(t+2R)L u(2R))kX 0,b (RRd ) .,b kF kX 0,b .

Thus it will suffice to control the remaining term. Applying Lemma 2.8, it thus
suffices to show that
ke2RL u(2R)kL2x .,b ku(0)kL2x + kF kX 0,b1 .
From Duhamels formula and the support of F we have
Z
e2RL u(2R) = u(0) + (s)1(,0] (s)esL F (s) ds
R

where is as before. Thus by the triangle inequality it suffices to show that

Z
k (s)1(,0] (s)esL F (s) dskL2x .,b kF kX 0,b1 .
R

Applying Parsevals identity, the left-hand side can be written as

Z Z
k ( (s)1(,0] (s)eis( h()) ds)F (, ) d kL2 .
R R

An integration by parts yields the bound

Z
(s)1(,0] (s)eis( h()) ds| . h h()i1
R

and hence by Cauchy-Schwarz and the hypothesis b > 1/2 we have

Z Z Z
is( h())
| ( (s)1(,0] (s)e ds)F (, ) d | .,b | h h(i2(b1) |F (, )|2 d )1/2 ,
R R R

and the claim follows. 

As observed earlier, X s,b spaces enjoy all the Strichartz estimates that free
solutions do. However, in some cases, particularly in periodic settings, it is not
always easy to obtain Strichartz estimates, as dispersive inequalities are typically
not available in periodic settings. (When the domain is compact, L x decay is
inconsistent with L2x conservation.) However, if one is interested in L4t,x or L6t,x
type inequalities, one can sometimes establish the Strichartz estimate by a direct
Fourier-analytic approach. A typical result, which is of application to the periodic
Schrodinger equation26 is as follows.
Proposition 2.13 (Periodic Schrodinger estimate). [Bou] We have
kukL4t L4x (RT) . kukX 0,3/8 (RT) .
=k2

for any u St,x (R T).

26Depending on the choice of normalisation used for the Schrodinger equation, the dispersion
relation = h(k) may differ from = k 2 by an absolute constant, but this makes no difference
to this Strichartz estimate. Note however that for bilinear estimates one needs to distinguish the
X s,b space associated to = k 2 from the conjugate X s,b space, associated with = k 2 .
 2.6. X s,b SPACES 105

P
Proof. We use an argument of Nikolay Tzvetkov. Split u = M uM , where
M ranges over integer powers of 2, and uM is the portion of u localised to the
spacetime frequency region 2M h k 2 i < 2M+1 . From Plancherels theorem we
have X
M 3/4 kuM k2L2 L2 (RT) . kuk2X 0,3/8 (RT)
t x
=k2
M
Squaring both sides of the desired inequality and using the triangle inequality, we
reduce to proving that
X X X
kuM uM kL2t L2x (RT) . M 3/4 kuM k2L2 L2 (RT) .
t x
MM M

Setting M = 2m M , it thus suffices by the triangle inequality to prove that

X X
kuM u2m M kL2t L2x (RT) . 2m M 3/4 kuM k2L2 L2 (RT)
t x
M M

for all m 0 and some absolute constant > 0; by Cauchy-Schwarz it thus suffices
to establish that
kuM u2m M kL2t L2x (RT) . 2m M 3/8 kuM kL2t L2x (RT) (2m M )3/8 ku2m M kL2t L2x (RT) .
Let us now normalise uM and u2m M to have L2t L2x norm one. We use Plancherel
and reduce to showing
X Z
k uM (1 , k1 )u2m M (2 , k2 ) d1 kL2 l2k (RZ) . 2(3/8)m M 3/4 .
k1 +k2 =k 1 +2 =

On the other hand, from the normalisation and Fubinis theorem we have
X Z
k( |uM (1 , k1 )|2 |u2m M (2 , k2 )|2 d1 )1/2 kL2 l2k (RZ) = 1
k1 +k2 =k 1 +2 =

so by Cauchy-Schwarz and the support of uM , u2m M it will suffice to show that

X Z
1 d1 . 2(3/42m) M 3/2
k1 +k2 =k 1 +2 = ;1 =k12 +O(M);2 =k22 +O(2m M)

for all k, .
Fix k, . Observe that for the integral to be non-empty, we must have =
k12 + k22 + O(2m M ), in which case the integral is O(M ). Thus it suffices to show
that X
1 . 2(3/42)m M 1/2 .
k1 +k2 =k; =k12 +k22 +O(2m M)

But if = k12 + k22 + O(2m M ) and k1 + k2 = k, then (k1 k2 )2 = 2 k + O(2m M ),

and hence k1 k2 is constrained to at most two intervals of length O(2m/2 M 1/2 ).
The claim then follows with = 1/8. 

In Section 4.1 we shall encounter some bilinear and trilinear X s,b estimates in
a spirit similar to the above (see also the exercises below).
Exercise 2.70 (X s,b vs. product Sobolev spaces). Let u S(R Rd ) be a
complex field and let h : Rd R be a polynomial. Let U (t) := exp(ith(/i))
be the linear propagators for the equation ut = Lu, where L = ih(/i). Let
106 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

v : R Rd C be the function v(t) := U (t)u(t), thus v is constant in time if

and only if u solves the equation ut = Lu. Show that
kvkX s,b (RRd ) = kukHtb Hxs (RRd )
=h()

for all s, b R.
Exercise 2.71 (Endpoint X s,b spaces). Show that Lemma 2.9, Corollary 2.10,
Lemma 2.11, and Proposition 2.12 all break down at the endpoint b = 1/2. (But
see the next exercise.)
Exercise 2.72 (Endpoint X s,b spaces, II). Let h : Zd R, and let s, b be
real numbers. Define the space Ys,b d
=h(k) (R T ) to be the closure of the Schwartz
functions under the norm
kukY s,b (RTd ) := khkis h h(k)ib ukl2k L1 (RZd ) .
=h(k)

Establish the embeddings

kukY s,b1/2 (RTd ) . kukX s,b (RTd )
=h(k) =h(k)

and
kukCt0 Hxs (RTd ) . kukY s,0 (RTd )
for all Schwartz functions u and all > 0; show that the former embedding breaks
down at = 0. With the notation of Proposition 2.12, establish the energy estimate
k(t)ukY s,0 (RTd ) + k(t)ukX s,1/2 (RTd )
=h() =h()

.,b (ku(0)kHxs (Td ) + kF kY s,1 (RTd ) + kF kX s,1/2 (RTd ) ).

=h() =h()

In the periodic theory, these estimates allow one to use the endpoint space X s,1/2
(which is otherwise very badly behaved, as the preceding exercise showed) by aug-
menting it with the additional space Y s,0 .
Exercise 2.73 (X s,b spaces for the wave equation). Let us work in Minkowski
space R1+d with c = 1. Define the norm kuks,b := kukX s,b (R1+d ) by
| |=||

kuks,b := khis h| | ||ib ukL2 (R1+d ) .

Also define the slightly stronger norm X s,b by
kukX s,b := kuks,b + kt uks1,b .
Develop analogues of Corollary 2.10 and Lemma 2.11. Establish that
kuks1,b1 .s,b,d kukX s,b
for all u St,x (R1+d ), and conversely that one has the energy estimate
k(t)ukX s,b .s,b, (ku(0)kHxs (Rd ) + kt u(0)kHxs1 (Rd ) + kuks1,b1 )

for all compactly supported bump functions , all u St,x (R1+d ), all s R, and
all b > 1/2. Typically, in applications one would place the solution u in the space
s1,b1
X s,b and the nonlinearity u in X| |=|| . What is the counterpart of Lemma 2.9?
 2.6. X s,b SPACES 107

Exercise 2.74. Let u St,x (R T2 ) solve the inhomogeneous Schrodinger

equation it u + u = F . Show that
k(t)ukCt0 L4x (RT2 ) + k(t)ukL4t,x (RT2 ) . ku(0)kL2x(T2 ) + kF kL4/3 (RT2 )
t,x

for all compactly supported cutoff functions . (Hint: use Proposition 2.13 and the
Christ-Kiselev lemma, Lemma 2.4.)
Exercise 2.75 (Bilinear refinement to Strichartz). [Bou9], [CKSTT11] Let
u, v St,x (R Rd ) be fields whose spacetime Fourier transforms u, v are supported
on the sets || M and || N respectively for some N, M 1. If d = 1, let us
impose the additional hypothesis N > 2M . Show that
M (d1)/2
kuvkL2t L2x (RRd ) .b kukX 0,b (RRd ) kvkX 0,b (RRd ) .
N 1/2 =||2 =||2

(Hint: use Lemma 2.9 twice to reduce u and v to free solutions of the Schrodinger
equation, and compute using Plancherel explicitly. In the case d = 2 and N 2M ,
one can use Strichartz estimates.)
Exercise 2.76 (Divisor bound). Show that a positive integer d has at most
O (n ) divisors for any > 0. (Hint: first show that if d is the power of a prime
p, then d has at most O (n ) divisors, and in fact has at most n divisors if p is
sufficiently large depending on . For the general case, factorise d into the product
of prime powers.)
Exercise 2.77 (Periodic Airy L6t,x estimate). [Bou] Using Exercise 2.76 and
the identity
(k1 + k2 + k3 )3 k13 k23 k33 = 3(k1 + k2 )(k2 + k3 )(k3 + k1 ),
show that for any integers k, t that the number of integer solutions to the system
k1 + k2 + k3 = k, k13 + k23 + k33 = t with k1 , k2 , k3 = O(N ) is at most O (N ) for
any N 1 and > 0. Use this to obtain the estimate
X 3 X
k ak eikx+ik t kL2t L2x (TT) . ( hki a2k )1/2
kZ kZ
for any complex numbers ak and any > 0, and use this in turn to conclude the
Strichartz estimate
k(t)ukL6t L6x (RT) .,b kukX ,b (RT)
=k3

for any > 0 and b > 1/2 and any field u. It would be of interest to know if this
estimate holds for = 0, or with the exponent p = 6 replaced by a larger exponent
such as p = 8.
Exercise 2.78 (Periodic Airy L6 estimate, II). [Bou] Show that for any in-
tegers k, t that the number of integer solutions to the system k1 + k2 k3 = k,
k12 + k22 k32 = t with k1 , k2 , k3 = O(N ) is at most O (N ) for any N 1 and > 0.
(Hint: use the first equation to eliminate k3 from the second, and then obtain an
identity of the form (k1 + a)(k2 + b) = c for some a, b, c given explicitly in terms of
k, t.) By arguing as in the preceding exercise, show that
k(t)ukL6t L6x (RT) .,b kukX ,b (RT)
=k2

for any > 0 and b > 1/2 and any field u. It is known that the cannot be set
to zero in this case, though perhaps if the exponent p = 6 were lowered slightly
108 2. CONSTANT COEFFICIENT LINEAR DISPERSIVE EQUATIONS

then this could be possible. See also [CKSTT3], [CKSTT12] for some trilinear
refinements of this estimate in which the loss can be eliminated.
 CHAPTER 3

Semilinear dispersive equations

Come what come may,

Time and the hour runs through the roughest day.
(William Shakespeare, Macbeth)
In this chapter we turn at last to the main subject of this monograph, namely
nonlinear dispersive equations. Specifically, we now study the local and global
low-regularity wellposedness of the following two Cauchy problems: the nonlinear
Schrodinger equation (NLS) 1
1
iut + u = |u|p1 u
(3.1) 2
u(t0 , x) = u0 (x) Hxs (Rd )

and the nonlinear wave equation (NLW)

u = |u|p1 u
(3.2) u(t0 , x) = u0 (x) Hxs (Rd )
t u(t0 , x) = u1 (x) H s1 (Rd ).

In this chapter we have normalised c = ~ = m = 1, so that  = t2 + . We

will often also take advantage of time translation invariance to normalise t0 = 0.
In both cases, the scalar field u : R1+d C (or u : I Rd C, if one only
seeks local solutions) is the desired solution, and the initial data u0 (and u1 , in the
case of NLW) is specified and lies in a given Sobolev space Hxs (Rd ) (or Hxs1 (Rd )).
The exponent 1 < p < denotes the power of the nonlinearity and is also given;
the sign {1, 0, +1} denotes whether the nonlinearity is defocusing, absent, or
focusing respectively2; we will see some reasons for this terminology later in this
chapter. The cases when p is an odd integer, and in particular the cubic case p = 3
and the quintic case p = 5, are particularly important in mathematical physics,
and have the advantage that the nonlinearity z 7 |z|p1 z is smooth, indeed it is a
polynomial in z and z. We shall refer to these instances of NLS and NLW as the
algebraic NLS and NLW respectively. The periodic analogues of these problems,
when the domain is the torus Td instead of Euclidean space Rd is also of interest,
though our main focus here shall be on the non-periodic case.

1
The factor of 21 can be easily eliminated by rescaling time by a factor of 2, and so can be
safely ignored. We retain it in order to make the dispersion relation (or de Broglie law) between
velocity and frequency as simple as possible, namely v = .
2The defocusing and focusing nonlinearities are sometimes called repulsive and attractive
nonlinearities in the literature.

109
110 3. SEMILINEAR DISPERSIVE EQUATIONS

For the NLW it is convenient to adopt the notation u[t] := (u(t), t u(t)), thus
for instance u[t0 ] = (u0 , u1 ). Thus u[t] describes the total state (both position and
velocity) of the solution u at time t.
The power-type nonlinearity function F (u) := |u|p1 u can be replaced by
other nonlinearities, and in many cases one obtains results similar to those stated
here. But the specific choice of power-type nonlinearity has a number of nice
properties that make it well-suited for exposition, in particular enjoying symmetries
such as the scaling and phase rotation symmetry F (zu) = |z|p F (u) for any complex
z, which will in turn lead to corresponding symmetries for NLS and NLW. It is
1
also naturally associated to a Hamiltonian potential V (u) := p+1 |u|p+1 via the
observation
d
V (u + v)|=0 = Re(F (u)v)
d
for any u, v C; this will lead to a Hamiltonian formulation for NLS and NLW
(Exercise 3.1).
We will be particularly interested in the low regularity problem: whether one
still has existence and uniqueness of solutions even when the initial datum only
lies in a very low Sobolev space. There are a number of reasons why one would
want to go beyond high-regularity (classical) solutions and consider low-regularity
ones3. Firstly, a good low-regularity theory gives more control on the nature of
singularities of a solution, if they do indeed form; generally speaking, if one has
a local wellposedness theory in Hxs , then that implies that a singularity can only
form by making the Hxs norm go to infinity (or to concentrate at a point, if the
norm Hxs is critical with respect to scaling). Secondly, many of the key structural
features of an equation - such as the conserved Hamiltonian, the symplectic form,
the scale invariance, or other conserved or monotone quantities - are typically as-
1/2
sociated to rather low regularities such as L2x , Hx , or Hx1 , and in order to fully
exploit these features it is often important to have a good local theory at those reg-
ularities. Thirdly, the technical challenge of working at low regularities (especially
near or at the critical regularity) enforces a significant discipline on ones approach
to the problem - requiring one to exploit the structural properties of the equation
as efficiently and as geometrically as possible - and has in fact led to the develop-
ment of powerful and robust new techniques and insights, which have provided new
applications even for smooth solutions (for instance, in clarifying the dynamics of
energy transfer between low and high frequencies). Finally, the task of extending a
local existence result to a global existence result can (somewhat paradoxically) be
easier if one is working at low regularities than high regularities, particularly if one
is working at the scale-invariant regularity, or a regularity associated to a conserved
quantity.
The nonlinear Schrodinger and wave models (3.1), (3.2) are among the simplest
nonlinear perturbations of the free (linear) Schrodinger and wave equations4. Both
equations are semilinear (the nonlinearity is lower order than the linear terms),
3Alternatively, one can continue to work exclusively with classical solutions so that there is
no difficulty justifying various formal computations, but demand that all estimates depend only
on low-regularity norms of the solution. In practice, the two approaches are essentially equivalent;
in most cases one can use limiting arguments to recover the former from the latter.
4Indeed, the NLS (together with the KdV equation) frequently arises in physics as the first
nonlinear approximation of a dispersive system, by performing a Taylor expansion of the nonlin-
earity and discarding all but the first term. See for instance [SSul].
 3. SEMILINEAR DISPERSIVE EQUATIONS 111

and furthermore the nonlinear term contains no derivatives or non-local terms.

Furthermore, these nonlinear equations retain many of the symmetries and struc-
ture of their linear counterparts; for instance, the NLS is a Hamiltonian equation
with conservation of mass (charge), momentum, energy, enjoys scaling, Galilean,
translation, and (partial) pseudoconformal symmetries, and enjoys several mono-
tonicity formulae, including some of virial and Morawetz type. The NLW is also
Hamiltonian with a conserved stress-energy tensor, with all its attendant conse-
quences such as Morawetz inequalities and finite speed of propagation, and also
enjoys scaling, Lorentz, translation and (partial) conformal symmetries. On the
other hand, these equations are not completely integrable (with the notable excep-
tion of the one-dimensional cubic (p = 3) NLS, as well as a variant of the NLW
known as the sine-Gordon equation), and so do not admit many explicit solutions
(beyond some standard solutions such as the ground state solitons). The large
number of parameters present in these equations (the dimension n, the power p,
the sign , the regularity s, and whether one wishes to consider periodic or non-
periodic solutions) means that these equations exhibit a wide range of phenomena
and behaviour, and in many ways are quite representative of the much larger class
of nonlinear dispersive and wave equations which are studied in the literature. Thus
while our understanding of these equations is somewhat better than for most other
nonlinear dispersive models (particularly for subcritical and critical regularities, for
small data, and for the defocusing regularity), they are still so rich in structure and
problems that there is still plenty to be understood.
Broadly speaking, there are two major classes of techniques one can use to ana-
lyze these equations. On the one hand, one has perturbative methods, which approx-
imate the nonlinear equations (3.1), (3.2) by more tractable and well-understood
equations such as5 the free (and linear) Schrodinger or wave equations. The er-
ror between the actual equation and the approximate equation is usually treated
by some sort of iteration argument (usually based on Duhamels formula) or by
a Gronwall inequality argument (usually based on energy estimates). Another re-
lated example of a perturbative method arises when constructing exact solutions
to NLS and NLW by first starting with an approximate solution (that solves the
equation up to a small error) and then constructing some sort of iterative scheme
or Gronwall inequality argument to convert the approximate solution to an exact
one.
As the name implies, perturbative methods only work when the solution is
very close to its approximation; typically, this requires the initial datum to be
small (or a small perturbation of a special initial datum), or the time interval to
be small (or perhaps some spacetime integral of the solution to be well controlled
on this time interval). When dealing with large solutions over long times, per-
turbative techniques no longer work by themselves, and one must combine them
with non-perturbative methods. Examples of such methods include conservation
laws, monotonicity formulae, and algebraic transformations of the equation. Such
methods are initially only justified for smooth solutions, but can often be extended
to rough solutions by means of the perturbative theory. Thus, global control of a
solution is often obtained via a collaboration between the perturbative techniques

5In some cases one will use a more complicated equation as the approximating equation. For
instance, if one is analyzing the NLS or NLW near a special solution such as a soliton solution,
one often uses the linearised equation around that soliton as the approximating equation.
112 3. SEMILINEAR DISPERSIVE EQUATIONS

and non-perturbative techniques; typically, the perturbative theory guarantees a

well-behaved solution provided that certain integrals of the solution stay bounded,
and the non-perturbative theory guarantees control of these integrals provided that
the solution remains well-behaved (see e.g. Figure 7.). This basic division of labour
already works remarkably well in many situations, although in some recent results
(most notably in those employing the induction on energy strategy, see Chapter 5)
one has had to apply a more advanced scheme.
We conclude this introduction by describing some special (and very explicit)
solutions to both (3.1) and (3.2), in order to build some initial intuition about
these equations, though we emphasise that for generic initial data we do not expect
any similarly explicit formula for the solution. Let us begin by using the classical
method of separation of variables, using solutions of ODE to construct special
solutions to PDE. For any Rd , the plane wave eix is an eigenfunction of the
Laplacian , and also has magnitude one, which leads one to consider the ansatz
(3.3) u(t, x) = eix v(t).
Simple calculation then shows that in order to solve the NLS (3.1), v must obey
the ODE
||2
t v = i( + |v|p1 )v
2
and to solve the NLW (3.2), v must obey the ODE
(3.4) t2 v = (||2 + |v|p1 )v.
In the case of NLS, the ODE for v can be explicitly solved, leading to the plane
wave solutions
2
t/2 i||p1 t
(3.5) u(t, x) := eix ei|| e
2
for any C and Zd . Note how the time oscillation of ei|| t/2 arising from
p1
the linear evolution is augmented by the additional time oscillation ei|| t . In
the defocusing case = +1, both time oscillations are anti-clockwise, so one can
view the defocusing nonlinearity as amplifying the dispersive effect of the linear
equation; in the focusing case the focusing nonlinearity is instead trying to cancel
the dispersive effect. If the amplitude is small compared the frequency then
the dispersive effect is stronger, but when the amplitude is large then the focusing
effect takes over. This already illustrates one useful heuristic: the focusing and
defocusing equations behave similarly when the initial data is small or when the
frequency is very large6.
As for the NLW (3.2), one can obtain a similar class of (complex) explicit
solutions
2 p1 1/2
u(t, x) := eix ei(|| +|| ) t
provided that ||2 +||p1 0. This latter condition is automatic in the defocusing
case = +1 or the linear case = 0, but requires either the frequency to be
large or the amplitude to be small in the focusing case = 1. This is again
consistent with the heuristic mentioned earlier. When the initial data is large and
6Actually, this heuristic is only valid in subcritical situations, in which the high frequencies
scale more favourably than the low frequencies. In critical cases, the high and low frequencies are
equally sensitive to the distinction between focusing and defocusing; the supercritical cases are
very poorly understood, but it is believed that the high frequency behaviour is radically different
in the focusing and defocusing cases.
 3. SEMILINEAR DISPERSIVE EQUATIONS 113

positive compared to the frequency, then the ODE (3.4) can blow up (this can be
seen for instance using Exercise 1.25, in the case when v is real); one explicit family
of blowup solutions in the focusing case = 1 (with = 0) is given by
(3.6) u(t, x) := cp (t0 t)2/(p1)
for t < t0 , where cp := ( 2(p+1)
(p1)2 )
1/(p1)
and t0 R is an arbitrary parameter. In
contrast, in the defocusing or linear cases = +1, 0 no blowup solution of the form
(3.3) is possible, because (3.4) enjoys a coercive Hamiltonian7
1 1
H(v, t v) = |t v|2 + ||2 |v|2 + |v|p+1
2 2 p+1
and thus (by Exercise 1.29) the solutions (3.3) will stay globally bounded for all
time. Similarly in the focusing case if the initial data is very small compared to the
frequency. Thus we see that the large data focusing behaviour is quite bad when
compared to the defocusing or linear cases.
The solutions of the form (3.3) have no decay in space and so will not lie in
Sobolev spaces such as Hxs (Rd ), although if the frequency lies in the integer lattice
Zd then we can view these solutions as lying in the periodic Sobolev spaces Hxs (Td )
for any s. In the (focusing) non-periodic case it is possible to create a different class
of solutions by choosing an ansatz which oscillates in time rather than in space:
(3.7) u(t, x) = Q(x)eit ,
where R. This leads to the ground state equation
(3.8) Q + |Q|p1 Q = Q
where (, ) := (2, 2 ) for NLS and (, ) := (, 2 ) for NLW. In the defocus-
ing case we can take , > 0 (choosing to be positive). From Appendix B we
then recall that if 1 < p < is energy-subcritical in the sense that d2 p12
< 1,
d
then there exists a smooth, positive, rapidly decreasing solution Q Sx (R ) to the
equation (3.8). This leads to the standard ground state soliton solution to either
NLS or NLW associated to the temporal frequency > 0; it lies in every spatial
Sobolev space Hxs (Rd ), and has a very simple behaviour in time. In the next section
we will apply the symmetries of NLW and NLS to generate further ground state
solitons. These solitons are only available in the focusing case; In Section 3.5 we
shall establish Morawetz inequalities which show that nothing remotely resembling
a soliton can occur in the defocusing equation.
Exercise 3.1. Obtain the analogue of Exercise 2.47 for the NLS, and Exercise
2.58 for the NLW, by adding the nonlinear potential energy term V (u) to the
Hamiltonian.
2
Exercise 3.2. Let u Ct,x,loc (R Rd V ) be a classical solution to a
2
d-dimensional NLS. Show that the field v Ct,x,loc (R Rd+1 V ) defined by
t xd+1
v(t, x1 , . . . , xd , xd+1 ) := ei(t+xd+1 ) u(
, x1 , . . . , xd )
2
is a classical solution to the corresponding d + 1-dimensional NLW (cf. Exercise
2.11). This correspondence may help explain why many of the algebraic expressions
7The case = = 0 is degenerate coercive, but this case can be treated by hand, leading to
solutions of linear growth in time.
114 3. SEMILINEAR DISPERSIVE EQUATIONS

defined below for the NLW have a counterpart for NLS, but with d replaced by
d + 1. This correspondence is less useful for the Hxs wellposedness theory, because
the functions v constructed above will not have finite Hxs norm. One should also
caution that this correspondence does not link periodic NLS solutions with periodic
NLW solutions.
Exercise 3.3. By taking formal limits of the Lax pair formulation of the
periodic Ablowitz-Ladik system as discussed in Section 1.7, discover a Lax pair
formalism for the one-dimensional cubic defocusing Schrodinger equation (in either
the periodic or nonperiodic settings).

3.1. On scaling and other symmetries

It has long been an axiom of mine that the little things are infin-
itely the most important. (Sir Arthur Conan Doyle, A Case of
Identity)
We now describe the concrete symmetries of NLS and NLW; to avoid technical-
ities let us just work with classical solutions for now (we will discuss more general
notions of solution in the next section). The NLS (3.1) enjoys the scaling symmetry
t x x
(3.9) u(t, x) 7 2/(p1) u( 2 , ); u0 (x) 7 2/(p1) u0 ( )

for any dilation factor > 0 (thus time has twice the dimensionality of space), and
the Galilean invariance
2
(3.10) u(t, x) 7 eixv eit|v| /2
u(t, x vt); u0 (x) 7 eixv u0 (x)
for any velocity v Rd (cf. Exercise 2.5). It also enjoys the more mundane sym-
metries of time translation invariance, space translation invariance, spatial rotation
symmetry, phase rotation symmetry u 7 ei u, as well as time reversal symmetry
u(t, x) 7 u(t, x); u0 (x) 7 u0 (x).
In the pseudo-conformal case p = pL2 := 1 + d4 , one also use Exercise 2.28 to verify
the pseudo-conformal symmetry
1 1 x 2
(3.11) u(t, x) 7 u( , )ei|x| /2t
(it)d/2 t t
for times t 6= 0. This symmetry is awkward to use directly (at least when t0 = 0)
because of the singularity at t = 0; one typically uses the time translation and time
reversal symmetries to move the singularity elsewhere (e.g. to the time t = 1).
Similarly, the NLW (3.2) enjoys the scaling symmetry
t x
u(t, x) 7 2/(p1) u( , );

x
(3.12) u0 (x) 7 2/(p1) u0 ( );

x
u1 (x) 7 2/(p1)1 u1 ( )

for any dilation factor > 0 (thus time and space have equal dimension), and (if
the solution exists globally in time) the Lorentz invariance
tvx xv vt
(3.13) u(t, x) 7 u( p , x xv + p )
1 |v|2 1 |v|2
 3.1. ON SCALING AND OTHER SYMMETRIES 115

for all sub-luminal velocities v Rd with |v| < 1 (cf. Exercise 2.6); note that the
effect of this invariance on the initial data u0 , u1 is rather complicated and requires
solving the equation (3.2). The NLW also enjoys spacetime translation invariance,
spatial rotation symmetry, phase rotation symmetry, conjugation symmetry, and
4
time reversal symmetry. In the conformal case p = pH 1/2 = 1 + d1 , one also has
x
the conformal symmetry
t x
u(t, x) 7 (t2 |x|2 )(d1)/2 u( 2 , )
t |x|2 t2 |x|2
inside the light cone |t| < |x|, thanks to Exercise 2.14.
Unlike the Galilean invariance (3.10), the Lorentz invariance (3.13) has the
effect of time dilation - solutions which would ordinarily exhibit some special be-
haviour (e.g. blowup)p at a time T will instead do so at a much later time, typically
of the order of T / 1 |v|2 . To compensate for this one can p compose the Lorentz
transformation with the scaling transformation with := 1/ 1 |v|2 , leading to
the normalised Lorentz invariance
p
(3.14) u(t, x) 7 (1 |v|2 )1/(p1) u(t v x, 1 |v|2 (x xv ) + xv vt).
Symmetries have many uses. Through Noethers theorem, they indicate what
conservation laws are available (though certain symmetries, particularly discrete
ones, do not necessarily yield a conservation law). They can give guidance as to
what type of techniques to use to deal with a problem; for instance, if one is trying to
establish wellposedness in a data class which is invariant under a certain symmetry,
this suggests the use of estimates and other techniques which are also invariant
under that symmetry; alternatively, one can spend the symmetry by normalising
the solution, for instance in making the solution centred or concentrated at the
origin (or some other specified location) in space, time, or frequency. If the data
class is subcritical with respect to scaling, one can use the scaling symmetry to
trade between time of existence and size of initial data; thus if one establishes a
local wellposedness at a fixed time (say up to time T = 1) for data with small
norm, then one can often also establish local wellposedness at a small time for
large data; typically the time of existence will be proportional to some negative
power of the norm of the data. Conversely, if the data class is supercritical with
respect to scaling (or more generally is lower than the invariant norm associated
to another symmetry), then it is likely that there is a significant obstruction to
obtaining a wellposedness theory below that regularity, and one also expects the
wellposedness theory at that regularity to be rather delicate. The reason for this
is that if the regularity is below the invariant regularity, then one can use the
symmetry to convert bad behaviour arising from large initial data at some time
t > 0 to bad behaviour arising from small initial data at some time less than or
equal to t, where large and small are measured with respect to the regularity
Hxs (Rd ). Since large initial data would be expected to display bad behaviour very
quickly, one then expects to give examples of arbitrary small initial data which
displays bad behaviour arbitrarily quickly. This can often be used to contradict
certain types of wellposedness that one could hypothesise for this regularity. See
also Principle 3.1 below.
Let us give some sample applications of the symmetry laws. The first is a
blowup result for the pseudoconformal focusing NLS (so = 1 and p = pL2x =
1+ d4 ). Recall that this equation has a soliton solution of the form u(t, x) = eit Q(x)
116 3. SEMILINEAR DISPERSIVE EQUATIONS

for any > 0, where the ground state Q is a nonnegative Schwartz solution to the
equation
Q + 2Q1+4/d = 2 Q.
Applying the pseudoconformal transformation (3.11), we obtain the solution
1 2
(3.15) d/2
ei /t ei|x| /2t Q(x/t).
(it)
This solves the NLS equation for all times t 6= 0, and in fact lies in every Sobolev
space Hxs (Rd ) for such times, but blows up in a rather dramatic way as t 0.
Thus the pseudoconformal focusing NLS can lead to blowup even from very smooth
decaying initial data, though we will later see that this is due to the initial datum
being large in an L2x (Rd ) sense. This blowup occurs despite the solution being
bounded in L2x , and despite the conservation of the L2x norm. Thus for PDE, a
positive definite conservation law is not always sufficient to prevent blowup from
occuring, in marked contrast to the situation for ODE; note that the solution (3.15)
demonstrates rather vividly the lack of compactness of bounded subsets of L2x (Rd ).
Now consider a general NLS. Applying the pseudoconformal transformation in
Exercise 2.28, one no longer expects to recover the original equation; instead, the
transformed field v(t, x) will now obey the equation
d
(ppL2 )
(3.16) it v + v = t 2 x |v|p1 v
for t 6= 0, where pL2x := 1 + d4 is the pseudoconformal power. We will analyze this
equation (3.16) in more detail later, but for now let us just extract a special class
of solutions to (3.16) (and hence to NLS), by considering solutions v which are
independent of the spatial variable and thus simply solve the ODE
d
(ppL2 )
(3.17) it v = t 2 x |v|p1 v.
This ODE can be solved explicitly as
i||p1 q
v(t, x) = exp( t )
q
for any C and with q := d2 (p pL2x ) + 1, though in the critical-range case q = 0
(so p = 1 + d2 ) we have instead the solution

v(t, x) = exp(i||p1 log |t|).

We can of course invert the pseudoconformal transformation and obtain explicit
solutions to the original NLS for t 6= 0, namely
1 i|x|2 i||p1
(3.18) u(t, x) = exp( + )
(it)d/2 2t qtq
when q 6= 0 and
1 i|x|2
(3.19) u(t, x) = exp( + i||p1 log |t|)
(it)d/2 2t
when q = 0. Of course when = 0 we recover the explicit solution u(t, x) =
2
1
(it)d/2
exp( i|x|
2t ) to the linear Schrodinger equation (essentially the fundamental
 3.1. ON SCALING AND OTHER SYMMETRIES 117

solution8 for that equation). Comparison of these solutions yields the following
heuristic: as t , the nonlinear Schrodinger equation should resemble the
linear Schrodinger equation when in the short-range case q > 0 (so p > 1 + d2 ), but
not in the long-range 9 case q < 0 or critical-range case q = 0 (though the divergence
between the two equations should consist primarily of a phase shift, which should
be somehow logarithmic in the critical-range case). We will see some justification
of this heuristic later in this chapter, though our understanding here is far from
complete.
Now we observe some applications of the Galilean invariance law (3.10). Let
us begin with a periodic NLS (with d, p and = 1 arbitrary). In this periodic
setting we have the plane wave solutions
2
t/2 i||p1 t
(3.20) u, (t, x) := eix ei|| e
d
for any 2Z and C; one can view this as the Galilean transform of the
p1
constant-in-space solutions ei|| t . Suppose one fixes the frequency parameter
to be large, and considers two distinct solutions u, , u , of the above type with
|| | |. At time zero we have
ku, (0)kHxs (Td ) , ku , (0)kHxs (Td ) ||||s ;
ku, (0) u , (0)kHxs (Td ) | |||s
while at any later time t we have
ku, (t)kHxs (Td ) , ku , (t)kHxs (Td ) ||||s ;
p1 p1
ku, (t) u , (t)kHxs |ei|| t
ei| | t
|||s .
Thus the Hxs norms of the solutions u, and u , do not change much in time, but
the Hxs separation of these solutions can change due to a phase decoherence effect.
Indeed we see that if || =6 | |, then there exists a time t ||1p for which the two
phases become completely decohered, and ku, (t) u , (t)kHxs (Td ) ||||s . If s
is negative, then by taking to be large and || to be comparable to (||/)1/s , we
can construct for any , > 0, a pair of solutions u, , u , to NLS of Hxs (Td ) norm
O() and Hxs (Td ) norm separation O() at time zero, such that at some later time
t = O() the Hxs (Td ) norm separation has grown to be as large as O(). This shows
that for negative s, a pair of solutions can separate in Hxs (Td ) norm arbitrarily
quickly; more precisely, the solution map u0 7 u is not uniformly continuous from
Hxs to Ct0 Hxs ([0, T ] Td ) even for arbitrarily small T and for arbitrarily small
balls in Hxs (Td ). This is a negative result that rules out certain types of strong
wellposedness results for the periodic NLS for negative Sobolev regularities.
8Indeed, one could view (3.18), (3.19) as the nonlinear fundamental solution for NLS.
However these solutions are nowhere near as useful as the fundamental solution is for the linear
equation, since we no longer have the principle of superposition in the nonlinear case and so we
cannot build general solutions by superimposing translates of the fundamental solution. Never-
theless, these explicit solutions provide some useful intuition as to the asymptotic behaviour of
the equation for general data.
9The terminology here signifies the long-term strength of the nonlinearity and can be justified
heuristically as follows. One can view the nonlinearity in NLS as a potential term with time-
dependent potential |u|p1 . Assuming that the nonlinear evolution decays at the same rate as
the linear one, dispersive estimates suggest that |u| should decay like td/2 . Thus we expect in the
short-range case we expect the potential to be integrable in time, which suggests by Gronwalls
inequality that the long-term effect of the nonlinearity is bounded.
118 3. SEMILINEAR DISPERSIVE EQUATIONS

One can run a similar argument for nonperiodic focusing NLS, by starting with
the ground state solution eit Q(x), rescaling it by and then applying a Galilean
transform to obtain the moving soliton solution
2
/2+it /2
(3.21) uv, (t, x) 7 2/(p1) ei(xv+it|v| Q((x vt)/)
for any v Rd and > 0; one can use these solutions to show that the solution map
to NLS (if it exists at all) is not uniformly continuous in Hxs for certain low s; see
Exercise 3.5. A similar result is also known for the defocusing case, replacing the
soliton solutions with another family of solutions that can be viewed as truncated
versions of the plane wave solutions (3.5); see Section 3.8.
Among all the symmetries, the scale invariance (3.9), (3.12) is particularly im-
portant, as it predicts a relationship between time of existence and regularity of
initial data. Associated to this invariance is the critical regularity sc := d2 p1 2
.
Note that the scaling (3.9) preserves the homogeneous Sobolev norm ku0 kH sc (Rd ) ,
and similarly (3.12) preserves ku0 kH sc (Rd ) + ku1 kH sc 1 (Rd ) . The relationship be-
tween scaling and the inhomogeneous counterparts to these Sobolev norms is a
little more complicated, of course. We refer to regularities s > sc above the critical
norm as subcritical, and regularities s < sc below the critical norm as supercrit-
ical. The reason for this inversion of notation is that higher regularity data has
better behaviour, and thus we expect subcritical solutions to have less pathologi-
cal behaviour than critical solutions, which in turn should be better behaved than
supercritical solutions. The other scalings also have their own associated regulari-
ties; the Galilean symmetry and pseudoconformal symmetry preserve the L2x (Rd )
norm, whereas the Lorentz symmetry and conformal symmetries are heuristically
associated to the H 1/2 (Rd ) H 1/2 (Rd ) norm (see Exercise 2.23).
In general, the relationship between the regularity Hxs of the initial data, the
scale-invariant regularity H sc of the equation, the frequencies of the solution, and
the evolution of the solution tends to follow the following informal principles10:
Principle 3.1 (Scaling heuristic). Let u be a solution to either the NLS (3.1)
or NLW (3.2), with initial position u0 in Hxs (and initial velocity u1 in H s1 , in
the case of the NLW).
(a) In the subcritical case s > sc , we expect the high frequencies of the solution
to evolve linearly for all time (unless a stronger obstruction than scaling
exists). The low frequencies of the solution will evolve linearly for short
times, but nonlinearly for long times.
(b) In the critical case s = sc , we expect the high frequencies to evolve linearly
for all time if their H sc norm is small, but to quickly develop nonlinear be-
haviour when the norm is large. (Again, we are assuming that no stronger
obstruction to linear behaviour than scaling exists.) The low frequencies of
the solution will evolve linearly for all time if their H sc norm is small, but

10This principle should be taken with a grain of salt. On the one hand, it gives a good
prediction for those equations in which the scaling symmetry is in some sense dominant, and
for which the worst types of initial data are given by bump functions. On the other hand, there are
other situations in which other features of the equation (such as Galilean or Lorentz symmetries,
or resonances) dominate, in which case instability can occur even when the scaling heuristic
predicts good behaviour. Conversely, some features of the equation, such as conservation laws or
monotonicity formulae, can provide more stability than the scaling heuristic predicts.
 3.1. ON SCALING AND OTHER SYMMETRIES 119

will eventually develop nonlinear behaviour (after a long period of time)

when the norm is large.
(c) In the supercritical case s < sc , the high frequencies are very unstable and
will develop nonlinear behaviour very quickly. The low frequencies are in
principle more stable and linear, though in practice they can be quickly
disrupted by the unstable behaviour in the high frequencies. (This rela-
tively good behaviour of the low frequencies is sometimes enough to obtain
a weak solution to the equation, however, by using viscosity methods to
suppress the high frequencies; see Exercise 3.56.)

Let us now briefly give a heuristic discussion that lends some support to Prin-
ciple 3.1. Let N > 0 be a frequency; frequencies N 1 correspond to high
frequencies, while frequencies N 1 correspond to low frequencies. A model ex-
ample of an initial datum u0 of frequency N is a function which is supported on
a ball B of radius 1/N , does not oscillate too much on this ball, and reaches an
amplitude A on this ball. (For the NLW, one would also need to similarly specify
an initial velocity.) We will assume that this rescaled bump function example is
the worst type of initial data in the given class (i.e. bounded or small functions
in Hxs or H s1 ); this assumption corresponds to the caveat given in the above prin-
ciple that no stronger obstructions to linear behaviour exist than the scaling one.
The L2x norm of such a datum is roughly AN d/2 , and more generally (from
the Fourier representation of the Hxs norm) we expect the Hxs norm of this datum
to be AN sd/2 ; thus if u0 is bounded in Hxs then A = O(N d/2s ), and if u0
is small in Hxs then A N d/2s . Now, both the NLS (3.1) and the NLW (3.2)
contain a linear term u and a nonlinear term |u|p1 u. On the ball B (at least
for times close to 0), the linear term has magnitude N 2 A, while the nonlinear
term has amplitude Ap . If N 2 A Ap , we thus expect the linear term to dom-
inate, and the solution should behave linearly (cf. Principle 1.37). If Ap N 2 A,
we expect the nonlinear term to dominate and so one eventually expects nonlinear
(and unstable) behaviour. The time in which this nonlinear behaviour becomes
apparent can be predicted by comparing u against its time derivative t u or its
second time derivative t2 u. For instance, suppose we have an NLS in which the
nonlinear behaviour dominates, thus t u will be dominated by the nonlinear term
|u|p1 u, which has amplitude Ap . Since u itself has amplitude A, we expect
the nonlinear behaviour to significantly affect the initial datum after time A/Ap .
Using these heuristics, one can give informal justification for all three components
(a), (b), (c) if Principle 3.1; see Exercise 3.4.
A particular interesting case is when the scale-invariant regularity coincides
with one of the other special regularities, such as the Hx1 norm (associated to
1/2
the energy or Hamiltonian), the Hx norm (associated to the momentum in NLS
and to the symplectic structure, Lorentz invariance, and conformal invariance in
NLW), and the L2x norm (associated to the Galilean invariance, pseudoconformal
invariance, and mass in NLS, and being the limiting regularity in NLW to even
make sense of (3.2) distributionally); see Table 1. Thus we isolate as special cases
4
the Hx1 -critical (or energy-critical ) case sc = 1 (thus d 3 and p = 1 + d2 ),
1/2 4
the Hx -critical case sc = 1/2 (thus d 2 and p = 1 + d1 ) and the L2x -critical
case sc = 0 (thus d 1 and p = 1 + d4 ). One can also discuss the Hx1 -subcritical
case sc < 1 and the Hx1 -supercritical case sc > 1, etc. Another relevant regularity
120 3. SEMILINEAR DISPERSIVE EQUATIONS

Table 1. The critical exponents for small dimension. The cases

when a critical exponent corresponds to an algebraic equation (i.e.
p is equal to 3 or 5) are of particular interest.

1/2
Dimension L2x -critical Hx -critical Hx1 -critical
1 5
2 3 5
3 7/3 3 5
4 2 7/3 3
5 9/5 2 7/3
6 5/3 9/5 2

sc
in the case of NLW is the Lorentz regularity sl := d+1 1 1
4 p1 = 2 + 4 , which is
the regularity which is heuristically associated to the normalised Lorentz invariance
(3.14), and is halfway between the scale-invariant regularity sc and the conformal
regularity 12 .

Exercise 3.4. Use the heuristic analysis of bump function initial data, as
described in this section, to give some informal justification to Principle 3.1. (Be
prepared to make a large number of hand-waving assumptions. The important thing
here is to develop the numerology of exponents; rigorous support for these heuristics
will be have to wait until later in this chapter.) In the subcritical case, develop a
heuristic relationship between the Hxs norm of the initial data and the predicted
2/(ss )
time T in which the linear behaviour dominates. (One should get T ku0 kH s c
1/(ss )
for NLS and T (ku0 kH s + ku1 kH s1 ) c
for the NLW.)

Exercise 3.5. [BKPSV] Let d, p be arbitrary, let = +1, and let s < 0 or
2
s < sc := d2 p1 . Using the solutions (3.21), show that for any , > 0 there exists
a pair of classical solutions u, u to (3.1) with Hxs (Rd ) norm O() and Hxs (Rd ) norm
separation O() at time zero, such that at some later time t = O() the Hxs (Rd )
norm separation has grown to be as large as O(). This shows that there is no
uniform wellposedness at this regularity, at least for the focusing regularity.

3.2. What is a solution?

For every complex problem, there is a solution that is simple, neat,
and wrong. (H.L. Mencken, The Divine Afflatus)
Before we begin the analysis of our model problems (3.1), (3.2), let us pause
to address a rather fundamental question, namely what it actually means for a
field u to be a solution of either of these two Cauchy problems. This question may
sound philosophical in nature, but the properties associated to making a solution
concept strong are well worth establishing rigorously, as they become important
in establishing many of the further properties of the equation, such as the global
existence and asymptotics of (classical) solutions.
The question of defining exactly what a solution is is more subtle than it may
first appear, especially at low regularities. The reason for this is that in order
for a solution to a PDE to actually be useful for applications, it is not merely
 3.2. WHAT IS A SOLUTION? 121

enough that it exist and solve the equation in some weak sense11 (e.g. in the sense
of distributions), though this is certainly a minimal requirement; one also often
desires additional properties on the solution, which do not automatically follow
from the fact that the equation is solved weakly. We informally describe some of
the most important of these properties12 as follows.
Existence: Is the solution guaranteed to exist (locally, at least) for all
initial data in a certain class (e.g. Hxs )?
Uniqueness: Is the solution the unique object in a certain solution class
(e.g. Ct0 Hxs (I Rd )) which solves the equation in a suitable sense (e.g.
in a distributional sense)? Is this solution concept compatible with other
notions of solution (i.e. if two solutions to the same equation exist in two
different senses, are they equal to each other)?
Continuous dependence on the data: Do small perturbations of the initial
datum (in some norm) lead to small perturbations in the solution (in
some other norm)? In other words, is the solution map continuous? One
can also ask for stronger continuity properties such as uniform continuity,
Lipschitz continuity, or analyticity.
Bounds: If the initial datum is in some class, say Hxs , can one control
the solution in some other class, e.g. Ct0 Hxs (I Rd )? In particular, does
one have persistence of regularity: is the solution always as smooth as the
initial datum (as measured in an Hxs sense)?
Lifespan estimates: Is there a lower bound on the lifespan of the solution
in terms of the initial data (or in terms of some norm of the initial data,
such as an Hxs (Rd ) norm)? Equivalently, is there a blowup criterion that
gives necessary conditions for the lifespan to shrink to zero? In some cases
one has global existence, which case the lifespan is infinite.
Approximability by smooth solutions: if the solution is rough, can it be
written as the limit (in some topology) of smoother solutions? If the
initial datum is approximated by a sequence of smooth initial data, do
the corresponding solutions necessarily converge to the original solution,
and in what sense?
Stability: If one perturbs the equation (thus considering fields which only
solve the original equation approximately), to what extent can these near-
solutions be approximated by the exact solution with the same (or a
nearby) initial datum?
11This is in marked contrast with the theory of linear differential equations, in which dis-
tributional solutions are very tractable, and can mostly be manipulated as if they were classical
solutions, in large part because they can be expressed as the weak limit of classical solutions. Since
weak convergence is often not preserved under basic nonlinear operations such as multiplication of
two functions, one generally requires in nonlinear applications that a solution be a strong limit of
classical solutions, which usually leads to the requirement that one work with wellposed solutions;
see below.
12For elliptic PDE, another important property that one often desires is that the solution u
is a minimiser, or at least a critical point, of the Lagrangian associated to the PDE, with respect
to various classes of perturbation. One could insist on something similar for nonlinear wave and
Schrodinger equations, but this has not proven to be as fruitful a property for these equations
as in the elliptic case, in large part because of the highly non-convex nature of the Lagrangians
involved. However, the Lagrangian formulation is (formally) linked to important properties such
as conservation laws and monotonicity formulae, which are very desirable properties for a solution
to obey.
122 3. SEMILINEAR DISPERSIVE EQUATIONS

Structures: Do the conservation laws of the equation, which can be rigor-

ously justified for classical (i.e. smooth and decaying) solutions, continue
to hold for the solution class being studied? Similarly for monotonicity
formulae, symmetries of the equation, etc.
Thus, instead of having a single unified concept of a solution class, one has
instead a multi-dimensional continuum of such classes, ranging from very weak
solution classes (in which the equation solves the equation in a weak sense, or
is perhaps a weak limit of smoother solutions or near-solutions, but not much
else), to very strong solution classes, in which one has many or all of the desirable
properties listed above. Generally speaking, it is fairly easy to show existence of
solution in a weak solution class by various limiting arguments (e.g. iteration or
weak compactness arguments), but non-trivial effort is then required to upgrade
those solutions to lie in stronger solution classes.
In this section we shall discuss five notions of solution, which in decreasing order
of strength are classical solution, wellposed Hxs solution, strong Hxs solution, weak
Hxs solution, and distributional solution respectively. In fact, in this monograph we
shall largely work with wellposed and classical solutions, in order to avoid a number
of subtleties involving the weaker notions of solution.
To fix the discussion let us just work with the NLS equation (3.1), and fix
our initial data class to be a Sobolev space Hxs (Rd ). The strongest notion of a
solution is that of a classical solution. These can (broadly speaking) be defined as
solutions which have so much differentiability and decay that there is no difficulty
interpreting the problem (3.1) in a classical sense (i.e. without requiring the theory
of weak derivatives). Furthermore, one has enough regularity and decay available13
to justify all the various formal manipulations associated to the equation, such as
conservation laws, monotonicity formulae, and so forth. The amount of regularity
required to do all this can be quite large; for instance, in order to justify conser-
vation of the Hamiltonian for NLS safely, one requires as much as three orders of
differentiability on the solution, as well as some additional uniformity and decay
conditions. Because of this, one occasionally runs into issues generating a classi-
cal solution theory when the nonlinearity |u|p1 u is not very smooth (which can
happen when p is close to 1); in such cases one may need to regularise the non-
linearity before discussing classical solutions. However this issue does not arise for
the algebraic equations, in which p is an odd integer.
It is also easy to establish uniqueness for classical solutions (essentially because
the proof of Theorem 1.14 carries over without difficulty). Here are two typical
such results, one for NLS and one for NLW.
Proposition 3.2 (Uniqueness for classical NLS solutions). Let I be a time
interval containing t0 , and let u, u Ct,x
2
(I Rd C) be two classical solutions
to (3.1) with the same initial datum u0 for some fixed and p. Assume also that we
have the mild decay hypothesis u, u L q d
t Lx (I R ) for q = 2, . Then u = u .

Proof. By time translation symmetry we can take t0 = 0. By time reversal

symmetry we may assume that I lies in the upper time axis [0, +). Let us write
u = u + v. Then v Ct,x
2
(I Rd C), v(0) = 0, and v obeys the difference

13This is a somewhat vague definition, but in practice we will always apply limiting arguments
to generalise classical solutions to a wider class of wellposed solutions, and so the exact notion of
a classical solution will not be important as long as it is dense in the class of wellposed solutions.
 3.2. WHAT IS A SOLUTION? 123

equation
it v + v = (|u + v|p1 (u + v) |u|p1 u).
Since v and |u+v|p1 (u+v)|u|p1u lie in L 2 d
t Lx (IR ), we may invoke Duhamels
formula (2.13) and conclude
Z t
v(t) = i ei(ts)/2 (|u + v|p1 (u + v) |u|p1 u)(s) ds
0

for all t I. By Minkowskis inequality, and the unitarity of ei(ts) , we conclude

Z t
kv(t)kL2x (Rd ) k(|u + v|p1 (u + v) |u|p1 u)(s)kL2x (Rd ) ds.
0

Since u and v are in L

t Lx (I Rd ), and the function z 7 |z|p1 z is locally
Lipschitz, we have the bound
k(|u + v|p1 (u + v) |u|p1 u)(s)kL2x (Rd ) .p (kukpL L (IRd ) + kvkpL L (IRd ) )
t x t x

kv(s)kL2x (Rd ) .
Applying Gronwalls inequality (Theorem 1.10) we conclude that kv(t)kL2x (Rd ) = 0
for all t I, and hence u = u as desired. 
Note that Exercise 2.24 shows that some sort of decay condition is necessary
in order to establish uniqueness, even when no nonlinearity is present. For NLW
one also has uniqueness of classical solutions, and moreover one can even localise
the uniqueness by exploiting finite speed of propagation:
Proposition 3.3 (Uniqueness and finite speed of propagation for classical
NLW solutions). Let t0 = 0. Let I be a time interval containing 0, and let u, u
2
Ct,x,loc (I Rd C) be two C 2 solutions to (3.2) such that the initial data u[0] =
(u(0), t u(0)) and u [0] = (u (0), t u (0)) agree on the ball {x Rd : |x x0 | R}.
Then we have u(t, x) = u (t, x) for all t I and x Rd with |x x0 | R |t|.
Proof. By spatial translation invariance we may take x0 = 0; by time reversal
symmetry we may restrict attention to times 0 t R. By shrinking I if necessary
we may take I to be compact. Write u = u + v, then v Ct,x 2
(I Rd C), v[0]
d
vanishes on the ball {x R : |x| R}, and v solves the difference equation
v = F
where F := (|u + v|p1 (u + v) |u|p1 u). Now let us define the local energy E(t)
for 0 t R by Z
E(t) := T00 (t, x) dx
|x|Rt

where T00 := 21 |t v|2 + 21 |x v|2 is the linear energy density, thus E(0) = 0. A
2
computation (which is justified when v is Ct,x ) shows that
t T00 + j T0j = Re(F t v)
where T0j := Re(j vt v) is the energy current. From Stokes theorem (and the
2
fact that v is Ct,x ) we conclude
Z Z
t E(t) = Re(F t v)(t, x) dx + T0j nj T00 dS
|x|Rt |x|=Rt
124 3. SEMILINEAR DISPERSIVE EQUATIONS

where dS is the surface element and nj := xj /|x| is the outward normal. From
Cauchy-Schwarz we see that |T0j nj | T00 , thus we have
Z
t E(t) |F (t, x)||t v|(t, x) dx.
|x|Rt

Now since u and v will be bounded on the compact region {(t, x) I Rd : 0

t R; |x| R t}, we see that F = Ou,v (|v(t, x)|). Applying Cauchy-Schwarz we
have
Z Z
t E(t) .u,v ( |v(t, x)|2 dx)1/2 ( |t v(t, x)|2 dx)1/2 .
|x|Rt |x|Rt
R
By definition of energy we have ( |x|Rt |t v(t, x)|2 dx)1/2 E(t)1/2 . Writing
Rt
v(t, x) = 0 t v(s, x) ds and using Minkowskis inequality and the fact that v(0, x) =
0 when |x| R, we also see that
Z Z t
( |v(t, x)|2 dx)1/2 E(s)1/2 ds.
|x|Rt 0
1/2
Dividing out by E(t) , we conclude that
Z t
t E(t)1/2 .u,v E(s)1/2 ds
0
which after integration in t (and recalling that E(0) = 0) yields
Z t
1/2
E(t) .u,v t E(s)1/2 ds.
0
Applying Gronwalls inequality (Theorem 1.10) we conclude E(t) = 0 for all 0
t R, and the claim easily follows. 
The classical theory is generally sufficient for very smooth initial data (and very
smooth nonlinearities u 7 |u|p1 u), but for rougher data one and nonlinearities
must adopt a different approach. Because the differential formulation of the prob-
lem (3.1) requires so much differentiability, it is often better to work instead with
the integral formulation of the equation,
Z t

(3.22) u(t) = ei(tt0 )/2 u0 i ei(tt )/2 (|u(t )|p1 u(t )) dt ;
t0
for NLS and

sin((t t0 ) )
u(t) = cos((t t0 ) )u0 + u1

(3.23) Z t
sin((t t0 ) )
(|u|p1 u(t )) dt
t0
for NLW; these equations can make sense even when u is a tempered distribution
which lies locally in Lpt Lpx . We refer to such solutions as distributional solutions
to the equation. When u has sufficient smoothness and regularity, these solutions
coincide with classical solutions, but are more general in the case when u is rough.
Typically, the initial datum u0 will also lie in a Sobolev space such as Hxs (Rd ).
Recall (from the Fourier transform) that if u0 Hxs (Rd ), then eit/2 u0 Ct0 Hxs (R
Rd ) L s d
t Hx (R R ). Inspired by this, we distinguish two subclasses of distribu-
tional solution:
 3.2. WHAT IS A SOLUTION? 125

A strong Hxs solution to (3.22) on a time interval I is a distributional

0
solution which also lies in Ct,loc Hxs (I Rd ).
s
A weak Hx solution to (3.22) on a time interval I is a distributional
solution which also lies in L s d
t,loc Hx (I R ).
Similarly, we can define a strong Hxs Hxs1 solution to (3.23) to be a distribu-
0
tional solution which also lies in Ct,loc Hxs Ct,loc
1
Hxs1 , while a weak solution lies

in Lt,loc Hx with one (weak) time derivative in Lt,loc Hxs1 .
s

These definitions correspond to the notions of strong and weak solutions for
ODE discussed in Section 1.1, though unfortunately in the PDE setting it is usually
not known whether these notions are equivalent to each other. Generally speaking,
the category of strong Hxs solutions is the broadest category of solution in which we
can hope to have a good existence and uniqueness theory; for weak Hxs solutions
one typically can hope to have existence but not uniqueness. In some cases it is
possible to use the formula (3.22) to show that all weak solutions are automatically
strong (as in Lemma 1.3) but this generally only happens when s is large (and one
also needs the nonlinearity to be fairly smooth); see for instance Exercise 3.12. As a
rule of thumb, perturbative methods such as Duhamel iteration tend to yield strong
solutions, whereas weak compactness methods such as viscosity methods tend to
only generate weak solutions (see Exercise 3.56).
With strong Hxs solutions, u(t) and ei(tt0 )/2 u0 varies continuously in t and
so one can make sense of (3.22) for all times t I (as opposed to almost every
time t, or in a weak distributional sense). In particular a strong Hxs solution obeys
the initial condition u(t0 ) = u0 in the usual classical sense. Also, the notion of a
strong solution is stable under time translation or time reversal, and one can glue
together two strong solutions with overlapping intervals of existence; see Exercises
3.10, 3.11.
Of course, with such a low level of regularity it is not obvious at all how to
use the equation (3.22) to justify other desirable properties of a solution, such as
conservation laws or uniqueness, even when the solution is known to be a strong
Hxs solution. To do this one often needs to strengthen the notion of a strong Hxs
solution even further, by adding some additional properties of the solution map
u0 7 u. One particularly successful such strengthening is the notion of a wellposed
solution.
Definition 3.4 (Wellposedness). We say that the problem (3.1) is locally well-
posed in Hxs (Rd ) if for any u0 Hxs (Rd ) there exists a time T > 0 and an open ball
B in Hxs (Rd ) containing u0 , and a subset X of Ct0 Hxs ([T, T ] Rd ), such that for
each u0 B there exists a strong unique solution u X to the integral equation
(3.22), and furthermore the map u0 7 u is continuous from B (with the Hxs topol-
ogy) to X (with the Ct0 Hxs ([T, T ] Rd)). We refer to this strong solution u as the
Hxs -wellposed solution to the Cauchy problem (3.1) with the specified initial datum
u0 . If we can take X = Ct0 Hxs ([T, T ] Rd ) then we say that the wellposedness
is unconditional ; if we can take T arbitrarily large14 we say the wellposedness is
global rather than local. If the time T depends only on the Hxs norm of the initial
14This is strictly weaker than as asking for T = +, which would be a uniformly global
wellposedness assertion which would imply, among other things, that the Hxs norm of u(t) stays
bounded as t (i.e. u lies in Ct0 Hxs (R Rd ) rather than just Ct,loc
0 Hxs (R Rd )). Obtaining
such uniformly global bounds is possible for certain defocusing equations, and is a subset of the
scattering theory developed in Section 3.6.
126 3. SEMILINEAR DISPERSIVE EQUATIONS

datum we say the wellposedness is in the subcritical sense, otherwise it is in the

critical sense. We say that the wellposedness is uniform if the solution map u0 7 u
is uniformly continuous from B to X; similarly we define the notion of Lipschitz
wellposedness, C k wellposedness for k = 1, 2, . . ., and analytic wellposedness.
Remark 3.5. One can of course adapt this definition to other equations. For
the nonlinear wave equation (3.2), the initial data class is Hxs (Rd ) Hxs1 (Rd )
instead of Hxs , and the solution should lie in Ct0 Hxs Ct1 Hxs1 instead of Ct0 Hxs ,
but otherwise the definition is the same. One can also easily replace the Sobolev
space Hxs with other variants such as the homogeneous Sobolev space Hxs , though
it is advisable to stick to spaces which are preserved by the linear evolution, since
otherwise there is very little chance that there will be any sort of wellposedness for
the nonlinear evolution. (This is a major reason why we work with the L2x -based
Sobolev spaces Hxs in the first place.)
Remark 3.6. In practice, the space X will be quite explicit; it is typically the
space of all fields in Ct0 Hxs ([T, T ] Rd ) which obey an additional integrability or
regularity condition (i.e. they lie in some additional function space). In some cases
one also imposes a smallness condition in X, though such conditions can usually
be removed by additional arguments (for instance, by shrinking the time interval
to ensure the smallness condition holds, and then using continuity arguments to re-
extend the time interval). The space X is useful for understanding the development
of singularities; typically, a solution needs to leave the space X in order for a
singularity to develop.
Wellposed solutions are highly compatible with classical solutions. If the initial
datum is smooth, then the wellposed solution and classical solution usually coincide;
this usually follows from the uniqueness theory, as well as persistence of regularity
results (which we shall discuss in the next section). If the initial datum is rough,
then by approximating this datum by smooth data and taking advantage of the
continuity properties of the solution one can usually represent the wellposed solution
as the strong limit of classical solutions15 in the Ct0 Hxs topology (and often in other
topologies also). Note that this shows that the wellposed solution is canonical - it
is the unique limit of the classical solutions generated by any sequence of smooth
data converging to the initial datum, and so two wellposed classes of solutions
corresponding to different regularities (or different spaces X) will automatically
coincide on their common domain of initial data. Furthermore, wellposed solutions
are often able to enjoy the conservation laws and other formal identities which
would normally be reserved for classical solutions, by taking appropriate limits. In
some cases one needs to regularise the nonlinearity in addition to the initial datum;
in such situations the continuity of the solution map is not quite sufficient, and
one needs to supplement it with some stability properties of the solution, so that
near-solutions to the equation can be well approximated by genuine solutions. Such
stability properties are of independent interest, both for theoretical reasons (such
as understanding the asymptotic behaviour of solutions), and for physical reasons
(because they help justify the heuristic assumptions that one used to arrive at that
model). We shall see some examples of these properties in Section 3.7.
15In some cases one has to regularise the nonlinearity |u|p1 u by smoothing it out at zero
or tempering its growth at infinity, in order to obtain good classical solutions; we will ignore these
technicalities.
 3.2. WHAT IS A SOLUTION? 127

Another common trick is to use the method of a priori estimates to obtain

control on wellposed solutions. Suppose one wants to show that all wellposed
solutions in a certain class and of a certain size are bounded in some norm Y
by some constant M . Since one can approximate wellposed solutions by classical
solutions, it typically suffices (using tools such as Fatous lemma) to obtain the
desired bound for classical solutions only. The Y norm then typically depends
continuously on the time interval I, and so by using a continuity argument in time
one can assume as a bootstrap hypothesis that the solution is bounded in Y by
a larger constant such as 2M . This reduces matters to establishing an a priori
estimate; the desired conclusion is the same, namely that the Y norm is bounded
by M , but now we can make the a priori assumptions that the solution is smooth,
and is already bounded in Y by 2M . These hypotheses can be immensely useful;
the former hypothesis allows one to make all formal computations rigorous, and the
latter hypothesis is often crucial in order to obtain control of nonlinear terms. Also,
the method of a priori estimates can also exploit various delicate cancellations (such
as energy cancellations) arising from the structure of the equation, which are not
picked up in some other methods such as iteration methods (because the iterates
do not solve the exact equation and so do not exhibit these cancellations).
One common way to construct wellposed solutions is to use iterative methods,
such as Proposition 1.38. Such methods tend to yield a fairly strong type of well-
posedness, and can reduce the task of constructing solutions to that of verifying a
single multilinear or nonlinear estimate. However, when the regularity of the data
is extremely low, or equation behaves in an extremely nonlinear fashion, then such
methods can break down; indeed there are examples known where solutions still
exist, but one does not have the strong type of wellposedness implied by a iterative
argument (see for instance the discussion on the Benjamin-Ono equation in Sec-
tion 4.4, or of the wave map equation in Chapter 6). In such situations one needs
to either augment the iterative argument (using for instance some sort of gauge
transformation), or else use a completely different approach. One such approach
is the viscosity method (also known as the penalisation, weak compactness, or reg-
ularisation method). In this approach, one approximates the equation (3.1) by a
smoother equation, in which the nonlinearity is smoothed out and bounded, and an
additional dissipation term is added to ensure global existence (forward in time, at
least). This gives a sequence of approximate solutions, which one can demonstrate
to be uniformly bounded in some norm (e.g. the energy norm); the establishment
of such a priori control on the regularised solutions is usually the most difficult
task. One can then use weak compactness to extract a weak limit of these approx-
imate solutions (see for instance Exercise 3.56). This procedure typically produces
a weak solution to the original equation without much difficulty, but it is often
significantly harder to upgrade this solution to a strong solution or to establish
wellposedness properties such as uniqueness, continuous dependence on the data,
or persistence of regularity; also, the conservation laws are often not preserved by
weak limits (though one can often obtain monotonicity of the conserved quantity,
at least, by tools such as Fatous lemma), and it often requires a non-trivial amount
of additional effort to establish such laws16.

16To give an example, the notorious global regularity problem for the Navier-Stokes equations
remains open, despite the construction of global weak solutions by Leray over seventy years ago,
in 1934!
128 3. SEMILINEAR DISPERSIVE EQUATIONS

Remark 3.7. It is of interest to search for other ways to build solutions beyond
the two standard methods of iteration and regularisation. One variant of the itera-
tion method which is occasionally useful is the Nash-Moser iteration method, which
is a PDE version of Newtons method for finding roots of equations. The iterates
in this method tend to lose regularity with each iteration, but this is counteracted
by the extremely rapid convergence of the iteration scheme. For other types of
PDE (notably elliptic PDE), variational and topological methods have been very
effective in constructing solutions, but so far these methods have not been partic-
ularly successful when applied to nonlinear dispersive or wave equations (though
the induction on energy method, which we discuss in Section 5.4, can be thought
of as a type of variational approach, while the continuity method from Section 1.3
is a crude example of a topological approach). Another speculative possibility is
that probabilistic constructions of solutions, valid for almost all initial data rather
than all initial data, may eventually be more powerful than the current determinis-
tic methods, especially for supercritical equations where the deterministic methods
appear to be useless. This may require utilizing ideas from thermodynamics, such
as the use of invariant Gibbs measures and similar devices.
Exercise 3.6 (Preservation of reality). Show that if a classical solution to a
NLW is real-valued at one time t0 , then it is real-valued for all other times for which
the classical solution exists. (Use uniqueness and conjugation invariance.)
Exercise 3.7 (Preservation of symmetry). Let I be a time interval and let t0
2
I. Suppose u Ct,x,loc (I Rd C) is a classical solution to a NLS (resp. NLW)
such that u(t0 ) (resp. u[t0 ]) is spherically symmetric. For NLS, we furthermore
require that u obey the boundedness decay conditions in Proposition 3.2. Prove
that u(t) is in fact spherically symmetric for all times t I.
Exercise 3.8 (Descent of NLS). Suppose that a periodic NLS on a torus Td+1
is locally wellposed in Hxs (Td+1 ) in either the subcritical or critical sense. Show
that the same NLS, but placed on the torus Td of one smaller dimension, is also
locally wellposed in Hxs (Td ) in the same sense. The same statement also holds for
global wellposedness, and with NLS replaced by NLW (but of course we replace Hxs
by Hxs Hxs1 in that case).
Exercise 3.9 (Localised blowup for focusing NLW). Show that for any focusing
NLW there exists smooth compactly supported initial data (u0 , u1 ) for which the
Cauchy problem (3.2) does not admit a global classical solution. (Hint: modify the
explicit solution (3.6) and use Proposition 3.3.)
Exercise 3.10 (Time shifting of strong solutions). Let I be a time interval
containing t0 , and let u be a strong Hxs solution to (3.1) with initial datum u(t0 ) =
u0 . Let t1 be any other time in I, and let u1 := u(t1 ). Show that u is also a strong
Hxs solution to (3.1) with initial datum u(t1 ) = u1 . Thus the notion of a strong
solution is independent of the initial time. Obtain a similar result for the NLW
(3.2). Also, show that the field u(t, x) := u(t, x) is a strong Hxs solution to (3.1)
on the interval I with initial datum u(t0 ) = u0 . (These results can fail for weak
solutions; see Exercise 3.15.)
Exercise 3.11 (Gluing of strong solutions). Let I, I be intervals which in-
tersect at a single time t0 . Suppose that u, u are strong Hxs solutions to (3.1) on
I Rd and I Rd respectively with initial data u(t0 ) = u (t0 ) = u0 . Show that
 3.3. LOCAL EXISTENCE THEORY 129

the combined field u on (I I ) Rd is also a strong solution to (3.1). Obtain the

similar result for the NLW equation where u and u have matching initial positions
and initial velocities. This exercise, combined with Exercise 3.10, shows that there
is no difficulty gluing together strong solutions on adjacent time intervals to create
a unified strong solution.
Exercise 3.12. Let p be an odd integer and s > d/2. Show that every weak
Hxs solution to (3.1) is also a strong Hxs solution. (You will need the fact that Hxs
is an algebra; see Lemma A.8.)
Exercise 3.13 (Local uniqueness implies global uniqueness). Fix p, d, , s and
suppose that one knows that for any time t0 and initial datum u0 Hxs (Rd ), there
exists an open time interval I containing t0 such that there is at most one strong
Hxs solution to (3.1) on I Rd (i.e. one has local uniqueness of strong solutions).
Show that this automatically implies global uniqueness of strong solutions, or more
precisely for any time interval J containing t0 that there is at most one strong Hxs
solution to (3.1) on J Rd . (Hint: prove by contradiction and use a continuity
method.)

3.3. Local existence theory

The greatest challenge to any thinker is stating the problem in a
way that will allow a solution. (Bertrand Russell)
We are now ready to construct solutions to NLS and NLW, and analyze the
strength of such solutions, in the senses discussed in Section 3.2. We will not
attempt to give the most complete results here, but instead give a sample of results
which illustrate the basic iteration method17. The underlying idea of this method is
simple - select spaces S and N in which to hold the solution u and the nonlinearity
|u|p1 u respectively, at which point the problem reduces to that of establishing
linear and nonlinear estimates in S and N . The selection of these spaces, however,
is something of an art rather than a science; there are some standard spaces that
work well in many situations, and one can analyze individual iterates to suggest
what spaces are likely to work, and which ones will not; however there is certainly
no universal iteration space that can cover all cases, and in one usually needs to
tailor the precise spaces to the equation at hand18.
17This method seems to be the best method for solving NLS and NLW, at least in the
subcritical and critical settings, with the Strichartz estimates (possibly with some Besov-type
augmentations) being the ideal tool to close the iteration. For the less semilinear equations
studied in later chapters, which contain derivatives in the nonlinearity, the iteration method often
requires more ingenious choices of spaces and estimates, as well as some additional tricks such
as gauge transformations, and thus face some nontrivial competition from other methods such as
viscosity methods or methods based on exploiting energy cancellation to obtain a priori estimates,
which have their own strengths and weaknesses.
18One common way to proceed here is to compute the first few nonlinear iterates in the
Duhamel iteration scheme, and see what spaces one can estimate them in; if one can place them
all in a common space then this suggests what choices of S and N to use. Another way is
to work in reverse, starting with the quantity kukC 0 H s (which one needs to control to obtain
t x
wellposedness) and estimating it in terms of other norms of u using the Duhamel formula; in
doing so it will become apparent what types of norms need to be controlled in order to have a
chance of closing the iteration. Typically, one needs to control u in those spaces in which the
linear solution is already known to be controlled in. We will use some of these heuristics when
studying the existence problem for other PDE in later chapters.
130 3. SEMILINEAR DISPERSIVE EQUATIONS

A systematic study of the wellposedness theory for NLS can be found in [Caz2],
and for NLW in [Sog]. A basic heuristic for NLS is that one has local wellposed-
ness in Hxs (Rd ) if and only if s max(sc , 0), where sc := d2 p1 2
is the scale-
invariant regularity (and 0 is the Galilean-invariant regularity); the corresponding
heuristic for NLW is that one has local wellposedness in Hxs (Rd ) if and only if
s max(sc , sl , 0), where sl := d+1 1
4 p1 is the regularity associated to the Lorentz
invariance. This heuristic is only partially accurate (wellposedness can break down
or become weaker when the nonlinearity becomes very rough compared to the reg-
ularity s, and in the case of the NLW there are still some very low regularities and
exponents for which the problem is not fully resolved, see [Tao].
To simplify the notation let us use the time translation invariance to fix the
initial time t0 to equal zero.
Let us begin with classical solutions to NLS. It turns out that to construct clas-
sical solutions it is more convenient to work in Sobolev spaces Hxs (Rd ) or weighted
Sobolev spaces Hxk,k (Rd ) (for suitably high values of s, k) than in more classical
spaces such as Cxk (Rd ); the main reason for this is that the linear propagator eit/2
preserves Hxs (Rd ) and are locally bounded on Hxk,k (Rd ) (see Exercise 2.52) but
does not preserve Cxk (Rd ) (cf. Exercise 2.35). To avoid some technicalities, let us
restrict attention for now to algebraic nonlinearities, so that p is an odd integer19.
Proposition 3.8 (Classical NLS solutions). Let p > 1 be an odd integer, let
k > d/2 be an integer, and let = 1. Then the NLS (3.1) is unconditionally
locally wellposed in Hxk,k (Rd ) in the subcritical sense. More specifically, for any R >
0 there exists a T = T (k, d, p, R) > 0 such that for all u0 in the ball BR := {u0
Hxk,k (Rd ) : ku0 kHxk,k (Rd ) < R} there exists a unique solution u Ct0 Hxk,k ([T, T ]
Rd ) to (3.1). Furthermore the map u0 7 u from BR to Ct0 Hxk,k ([T, T ] Rd ) is
Lipschitz continuous.
The same statements hold if Hxk,k is replaced by Hxs for any s > d/2 (not
necessarily an integer).

Remark 3.9. This proposition implies that for a Schwartz initial datum u0

Sx (Rd ) and an odd integer p one has a maximal Schwartz solution u Ct,loc Sx (I
d
R ) to any given algebraic NLS for some open interval I containing 0, which is
unique by Proposition 3.2. Note that one can use the equation (3.1) to trade
regularity in space for regularity in time (at a two-for-one conversion ratio), and so
solutions which are Schwartz in space will also be smooth in time.
Proof. The key observations20 are Exercise 2.52, and the fact that the space
Hxk,k (Rd )
is a Banach space algebra:
(3.24) kf gkHxk,k (Rd ) .k,d kf kHxk,k (Rd ) kgkHxk,k (Rd ) .

19When the nonlinearity is rough, it is often necessary to regularise it, for instance by re-
placing |u|p1 u by (2 + |u|2 )(p1)/2 u for some > 0 and then setting 0, in order to have
a concept of a smooth solution that one can use to approximate rough solutions to the original
equation; see for instance Exercise 3.55. In some cases one can use Schauder estimates (Lemma
A.9) as a substitute for product estimates such as (3.24). As these technical issues are rather dull,
we shall try to avoid them as much as possible.
20Indeed, this argument is quite abstract, and applies to any Banach algebra which is pre-
served by the linear flow. This is known as the semigroup method and has been extensively
developed, see for instance [Kat7].
 3.3. LOCAL EXISTENCE THEORY 131

Forcing term
k,k
F(u) in C0H
Nonlinearity F
Duhamel (loses O_R(1))
operator
(gains O(T)) [Schauder or
product estimates]
[energy estimates]

Initial datum Solution

u in Hk,k 0 k,k
u in C H
0 Linear evolution
size O(R) size O(R)

Figure 1. The classical energy method iteration scheme; the it-

eration closes if T is sufficiently small depending on R. One can
also replace Hxk,k by Hxs for s > d/2 without difficulty.

We leave this estimate (a variant of (A.18)) as an exercise.

Let us now fix R, and let 0 < T < 1 be a small time to be chosen later. We shall
use Proposition 1.38 with S = N = Ct0 Hxk,k ([T, T ] Rd ), with linear operator
D : N S set equal to the Duhamel operator
Z t
DF (t) := i ei(ts)/2 F (s) ds
0

and the nonlinear operator N : S N set equal to

N u(t) := |u(t)|p1 u(t).
(See Figure 1.) From Minkowskis inequality and Exercise (2.52) we verify the
bound (1.51) with C0 = Ok,d (T ). From the algebra property we see that
kN ukN .k,d,p,R kukS ; kN u N vkN .k,d,p,R ku vkS
whenever u, v S are such that kukS , kvkS .k,d R. If we choose T sufficiently
small depending on k, d, p, R, we can thus apply Proposition 1.38 and conclude
that for all ulin S with kulinkS . R there exists a unique solution u S to (1.50)
with kukS . R. Applying this in particular to ulin := eit/2 u0 (and using Exercise
2.52) we obtain a solution to (3.1) (in the Duhamel integral form), with the map
u0 7 u being Lipschitz continuous from the ball in Hxk,k (Rd ) of radius O(R) to
Ct0 Hxk,k ([0, T ] Rd ).
The above argument establishes uniqueness so long as we restrict the S norm
of solutions S to be O(R). But since the H k,k norm of u0 is at most R at time
zero, one can in fact obtain unconditional uniqueness in S by a standard continuity
argument. Specifically, let u S be the solution constructed by the above method,
then we have kukS C1 R for some absolute constant C1 . Let u S be another
solution. For any 0 t T let H(t) be the property that ku kC 0 Hxk,k ([t,t]Rd )
t
2C1 R, and let C(t) be the property that ku kC 0 Hxk,k ([t,t]Rd ) C1 R. Then the
t
assumptions (b),(c),(d) of Proposition 1.21 are clear, and property (a) follows from
132 3. SEMILINEAR DISPERSIVE EQUATIONS

the uniqueness theory already established (if T is suitably small). This gives the
unconditional uniqueness.
The same argument works with Hxk,k replaced by Hxs since one still has the
crucial algebra property
kf gkHxs (Rd ) .k,d kf kHxs (Rd ) kgkHxs (Rd ) ;
see Lemma A.8. 

Remark 3.10. This argument was completely insensitive to the sign of the
nonlinearity; this is a typical feature of the local existence theory. The global exis-
tence theory, however, will be much more sensitive to the sign of the nonlinearity.
The above result shows that one has unconditional local wellposedness in
Hxs (Rd ) for an algebraic NLS equation for any s > d/2. This shows (using the
argument in the proof of Theorem 1.17) that given any u0 Hxs (Rd ), there exists
a unique maximal interval of existence I and a unique solution u Ct0 Hxs (I Rd ).
The size of this interval can only shrink to zero if the Hxs (Rd ) norm of the data
goes to infinity. Hence if I has a finite endpoint T , then the Hxs (Rd ) norm of u(t)
will go to infinity as t approaches T . Thus the maximal interval is necessarily open,
as one cannot possibly continue a solution in Ct0 Hxs at a point where the Hxs norm
is going to infinity. An identical result also holds in the periodic case Tn .
To rephrase the above discussion, if a solution to NLS (with algebraic nonlin-
earity) is initially in Hxs (Rd ) with s > d/2, then it can be continued in a unique
continuous manner in Hxs (Rd ) so long as the Hxs (Rd ) stays bounded. Let us infor-
mally call a norm X a controlling norm 21 for this equation if the boundedness of
this X norm is enough to ensure continuation of smooth solutions. Thus we now
know that any sufficiently high regularity Sobolev norm is a controlling norm for
any algebraic NLS. It is of interest to obtain controlling norms which are as low
regularity as possible. As a rule of thumb, any reasonable norm which is subcritical,
or which is critical and involves some integration in time, has a chance of being a
controlling norm. For instance, we have
Proposition 3.11 (Persistence of regularity). Let I be a time interval con-
taining t0 = 0, let s 0, and let u Ct0 Hxs (I Rd ) be a strong Hxs solution to
an algebraic NLS equation. If the quantity kukLp1L (IRd ) is finite, then u(t) is
t x
uniformly bounded in Hxs , indeed we have
p1
(3.25) kukL s d ku(0)kH s exp(Cp,s,d kuk p1
t Hx (IR ) x L L (IRd )
).
t x

In particular, if I has finite length |I|, then we have

p1
kukL s d ku(0)kH s exp(Cp,s,d |I|kuk
t Hx (IR ) x L L (IRd )
).
t x

Proof. We use the energy method. By time reversal symmetry we may take
I = [0, T ] for some T > 0. From the Duhamel formula
Z t

u(t) = eit/2 u(0) i ei(tt )/2 |u(t )|p1 u(t ) dt
0

21We shall be somewhat vague as to whether X is a spatial norm or a spacetime norm. Both
types of norms are useful types of controlling norms.
 3.3. LOCAL EXISTENCE THEORY 133

and the unitary nature of eit/2 on Hxs (Rd ), we conclude from Minkowskis in-
equality that
Z t
ku(t)kHxs (Rd ) ku(0)kHxs (Rd ) + k|u(t )|p1 u(t )kHxs (Rd ) dt .
0
We expand |u(t )|p1 u(t ) as a polynomial of degree p in u(t ) and its complex
conjugate u(t ). Applying Lemma A.8 repeatedly we have
k|u(t )|p1 u(t )kHxs (Rd ) .p,s,d ku(t )kHxs ku(t )kp1
L
x
.
The claim now follows from Gronwalls inequality. 
Remark 3.12. In the converse direction, any field in Ct0 Hxs will be locally in
L
t Lx and hence in Lp1 t Lx by Sobolev embedding. Thus one can continue a
solution in Hx for s > d/2 if and only if the Lp1
s
t Lx norm remains locally finite;
in particular, if the solution blows up (fails to remain smooth) at some time T ,
then the solution must become unbounded near the blowup time T (which justifies
the terminology of blowup). Since these blowup criteria are independent of Hxs ,
we thus observe if an initial datum u0 lies both in a lower regularity Sobolev space
Hxs1 and a higher regularity Sobolev space Hxs2 , where s2 > s1 > d/2, then the
solution can be continued in one regularity for precisely the same amount of time
as it can be continued in another; thus it is not possible to develop a singularity
which causes the Hxs2 norm to blow up while the Hxs1 norm remains bounded. This
phenomenon (known as persistence of regularity - if a solution map preserves rough
regularities, then it also preserves smooth regularities) is typical of all regularities
for which one has a strong wellposedness theory, but can fail22 for regularities that
are excessively low (see Exercise 3.15). Note that from the time reversal symmetry
(and uniqueness) we also see that the regularity cannot spontaneously increase: if
a solution lies in Ct0 Hxs1 (I Rd ) and is not in Hxs2 at some initial time t0 , then it
will also not be in Hxs2 for any later (or earlier) time. Thus regularity is neither
created nor destroyed in the Sobolev scale, so long as the solution persists. This is
in contrast to dissipative equations such as the heat equation, which is smoothing
when evolved forwards in time and illposed when evolved backwards in time.
Remark 3.13. Observe that the Lp1 t L
x norm that controls the persistence
of regularity here is invariant under the scaling (3.9). This is closely related to the
fact that no factor of |I| appears in (3.25). It has the consequence that the bound
(3.25) holds even for unbounded intervals I, and thus shows that one can keep
the Hxs norm of a solution u(t) bounded even as t , provided that one can
somehow keep the global Lp1 t L
x norm bounded also. This result is an instance of
a general principle, that scale-invariant global spacetime integrability bounds imply
good asymptotic behaviour at infinity; this philosophy will be particularly apparent
in the scattering theory in Section 3.6.
Remark 3.14. The scale-invariance of the controlling norm is a general phe-
nomenon; controlling norms are either critical (invariant with respect to scaling) or
22More precisely, persistence of regularity can fail when there is no control of critical or
subcritical type on the solution. Note that in Proposition 3.11, the regularity H s that was being
controlled could be subcritical, critical, or supercritical; the important thing is that the controlling
norm Lp1
t Lx is critical. More generally, it is permissible in a scale invariant argument to use one
quantity which is subcritical or supercritical, so long as all the bounds are linear in that quantity.
134 3. SEMILINEAR DISPERSIVE EQUATIONS

subcritical (they only become scale-invariant if multiplied by some positive power

of the length |I| of the time interval, as is the case for instance with the L t Lx

norm here). All other things being equal, it is preferable to use a critical controlling
norms than a subcritical one (provided of course that a critical controlling norm
can be located in the first place) as they are generally smaller, and can yield global
control on solutions rather than just local control. Norms which are supercritical,
on the other hand, cannot possibly be controlling norms (this would lead to absurd
results, such as the spacetime bounds for large time intervals being smaller than
the bounds for small time intervals). The most famous example of this is the three-
dimensional Navier-Stokes equations, which enjoy boundedness of kinetic energy
but for which global existence of smooth solutions is a major unsolved problem,
in large part because the kinetic energy turns out to be a supercritical quantity
in three spatial dimensions and thus cannot be a controlling norm. In practice,
possession of a bound on a supercritical quantity has proven to be of little use in
the global regularity theory, unless combined with additional information such as
a bound on a subcritical quantity (so that one can interpolate between the two to
obtain critical controlling quantities). More recently, techniques have been devel-
oped to combine supercritical control with existing critical control, to obtain even
better critical control; in particular, in the energy-critical defocusing NLS, the mass
and momentum conservation laws (which are supercritical in this case) can be used
to limit the concentration behaviour of energy towards higher frequencies and thus
yield control of other critical quantities such as certain spacetime Lebesgue norms.
See Chapter 5.
We now turn from classical solutions to less regular solutions, in particular
considering solutions in Hxs for s d/2. In this case, we no longer expect the
solution to lie in L d s
x (R ) for all time, since Hx no longer embeds into Lx . However,
the Strichartz estimates in Theorem 2.3 suggest that one can still lie in time-
p1
averaged L x spaces such as Lt Lx (Rd ) for regularities lower than d/2; intuitively,
this reflects the fact that while an Hxs function can focus much of its energy at
one spatial point to create a large L x norm (cf. Proposition A.4), the dispersive
effects of the Schrodinger evolution imply that this focus cannot be maintained for
more than a short period of time. Of course, this is only a heuristic, because the
Strichartz estimates only apply directly to the linear Schrodinger evolution rather
than the nonlinear one, however it does suggest that some sort of iterative argument,
using the Strichartz estimates to treat the nonlinear equation as a perturbation of
the linear one, can work.
To do this, it is convenient to create a single space S s which captures all the
Strichartz norms at a certain regularity Hxs simultaneously. We illustrate this first
with the L2x theory. We introduce the Strichartz space S 0 (I Rd ) for any time
interval I, defined as the closure of the Schwartz functions under the norm23
kukS 0 (IRd ) := sup kukLqt Lrx (IRd ) ,
(q,r) admissible

23In the case d = 2 case, the set of admissible exponents is not compact, and so one has
to truncate the supremum, for instance restricting q 2 + for some > 0, in order for the
Strichartz constants to be uniform in the exponent. Also, in some endpoint applications it is more
convenient to strengthen the norms S 0 , N 0 to a certain Besov-space version of themselves. We
ignore these technicalities to simplify the exposition.
 3.3. LOCAL EXISTENCE THEORY 135

where admissibility was defined in Theorem 2.3. In particular the S 0 norm controls
the Ct0 L2x norm. This norm is a Banach space and has a dual N 0 (I Rd ) :=
S 0 (I Rd ) ; by construction we see that
kF kN 0 (IRd ) kF kLq Lr (IRd )
t x

whenever the right-hand side is finite. The Strichartz estimates in Proposition 2.3
can then be combined into a unified estimate
(3.26) kukS 0 (IRd ) .d ku(t0 )kL2x (Rd ) + kF kN 0 (IRd )
whenever t0 I and iut + 12 u = F . Because of this estimate, one often expects to
place L2x solutions of NLS in the space S 0 , at least provided that one has some hope
of placing the nonlinearity F = |u|p1 u in the companion space N 0 . A typical
application of this estimate is
Proposition 3.15 (Subcritical L2x NLS solutions). [Tsu] Let p be an L2x -
subcritical exponent (so 1 < p < 1 + d4 ) and let = 1. Then the NLS (3.1)
is locally wellposed in L2x (Rd ) in the subcritical sense. More specifically, for any
R > 0 there exists a T = T (k, d, p, R) > 0 such that for all u0 in the ball
BR := {u0 L2x (Rd ) : ku0 kL2x (Rd ) < R} there exists a unique strong L2x solu-
tion u to (3.1) in the space S 0 ([T, T ] Rd ) Ct0 L2x ([T, T ] Rd ). Furthermore
the map u0 7 u from BR to S 0 ([T, T ] Rd ) is Lipschitz continuous.
Remark 3.16. One can weaken the space S 0 ([T, T ] Rd) somewhat and still
obtain uniqueness (see [CWeis], [CWeis2]). However, it is not known if one can
replace S 0 by Ct0 L2x and thus obtain an unconditional wellposedness result. In the
next section we shall extend this local existence result to a global existence result.

Proof. We modify the proof of Proposition 3.8. Again we fix R and choose
T > 0 later. We will apply Proposition 1.38 for a suitable choice of norms S, N and
some > 0; a specific instance of our scheme in the case d = 1, p = 3 is described in
Figure 2. One such choice is to set S = S 0 ([T, T ]Rd) and N = N 0 ([T, T ]Rd).
In order to place the ulin in B/2 , we see from the Strichartz estimate (3.26) that
we need to take = C1 R for some large constant C1 > 0 (depending only on d).
The estimate (1.51) also follows from (3.26) (for some large C0 > 0 depending on
d), so it remains to verify (1.52). In other words, we need to show that
1
k|u|p1 u |v|p1 vkN 0 ([T,T ]Rd ) ku vkS 0 ([T,T ]Rd )
2C0
whenever kukS 0 ([T,T ]Rd ) , kvkS 0 ([T,T ]Rd ) C1 R. It is convenient to introduce
the exponent pair (q, r) by solving the equations
2 d d p 1
+ = ; = .
q r 2 r r
One can easily check using the hypothesis 1 < p < 1 + d4 that we have 2 < r < q <
, so in particular (q, r) is admissible. In particular, we can estimate the N 0 norm

by the Lqt Lrx norm. Since q > r, we see that pq < q1 , so we may replace the Lqt
q/p
norm by the Lt norm by paying a factor of T for some > 0. If we then use
the elementary estimate
(3.27) ||u|p1 u |v|p1 v| .p |u v|(|u|p1 + |v|p1 )
136 3. SEMILINEAR DISPERSIVE EQUATIONS

Forcing term Forcing term Forcing term

0 8/7 4/3 8/3 4/3
F(u) in N
F(u) in L L [Holder in F(u) in L L
time]
(gain power of T)
F
[Holder]
[Strichartz estimates]
Initial datum Solution Solution
u in L2 u in S0
8 4
0 u in L L

Solution
0 2
u in C L

Figure 2. An iteration scheme in L2x (R) for the the one-

dimensional cubic NLS d = 1, p = 3 (which is L2x -subcritical),
so that q = 8 and r = 4. In all the iteration schemes presented
here, the sign of the nonlinearity is irrelevant. The subcritical
nature of the equation allows for a gain of a power of T at some
stage. This is not the simplest iteration scheme available for this
equation, but it is rather representative. Note that the fact that
u was a strong solution (i.e. u Ct0 L2x ) is a byproduct of the
argument but is not otherwise used in an essential way in the proof.

and Holders inequality, we conclude

k|u|p1 u |v|p1 vkN 0 ([T,T ]Rd )
.p T ku vkLqt Lrx ([T,T ]Rd ) (kukLqt Lrx ([T,T ]Rd ) + kvkLqt Lrx ([T,T ]Rd ) )p1
.p,C1 ,R T ku vkLqt Lrx ([T,T ]Rd )
T ku vkS 0 ([T,T ]Rd )

Thus we obtain (1.52) if T is chosen sufficiently small depending on p, C1 , R. We

can then apply Proposition 1.38 to construct a solution in S to (3.1) with norm
at most C1 R/2, which is unique among all solutions with norm at most C1 R, and
the map u0 7 u will be Lipschitz continuous. The requirement that the norm be
at most C1 R can be dropped from the uniqueness conclusion by using a continuity
argument as in Proposition 3.8. 
In the critical case p = 1 + d4 one still has wellposedness, but in the critical
sense (so that the time of existence T depends on the profile of the datum and not
just on the norm). More precisely, we have
 3.3. LOCAL EXISTENCE THEORY 137

(small norm) (small norm)

Forcing term Forcing term
0 4/3 4/3
F(u) in N
F(u) in L L

F
[Holder]
[Strichartz estimates]
(large norm)
Initial datum Solution Solution
u in L2 u in S0
4 4
0 u in L L
(large norm) (small norm)

Solution
0 2
u in C L

(large norm)

Figure 3. An iteration scheme in L2x (R2 ) for the two-dimensional

cubic NLS d = 2, p = 3 (which is L2x -critical). For small data one
can simplify the scheme somewhat, but for large data it is impor-
tant that the S 0 is allowed to be large, while the L4t,x component
of the S 0 kept small; thus the main loop of the iteration should
involve the L4t,x norm more than once in order to close the argu-
ment, because no gain of a power of T is available in the critical
setting. This also makes the L4t,x norm a controlling norm for the
evolution.

Proposition 3.17 (Critical L2x NLS solutions). [Tsu] Let p be the L2x -critical
exponent p = 1 + d4 and let = 1. Then the NLS (3.1) is locally wellposed
in L2x (Rd ) in the critical sense. More specifically, given any R > 0 there exists
0 = 0 (R, d) > 0, such that whenever u L2x (Rd ) has norm at most R, and I is
a time interval containing 0 such that
keit/2 u kL2(d+2)/d (IRd ) 0
t,x

then for any u0 in the ball B := {u0 L2x (Rd ) : ku0 u kL2x (Rd ) 0 } there exists
a unique strong L2x solution u S 0 (I Rd ) to (3.1), and furthermore the map
u0 7 u is Lipschitz from B to S 0 (I Rd ) (of course, the Lipschitz constant will
depend on u ).
This proposition is proven similarly to Proposition 3.15 and is left to Exercise
3.18. Note that if the initial datum is sufficiently small in L2x norm, then this Propo-
sition, combined with Strichartz estimates, yields global existence in time. If the
initial datum is instead large, the Proposition combined with Strichartz estimates
138 3. SEMILINEAR DISPERSIVE EQUATIONS

2(n+2)/n
will still give local existence, because the global Lt,x norm of eit/2 u will be
finite, and hence can be localised to be small by choosing a sufficiently small time
interval.
Remark 3.18. This proposition and the preceding one should be compared
against Principle 3.1. It turns out that the L2x theory becomes bad for supercritical
powers p > 1 + d4 ; see Section 3.8 for further discussion and results.
Similar results hold for other regularities, such as Hx1 . Here it is convenient to
use the norms
kukS 1 (IRd ) := kukS 0 (IRd ) + kukS 0 (IRd )
and
kukN 1 (IRd ) := kukN 0 (IRd ) + kukN 0 (IRd )
Note that as the Schrodinger equation commutes with derivatives, we see from
(3.26) that
(3.28) kukS 1 (IRd ) .d ku(t0 )kHx1 (IRd ) + kF kN 1 (IRd ) .
Let us give two sample results in dimension d = 3, in which the Hx1 -critical exponent
is the quintic one p = 5:
Proposition 3.19 (Hx1 (R3 ) subcritical NLS solutions). Let = 1. If 2
p < 5, then the NLS (3.1) is locally wellposed in Hx1 (R3 ) in the subcritical sense.
Remark 3.20. For this Proposition and the next, the reader may wish to
refer back to the Strichartz game board for Schrodinger equations on Hx1 (R3 )
from Figure 1 of Chapter 2, and see how the various moves of Leibnitz, Holder,
Sobolev, and Strichartz affect the game pieces u, u, F (u), etc. on this board.
(The objective of the iteration game is to construct a set of assumptions (thus
placing game pieces in various spaces with various norm bounds) on the solution,
such that it is possible to apply a legal sequence of moves and end up with all the
game pieces returning to the same spaces but with better estimates.)

Proof. (Sketch) We apply Proposition 1.38 with S = S 1 ([T, T ] R3 ), N =

N ([T, T ] R3 ). By arguing as in Proposition 3.15, we will be done as soon as
1

we show that
k|u|p1 u |v|p1 vkN 1 ([T,T ]R3 ) .p,R T ku vkSx1 ([T,T ]R3 )
for some = (p) > 0, whenever the Sx1 norms of u, v are O(R) for some R > 0.
Let us omit the domain [T, T ] R3 from the notation for brevity. Choosing the
admissible exponents (10, 30/13) for the Sx1 norm and (2, 6) for the N 1 norm, it
suffices to show that
kk (|u|p1 u |v|p1 v)kL2 L6/5 .p,R T ku vkL10 W 1,30/13
t x t x

for k = 0, 1. Let us just deal with the higher order case k = 1, which is the
more difficult of the two24. Observe that the gradient of |u|p1 u can be written
as F1 (u)u + F2 (u)u, where F1 , F2 : C C are functions which grow like
Fj (z) = Op (|z|p1 ), and which have the Lipschitz-type bound Fj (z) = Op (|z|p2 ).
24A general principle in the local-in-time theory is that the highest order terms are always
the most difficult to estimate, so that once those are dealt with the lower order terms should be
treatable by a modification of the argument.
 3.3. LOCAL EXISTENCE THEORY 139

Forcing term Forcing term Forcing term

1 2 1,6/5 10/3 1,6/5
F(u) in N F(u) in L W F(u) in L W
[Holder in time]
(gain power of T)
F
[Leibnitz &
[Strichartz estimates] Holder]

Initial datum Solution Solution Solution

u in H1 u in S
1 10 1,30/13 10 5
0 u in L W
[Sobolev] u in L L

Solution
0 1
u in C H

Figure 4. An iteration scheme in Hx1 (R3 ) for the three-

dimensional cubic NLS d = 3, p = 3 (which is Hx1 -subcritical).
This is similar to Figure 2; besides the changes in numerology,
the main new feature is the appearance of the Leibnitz rule and
(non-endpoint) Sobolev embedding to handle the additional deriva-
tives.

Let us just consider the contribution of the F1 terms to the above expression, so
we need to show
kF1 (u)u F1 (v)vkL2 L6/5 .p,R T ku vkL10 W 1,30/13 .
t x t x

From the Lipschitz bound (and the hypothesis p 2) we have

F1 (u)uF1 (v)v = Op ((|u|p1 +|v|p1 )(uv))+Op ((|u|p2 +|v|p2 )(uv)u).
6/5 10/p 6/5
We estimate the L2t Lx norm by the Lt Lx norm, gaining a power of T (here
we use the fact that p < 5), and use Holder to estimate
kF1 (u)u F1 (v)vkL2 L6/5 . T (kukp1
10 5(p1)/2 k(u v)kL10 L30/13
t x Lt Lx t x

+ kukp2 5(p1)/2 ku vkL10 L5/2p kukL10L30/13 ).

L10
t Lx t x t x

From Sobolev embedding and the hypothesis 2 p < 5 we have

kukL10L5(p1)/2 . kukL10 W 1,30/13
t x t x

and the claim then follows from the hypothesised bounds on u, v. (See Figure 4 for
an illustration of the scheme in the case p = 3.) 

There is a version of this argument available in the limit p = 5:

Proposition 3.21 (Hx1 (R3 ) critical NLS solutions). Let = 1 and p = 5.
Then the NLS (3.1) is locally wellposed in Hx1 (R3 ) in the critical sense. More
140 3. SEMILINEAR DISPERSIVE EQUATIONS

(small norm) (small norm)

Forcing term Forcing term

.1 2 . 1,6/5
F(u) in N F(u) in L W

F
[Leibnitz &
Holder]
[Strichartz estimates] (large norm)
(large norm)
Initial datum Solution Solution Solution
. .1
u in H1 u in S
10 . 1,30/13 10 10
0 u in L W
[Sobolev] u in L L
(large norm) (small norm)

Solution
0 .1
u in C H
(large norm)

Figure 5. An iteration scheme in Hx1 for the three-dimensional

quintic NLS d = 3, p = 5 (which is Hx1 -critical). This rather tricky
scheme is similar to Figure 4 but with homogeneous norms, and
with certain norms identified as being small to compensate for the
lack of a Holder in time (cf. Figure 3). It will be important that
the smallness in L10
t,x is exploited more than once in order to close
the iteration properly.

precisely, given any R > 0 there exists 0 = 0 (R) > 0, such that whenever u
Hx1 (R3 ) has norm at most R, and I is a time interval containing 0 such that
keit/2 u kL10 3 0
t,x (IR )

then for any u0 in the ball B := {u0 Hx1 (R3 ) : ku0 u kH 1 (R3 ) 0 } there
x

exists a unique strong Hx1 solution u S 1 (I R3 ) to (3.1), and furthermore the

map u0 7 u is Lipschitz from B to S 1 (I R3 ). Here kukS 1 := kukS 0 is the
homogeneous counterpart to the S 1 norm.
We leave the proof of this to the exercises. Note that the L10
t,x norm is controlled
1,30/13
(via Sobolev embedding) by the L10
t Wx norm, which in turn is controlled by
1
the S norm. From Strichartz estimates we thus conclude that
keit/2 u kL10 3 . ku k 1
t,x (RR ) H (R3 ) ,
x

and thus Proposition 3.21 implies global wellposedness for quintic NLS on R3 with
small energy, and local wellposedness (in the critical sense) for large energy. Again,
this proposition and the preceding one should be compared against Principle 3.1.
In the supercritical case p > 5, the Hx1 perturbation theory breaks down com-
pletely; again, see Section 3.8. However in the defocusing case with 5 < p < 6 one
 3.3. LOCAL EXISTENCE THEORY 141

can still construct global weak Hx1 solutions by a weak compactness method; see
Exercise 3.56.
Similar wellposedness results exist for the NLW equation, and for the periodic
NLS equation; we leave this to the exercises. One can briefly summarise (and
oversimplify) the known results for local wellposedness as follows. For NLS, one
has local wellposedness in Hxs (Rd ) as long as s 0 and the nonlinearity25 is Hxs -
subcritical or Hxs -critical, though in the latter case the wellposedness is in the
critical sense (the time of existence depends on the profile of the datum rather than
the norm, but for small norm one has global existence). See [CWeis2], [Caz2]. For
NLW, one has a similar result but with an additional constraint26 s sl , where sl is
1/2
the Lorentz regularity sl := d+1 1
4 p1 ; this constraint is only relevant in the Hx -
4
subcritical cases p < 1 + d1 ; see [LSog], [Sog]. For periodic NLS, the situation
is much less well understood, due to the lack of sharp Strichartz estimates in this
setting; see [Bou]. (The local theory for periodic NLW is essentially identical to
non-periodic NLW; see Exercise 3.24.)
Exercise 3.14 (H k,k is a Banach algebra). Prove (3.24). (Hint: use the Leib-
nitz rule, Holder, Sobolev, and Gagliardo-Nirenberg, controlling the lower order
terms before moving on to the higher ones. A Littlewood-Paley approach is possible
but somewhat lengthy because of the need to continually commute the Littlewood-
Paley operators with weights such as hxik .)
Exercise 3.15. Using the Fourier transform, show that the solution (3.15) to
the pseudoconformal focusing NLS blows up in Hxs for any s > 0 as t 0, but
stays bounded in L2x , and even goes to zero in Hxs for s < 0. (This reflects the fact
that this equation is locally wellposed in the subcritical sense in Hxs for s > 0, is
locally wellposed in the critical sense in L2x , and is illposed in Hxs for s < 0.) Using
this, show that when s < 0, one no longer has uniqueness for weak Hxs solutions,
and that Exercise 3.10 also breaks down for weak Hxs solutions.
Exercise 3.16. Let u S 1 (I R3 ) be an Hx1 -wellposed solution to quintic
NLS (so p = 5 and d = 3), and suppose that u(t0 ) H k (R3 ) for some t0 I and
some integer k 0. Show that u(t) H k (R3 ) for all t I, and in fact
ku(t0 )kH k (R3 ) .kukS1 (IR3 ) ku(t0 )kH k (R3 ) .
(Hint: Let M := kukS 1 (IR3 ) . Subdivide I into time intervals Ij where the L10
t,x
norm on Ij is small compared with M . Then use Strichartz estimates and continuity
arguments to establish S k control on each Ij .)
Exercise 3.17 (Unconditional uniqueness). Let u, v Ct0 Hx1 (I R3 ) be strong
Hx1 solutions to quintic NLS (so p = 5 and d = 3) with u(t0 ) = v(t0 ) for some
t0 I. Show that u = v. (Hint: Let J be a small time interval containing t0 ,
and use Strichartz estimates to control ku vkL2t L6x (JR3 ) in terms of itself and
1 6
ku vkL 6 3 . Then use the continuity of u v in H
t Lx (JR ) x and hence in Lx to
close the argument. To extend from J back to I, use the continuity method.) In
particular, this shows that the Hx1 -wellposed solution given by Proposition 3.21 is
25There is also a technical smoothness condition required on the nonlinearity in the non-
algebraic case; see [CWeis2], [Caz2].
26Again, we need a smoothness condition on the nonlinearity. Also this result is not fully
established in high dimension n 4 when s is very close to zero, for technical reasons; see [Tao].
142 3. SEMILINEAR DISPERSIVE EQUATIONS

the only strong Hx1 solution. One in fact has unconditional uniqueness for strong
Hx1 solutions for all Hx1 -critical and Hx1 -subcritical equations; see [Kat9], [Caz2],
[TV].
Exercise 3.18 (L2x -critical wellposedness). Prove Theorem 3.17. (Hint: there
are now two norms one wishes to place the solution u in: the S 0 norm, and the
2(d+2)/d
Lt,x norm. The solution will be small in the latter norm but can be large in
the former norm. To account for this, one either has to apply Proposition 1.38 with
an artificial norm such as
kukS := kukS 0 + kukL2(d+2)/d
t,x

for some small , or else use work with the iteration scheme directly and establish
bounds on all the various norms at each stage of the iteration. See also Figure 3.)
Exercise 3.19 (Hx1 -critical wellposedness). Prove Proposition 3.21. (You may
find Figure 5 to be helpful.)
Exercise 3.20. Show that the cubic NLS on the circle T is locally wellposed
in L2x (T) in the subcritical sense. (Hint: use Exercise 2.74.) Also, show persistence
of regularity, or more precisely if the initial datum lies in Hxk (T) for some positive
integer k, then the local L2x (T) solution constructed by the iteration method is in
fact a strong Hxk (T) solution.
Exercise 3.21 (Classical local wellposedness of NLW). Show that an algebraic
NLW is unconditionally locally wellposed in Hxs Hxs1 for s > d/2 in the subcritical
sense, thus for each R > 0 there exists a T = T (k, d, p, R) > 0 such that for all initial
data (u0 , u1 ) in the ball BR := {(u0 , u1 ) Hxs (Rd ) Hxs1 (Rd ) : ku0 kHxs (Rd ) +
ku1 kHxs (Rd ) < R} there exists a unique classical solution u Ct0 Hxs ([T, T ] Rd)
Ct1 H s1 ([T, T ] Rd ) to (3.1). Furthermore the map (u0 , u1 ) 7 u is Lipschitz
continuous. (Hint: adapt the proof of Proposition 3.8, and use (2.29).) Show also
that the solution can be continued in time as long as u stays bounded in spacetime.

Exercise 3.22 (Hx1 (R3 ) subcritical NLW solutions). Let = 1 and 2

p < 5. Show that the NLW (3.2) is locally wellposed in Hx1 (R3 ) L2x (R3 ) in the
subcritical sense. (Hint: there are many schemes available. The simple scheme
Figure 6, that does not use Strichartz estimates and which dates back to [Jor],
only works up to p 3; for higher p one needs to use spaces that involve some
integration in time. You may also wish to review Figure 2 from Chapter 2, and
peek at Figure 4 from Chapter 5.) For the critical case p = 5, see Exercise 5.1.
4
Exercise 3.23. Let d 3, = 1, and let p := 1 + d2 be the Hx1 -critical
power. Show that for any u0 Hx1 (Rd ) with sufficiently small norm, there exists a
unique global solution u S 1 (R Rd ) to the NLS (3.1) with the specified initial
datum. (Hint: for 3 d 6, one can modify the proof of Proposition 3.21 to
accomodate the inhomogeneous Sobolev and Strichartz norms. For d > 6, we have
p < 2 and there is now a difficulty obtaining a contraction mapping. However, one
can still construct iterates, show that they are bounded in S 1 (RRd), and converge
locally in time in S 0 (I Rd ). A variant of this argument then gives uniqueness, at
least locally in time, which can then be extended to be global in time by continuity
arguments. See [CWeis2], [TV] for a more thorough treatment of this equation.)
 3.4. CONSERVATION LAWS AND GLOBAL EXISTENCE 143

Forcing term Forcing term

1 2 0 2
F(u) in L L F(u) in C L
[Holder in time]
(gain power of T)
[energy estimates] F
[Holder]

Initial data Solution Solution

1 1 0 6
u0 in H u in C0 H u in C L
[Sobolev]
u1in L2 1 2
u in C L

Figure 6. A simple iteration scheme in Hx1 L2x , based on the

energy estimate, for the three-dimensional cubic NLW d = 3, p = 3.
For higher powers p, Strichartz estimates and spaces are needed.

Exercise 3.24. Suppose that an NLW on Rd is known to be locally wellposed

in Hxs Hxs1 in the subcritical sense for some s 0. Assume also that one has the
finite speed of propagation result, Proposition 3.3, for Hxs Hxs1 -wellposed solu-
tions. (In practice, this hypothesis can either be deduced from the corresponding
result for classical solutions by a limiting argument, or else by direct inspection of
the iterates used to construct the wellposed solution.) Show that the corresponding
periodic NLW on Td is also locally wellposed in Hxs Hxs1 . (You may find Exer-
cise A.18 to be useful.) This type of descent argument does not always work in the
nonperiodic setting, especially for large times; indeed, it is quite possible for the
behaviour of an equation for large, localized data to be better in higher dimensions
than in lower ones, due to the increased dispersion available.
Exercise 3.25 (Analytic wellposedness). Consider an algebraic NLS, and let
s > d/2 and R > 0. By the Hxs version of Proposition 3.8, we know that there
exists T = Td,p,s > 0 such that every u0 Hxs (Rd ) with norm at most R extends
to a strong Hxs solution u Ct0 Hxs ([0, T ] Rd). Show that if T is small enough, the
map P u0 7 u is in fact an real analytic map, thus there is a power series expansion

u = k=0 Nk (u0 , . . . , u0 ) which converges absolutely in Ct0 Hxs ([0, T ] Rd ), where
for each k, Nk is a k-multilinear operator from Hxs (Rd ) to Ct0 Hxs ([0, T ]Rd). (Hint:
you can adapt the proof of the Cauchy-Kowalevski theorem, see Exercise 1.1.) One
consequence of analyticity is that the solution map is also infinitely differentiable
from Hxs to Ct0 Hxs .

3.4. Conservation laws and global existence

The journey of a thousand miles begins with one step. (Lao Tzu)
The wellposedness theory of the previous section allowed us to use iterative
methods to construct local-in-time solutions to NLS and NLW given suitable reg-
ularity assumptions on the initial data; if the datum was also small in a critical
144 3. SEMILINEAR DISPERSIVE EQUATIONS

norm (e.g. small in L2x norm for the L2x -critical NLS, or small in Hx1 norm for the
Hx1 -critical NLS) then one obtained a global solution also. These methods were
perturbative in nature (using the Duhamel formula to approximate the nonlinear
evolution by the linear evolution) and thus do not work directly for large data and
long times (since one expects the evolution to be genuinely nonlinear in this case).
However, in these cases one can use non-perturbative tools to gain enough control
on the equation to prevent the solution from blowing up. In this section we describe
the most important tool for doing this, namely the conservation laws.
As before, we begin by discussing the nonlinear Schrodinger equation (3.1). In
this case we have two independent conservation laws, namely energy conservation
and mass/momentum conservation27. The latter can be most effectively described
by pseudo-stress-energy tensor T , defined for , = 0, 1, . . . , d by by
T00 = |u|2
T0j = Tj0 = Im(uxj u)
1 (p 1)
Tjk = Re(xj uxk u) jk (|u|2 ) + jk |u|p+1 .
4 p+1
If the solution is sufficiently smooth, one easily verifies the local conservation laws
(2.35); see Exercise 3.26. In particular, for smooth decaying solutions, this leads to
conservation of total mass M [u(t)], defined by
Z Z
(3.29) M [u(t)] := T00 (t, x) dx = |u(t, x)|2 dx
Rd Rd
and total momentum p~[u(t)] = (p1 [u(t)], . . . , pd [u(t)]), defined by
Z Z
pj [u(t)] := T0j (t, x) dx = Im(uxj u) dx;
Rd Rd
again, see Exercise 3.26. As in Section 2.4, the pseudo-stress-energy tensor also
yields many other important conservation laws and montonicity formulae; we will
develop some of these later in this chapter.
The above conservation laws are initially only justified for smooth decaying
solutions. However, if the conservation law is controlled by an Hxs norm, and one
has a satisfactory wellposedness theory at Hxs , then there is a standard limiting
argument that allows one to extend these conservation laws to these Hxs -well posed
solutions. We illustrate this principle with the example of the one-dimensional
cubic NLS:
Proposition 3.22 (Conservation law for a wellposed solution). Let d = 1,
p = 3, = 1. Let u S 0 (I Rd ) Ct0 L2x (I Rd ) be a strong solution to (3.1)
defined on some time interval I. Let the total mass M [u(t)] be defined by (3.29).
Then M [u(t)] = M [u(t0 )] for all t0 , t I.
Proof. Since I is the increasing union of open intervals we may assume with-
out loss of generality that I is open. Fix t0 I. Since u Ct0 L2x (I R), the set
of times {t I : M [u(t)] = M [u(t0 )]} is closed in I, and it clearly contains t0 , so it
suffices by the continuity method to show that this set is open. In other words, it

27In the special case of the one-dimensional cubic NLS (d = 1, p = 3), it turns out that
the equation is completely integrable and there are in fact infinitely many conservation laws, see
Exercise 3.36.
 3.4. CONSERVATION LAWS AND GLOBAL EXISTENCE 145

suffices to show that for any t1 I, that we have M [u(t)] = M [u(t1 )] for all times
t in a neighbourhood of t1 .
By time translation invariance we can take t1 = 0. Set u0 := u(0). We
(n)
can approximate u0 as the limit in L2x (Rd ) of a sequence u0 Hx100 (Rd ) (say)
of smooth initial data, which will be uniformly bounded in L2x (Rd ). Applying
Proposition 3.15, we can obtain a time interval [T, T ] I independent of n, and
(n)
strong solutions u(n) S 0 ([T, T ] Rd ) to (3.1) with initial data u(n) (0) = u0 .
(n)
Since u0 converges to u0 in L2x , the uniqueness and continuity conclusions of
Proposition 3.15 guarantee that u(n) will converge to u in S 0 ([T, T ] Rd ). Next,
since the u(n) are uniformly bounded in S 0 ([T, T ] Rd ), they are also uniformly
bounded in L4t L d 2 d
x ([T, T ] R ) and hence in Lt Lx ([T, T ] R ). We may thus
apply Proposition 3.11 (and the remarks following that proposition) and conclude
that u(n) Ct0 Hxs ([T, T ] Rd ) for any s > d/2. From this, Sobolev embedding,
and the equation (3.1) it is easy to see that u(n) is smooth on [T, T ] Rd . This
is enough regularity for us to apply the classical mass conservation law in Exercise
3.26 and conclude that
Z Z
|u(n) (t, x)|2 dx = |u(n) (0, x)|2 dx
Rd Rd

for all t [T, T ]. Since u(n) converges to u in S 0 ([T, T ] Rd ), it also converges

in Ct0 L2x ([T, T ] Rd ), and hence on taking limits as n we have
Z Z
2
|u(t, x)| dx = |u(0, x)|2 dx
Rd Rd

and hence M [u(t)] = M [u(0)] for all t in a neighbourhood of t, as desired. 

Note that the full power of the wellposedness theory was used here; not only the
existence aspect of the theory, but also the uniqueness, persistence of regularity, and
continuous dependence on the data. This basic argument - obtaining conservation
laws for rough solutions by approximating them by smooth solutions - is quite
general and extends to many other equations. There is an additional twist however
in the case when the nonlinearity is not algebraic (i.e. p is not an odd integer),
because it is often not possible in such cases to obtain an approximating solution
that is sufficiently smooth that one can justify the conservation law classically. In
such cases one must not only regularise the initial data, but also regularise the
equation as well; this requires a further aspect of the wellposedness theory, namely
the stability theory of the equation, which we will address in Section 3.7.
A conservation law can often, but not always, be combined with a local well-
posedness result to obtain a global wellposedness result. Let us illustrate this with
a simple example:
Proposition 3.23. [Tsu] Let d = 1, p = 3, = 1, and t0 R. Let
u0 L2x (Rd ), and let I be any bounded time interval containing t0 . Then there is
a unique strong solution u S 0 (I Rd ) Ct0 L2x (I Rd ) to (3.1) defined on some
time interval I. Furthermore, the map u0 7 u is a continuous map from L2x to
S 0 (I Rd ).
In particular, this proposition gives a global strong solution u Ct0 L2x (R Rd )
to (3.1). However this solution is only in the S 0 space locally in time. (Indeed, in
146 3. SEMILINEAR DISPERSIVE EQUATIONS

the focusing case = 1, soliton solutions to this equation exist which do not lie
globally in S 0 .)

Proof. we first use time translation invariance and time reversal symmetry
to reduce to the case when t0 = 0 and I = [0, T ] for some T > 0. (Note that by
Exercise 3.11 one can easily glue a strong solution on an interval such as [0, T ] to
a strong solution to an interval such as [T, 0], and stay in S 0 .)
We give two proofs of this result; they are equivalent, but offer slightly different
perspectives as to how a conservation law extends a local existence result to a global
existence result. For the first proof, we divide the long time interval [0, T ] into
shorter time steps, where on each smaller interval the perturbative theory gives a
local solution. More precisely, let M [u(0)] := ku0 k2L2 (Rd ) denote the initial mass,
x
and observe from Proposition 3.15 that there exists a time = (M [u(0)]) > 0
such that the equation (3.1) will have a local strong solution in S 0 ([t0 , t0 + ] Rd )
whenever the initial datum u(t0 ) has mass less than or equal to M [u(0)]. We now
split the time interval [0, T ] as a finite union of intervals [tn , tn+1 ], where each
interval has length less than or equal to , and t0 = 0. By applying Proposition
3.15 followed by Proposition 3.23 (and Exercise 3.11), an easy induction shows
that for every n we can construct a strong solution u to (3.1) in S 0 ([0, tn ] Rd ),
and thus we eventually obtain a strong solution to S 0 ([0, T ] Rd ); see Figure 7.
Uniqueness follows from Proposition 3.15 and a continuity argument similar to that
used to prove Exercise 3.13. The continuous dependence follows by concatenating
the continuous dependence results on each of the intervals [tn , tn+1 ], using the fact
that the S 0 ([tn , tn+1 ] Rd ) norm of u controls the L2x norm of u(tn+1 ), and using
the fact that the composition of continuous maps is continuous.
The second proof proceeds by contradiction; it is quicker but is more indirect
(and does not give the continuous dependence as easily). We sketch it as follows.
Just as the Picard existence and uniqueness theorems imply a blowup criterion for
ODE (Proposition 1.17), the existence and uniqueness theory in Proposition 3.15
gives a blowup criterion for L2x strong solutions to (3.1), namely that these solutions
will exist globally unless the L2x norm goes to infinity in finite time. However, Propo-
sition 3.22 clearly implies that the L2x norm of a strong solution stays bounded.
Thus blowup cannot occur, and one must instead have global existence. 

One can combine this global existence result with persistence of regularity the-
ory (e.g. Proposition 3.11) to show that the global solution constructed in Propo-
sition 3.23 preserves regularity; see Exercise 3.28. In particular, with a smooth
decaying initial datum we have a global smooth solution to the one-dimensional
cubic NLS.
Similar arguments give global L2x -wellposedness for any L2x -subcritical equation.
The situation is remarkably different when one considers the two-dimensional cubic
NLS (d = 2, p = 3, = 1). The key difference is that whereas the one-dimensional
cubic NLS was L2x -subcritical, the two-dimensional cubic NLS is L2x -critical. This
is reflected in the local wellposedness theory for this equation, given by Proposition
3.17. If the initial datum has a sufficiently small L2x norm, then this proposition
already gives a global existence result without any need for a conseration law. How-
ever, when the L2x norm is large, one cannot simply combine the conservation law
with the local existence theory to obtain a global existence theory; the problem is
that the time of existence given by the local wellposedness theory does not depend
 3.4. CONSERVATION LAWS AND GLOBAL EXISTENCE 147

2 local theory 2 local theory 2

u(t0) in L u(t1) in L u(t2) in L
on [t 0, t 1] on [t 1, t 2]

M(t0) bounded M(t1 ) bounded M(t2 ) bounded

[conservation law] [conservation law]

Figure 7. The first proof of Proposition 3.23. The nonlinear iter-

ation in the local theory could potentially increase the L2x norm as
one advances from one time step tn to the next tn+1 , thus leading
to a collapse of the lifespan tn+1 tn of that theory; however if one
uses instead the conservation law to control the L2x norm then no
collapse occurs (cf. the global wellposedness of the ODE in Figure
9 from Chapter 1). Indeed the local theory plays a mostly quali-
tative role in the global argument, justifying the local existence of
the solution as well as the conservation law, but not providing the
key quantitative bounds.

only on the L2x norm on the datum, but also on the profile of the datum itself (and
more specifically on the spacetime behaviour of the free evolution of the datum).
Because of this, the conservation law is insufficient by itself to make either of the
arguments used in the proof of Proposition 3.23 extend to this case; iterating the
local wellposedness theorem can lead to a shrinking interval of existence, which can
lead to blowup in finite time. Indeed, in the focusing case = 1, the explicit
blowup solution given in (3.15) shows that even for smooth L2x (Rd ) initial data one
can have finite time blowup for this equation. (Note that the classical uniqueness
theory shows that this failure of global existence cannot be avoided by strengthening
the notion of solution.) In the defocusing case = 1, global existence (and well-
posedness) for L2x (Rd ) initial data is suspected to be true for the two-dimensional
cubic NLS, but this is not known, even for radially symmetric initial data, and is
considered a major open problem in the field; a similar open question exists for any
other L2x -critical defocusing NLS. However, the situation improves when the initial
data is assumed to be in the energy class Hx1 (Rd ), rather than merely in L2x (Rd ),
because a new conservation law becomes available, namely energy conservation28
For a general NLS, the total energy E[u(t)] takes the form
Z
1 2
(3.30) E[u(t)] := |u(t, x)|2 + |u(t, x)|p+1 dx.
Rd 2 p+1
R
We refer to the first term 12 Rd |u(t, x)|2 as the kinetic energy or linear energy,
2
and the second term p+1 |u(t, x)|p+1 dx as the potential energy or nonlinear energy.
Note that in the defocusing case = +1 these two terms have the same sign,
whereas in the focusing case = 1 they have opposite signs. In practice, this

28In principle, momentum conservation should become useful also once one reaches the reg-
1/2
ularity Hx or higher, thanks to Lemma A.10. However, because the momentum is a vector
rather than a positive quantity, the momentum is in practice not sufficiently coercive to obtain
any useful control of the solution. See however the Morawetz arguments in the next section.
148 3. SEMILINEAR DISPERSIVE EQUATIONS

means that the energy conservation law is more coercive (gives more control on the
solution) in the defocusing case than in the focusing case.
The following heuristic principle, related to Principle 1.37 and Principle 3.1,
can be helpful in predicting behaviour of equations in the energy class:
Principle 3.24 (Energy principle). Suppose a solution has finite energy. If
the linear energy dominates the nonlinear energy, we expect linear behaviour; if the
nonlinear energy dominates the linear energy, we expect nonlinear behaviour.
This heuristic has very little rigorous justification to back it up, yet is surpris-
ingly accurate in many cases, as we shall see in several places in this text.
For sufficiently classical solutions one can justify conservation of energy E[u(t)]
by integration by parts; see Exercise 3.31. In the energy subcritical cases sc 1, the
energy functional u 7 E[u] is continuous in Hx1 (Rd ) (Exercise 3.32). Combining
this with the local Hx1 wellposedness theory (such as Proposition 3.19 and 3.21) as
in Proposition 3.22, one can extend the energy conservation law to all Hx1 -critical
and Hx1 -subcritical wellposed solutions. (In fact one has this for strong solutions
also, thanks to uniqueness results such as Exercise 3.17. The high-dimensional cases
n > 6 are a little tricky; see [Caz2], [TV].)
Let us now return to the two-dimensional cubic NLS (d = 2, p = 3), and see
what this new conservation law gives us. The focusing and defocusing cases are
now rather different (as one can already see from the blowup solution (3.15)). In
the defocusing case, it is clear that energy and mass together will control the Hx1
norm of the solution:
ku(t)k2Hx1 . E[u(t)] + M [u(t)].
Conversely, the Gagliardo-Nirenberg inequality (Proposition A.3) shows that the
energy and mass are controlled by the Hx1 norm:
E[u(t)] . ku(t)k2Hx1 (1 + ku(t)k2L2x ) . ku(t)k2Hx1 (1 + ku(t)k2Hx1 )
M [u(t)] = ku(t)k2L2x . ku(t)k2Hx1 .
From these bounds and the energy and mass conservation laws we see that for
any Hx1 -wellposed solution, the Hx1 norm of the solution u(t) at some later time
t is bounded by a quantity depending only on the Hx1 norm of the initial datum.
On the other hand, Proposition 3.19 shows that an Hx1 wellposed solution can be
continued in time as long as the Hx1 norm does not go to infinity. Combining these
two statements we obtain
Proposition 3.25. The defocusing two-dimensional cubic NLS (d = 2, p =
3, = +1) is globally wellposed in Hx1 . Indeed for any u0 Hx1 and any time
interval I containing t0 , the Cauchy problem (3.1) has an Hx1 -wellposed solution
u S 1 (I R2 ) Ct0 Hx1 (I R2 ).
The reader should see how the scheme in Figure 7 is modified to accomodate
the utilisation of two conservation laws (mass and energy) rather than just one.
Remark 3.26. This argument is in fact quite general and works for any Hx1 -
subcritical defocusing NLS; see for instance [Caz2] or Exercise 3.35. One also has
global wellposedness in Hx1 for the Hx1 -critical equation but this is significantly more
difficult and will be discussed in Chapter 5. This fits well with Principle 3.24, since
in the Hx1 -critical equation, Sobolev embedding only barely manages to control the
nonlinear energy in terms of the linear energy.
 3.4. CONSERVATION LAWS AND GLOBAL EXISTENCE 149

Remark 3.27. In the focusing case, a problem arises because of the negative
sign of the nonlinear component of the energy E[u(t)]. This means that while the
energy is still controlled by the Hx1 norm, the Hx1 norm is not necessarily controlled
by the energy. Indeed, (3.15) shows that global wellposedness fails for some Hx1
data. However this turns out to be the borderline case: for any Hx1 data with mass
strictly less than that of the blowup solution in (3.15), the Gagliardo-Nirenberg
inequality becomes strong enough again to control the Hx1 norm by the energy and
thus regain global wellposedness. See Exercise 3.33.

The above discussion was for the two-dimensional cubic NLS, which was L2x -
critical. For equations which are L2x -subcritical, it turns out that the Gagliardo-
Nirenberg inequality is now so strong that the sign of the nonlinearity plays
essentially no role in the global theory; see Exercise 3.34. For equations which are
L2x -supercritical but Hx1 -subcritical or Hx1 -critical, the defocusing equation enjoys
global existence in Hx1 as discussed above, but blowup can occur for the focusing
equation unless a suitable smallness condition is met; see Exercise 3.35 and Section
3.8.
Having discussed the conservation laws for the NLS, we now turn to the NLW
(3.2), which we write using the Minkowski metric as

(3.31) u = |u|p1 u.

For this equation there is no mass conservation law, and the energy/momentum
conservation laws can be unified via the stress-energy tensor

1 2
(3.32) T := Re( u u) g Re( u u + |u|p+1 )
2 p+1

(compare with (2.45)). In coordinates,

1 1
T00 = T00 = |t u|2 + |u|2 + |u|p+1
2 2 p+1
T0j = Tj0 = Re(t uxj u)
jk 2
Tjk = Tjk = Re(xj uxk u) (|u|2 |t u|2 + |u|p+1 ).
2 p+1

From (3.31) we easily verify the divergence-free property

(3.33) T = 0

or in coordinates

(3.34) t T00 + xj T0j = 0; t T0j + xk Tjk = 0,

for classical solutions at least. This leads (for classical, decaying solutions) to
conservation of the total energy
Z Z
00 1 1
(3.35) E[u[t]] := T (t, x) dx = |t u|2 + |u|2 + |u|p+1 dx
R d R d 2 2 p + 1
150 3. SEMILINEAR DISPERSIVE EQUATIONS

and total momentum29

Z Z
0j
pj (u[t]) := T (t, x) dx = Re(t uxj u) dx.
Rd Rd
By the limiting arguments as before, these conservation laws can be extended to
Hx1 -wellposed solutions, as long as the equation is either Hx1 -subcritical or Hx1 -
critical.
For defocusing Hx1 -subcritical equations, the energy conservation can lead to
global existence in even for large initial data. Let us illustrate this with the three-
dimensional cubic NLW (d = 3, p = 3, = +1). From Exercise 3.22 we have a local
wellposedness result for initial data in Hx1 L2x in the subcritical sense, which easily
implies a blowup criterion, namely that the Hx1 L2x wellposed solutions to this
equation can be continued in time as long as the quantity ku(t)kHx1 + kt u(t)kL2x
does not go to infinity in finite time. To bound this quantity, we observe from
(3.35) and energy conservation that we can bound the homogeneous component of
this quantity easily:
ku(t)kH 1 + kt u(t)kL2x . E[u[t]]1/2 = E[u[0]]1/2 .
x

To control the lower order term ku(t)kL2x we use the fundamental theorem of cal-
culus and Minkowskis inequality:
Z t
ku(t)kL2x ku(0)kL2x + | kt u(s)kL2x ds|
0
Z t
. ku(0)kL2x + | E[u[s]]1/2 ds|
0
= ku(0)kL2x + |t|E[u[0]]1/2 .
Thus we can control the Hx1 L2x norm of u[t] by a quantity that depends mostly
on the initial data:
ku(t)kHx1 + kt u(t)kL2x . ku(0)kL2x + (1 + |t|)E[u[0]]1/2 .
In particular, if u[0] Hx1 L2x , then E[u[0]] is finite and the quantity ku(t)kHx1 +
kt u(t)kL2x cannot go to infinity in finite time for any Hx1 L2x -wellposed solution.
Comparing this with the blowup criterion30 we see that we have a global Ct0 Hx1
Ct1 L2x strong solution to this equation for any initial data in the energy space
Hx1 L2x . Once one has this global existence result in the energy class, we also
obtain it for smoother classes; see Exercise 3.38.
The above argument in fact works for any Hx1 -subcritical NLW; see [GV2],
[Sog]. The case of the Hx1 -critical equation is again more delicate (because the
lifespan given by the local existence theorem depends on the profile of the data as
29Note the presence of the time derivative, which is absent for the NLS momentum. Indeed,
1/2
while the NLS momentum is naturally associated to the regularity Hx , the NLW momentum is
associated to the regularity Hx1 , and thus is of the same order as the energy. Thus the relationship
between momentum and energy in NLW is different from that in NLS, which turns out to be a
crucial difference in the critical scattering theory, see Chapter 5 below.
30It is instructive to use the other approach to global existence from Proposition 3.23, namely
dividing up a long time interval into short ones. Here, because the quantity ku(t)kH 1 +kt u(t)kL2
x x
could grow linearly in t, the interval of existence guaranteed by the local theory could decay
polynomially in t. However, the length of this interval will not go to zero in finite time, which is
all one needs to establish global existence.
 3.4. CONSERVATION LAWS AND GLOBAL EXISTENCE 151

well as on the H 1 L2x norm, and so the blowup criterion is more subtle), and will be
treated in Section 5.1; again, this is in accordance with Principle 3.24. For focusing
NLW, there are no Gagliardo-Nirenberg tricks available to control the nonlinearity
(since there is no mass conservation law), and indeed for such equations large data
can lead to blowup no matter what the power is; see Exercise 3.9.
3
Exercise 3.26. Verify (2.35) for the pseudo-stress-energy tensor for Ct,x,loc
solutions to (3.1). These conservation laws may seem somewhat miraculous, but
can be explained using Lagrangian methods; see [SSul], as well as Exercises 3.30,
3
3.40. Conclude that if u Ct,x,loc (I Rd ) is a solution to (3.1) which also lies in
0 s d
Ct Hx (I R ) for a sufficiently large s, then we have mass conservation M [u(t)] =
M [u(t0 )] and momentum conservation pj (t) = pj (t0 ) for all t, t0 I. (In order to
justify the integration by parts, one needs to apply a smooth cutoff in space, and
then let the cutoff go to infinity, using the Hxs control on u to show that the error
incurred by the cutoff goes to zero.) These rather restrictive regularity conditions
can usually be removed by a limiting argument.
Exercise 3.27. Obtain the analogue of Proposition 3.22 but with d = 2 instead
of d = 1. (The challenge here is that the equation is now L2x -critical instead of L2x -
subcritical, and one has to use Proposition 3.17 instead of Proposition 3.15.)
Exercise 3.28. Let p = 3, d = 1, = 1, and let u0 Hxs (R) for some s 0.
Show that the solution u constructed in Proposition 3.23 is a strong Hxs solution,
and we have the bound ku(t)kHxs . exp(Ct)ku0 kHxs , where C = C(s, ku0 kL2x ) > 0
depends only on s and the initial mass.
Exercise 3.29. Show that the energy (3.30) is formally the Hamiltonian for
the NLS (3.1) using the symplectic structure from Exercise 2.47. Also use Noethers
theorem to formally connect the mass conservation law to the phase invariance of
NLS, and the momentum conservation law to the translation invariance of NLS.
Exercise 3.30. Use Exercise 3.2 to link the pseudo-stress-energy tensor and
energy conservation law for d-dimensional NLS to the stress-energy tensor for d+ 1-
dimensional NLW. (In d + 1 dimensions, you may find it convenient to introduce a
null frame (t , x , x1 , . . . , xd ) where t := t xd+1 and x := t + xd+1 , and compute
the coordinates of the stress-energy tensor in that frame.)
3
Exercise 3.31. Let u Ct,x (I Rd ) be a classical solution to (3.1). Verify
the identity
1 2
t ( |u(t, x)|2 + |u(t, x)|p+1 ) =
2 p+1
1
j ( Im(jk u(t, x)k u(t, x)) + |u(t, x)|p1 Im(u(t, x)j u(t, x))).
2
If u is also in Ct0 Hxk,k (I Rd ) for some sufficiently large k, deduce the energy
conservation law E[u(t1 )] = E[u(t0 )] for all t0 , t1 I (by arguing as in Exercise
3.26).
Exercise 3.32. If sc 1, show that the energy functional u 7 E[u] is well-
defined and continuous on the space Hx1 (Rd ). (Hint: use Sobolev embedding and
an estimate similar to (3.27).) When sc > 1, show that the energy is not always
finite for Hx1 (Rd ) data.
152 3. SEMILINEAR DISPERSIVE EQUATIONS

Exercise 3.33. [Wei] Let u0 Hx1 (R2 ) have mass strictly less than the mass
of the blowup solution in (3.15). Show that there is a global strong Hx1 solution to
the cubic defocusing two-dimensional NLS (i.e. (3.1) with d = 2, p = 3, = 1)
with initial datum u0 . (Hint: you will need the relationship between Q and Wmax
given by Lemma B.1 in order to control the Hx1 norm by the energy.) Note that
this is consistent with Principle 3.24. There has been much recent analysis of the
case when the mass is exactly equal to, or slightly higher than, the mass of the
blowup solution (3.15): see [Mer2], [Mer3], [MR], [MR2], [MR3].
Exercise 3.34. Let u be an Hx1 -wellposed solution to the one-dimensional
cubic NLS (d = 1, p = 3, = 1) with initial datum u0 ; this is a global so-
lution by Proposition 3.23 and persistence of regularity. Establish the bound
ku(t)kHx1 (R) .ku0 kH 1 (R) 1 for all times t, regardless of whether the equation is
x
focusing or defocusing. (The point here is that in the L2x -subcritical equations, the
Gagliardo-Nirenberg inequality allows one to control the nonlinear component of
the energy by a fractional power of the linear component of the energy, times a
factor depending only on the conserved mass; cf. Exercise 1.22 and Principle 3.24.)
Exercise 3.35. Consider the defocusing three-dimensional cubic NLS (d =
3, p = 3). Show that one has global Hx1 -wellposed solutions if the initial datum
u0 is sufficiently small in Hx1 (R3 ) norm, and in the defocusing case one has global
Hx1 -wellposedness for arbitrarily large Hx1 (R3 ) initial data. (Again, one can rely
primarily on the Gagliardo-Nirenberg inequality and the conservation laws. An
alternate approach is to develop a small data global existence theory at the crit-
1/2
ical regularity Hx (R3 ) by perturbative arguments, and then use persistence of
1/2
regularity to move from Hx to Hx1 .)
Exercise 3.36. Consider the one-dimensional cubic NLS (d = 1, p = 3). It
turns out that there is a conserved quantity (at least for classical solutions to NLS)
of the form
Z
E2 (u) := |xx u|2 + c1 |x u|2 |u|2 + c2 Re((ux u)2 )) + c3 2 |u|6 dx
R
for certain absolute constants c1 , c2 , c3 whose exact value is unimportant here. (The
verification of this conservation law is extremely tedious if done directly, though
the machinery in Section 4.2 can expedite some of the algebra; one can also proceed
via the Lax pair formulation of this equation.) Assuming this, conclude the bound
ku(t)kHx2 (R) .ku0 kH 2 (R) 1 for all times t, at least for classical solutions to (3.1).
x
(The same bound in fact holds for all Hx2 solutions, and one has a similar result
with Hx2 replaced by Hxk for any integer k 0.)
Exercise 3.37. Show that the 1D cubic periodic NLS (with either sign of
nonlinearity) is globally wellposed in L2x (T). Also show that if the initial datum is
smooth, then the solution is globally classical. (One should of course use the local
theory from Exercise 3.20.)
Exercise 3.38. Show that for every smooth initial data u[0] to the three-
dimensional cubic defocusing NLW (d = 3, p = 3, = +1), there is a unique
classical solution. (In the text we have already established global wellposedness in
Hx1 L2x for this equation; it is now a matter of applying the persistence of regularity
theory (to ensure the solution is smooth) and the finite speed of propagation and
uniqueness theory (to localise the initial data to be compactly supported).)
 3.5. DECAY ESTIMATES 153

Exercise 3.39. Consider a global Hx1 L2x -wellposed solution u to the three-
dimensional cubic defocusing NLW (d = 3, p = 3, = +1), as constructed in the
text. If (u0 , u1 ) Hxk Hxk1 for some k 1, show that the quantity ku(t)kHxk +
ku(t)kHxk1 grows at most polynomially in time t, in fact we have the bound

ku(t)kHxk + ku(t)kHxk1 .ku0 kH k +ku1 k k1

(1 + |t|)Ck
x Hx

for some Ck > 0. (Hint: induct on k. This result should be compared with the
exponential bounds in Exercise 3.28. The difference is that for the wave equation,
the energy estimate (2.28) or (2.29) gains one degree of regularity, which is not
the case for the Schrodinger equation. However, in many cases it is possible to use
additional smoothing estimates and almost conservation laws to recover polynomial
growth of Sobolev norms for the Schrodinger equation; see [Sta], [CKSTT8], and
the scattering estimates we give in Section 3.6 can eliminate this growth altogether).
Exercise 3.40. Show that theRNLW (3.2) is the (formal) Euler-Lagrange equa-
tion for the Lagrangian S(u, g) = R1+d L(u, g) dg, where L(u, g) := g u u +
2 p+1
p+1 |u| . Conclude that the stress-energy tensor given here coincides with the
one constructed in Exercise 2.60.
Exercise 3.41 (Positivity of stress-energy tensor). Let u be a classical solution
to a defocusing NLW, and let T be the associated stress energy tensor. Let x , y
be forward timelike or forwardlike vectors, thus x0 , y 0 > 0 and x x , y y 0.
Show that T x y 0. In particular we have the positivity property (2.47).
(Hint: first establish this when x is the standard timelike vector e0 , then use
Lorentz invariance to handle the case when x is a general timelike vector, then
use limiting arguments to handle the lightlike case.) This positivity is related to
the finite speed of propagation property for NLW but is not identical; indeed,
in the focusing case = 1, the positivity fails but one still has finite speed of
propagation.
Exercise 3.42. Consider a Hx1 -subcritical focusing NLS. Show that the ground
state eit Q(x) has positive energy in the L2x -supercritical case p > 1+ d4 , zero energy
in the L2x -critical case p = 1 + d4 , and negative energy in the L2x -subcritical case
p < 1 + d4 . Similarly for all translates, rescalings, and Galilean transforms of the
ground state. (Hint: Use Exercise B.3.)
Exercise 3.43 (Orbital stability of NLS). [Wei3] Consider a L2x -subcritical
focusing NLS. Define the ground state cylinder as in Exercise B.14. Show that if
the initial datum u0 Hx1 (Rd ) is sufficiently close to in Hx1 (Rd ), then the global
Hx1 solution u to the Cauchy problem is such that distHx1 (u(t), ) distHx1 (u0 , )
for all t R. (Hint: first rescale so that u0 has the same mass as , then use
Exercise B.14.) An earlier result of this type appears in [CL]. The theory of
orbital stability of ground states for much more general equations has been studied
extensively; see for instance [GSS], [GSS2] for a systematic approach.

3.5. Decay estimates

Things fall apart; the centre cannot hold. (WB Yeats, The second
coming)
154 3. SEMILINEAR DISPERSIVE EQUATIONS

The conservation laws of the preceding section can give global bounds on a
solution u(t) to NLS or NLW that are either uniform in time, or grow at some
controlled rate (polynomial or exponential in time). We have already seen that such
bounds can be sufficient to obtain global existence of the solution. However, one is
not just interested in whether solutions exist globally in time; one is also interested
in the asymptotic behaviour of these solutions as t . The conservation laws
show that these solutions stay bounded in certain norms, but this still leaves a lot
of possibilities for the asymptotic development. For instance, consider the following
two rather different (and informally described) modes of behaviour:
(Linear-type behaviour) The solution u behaves like a solution to the linear
equation; thus the nonlinear effects become asymptotically negligible. In
particular, we expect the solution to obey the same type of dispersive
and Strichartz estimates as the linear equation, thus for instance we may
expect the L p
x or other Lx norms of the solution to go to zero as t .
We also expect Sobolev norms such as ku(t)kHxs to stabilise as t ,
as opposed to growing polynomially or exponentially in time.
(Soliton-type behaviour) The solution refuses to disperse, and for every
time t the solution has a significant presence
R at some location x(t) de-
pending on t, for instance the local mass |xx(t)|R |u(t, x)|2 dx might be
bounded away from zero for some fixed R. In particular the L x or Lx
p

norms of u(t) will not go to zero as t . This is the type of behaviour

exhibited by soliton solutions such as (3.7) (possibly after applying some
symmetries such as Galilean or Lorentz invariance). One can also con-
sider more complex behaviour when for each fixed time t, the solution has
significant presence at multiple points x1 (t), . . . , xk (t); this is the case for
multi-soliton solutions, which are essentially a nonlinear superposition of
several separated solitons.
There is some evidence (both theoretical and numerical) that for most31
global solutions to an NLS or NLW, that the asymptotic behaviour eventually
decouples into the above two extremes: most solutions should eventually resolve
into a localised component which behaves like a soliton or multi-soliton solution,
plus a radiation component which disperses like a linear solution. Making this
rather vaguely worded soliton resolution conjecture a rigorous theorem is a major
open problem in the field, and somewhat out of reach of current technology except
in special cases (e.g. small perturbations of a special solution such as a soliton,
multisoliton, or the vacuum state 0, or the one-dimensional cubic NLS, which is
completely integrable). However, significantly more is known in the defocusing
case = 1. In many defocusing cases it is known that soliton-type behaviour is
excluded, and all solutions in fact disperse like a linear equation. These results are
part of the scattering theory for these equations and will be discussed more fully in
the next section. For now, let us just say that the question of whether a solution

31One of the many difficulties with establishing this conjecture is that we expect there to
be a small class of exceptional solutions which exhibit more exotic (and unstable) behaviour,
such as periodic breather solutions, or clusters of solitons which diverge from each other only
logarithmically. Almost all of the known tools in the subject are deterministic in the sense that if
they work at all, they must work for all data in a given class, while to settle this conjecture it may
be necessary to develop more stochastic techniques that can exclude small classes of exceptional
solutions.
 3.5. DECAY ESTIMATES 155

disperses or not is intimately tied to whether there is some sort of decay estimate
for the solution in various norms, such as an Lpx norm for some p > 2; in many
cases, knowing that such an Lpx norm goes to zero as t (either in a classical
sense, or in some time-averaged sense) is sufficient to establish that the solution
scatters to a linear solution, while conversely estimates such as Strichartz estimates
assure us that the Lpx norms of such solutions do indeed go to zero. (This should
also be compared with Principle 3.24.)
Thus it is of interest to obtain decay estimates on solutions to defocusing equa-
tions. The conservation laws establish boundedness in L2x -based spaces such as L2x
and Hx1 , but do not yield any decay estimates in higher Lpx norms. There are a num-
ber of known ways to establish such a decay estimate; in this section we shall discuss
three such, namely the Morawetz inequality approach, the (pseudo)conformal iden-
tity approach, and the vector fields approach.
We begin with the Morawetz inequality approach. This method is based on
monotonicity formulae, as discussed in Section 1.5. It is here that we can begin to
usefully exploit the momentum conservation laws. As momentum is a vector, these
laws are not of the coercive type needed to obtain uniform bounds on a solution as
in the preceding section, but the vector structure does permit32 one to construct
various quantities based on the momentum density which are monotone in time,
and so the fundamental theorem of calculus will provide some spacetime bounds
that force some decay in the solution.
In the linear setting, Morawetz inequalities for the NLS and NLW were already
introduced in Section 2.4 and Section 2.5, using the pseudo-stress-energy and stress-
energy tensors respectively. The NLS and NLW also have such tensors, and in the
defocusing case the sign of the nonlinearity will be favourable for preserving the
desired monotonicity. In the case of the NLS, we can repeat the derivation of
(2.37) (taking into account the new nonlinear term in the Tjj components of the
pseudo-stress-energy tensor) and obtain the identity

Z Z
t2 a(x)|u(t, x)|2 dx = t xj a(x)Im(u(t, x)xj u(t, x)) dx
Rd Rd
Z
= (xj xk a(x))Re(xj uxk u) dx
Rd
(3.36) Z
(p 1)
+ |u(t, x)|p+1 a(x) dx
p+1 Rd
Z
1
|u(t, x)|2 2 a(x) dx,
4 Rd

at least for smooth a of polynomial growth, and for classical, decaying solutions u
to (3.1). One can specialise this to a = |x| and d 3 (justifying this as in Exercise

32Indeed, it is not possible for an (autonomous) quantity based on the mass or energy density
to have a non-trivial monotonicity in time, as this would conflict with time reversal symmetry.
The momentum density is not subject to this problem, since reversing time also reverses the sign
of the momentum density.
156 3. SEMILINEAR DISPERSIVE EQUATIONS

2.57) to conclude that

Z Z
x |/ u(t, x)|2
t Im(u(t, x) u(t, x)) dx = dx
Rd |x| Rd |x|
Z
2(p 1) |u(t, x)|p+1
(3.37) + dx
p+1 Rd |x|
Z
1
|u(t, x)|2 2 a(x) dx.
4 Rd
When d 3 the tempered distribution 2 a is non-negative. In particular, in the
defocusing case this time derivative is non-negative, and we have the monotonicity
formula Z Z
x |u(t, x)|p+1
t Im(u(t, x) u(t, x)) dx &p dx.
Rd |x| Rd |x|
Integrating this along a time interval [t0 , t1 ] and using Lemma A.10, we obtain the
Morawetz inequality
Z t1 Z
|u(t, x)|p+1
(3.38) dx .p,d sup ku(t)k2H 1/2 (Rd )
t0 R d |x| t=t0 ,t1

for any classical solution to defocusing NLS on [t0 , t1 ] Rd . In practice, the re-
quirement that this solution is classical can be dropped by the usual limiting ar-
guments33, provided that one is working with a wellposed solution at a regularity
strong enough to control both sides of (3.38).
Suppose we are working with an Hx1 -wellposed solution, with a defocusing equa-
tion which is Hx1 -subcritical or Hx1 -critical (this turns out to be sufficient to justify
the bound (3.38)). As we saw in the preceding section, the conservation laws of
mass and energy allow one in this case to show that the solution is global, and
1/2
bound the Hx1 norm (and hence Hx norm) of u(t) by a quantity depending on
the Hx1 norm of the initial datum u0 . Applying this to (3.38) and letting the time
interval [t0 , t1 ] go to infinity, we obtain the global spacetime bound
Z Z
|u(t, x)|p+1
(3.39) dxdt .p,d,ku0 kH 1 1,
R Rd |x| x

first observed in [LStr] (and inspired by a similar result in [Mor] for nonlinear wave
R p+1
equations). This is a decay estimate, as it shows that the quantity Rd |u(t,x)|
|x| dx
1
must go to zero, at least in some time-averaged sense. Because the weight |x| is large
at the origin, this means (roughly speaking) that the solution cannot maintain a
significant presence near the origin for extended periods of time. This is a nonlinear
effect caused by the defocusing nature of the nonlinearity; it fails utterly in the
focusing case = 1 (as one can see by inspecting the soliton solution (3.7)), and
also behaves strangely in the linear case = 0 (see Exercise 3.44). It is especially
useful for spherically symmetric solutions, as such solutions already decay away
from the origin (Exercise A.19). However, this estimate is not as effective for
general solutions, which can be located arbitrarily in space. This problem can be
alleviated to some extent by exploiting spatial translation invariance. For sake of
simplicity we discuss the three-dimensional case d = 3, in which the formulae are
33In the case where the NLS is not algebraic, one also needs to regularise the nonlinearity in
order to create an approximating sequence of classical solutions, and exploit some stability theory
as in Section 3.7; we ignore this rather tedious detail.
 3.5. DECAY ESTIMATES 157

cleanest (for higher dimensions, see [Vis], [TVZ], and for lower dimensions, see
[Gri6]). By translating (3.37) by y we obtain
Z Z
xy / y u(t, x)|2
|
t Im(u(t, x) u(t, x)) dx = dx
R3 |x y| R3 |x y|
Z
(3.40) 2(p 1) |u(t, x)|p+1
+ dx
p+1 R3 |x y|
+ |u(t, y)|2
where / y is the angular component of the gradient using y as the origin. This
estimate can then be used to obtain a translated version of (3.39) which prevents
the solution u from concentrating at the point y for long periods of time. The
freedom afforded by this additional parameter y can be exploited by integrating
(3.40) against a suitable weight in y. It turns out that the best weight to achieve
this with is the mass density T00 (t, y) = |u(t, y)|2 . This gives
(3.41)
Z Z
xy
t |u(t, y)|2 Im(u(t, x) u(t, x)) dxdy
R3 R3 |x y|
Z Z
dxdy
= |u(t, y)|2 |/ y u(t, x)|2
R3 R3 |x y|
Z Z
2(p 1) dxdy
+ |u(t, y)|2 |u(t, x)|p+1
p+1 R3 R3 |x y|
Z
+ |u(t, y)|4 dy
R3
Z Z
xy
+ (t |u(t, y)|2 )Im(u(t, x) u(t, x)) dxdy.
R 3 R 3 |x y|
To deal with the final term of (3.41), we rewrite it in terms of the pseudo-stress-
energy tensor as
Z Z
xk yk
(t T00 (t, y)) T0k (t, x) dxdy
R 3 R 3 |x y|
and then use (2.35) and integration by parts to write this as
Z Z
(xj yj )(xk yk ) dxdy
T0j (t, y)(1 2
)T0k (t, x) ;
R3 R3 |x y| |x y|
if u is smooth and decaying then there is no difficulty justifying the integration by
parts. However, an application of Cauchy-Schwarz shows that
(xj yj )(xk yk )
|T0j (t, y)(1 )T0k (t, x)|
|x y|2
|u(t, x)||
/ y u(t, x)||u(t, y)||/ x u(t, y)|
1
(|u(t, y)|2 | / y u(t, x)|2 + |u(t, x)|2 |/ x u(t, y)|2
2
whenever x 6= y; this can be seen for instance by rotating x y to be a multiple
of the unit vector e1 and working in coordinates. From this pointwise bound and
symmetry we thus see that we can bound the last term in (3.41) by the first. If we
158 3. SEMILINEAR DISPERSIVE EQUATIONS

are in the defocusing or linear cases 0, we can also drop the second term as
being non-negative, and we conclude
Z Z Z
xy
t |u(t, y)|2 Im(u(t, x) u(t, x)) dxdy |u(t, x)|4 dx.
R3 R3 |x y| R3
But by (a translated version of) Lemma A.10 we have the pointwise bound
Z Z
xy
| |u(t, y)|2 Im(u(t, x) u(t, x)) dxdy| . ku(t)k2L2x ku(t)k2H 1/2 .
R3 R3 |x y| x

From the fundamental theorem of calculus, we thus obtain the interaction Morawetz
inequality
Z t1 Z
(3.42) |u(t, x)|4 dxdt . sup ku(t)k2L2x ku(t)k2H 1/2
t=t0 ,t1 x
t0 R3

whenever u is a classical solution to a defocusing or linear NLS on [t0 , t1 ] R3 ;

this was first observed in [CKSTT10], and should also be compared with Example
1.34. There is no difficulty applying a limiting argument to extend this inequality
to Hx1 -wellposed solutions when the NLS is Hx1 -subcritical or Hx1 -critical. Using the
energy and mass conservation laws, we see in particular that we have the spacetime
bound
Z t1 Z
(3.43) |u(t, x)|4 dxdt .ku0 kH 1 1
x
t0 R3
in this case, where u0 is the initial datum of this solution. This bound resembles
(3.39), but is a linear phenomenon rather than a nonlinear one (it holds true even
1
when = 0), and does not involve the weight |x| and so is not tied to any particular
34
spatial origin . This makes it a more useful decay estimate for controlling solutions
to NLS when there is no assumption of spherical symmetry.
The ordinary Morawetz estimates for NLS have a counterpart for NLW; see Ex-
ercise 3.46. However, it seems difficult to locate a useful analogue of the interaction
Morawetz inequality for the NLW; the somewhat miraculous positivity properties
of the time derivative of the interaction functional do not seem to be present in
the wave equation setting, even if one drops the nonlinearity. Fortunately, for these
equations the ordinary Morawetz estimate is already quite powerful, especially when
combined with finite speed of propagation.
The Morawetz inequalities are based on the monotonicity formulae method.
Another way to obtain decay is to find a conserved (or almost conserved) quantity
which is non-autonomous (depending explicitly on t). Often, such laws arise by
conjugating an autonomous conservation law with a symmetry (or approximate
symmetry) of the equation. For instance, for the free Schrodinger equation we have
already seen that the pseudoconformal symmetry from Exercise 2.28 conjugates the
usual energy conservation law to the conservation of the pseudoconformal energy
(2.33). Turning now to (classical) solutions u of the nonlinear Schrodinger equation,
34While the original Morawetz inequality controls the extent to which a solution can con-
centrate near a fixed point y, the trick of integrating that inequality against the mass density
means that the interaction Morawetz inequality controls the extent
R to which the solution con-
centrates against itself. In this perspective, the L4x quantity R3 |u(t, x)|4 dx can be rewritten
R R
as R3 R3 |u(t, x)|2 |u(t, y)|2 (x y) dxdy and thought of as a self-interaction of the mass den-
sity. The interaction inequality can also be thought of as an ordinary Morawetz inequality for a
six-dimensional (or two-particle) Schrodinger equation; see Exercise 3.45.
 3.5. DECAY ESTIMATES 159

recall that the pseudoconformal transform v of u, as defined in Exercise 2.28, obeys

the modified equation (3.16). In particular, in the L2x -critical case p = pL2x the
pseudoconformal transformation is a symmetry of the equation; we have already
used this fact to construct the blowup solution (3.15). But this transform is still
useful even in the non-L2x -critical case; one can view the equation (3.16) as an NLS
d
(ppL2 )
in which the degree of focusing or defocusing, as quantified by t 2 x , is now

time-dependent. In analogy with the usual NLS, we can define a (non-autonomous)

energy E[v(t), t] for t 6= 0 by
Z p+1
1 d
(ppL2 ) |v(t)|
E[ v(t), t] := |v(t)|2 + 2t 2 x dx.
Rd 2 p+1
Because (3.16) is not time translation invariant in general, we do not expect this
energy to be perfectly conserved in time (except when p = pL2x ). Nevertheless, it
should be almost conserved in that its time derivative should be small. Indeed,
a computation (essentially the one in Exercise 3.31) shows that
Z
d d
(ppL2 )1 |v(t)|p+1
t E[v(t), t] = (p pL2x )t 2 x dx;
2 Rd p+1
in other words, the only source for the time variation of the energy of v is the
explicit dependence on the time variable t.
This formula can be used to control the time evolution of the energy of v, which
in turn gives control on the original solution by means of the easily verified identity
(3.44) E[v(t), t] = Epc [u(1/t), 1/t]
where Epc is the pseudoconformal energy
Z Z
1 2 t2
(3.45) Epc [u(t), t] := |(x + it)u(t)| dx + 2 |u(t)|p+1 ,
2 Rd p + 1 Rd
which is defined for all t including 0. For instance, in the L2x -critical case p = pL2x ,
the quantity E[v(t), t] is conserved in t, and hence the pseudoconformal energy
Epc [u(t), t] is also conserved in t. (This was only established for t 6= 0, but can
be verified for t = 0 also, either by a limiting argument or by establishing the
conservation of Epc [u(t), t] directly; see Exercise 3.47.) In the L2x -critical defocusing
case (p = pL2x , = +1), we obtain in particular that
Z
t2 1
2 |u(t)|p+1 Epc [u(t), t] = Epc [u(0), 0] = kxu(0)k2L2x
p + 1 Rd 2
which leads to the decay bound
d d
ku(t)kL2(d+2)/2 .d |t| d+2 kxu(0)kLd+2
2
x x

for all times t 6= 0 for which the (classical) solution exists; this bound can be
extended to more general classes of solution by the usual limiting arguments. Some
further examples of decay estimates of this type are given in the exercises; they
can give quite strong decay for a wide range of powers, but have the drawback that
they require some spatial decay on the initial datum (in this case, one needs xu(0)
to lie in L2x ).
For the NLW, the analogue of the pseudoconformal energy is the conformal
energy Q[u(t), t], which was already introduced in (2.54) for the linear wave equa-
tion and is defined the same way for the NLW (using the nonlinear stress-energy
160 3. SEMILINEAR DISPERSIVE EQUATIONS

tensor defined in (3.32), of course). For classical solutions to NLW one can use the
properties of the stress energy tensor to verify the identity
Z
p pc
t Q[u(t), t] = t(d 1) |u(t, x)|p+1 dx
p + 1 Rd
4
where pc := 1 + d1 is the conformal (or Hx1 -critical) exponent. This can be utilised
to obtain decay estimates in analogy with the pseudoconformal energy and the NLS,
especially for the conformal power p = pc .
The pseudoconformal and conformal energy methods provide decay of the so-
lution in an Lpx sense. In some cases one wishes to also establish decay of the
solution in an L sense. This can be done via Sobolev embedding but requires one
to control quantities that involve more than one derivative. One way to do this
(assuming a sufficiently small, smooth, and localised initial datum) is by the vector
field method, which we introduced in Section 2.5. For technical reasons it is a little
simpler to work with a derivative nonlinear wave equation rather than a semilinear
NLW; for sake of illustration and concreteness we shall work with classical solutions
to the (rather artificial) three-dimensional scalar equation u = (t u)3 . We recall
the Killing vector fields X defined in Section 2.5. We use this to define the higher
order energies En (t) for any n = 0, 1, . . . as
n
X X
En (t) = kK1 . . . Km u(t)k2H 1 + kt K1 . . . Km u(t)k2L2x
m=0 K1 ,...,Km

where K1 , . . . , Km range over the vector fields t , xj , xj xk xk xj , or txj + xj t .

Since the Killing vector fields commute with , we have
(K1 . . . Km u) = K1 . . . Km ((t u)3 )
and hence by the energy estimate (2.28) we have
n
X X Z t
En (t) . En (0) + kK1 . . . Km ((t u)3 )(t )kL2x dt
m=0 K1 ,...,Km 0

for all t 0. Let us apply this with n = 5 (this is far more regularity than
strictly necessary, but serves to illustrate the general idea). We have at most five
Killing vector fields K1 , . . . , Km applied to (t u)3 on the right-hand side. Using
the Leibnitz rule repeatedly, we can distribute these derivatives and observe that
at most one of the factors t u will receive more than two of these Killing vector
fields. We place that factor in L2x and the other two in L , and obtain
kK1 . . . Km ((t u)3 )(t )kL2x . ( sup sup kK1 . . . Km t u(t )kL2x )
m 5 K1 ,...,Km

( sup sup kK1 . . . Km t u(t )kL

x
)2 .
m 2 K1 ,...,Km

Now one observes that the commutator of t or xj with one of the vector fields in
the list t , xj , xj xk xk xj , or txj + xj t is a linear combination of the vector
fields t and xj . Using that fact repeatedly, we can bound
kK1 . . . Km t u(t )kL2x . E5 (t)
whenever m 5. Applying the Klainerman-Sobolev inequality (see Exercise 3.48)
one obtains
kK1 . . . Km t u(t )kL 1
3 . ht i
x (R )
E5 (t).
 3.5. DECAY ESTIMATES 161

Combining all these estimates together we obtain the integral inequality

Z t
E5 (t) . E5 (0) + ht i2 E5 (t )3 dt .
0
From this and a standard continuity argument we see that if E5 (0) . for some
suffiicently small absolute constant > 0, then we have E5 (t) . for all t 0 for
which the classical solution u exists. Applying the Klainerman-Sobolev inequality,
this leads to decay bounds such as kt,x u(t)kL x
. hti1 . See [Sog] for a more
detailed and general treatment of this vector fields approach.
Exercise 3.44. Let d = 3. Show that the estimate (3.39) continues to hold
for the linear equation = 0 when
R R 5/3 < p < 7, but fails for p < 5/3 or p > 7.
(Hint: first obtain bounds for R |x|R |u(t, x)|p+1 dxdt for R a power of two by
using Strichartz estimates, and then sum in R. For the negative results, start with
a bump function initial datum (or a Gaussian) and rescale it as in Exercise 2.42.)
The estimate is also true at the endpoints p = 5/3 and p = 7 but requires a Lorentz
space refinement of the Strichartz estimates, observed in [KTao].
Exercise 3.45. For simplicity let us work with a global classical solution u :
RR3 C to the three-dimensional linear Schrodinger equation (so d = 3 and =
0). Define the two-particle field U : R R6 C by U (t, x, y) := u(t, x)u(t, y).
Show that U solves the six-dimensional linear Schrodinger equation. Apply (2.37)
to the solution U with the weight a(x, y) := |x y| (using limiting arguments as
necessary to deal with the fact that a is not smooth) and deduce another proof of
(3.42) in the linear case = 0. How does the argument change when one places a
defocusing nonlinearity in the equation?
Exercise 3.46 (Morawetz inequality for the wave equation). [Mor] Let u :
I R3 C be a classical solution to a three-dimensional defocusing NLW (thus
d = 3 and = 1), and let T be the associated stress-energy tensor. Using (3.34)
and the identity
jk pjk
Tjk = Re(xj uxk u) (|u|2 ) + |u|p+1
4 2(p + 1)
for the spatial component of the stress-energy tensor, establish the identity
Z Z
xj 0j / u|2
| p |u|p+1 1
t T dx = + (|u|2 ).
R 3 |x| R 3 |x| p + 1 |x| 2|x|
Integrate this in time and use the Hardy inequality (Lemma A.2) to establish the
Morawetz inequality
Z Z Z Z Z
/ u|2
| |u|p+1
dxdt + dxdt + |u(t, 0)|2 dt .p E[u]
I R3 |x| I R3 |x| I
where E[u] = E[u[t]] is the conserved energy; compare this with (3.38).
Exercise 3.47. Let u be a classical solution to a NLS. Verify the identity
Z Z
2 1 2
Epc [u(t), t] = t E[u(t)] t xj T0j (t, x) dx + |x| T00 (t, x) dx
Rd Rd 2
which connects the pseudoconformal energy to the ordinary energy and the pseudo-
stress-energy tensor. Use this to verify the evolution law
Z
dt(p pL2x )
t Epc [u(t), t] = |u(t, x)|p+1 dx
p+1 Rd
162 3. SEMILINEAR DISPERSIVE EQUATIONS

directly, without recourse to the pseudoconformal transformation. From this and

Gronwalls inequality, deduce the estimate
ku(t)kp+1
Lp+1
.d,p t2 kxu(0)k2L2x ,
x

in the defocusing, L2x -supercritical case = +1, p > pd and all t 6= 0, as well as the
estimate
Z t0 Z
ku(t)kp+1
Lp+1
. d,p,t0 t d(p1)/2
(kxu(0)k 2
Lx2 + |u(t, x)|p+1 dxdt),
x
0 Rd
in the defocusing, L2x -subcritical case = +1, p < pd and all t t0 > 0.
Exercise 3.48. Let f Ct Sx (I R3 ) for some time interval I. By repeating
the arguments used to deduce (2.57) from (2.56), derive the Klainerman-Sobolev
inequality
X X
1
(3.46) kt,x f (t)kL 3 . hti
x (R )
kt,x K1 . . . Km f (t)kL2x (R3 )
m3 K1 ,...,Km

for all t I, where K1 , . . . , Km ranges over all the vector fields t , xj , xj xk

xk xj , or txj + xj t .

3.6. Scattering theory

To know the road ahead, ask those coming back. (attributed to
Confucius)
The decay estimates of the preceding section give asymptotic control for global
solutions to NLS or NLW. It turns out in many cases, these estimates can be boot-
strapped to provide quite strong control on these solutions, in particular establishing
that they scatter to a linear solution. Intuitively, the reason for this is that if u(t)
decays to zero as t , then the nonlinearity |u(t)|p1 u(t) decays even faster,
and so the nonlinear component of the NLS or NLW equation will vanish asymp-
totically (in a relative sense), and thus (by Principle 1.37) we expect the evolution
to behave linearly as t . The main tool for making these heuristics rigorous
is the Duhamel formula ((3.22) or (3.23)), applied for various values of t and t0 ; the
arguments often bear some similarity with the Duhamel iteration arguments used to
establish local existence, though with some subtle differences. For instance, in the
local theory, large exponents p are more difficult to deal with than small exponents
(because they exacerbate the large values of the solution, which are the main source
of difficulty in closing a local iteration argument), but in the asymptotic theory, the
small exponents tend to be the most difficult35 (because they do not attenuate the
small values of the solution as much as the large exponents, and so the nonlinearity
does not decay as fast asymptotically). Since one needs to combine the local and
asymptotic theories to understand scattering, it should thus be unsurprising that
most scattering results only hold for exponents p that are neither too large nor too
small. (For instance, recall from our discussion of the exact solutions (3.18), (3.19)
that we do not expect scattering results when p 1 + 2d .) It is sometimes useful
to ensure that ones arguments are as scale-invariant as possible, as this can allow
one to treat both the local theory and the asymptotic theory in a unified manner.
35For similar reasons, the asymptotic theory sometimes gets easier when there are derivatives
in the nonlinearity, despite the fact that these derivatives can make the local theory significantly
harder.
 3.6. SCATTERING THEORY 163

We begin by discussing the scattering theory for NLS in the energy class Hx1 . 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 and t = , or between t = and t = +, but the theory is more
or less the same in each of these cases. We will also assume that the nonlinearity
is either Hx1 -subcritical or Hx1 -critical, so that we have a good Hx1 -wellposedness
theory (locally in time, at least).
A solution to the linear Schrodinger equation in this class takes the form
eit/2 u+ for some u+ Hx1 . We say that a global strong Hx1 solution u to the
nonlinear equation (3.1) with initial datum u(0) = u0 scatters in Hx1 to a solution
eit/2 u+ to the linear equation as t + if we have
ku(t) eit/2 u+ kHx1 0 as t +,
or equivalently (by using the unitarity of eit/2 )
keit/2 u(t) u+ kHx1 0 as t +.
In other words, we require the functions36 eit/2 u(t) converge in Hx1 as t +.
From the Duhamel formula (3.22) we have
Z t

eit/2 u(t) = u0 i eit /2 (|u(t )|p1 u(t )) dt
0
and so u scatters in Hx1 as t + if and only if the improper integral
Z
eit/2 (|u(t)|p1 u(t)) dt
0
is conditionally convergent in Hx1 , in which case the asymptotic state u+ is given
by the formula
Z
(3.47) u+ = u0 i eit/2 (|u(t)|p1 u(t)) dt.
0
Thus one can view the asymptotic state u+ as a nonlinear perturbation of the initial
state u0 . If we compare (3.47) with (3.22) and eliminate u0 we obtain the identity
Z

(3.48) u(t) = e it/2
u+ + i ei(tt )/2 (|u(t )|p1 u(t )) dt
t
which can be viewed as the limiting case t0 = + of (3.22).
Suppose that for every asymptotic state u+ Hx1 there exists a unique initial
datum u0 Hx1 whose corresponding Hx1 -wellposed solution is global and scatters
to eit/2 u+ as t +. Then we can define the wave operator + : Hx1 Hx1
by + u+ := u0 . Note that the uniqueness aspect of the Hx1 -wellposedness theory
ensures that the wave operator is injective. If it is also surjective - in other words,
if every Hx1 -wellposed solution is global and scatters in Hx1 as t +, we say that
we also have asymptotic completeness.
In general, the existence of wave operators is relatively easy to establish (as long
as the power p is not too small or too large, and especially if a smallness condition
is assumed), both in focusing and defocusing cases. The asymptotic completeness,
36It is instructive to write eit/2 u(t) = S (t)1 S(t)u , where S(t) : u 7 u(t) is the prop-
lin 0 0
agator for the nonlinear Schrodinger equation, and Slin (t) is the corresponding linear propagator.
Thus scattering is an assertion that the gap between S(t) and Slin (t) converges to something
bounded in Hx1 as t .
164 3. SEMILINEAR DISPERSIVE EQUATIONS

Forcing term
F(u)
nonlinearity [linear evolution]
F
(backwards inhomogeneous)

Solution Asymptotic state

u u
+
[linear evolution]
(homogeneous)

Figure 8. The iteration scheme used to construct a solution from

an asymptotic state at late times; it is essentially a backwards-in-
time version of the local existence scheme, but on an unbounded
time interval.

however, is a bit harder, is restricted to the defocusing case (since soliton solutions
clearly do not scatter to linear solutions), and requires the decay estimates. We
will not attempt a complete theory here, but just illustrate with a single example,
namely the cubic defocusing three-dimensional NLS (d = 3, p = 3, = +1). Note
that global wellposedness for Hx1 for this equation (in the subcritical sense) was
already established in Exercise 3.35.

Proposition 3.28 (Existence of wave operators). Let d = 3, p = 3, and

= +1. Then the wave operator + : Hx1 Hx1 exists and is continuous.

Proof. (Sketch) To construct the wave operator + , we need to evolve a state

at t = + to t = 0. We shall factor this problem into two sub-problems; first we
shall solve the asymptotic problem, getting from t = + to some finite time
t = T > 0, and then we will solve the local problem of getting from t = T
to t = 0. The latter problem will be an immediate consequence of the global
wellposedness problem, so we focus on the former. We shall use the same Duhamel
iteration method used to prove Proposition 3.19, but with (3.48) being used instead
of the usual Duhamel formula (3.22). Fix u+ Hx1 ; we will assume the bound
ku+ kHx1 A for some A > 0. From the Strichartz estimate (3.28) we have

keit/2 u+ kS 1 (RR3 ) .A 1.

We would like to make this norm not only bounded, but small, by restricting the
time variable. This is not possible at present because the S 1 norm contains some
components of L t type, which do not necessarily shrink upon restricting time. To
fix this we shall pass from S 1 to a smaller controlling norm; a convenient choice
here is the norm
kukS0 := kukL5t,x + kukL10/3 W 1,10/3 .
t x
 3.6. SCATTERING THEORY 165

From Sobolev embedding we have

keit/2 u+ kS0 (RR3 ) . keit/2 u+ kL5 W 1,30/11 (RR3 ) + keit/2 u+ kL10/3 W 1,10/3 (RR3 )
t x t x

. keit/2 u+ kS 1 (RR3 )
.A 1.
Let > 0 be a small absolute constant to be chosen later. If we set T = T (u+ )
large enough, we see from monotone convergence that
keit/2 u+ kS0 ([T,+)R3 ) .
We now solve (3.48) in the spacetime slab [T, +) R3 by iteration, keeping
the iterates bounded in S 1 ([T, +) R3 ) and small in S0 , and the nonlinearity
10/7 1,10/7
|u|p1 u small in Lt Wx ([T, +)R3 ). This constructs a unique solution u
S 1 ([T, +) R3) to (3.48), which can be shown to be a strong Hx1 solution to (3.1)
in this interval by a variant of Exercise 3.10. Using the global Hx1 -wellposedness
theory, one can then extend this solution uniquely to S 1 ([0, +) R3 ), and in
particular u will take some value u0 = u(0) Hx1 at time t = 0. This gives
existence of the wave map; continuity can be established by concatenating the
continuity given from the above iteration scheme in the interval [T, +) with the
continuity arising from the global wellposedness in the interval [0, T ]; note that the
time T can be chosen to be uniform under small Hx1 perturbations in u+ thanks to
the Strichartz estimates. The uniqueness can be made to be unconditional (in the
category of strong Hx1 solutions) by arguing as in Exercise 3.17. 
Remark 3.29. The above argument shows that (perhaps unintuitively) it is
in fact easier to evolve from an asymptotic state at t = + to a large finite time
t = T , than it is to evolve from t = T down to t = 0, as the former does not
even require energy conservation or the defocusing sign of the nonlinearity. The
reason for this is that in the asymptotic regime t +, the asymptotic state is
so dispersed that the nonlinear effects are extremely weak; it is only at time T and
below that the solution reaches sufficient levels of concentration that one must start
paying more serious attention to the nonlinearity.
Now we establish asymptotic completeness. For pedagogical purposes we shall
split the argument into three parts. First we begin with a conditional result, that
shows that asymptotic completeness is implied by a certain spacetime bound; this
is a purely perturbative argument that does not require any decay estimates. Then,
we show that this rather strong spacetime bound is implied by a seemingly weaker
spacetime bound. Finally, we use the decay estimates of the previous section to
establish that spacetime bound.
Proposition 3.30 (Spacetime bound implies asymptotic completeness). Let
d = 3, p = 3, and = +1. Suppose that there exists a bound of the form
(3.49) kukS 1 (RR3 ) .ku0 kH 1 1
x

for all Hx1 -wellposed solutions to (3.1) (thus we assume that the nonlinear equation
obeys the same type of global Strichartz estimate as the linear equation). Then the
wave operator + is surjective from Hx1 to Hx1 , and the inverse 1+ is continuous.
(In conjunction with Proposition 3.28, this implies that + is a homeomorphism
from Hx1 to itself.)
166 3. SEMILINEAR DISPERSIVE EQUATIONS

Proof. We shall demonstrate the surjectivity here, and leave the continuity
to an exercise. We need to show that for any u0 Hx1 , the global Hx1 -wellposed
solution u to (3.1) scatters in Hx1 ; by the precedingR discussion, this is equiva-
lent to the conditional convergence of the integral 0 eit/2 (|u(t)|2 u(t)) dt in
Hx1 . By Strichartz estimates (e.g. (3.28)), it will suffice to show that |u|2 u lies
in N 1 (R R3 ). But from the Leibnitz rule and Holders inequality, followed by
Sobolev embedding we have
1
X
k|u|2 ukN 1 (RR3 ) . kk (|u|2 u)kL10/7 (RR3 )
t,x
k=0
1
X
. k|u|2 |k u|kL10/7 (RR3 )
t,x
k=0
. kuk2L5 kukL10/3W 1,10/3 (RR3 )
t,x t x

. kuk3S 1 (RR3 )
and the claim follows by (3.49). 
Proposition 3.31 (Weak spacetime bound implies strong spacetime bound).
Let d = 3, p = 3, and = +1. Suppose that there exists a bound of the form
(3.50) kukLqt,x(RR3 ) .ku0 kH 1 1
x

for all Hx1 -wellposed solutions to (3.1) and some fixed 10/3 q 10. Then we
have (3.49).
Note that Sobolev embedding shows that the Lqx norm is controlled by the S 1
norm. The S 1 norm is ostensibly a stronger norm as it also controls one derivative
of the solution, but the point is that Strichartz estimates will allow one to control
this stronger norm by the weaker norm (and the energy). This bootstrapping
phenomenon is typical for any subcritical equation (reflecting a certain amount
of room in the iteration argument); for critical equations, the situation is more
delicate as the relevant Strichartz norm is now scale-invariant and thus can only
be controlled by other scale-invariant quantities; see Exercise 3.51. One can also
combine this result with persistence of regularity results such as Proposition 3.11,
giving in particular the bound
kukCt0 Hxs (RR3 ) .s,ku0 kH 1 ku0 kHxs (R3 )
x

for any s 0 for which the right-hand side is finite.

Proof. Let u be an Hx1 -wellposed solution to (3.1). We shall apply a pertur-
bative argument; to do this, we need the solution u to be made small in some sense.
This shall be accomplished by partitioning the time axis37.
Let = (ku0 kHx1 ) > 0 be a small number to be chosen later. Using (3.50), we
can divide the time axis R into O,q,ku0 kH 1 (1) intervals I, such that on each such
x

37This is very similar to how one iterates a local existence result to a global one, as in Figure
7. A key difference is that the time intervals considered here can be arbitrarily large or even
infinite. In practice, this means that we are no longer permitted to use Holder in time (except
perhaps on some exceptionally short intervals), as we generally cannot afford to lose a power of
the length of the time interval.
 3.6. SCATTERING THEORY 167

interval we have
(3.51) kukLqx(IR3 ) .
Now fix one of these intervals I, say I = [t0 , t1 ]. From (3.28) we have
kukS 1(IR3 ) . ku(t0 )kHx1 (R3 ) + k|u|2 ukN 1 (IR3 ) .
From energy conservation we have ku(t0 )kHx1 (R3 ) = Oku0 kH 1 (1). Now we argue as
x
10/7 1,10/7
in the proof of Proposition 3.30. Estimating the N 1 norm by the Lt Wx
norm and using the Leibnitz rule and Holder inequality, we see that
1
X
k|u|2 ukN 1 (IR3 ) . k|u|2 |k u|kL10/7 (IR3 )
t,x
k=0
. kuk2L5 3) kukL10/3W 1,10/3 (IR3 )
t,x (IR t

. kuk2L5 (IR3 ) kukS 1 (IR3 ) .

t,x

1
Now from the definition of S and Sobolev embedding we have
kukLrt,x (IR3 ) . kukS 1 (IR3 )
for all 10/3 r 10. Interpolating this with (3.51) we obtain
kukL5t,x (IR3 ) . kuk1
S 1 (IR3 )
for some 0 < < 1 depending on q. Combining all these estimates we obtain
kukS 1 (IR3 ) . Oku0 kH 1 (1) + kuk1
S 1 (IR3 ) .
x

If we choose sufficiently small depending on ku0 kHx1 , then standard continuity

arguments (see Exercise 1.21) yields
kukS 1 (IR3 ) = Oku0 kH 1 (1).
x

Summing this over all of the intervals I we obtain (3.49) as desired. 

It thus remains to establish the spacetime bound (3.50). In the case of spher-
ically symmetric solutions, one can combine the ordinary Morawetz inequality
(3.39), which in this case gives
Z Z
|u(t, x)|4
dxdt .ku0 kH 1 1,
R R3 |x| x

with the radial Sobolev inequality (Exercise A.19), which when combined with the
conservation of mass and energy give
ku(t, x)|x|kL 3 .ku k
t Lx (RR ) 0 H1
1.
x

Multiplying the two gives

kukL5t,x (RR3 ) .ku0 kH 1 1,
x

which is of the desired form (3.50) with q = 5. Note how the Morawetz inequality
provides the decay near the origin, while the radial Sobolev inequality provides the
decay away from the origin. In the non-radial case, we cannot run this argument
so easily (though see [Bou6]); however the interaction Morawetz inequality (3.43)
yields (3.50) immediately (with q = 4).
The above types of arguments are known to give scattering results in Hx1 for
defocusing NLS equations which are strictly between the Hx1 -critical and L2x -critical
168 3. SEMILINEAR DISPERSIVE EQUATIONS

powers; see [Caz2]. A scattering theory at the Hx1 -critical endpoint (based upon a
spacetime bound such as (3.50)) has recently been established but is significantly
more difficult; see Chapter 5. The Hx1 -scattering theory for the L2x -critical equation
remains open, even in the spherically symmetric defocusing case. Similar remarks
also apply to the NLW, but with the role of the L2x -critical exponent now played
1/2
by the Hx -critical (conformal) exponent.
For NLS equations below the L2x -critical exponent, no scattering theory is
known in Hx1 , but one can extend the range of exponents for which a scatter-
ing result is known by assuming more spatial decay on the solution. For instance,
one can work in the pseudoconformal space
:= {u0 Hx1 (Rd ) : xu0 L2x (Rd )},
as one can now utilise the pseudoconformal decay laws for such initial data (such as
those in Exercise 3.47). It turns out that the exponent p still needs to be above a
certain threshold in order for that decay law to be strong enough to give scattering;
more precisely, if we have a defocusing NLS with

2 + d + d2 + 12d + 4 4
<p1+
4d d2
and u0 , then there is a global Hx1 -wellposed solution u, and eit/2 u(t) con-
verges in to some asymptotic state u+ . See [Caz2], [TVZ]. On the other
hand, for NLS equations in which the power p is less than or equal to 1 + d2 , the
asymptotic effects of the nonlinearity are not negligible, and it is known that the
solution does not in general scatter to a free solution; see Section 3.8.
The pseudoconformal transformation is a useful tool for analyzing the asymp-
totic behaviour of NLS, because it swaps the asymptotic regime t + with
the local regime t 0+ (though at the possible cost of introducing a singular-
ity at t = 0). This transformation should also (heuristically at least) swap the
initial datum with its Fourier transform, or something resembling its Fourier trans-
form; see Exercise 2.28. The Fourier transform swaps Hx1 to the weighted space
Hx0,1 := {u0 : hxiu0 L2x }, and so one might expect to be able to use this transfor-
mation to somehow intertwine the Hx1 theory with an Hx0,1 theory. A sample result
is as follows.
Proposition 3.32. Consider the two-dimensional defocusing cubic NLS (thus
d = 2, p = 3, = +1, and the equation is L2x -critical). Let u0 Hx0,1 . Then there
exists a global L2x -wellposed solution to (3.1), and furthermore the L4t,x (R R3 )
norm of u0 is finite.
The L4t,x bound is sufficient to yield a scattering result in L2x ; see Exercise 3.54.
In contrast, for Hx1 data, no scattering result is known (the Morawetz inequalities
do some decay here, but it is not scale-invariant), while for L2x data, not even global
existence is known (unless the mass is small).
Proof. We shall use an argument from [BC]. By time reversal symmetry and
gluing arguments we may restrict attention to the time interval [0, +). Since u0
lies in Hx0,1 , it also lies in L2x . Applying the L2x wellposedness theory (Proposition
3.17) we can find an L2x -wellposed solution u S 0 ([0, T ]R3) on some time interval
[0, T ], with T > 0 depending on the profile of u0 . In particular the L4t,x ([0, T ] R3)
norm of u is finite. Next we apply the pseudoconformal law (which is exact in the
 3.6. SCATTERING THEORY 169

4 4 4 4
u in L L on [T, ]

8
v in L L on [1/T, ]

8
u(T) has finite pcenergy v(1/T) has finite energy
1
[pc conservation] [pc transform] [global theory in H ]
4 4 4 4
u in L L on [0,T] v in L L on [0, 1/T]
2
[local theory in L ]

0,1
u(0) in H

Figure 9. The scheme used to establish a global L4t,x bound from

Hx0,1 initial data. Notice how the original and pseudoconformal
viewpoints together form a coordinate chart for the compactified
time interval [0, +], thus reducing a global problem to two local
ones.

L2x -critical case, and can be justified by the usual limiting arguments) to conclude
that Z
1
Epc [u(T ), T ] = Epc [u0 , 0] = |xu0 |2 dx <
2 R2
since u0 Hx0,1 .
We have obtained a solution from t = 0 to t = T . To go all the way to
t = + we apply the pseudoconformal transformation (Exercise 2.28) at time
t = T , obtaining an initial datum v(1/T ) at time 1/T by the formula
1 2
v(1/T, x) := u(T, T x)eiT |x| /2 .
i/T
From (3.44) we see that v has finite energy:
Z Z
1 1
|v(1/T, x)|2 dx + |v(1/T, x)|4 dx = Epc [u(T ), T ] < .
2 R2 2 R2
Also, the pseudoconformal transformation conserves mass and hence
Z Z Z
|v(1/T, x)|2 dx = |u(T, x)|2 dx = |u0 (x)|2 dx < .
R2 R2 R2

We thus see that v(1/T ) has a finite Hx1 norm. We can thus use the global Hx1 -
wellposedness theory (from Exercise 3.35) backwards in time to obtain an Hx1 -
wellposed solution v S 1 ([0, 1/T ] R2 ) to the equation (3.16), which in this
case is identical to the original NLS: it v + 12 v = |v|2 v. In particular, v
L4t,x ([0, 1/T ]R2). We now invert the pseudoconformal transformation, which now
defines the original field u on the new slab [1/T, ) R2 . From Exercise 2.38 we
see that the L4t,x ([1/T, ) R2 ) and Ct0 L2x ([1/T, ) R2 ) norms of u are finite.
This is enough to make u an L2x -wellposed solution to NLS on the time interval
[1/T, ); for v classical this is an immediate consequence of Exercise 2.28, and for
general v S 1 ([0, 1/T ] R2 ) the claim follows by a limiting argument using the
Hx1 -wellposedness theory. Gluing together the two intervals [0, 1/T ] and [1/T, ),
170 3. SEMILINEAR DISPERSIVE EQUATIONS

we have obtained a global L4t,x ([0, +) R2 ) solution u to (3.1) as desired. We

summarise the above argument in Figure 9. 
Remark 3.33. One can go through the above argument and extract an explicit
bound on the global L4t,x norm of the solution u, but it depends on the profile of the
initial datum u0 and not just on its Hx0,1 norm (as this is what determines how small
T is). Indeed, if one could obtain a bound depending only in the Hx0,1 norm then
the scaling invariance and a limiting argument would allow one to replace Hx0,1 with
L2x , which would lead to the (still open) result that one has global wellposedness
and scattering in L2x for this equation. The above argument can also be generalised,
linking a wellposedness theory in Hxs to a scattering theory in H 0,s for any s 0
and any L2x -critical equation; see [BC].
Observe how in the above argument, the pseudoconformal transformation was
used to convert an asymptotic time horizon t = + to a finite time horizon t = 0,
thus allowing one to use local theory to obtain asymptotic control of the solu-
tion. There is a somewhat similar trick for wave equations known as conformal
compactification, in which one applies a conformal transformation of Minkowski
spacetime to a pre-compact Lorentzian manifold (the Penrose diamond). In
one dimension d = 1, this compactification is especially simple in null coordinates
u := t + x, v := t x, as it is given simply by (u, v) 7 (tan1 u, tan1 v), thus map-
ping R1+1 to the diamond-shaped region {(t, x) R1+1 : |t + x|, |t x| < /2}. If
the equation is of a suitable type (it typically must obey some sort of null condi-
tion, or the nonlinearity must be sufficiently high order), then this transformation
does not introduce any severe singularities at the boundary of the diamond (cf.
(3.16) in the case p < pL2x ), and one can use local theory on the Penrose diamond
to obtain a transformed solution on the entire diamond (if the datum is sufficiently
small, smooth, and decaying), and then by inverting the conformal compactifica-
tion one obtains a global solution in Minkowski space. Typically the transformed
solution extends to the boundary of the Penrose diamond and beyond, which often
leads to scattering-type behaviour for the original solution. See [Chr].
Several of the above methods can also be used to establish various scattering
results for NLW; the conformal conservation law, Morawetz estimates, the vector
fields method, and the conformal compactification methods are particularly useful.
Because of finite speed of propagation, one can often reduce the case of compactly
supported data. We will not discuss these results here, except in the energy-critical
case which we treat in see Section 5.1, and refer the reader to [Stra], [GV6], [BZS],
[GV5], [Nak3], [Hid].
Exercise 3.49. Complete the proof of Proposition 3.28. (Full details can also
be found in [Caz2].)
Exercise 3.50. Establish the continuity component of Proposition 3.30. (One
may need to divide the time axis into intervals on which certain spacetime norms are
small, in order that the perturbative argument can apply to give local continuity.
Then concatenate the results to obtain global continuity.)
Exercise 3.51. Suppose one replaces the Hx1 -subcritical cubic NLS in Propo-
sition 3.31 with the Hx1 -critical quintic NLS (so d = 3, p = 5, = +1). Show that
one can still prove this Proposition if one fixes q = 10 (this is the unique value of
q which is invariant under the scaling symmetry of the equation).
 3.7. STABILITY THEORY 171

Exercise 3.52. Suppose one is working with a global Hx1 (R3 )-wellposed so-
lution u to either the cubic or quintic three-dimensional NLS (with R either sign of
1
nonlinearity). Suppose it is known that the potential energy p+1 R 3 |u(t, x)|p+1 dx
goes to zero as t . Conclude that the solution scatters in Hx1 to an asymptotic
state eit/2 u+ . (This is yet another affirmation of Principle 3.24.)
Exercise 3.53 (Blowup criterion for Hx1 -critical NLS). Suppose that u
0
Ct,loc Hx1 ([0, T ) R3 ) is a strong Hx1 solution to quintic NLS (so d = 3 and p = 5)
which cannot be continued beyond a finite time T as a strong solution. Show that
the L10 3
t,x ([0, T )R ) norm of u is infinite. (Hint: argue by contradiction and obtain
1 3
an S ([0, T ) R ) bound on u. Conclude that for times t close to T , both the
linear and nonlinear evolution of u(t) will be small in L10 3
t,x ([t, T ) R ), and hence
10 3
for Lt,x ([t, T + ) R ) for some > 0, contradicting the hypothesis that T is the
maximal time of existence.)
Exercise 3.54. Consider the two-dimensional defocusing cubic NLS (thus d =
2, p = 3, = +1). Show that if a global L2x -wellposed solution u is known to have
finite L4t,x (R R2 ) norm, then eit/2 u(t) converges in L2x to an asymptotic state
u+ L2x as t +.

3.7. Stability theory

True life is not lived where great external changes take place - where
people move about, clash, fight, and slay one another - it is lived
only where these tiny, tiny, infinitesimally small changes occur.
(Leo Tolstoy, Why Do People Intoxicate Themselves?)
The differential equations that one studies in mathematics, such as NLS and
NLW, often arise from physics as simplified models for more complicated systems.
In reality, the actual equations that govern a physical system will not evolve by
these model equations exactly, but will contain some additional terms. For sake
of discussion let us fix the model equation to be the NLS (3.1). Instead of solving
NLS, the true system may be governed by a field u which obeys a perturbed NLS
1
(3.52) it u + u = |u|p1 u + e; u(t0 ) = u0 + e0
2
where the forcing term e = e(t, x) and the initial datum error e0 = e0 (x) are small,
and possibly depending on u and on some external forces38. It is thus of interest to
develop a stability theory for equations such as NLS, which would guarantee that the
solution to a perturbed NLS does not deviate too far from the solution to the actual
NLS if e and e0 are small some suitable norms. Note that this would generalise
the property of continuous dependence of the data, which is already given by the
wellposedness theory and corresponds to the special case e = 0. It also generalises
the uniqueness theory, which can be viewed as the case e = e0 = 0. A strong
stability theory lends confidence as to the robustness of the results obtained for the
model equation. Conversely, if a PDE is known to be very unstable then this would
38The field of stochastic partial differential equations studies such equations with the as-
sumption that e is some stochastic field, e.g. Gaussian white noise. These random fluctuations
often serve to regularise the behaviour of the equation and keep it well-behaved even when the
deterministic equation is not known to be wellposed. However, we shall focus on the determin-
istic theory in which e is fixed, which is the case needed for applications such as construction of
solutions via approximate solutions.
172 3. SEMILINEAR DISPERSIVE EQUATIONS

cast doubt on the ability of that PDE to accurately simulate (either numerically or
theoretically) a real-life system, except perhaps in some stochastic sense.
A stability theory is also useful to have in the analysis of PDE. It opens up a
very useful strategy for constructing solutions u to an equation such as NLS, by
first constructing an approximate solution u, for which it u + 12 u is very close
to |u|p1 u and u(t0 ) u0 . In other words, an approximate solution to NLS is
nothing more than an exact solution to the perturbed NLS (3.52) for some small e
and e0 . Stability theory can then let us pass from the approximate solution u to a
nearby exact solution u to the unperturbed NLS. This approach is quite powerful,
because it is much easier to construct approximate solutions than exact solutions,
for instance by asymptotic expansions39, or by omitting certain terms from an
equation that one believes to be negligible and then solving the reduced equation;
see Section 3.8 below. To give another example, while the superposition of two
solutions to a nonlinear equation will not in general yield another solution to that
equation, such a superposition is often an approximate solution to the equation if the
two component solutions are sufficiently separated, either in space or frequency.
This strategy can be used for instance to construct multisoliton solutions, and is
also the main reason why the induction on energy strategy that we shall introduce
in Chapter 5 is so powerful.
Fortunately, any equation with a good wellposedness theory is also likely to
have a good stability theory, by modifying the arguments used to prove wellposed-
ness suitably; indeed the stability theory is in many ways the culmination of the
perturbation theory methods. The main trick (which we have already seen with the
uniqueness and continuity theory) is to look at the difference equation between the
approximate solution u and the exact solution40, and then solve this difference equa-
tion using the same types of techniques and estimates used for the wellposedness
theory. Specifically, if we set u = u + v, then v solves the difference equation
1
(3.53) it v + v = (|u + v|p1 (u + v) |u|p1 u) + e; v(t0 ) = e0 .
2
Thus the initial datum of v is small. As for the nonlinearity, we can use Taylor
expansion to expand
(|u+v|p1 (u+v)|u|p1u)+e = O(|u|p1 |v|)+O(|u|p2 |v|2 )+. . .+O(|v|p )+O(|e|).
assuming for simplicity that we are in the algebraic case where p is an odd integer.
In practice, if e and e0 are both small, then we expect v to be small also, and
the dominant terms in the nonlinearity will be the terms O(|u|p1 |v|) which are
linear in v. These terms can be dealt with for short times by iterative arguments
based on the Duhamel formula, as well as estimates such as Strichartz estimates; for

39In particular, the theory of nonlinear geometric optics proceeds in this fashion, constructing
solutions to an equation by first creating an ansatz consisting of an asymptotic series with certain
amplitude and phase parameters. One then solves for these amplitudes and phases in order to
make the partial sums of this series an approximate solution to the original equation, and then uses
some stability theory to pass to an exact solution. These methods are very useful in constructing
large classes of interesting solutions to many PDE, though they tend to require the initial data to
be of a special form and are unsuited for the Cauchy problem with generic Hxs initial data. Due
to limitations of space we will not be able to discuss this important technique in this text.
40This assumes that the exact solution u exists for at least as long as the approximate solution
u. In practice one can establish this by a continuity argument or by a suitable iteration of the
wellposedness theory.
 3.7. STABILITY THEORY 173

Error in approximate Difference in

solution forcing terms
Nonlinearity
e F(u+v) F(u)
F

Exact solution
Linear evolution u
(inhomogeneous)

Initial error Difference in

solutions
e
0 Linear evolution v
(homogeneous)

Figure 10. The difference scheme for an approximate perturba-

tion u = u + v to an exact solution u. This scheme can also be
reversed to convert an approximate solution u to an exact solution
u = u v (by replacing F (u + v) F (u) with F (u) F (u v)).

longer times, one can use energy methods41, combined with tools such as Gronwalls
inequality, to try to keep control of the solution.
To illustrate the method, we shall consider asymptotics of one-dimensional
defocusing NLS in the short range case p > 3. (The critical range case p = 3
and the long range case p < 3 are significantly more interesting, but also more
difficult technically.) Applying the pseudoconformal transformation as in (3.16),
we obtain the equation
1 1
(3.54) it v + xx v = (5p)/2 |v|p1 v
2 t
for 3 < p < 5, which is obtained from the one-dimensional defocusing NLS via the
pseudoconformal transformation (see (3.16)). To construct solutions near t = 0,
we first omit the dispersive term 12 xx v (using the intuition that this term will
1
be dominated by the singular nonlinearity t(5p)/2 |v|2 v for very small times t) and
solve the simpler equation
1
(3.55) it v =|v|2 v.
t(5p)/2
This equation just the ODE (3.17), and can be solved explicitly as
2
2
|(x)|2 t(p3)/2
(3.56) v(t, x) = ei p3 (x)

41In some cases, when the exact solution u is an explicit form such as a soliton, one can use
more advanced spectral analysis of the linearised equation it v + 12 v = O(|u|p1 |v|) to obtain
long-time control of the solution; this is an important tool in the theory of stability of solitons and
multisolitons. However, such spectral methods are currently unavailable for more general classes
u of solution.
174 3. SEMILINEAR DISPERSIVE EQUATIONS

for any complex-valued function (x), and 0 < < 1 is a small parameter we have
introduced to allow v to be small (compare with (3.18)). Thus, if is small, we
expect the original PDE (3.54) to have solutions which are approximately of the
form (3.55). This can be established as follows.
Proposition 3.34. Let Sx (R) and 0 < 1. If is sufficiently small
depending on , then we have a solution v to (3.54) on the slab (0, 1) R obeying
the bounds
kv(t) v(t)kHx1 . t
for all 0 < t < 1, where v was defined in (3.56).
Proof. To construct v, we use the ansatz v = v + w. Subtracting (3.55) from
(3.54), we see that w needs to solve the equation
1 1
it w + xx w = (5p)/2 (F (v + w) F (v)) xx v
2 t
p1
where F (z) := |z| z. We set initial datum w(0) = 0, and write the equation in
integral form as w = (w), where is the nonlinear operator
Z t
1
(w) = )(5p)/2
(F (v(t ) + w(t )) F (v(t ))) xx v(t ) dt .
0 (t
One can use energy estimates to verify that is a contraction on the set {tu :
kukCt0Hx1 ([0,1]R) . }, if is sufficiently small depending on ; we leave this as
an exercise. The claim now follows from the contraction mapping principle. 
Informally, the above proposition gives the approximation
2
2
|(x)|2 t(p3)/2
v(t, x) = ei p3 (x) + l.o.t.
for 0 < t < 1, where the lower order terms go to zero in a suitable sense as t 0.
Inverting the pseudoconformal transformation, one obtains
1 i|x|2 2 2
(3.57) u(t, x) = exp( + |(x/t)|2 t(p3)/2 )(x/t) + l.o.t.
(it)d/2 2t p3
for 1 < t < , where the lower order terms go to zero in a suitable sense as t
+. A similar argument applied to the linear Schrodinger equation (or using the
fundamental solution) allows one to construct a solution ulin to the linear equation
with the asymptotics
1 i|x|2
ulin (t, x) = d/2
exp( )(x/t) + l.o.t.
(it) 2t
Because we are in the short-range case p > 3, we can thus conclude that u(t)ulin(t)
converges to zero in certain norms (for instance, it converges in Hxs (R) for any s).
This suggests that the short-range case, one has scattering, at least for certain types
of initial data. In the critical-range case p = 3 or the long-range case p > 3, it turns
out that one can still construct solutions to NLS of the form (3.57); the arguments
are similar though the singular nature of (3.54) now presents some delicate issues
(cf. Exercise 1.19); see [Oza], [GO], [HN], [CCT] for some resolutions of this issue.
These solutions fail to scatter to a solution to the linear Schrodinger equation in any
Hxs norm; thus long-range and critical-range equations do not exhibit scattering to
the linear solution (this was first observed in [Gla]). However one can still hope to
 3.7. STABILITY THEORY 175

establish a modified scattering result, in which the approximating solution is not a

linear solution, but rather a phase-shifted linear solution; see the above references.
Next, we illustrate how Gronwall type inequalities can be used to obtain sta-
bility for longer times than a simple iteration method (such as that given above)
would give. The time interval on which one has non-trivial control is only extended
by a logarithmic factor, but this is sometimes sufficient for applications. It would
be of great interest to derive stability estimates on even longer intervals, perhaps
by adapting the theory of Nekhoroshev stability from ODE, but this seems to be a
difficult task (see [BK]).
Proposition 3.35. Let Sx (R) and let 0 < 1 be a small number. Then
there exists a time T log1/3 1 and a strong Hx1 solution u Ct0 Hx1 ([0, T ] R)
to the small dispersion NLS
2
(3.58) it u + xx u = |u|2 u; u(0) =
2
such that ku ukCt0 Hx1 ([0,T ]R) . , where
2
u(t, x) := ei|(x)| t (x)
is the explicit solution to the ODE
it u = |u|2 u; u(0) = .
Proof. From (a rescaled version of) Proposition 3.23 we know that a strong
Hx1 solution u to (3.58) exists globally in time42 Writing u = u + w, we see that w
solves the equation
2
it w + 2 xx w = (|(u + w)|2 (u + w) |u|2 u) xx u; w(0) = 0
2
which we write in Duhamel form as
Z t
2 2
w(t) = ei (tt )xx [(|(u + w)|2 (u + w) |u|2 u) xx u](t ) dt .
0 2
2
We take Hx1 norms of both sides, and use the fact that ei (tt )xx is bounded in
Hx1 , to obtain
Z t
2
kw(t)kHx1 k|(u + w)|2 (u + w)(t ) |u|2 u(t )kHx1 + kxx u(t )kHx1 dt .
0 2
A direct computation shows that
ku(t )kHx1 . ht i; kxx u(t )kHx1 . ht i3
while a computation using the algebra property of Hx1 (see Lemma A.8) gives
k|(u + w)|2 (u + w)(t ) |u|2 u(t )kHx1 . kw(t )kHx1 (ku(t )kHx1 + kw(t )kHx1 )2 .
Putting this all together, we obtain the bound
Z t
kw(t)kHx1 . 2 hti4 + hti2 kw(t )kHx1 + kw(t )k3Hx1 dt .
0

42This global wellposedness is convenient for the argument, but not absolutely necessary;
the energy bounds we obtain in the proof, combined with the local Hx1 wellposedness theory,
are sufficient (via a standard continuity argument) to construct the solution u on the given time
interval [0, T ].
176 3. SEMILINEAR DISPERSIVE EQUATIONS

[local theory]
(small norm)
2 . 1,6/5
e in L W ~ 1 ~ 0 .1
u in S u in C H
~ ~ 2 . 1,6/5
(small norm) F(u+v)F(u) in L W
(large norm) (large norm)
[Leibnitz, Holder,
Net forcing
. Sobolev] 10 . 1,30/13
~
term in N1 u in L W
(small norm) (small norm)
[Strichartz estimates]

1 .1 10 . 1,30/13 10 10
v(0) in H v in S v in L W v in L L
(large norm) (large norm) (small norm) [Sobolev]
(small norm)

10 . 1,30/13
v lin in L W

(small norm)

Figure 11. The scheme for estimating the difference v = u u

in Lemma 3.36; it is thus a rather complex variation of the usual
Strichartz iteration loop for this equation (see Figure 5).

If is sufficiently small depending on t, a continuity argument then gives

kw(t)kHx1 . 2 hti4 exp(Chti3 )

for all 0 < t log1/3 1 (cf. what one would obtain by Gronwalls inequality by
dropping the nonlinear term kw(t )k3H 1 ), and the claim follows. 
x

In the next section we will use this proposition to obtain some illposedness
results for NLS.
Our final example of a stability theory result comes from the defocusing energy-
critical three-dimensional NLS (d = 3, p = 5, = +1). We shall show that Hx1 -
wellposed solutions to this equation are stable as long as the L10 t,x norm stays
bounded. We first state a preliminary result in which we assume that a certain
spacetime norm on the solution is small.

Lemma 3.36 (Short-time perturbations). [CKSTT11] Let d = 3, p = 5, =

+1. Let I be a compact interval, and let u be a field on I R3 which is a near-
solution to NLS in the sense that
1
(3.59) (it + )u = |u|4 u + e
2
for some field e. Suppose that we also have the energy bound

kukL Hx1 (IR3 ) E

t

for some E > 0. Let t0 I, and let u(t0 ) be close to u(t0 ) in the sense that

(3.60) ku(t0 ) u(t0 )kH 1 E

x
 3.7. STABILITY THEORY 177

(large norm) (large norm) (large norm) (small norm)

0 .1 ~ 1 ~ 10 . 1, 30/13 ~ 10 . 1, 30/13
~ u in S u in L W
u in C H u in L W
[subdivide time]
[local theory]

~ 10 10 ~ 10 10
u in L L [subdivide time] u in L L
(large norm) (small norm)

Figure 12. The additional time decompositions necessary in or-

der to deduce Lemma 3.37 from Lemma 3.36.

for some E > 0. Assume also that we have the smallness conditions
(3.61) kukL10 W 1,30/13 (IR3 ) 0
t x

i(tt0 )/2
(3.62) ke (u(t0 ) u(t0 ))kL10 W 1,30/13 (IR3 )
t x

(3.63) kekL2W 1,6/5

t x

for some 0 < < 0 , where 0 is some constant 0 = 0 (E, E ) > 0.

We conclude that there exists a solution u to (3.1) on I R3 with the specified
initial datum u(t0 ) at t0 obeying the bounds
(3.64) ku ukS 1 (IR3 ) . E
(3.65) kukS 1 (IR3 ) . E + E
(3.66) ku ukL10 3 . ku uk 10
t,x (IR ) L W
1,30/13
(IR3 )
.
t x

(3.67) k(it + )(u u)kL2 W 1,6/5 (IR3 ) . .

t x

Note that u(t0 ) u(t0 ) is allowed to have large energy, albeit at the cost of
forcing to be smaller, and worsening the bounds in (3.64). From Strichartz esti-
mates and (3.60) we see that the hypothesis (3.62) is redundant if one is willing to
take E = O().
We leave the proof of Lemma 3.36 to the exercises. One can amplify this lemma
to deal with the more general situation of near-solutions with finite but arbitrarily
large L10
t,x norms.

Lemma 3.37 (Long-time perturbations). [CKSTT11] Let d = 3, p = 5, =

+1. Let I be a compact interval, and let u be a field on I R3 which obeys the
bounds
(3.68) kukL10 3 M
t,x (IR )

and
(3.69) kukL Hx1 (IR3 ) E
t

for some M, E > 0. Suppose also that u is a near-solution to NLS in the sense that
it solves (3.59) for some e. Let t0 I, and let u(t0 ) be close to u(t0 ) in the sense
that
ku(t0 ) u(t0 )kH 1 E
x
178 3. SEMILINEAR DISPERSIVE EQUATIONS

for some E > 0. Assume also that we have the smallness conditions,

(3.70) kei(tt0 )/2 (u(t0 ) u(t0 ))kL10 W 1,30/13 (IR3 )

t x

kekL2 W 1,6/5 (IR3 )

t x

for some 0 < < 1 , where 1 is a small constant 1 = 1 (E, E , M ) > 0. We

conclude there exists a solution u to (3.1) on I R3 with the specified initial datum
u(t0 ) at t0 , and furthermore

ku ukS 1 (IR3 ) .M,E,E 1

kukS 1(IR3 ) .M,E,E 1
ku ukL10 3 . ku uk 10
t,x (IR ) L W
1,30/13
(IR3 )
.M,E,E .
t x

Again, we leave the details to the exercises. This stability lemma is quite
powerful; it shows that approximate solutions can be adjusted to become exact
solutions even when the energy of both initial data and their difference are large,
as long as the approximate solution is bounded (but not necessarily small) in L10 t,x
norm, and the error e is very small. It will play an important role in the large
energy theory of this equation in Chapter 5.
In the preceding examples of stability theory, we approximated an exact solu-
tion u by an explicit approximate solution u. In some cases, most notably in the
stability theory of solitons and multisolitons, it is better to approximate u by a par-
tially explicit approximate solution, which involves some free parameters that one
has some freedom to choose in order to make the analysis of the error terms as easy
as possible. For instance, if considering perturbations u of a soliton solution such as
eit Q(x), the ansatz u = eit Q(x)+w turns out to not be very effective (the bounds
on w will grow exponentially in time if one applies perturbation theory naively).
Instead, a better procedure is to perform an ansatz u = eit +(t) Q(x x(t)) + w,
where : R R and x : R Rd are parameters that one can choose. Typically,
one chooses these parameters in order to obtain some moment conditions on w
(for instance, one could try to force w to be orthogonal to functions such as iQ or
Q), which can improve the behaviour of the equation for w (by eliminating some
degeneracies in the linearised operator associated to Q). This reflects the fact that
perturbations to a soliton can cause that soliton to move in a significant manner
along the directions given by the symmetries of the equation, namely phase rotation
and spatial translation; these are major channels of propagation for the equation as
motion in these directions does not conflict with any of the conservation laws. (In
the case of the L2x -critical equation, motion in the scaling direction is also possible
as it does not contradict conservation of mass.) We will not have space to devote
attention to these tools, which are fundamental in the stability theory of solitons,
but see [Wei2] and many subsequent papers (e.g. [MR], [MR2], [MR3] and the
references therein).

Exercise 3.55 (Justification of energy conservation). Let d = 3 and 1 < p < 5,

= +1, t0 = 0, and u0 Hx1 (Rd ). For each > 0, show that there exists a global
Hx1 -wellposed solution u() solution to the regularised NLS
1
it u() + u() = (|u() |2 + 2 )(p1)/2 u() ; u() (0) = u0
2
 3.7. STABILITY THEORY 179

with a conserved energy

Z
() () 1 2
E [u (t)] := |u() |2 + (|u() |2 + 2 )(p+1)/2 dx.
Rd 2 p+1
Then show that for any compact time interval I containing 0, u() converges in
S 1 (I R3 ) to a strong Hx1 solution u S 1 (I R3 ) to (3.1) with the conserved
energy
Z
1 2
E[u(t)] := |u() |2 + |u|p+1 dx.
Rd 2 p+1
This is one way in which to justify the conservation of energy for fractional-power
NLS.
Exercise 3.56 (Weak solutions). Let d = 3 and 1 < p < 6, = +1, t0 = 0,
and u0 Hx1 (Rd ). Show that for any > 0 there exists a global Hx1 -wellposed
solution u() to the tempered NLS
1
it u() + u() = max(|u() |p1 , |u() |4 )u() ; u() (0) = u0
2
R
with a conserved mass Rd |u() |2 dx and conserved energy
Z
1
E () [u() (t)] := |u() |2 + V (|u() |) dx,
Rd 2
Ry
where V (y) := 0 max(wp , w5 ) dw. Using weak compactness, show that there
exists a sequence n such that the solutions u(n ) converges weakly in
L 1 3 1 1
t Hx (R R ) to a global weak Hx solution u Lt Hx (R R ) to the NLS
3

(3.1). Thus for certain supercritical equations it is still possible to construct global
weak solutions. Existence or uniqueness of global strong Hx1 solutions for these
equations is a major unsolved problem (sharing many difficulties with the notori-
ous global regularity problem for Navier-Stokes). Even energy conservation for the
global weak solution is not known (the above construction, combined with Fatous
lemma, only shows that the energy at time t is less than or equal to the energy at
time 0). The analogous construction for global weak solutions for NLW dates back
to [Seg2].
Exercise 3.57. Complete the proof of Proposition 3.34. (Hint: use energy
estimates as in the proof of Proposition 3.8).
Exercise 3.58. [CKSTT11] Prove Lemma 3.36. (Hint: first establish L10 t,x
and S 1 control on u, then write v := u u and S := k(it + )vkL2 W 1,6/5 (IR3 ) ,
t x
and use the Leibnitz rule, Holders inequality, Strichartz, and Sobolev to bound
S in terms of itself and . Then use a continuity method argument to obtain an
unconditional bound on S. See also Figure 11.)

Exercise 3.59. [CKSTT11] Prove Lemma 3.37. (Hint: first establish S 1

30/13
control on u. Then divide up I into intervals where the L10
t Lx norm of u is
small, and apply Lemma 3.36 inductively on these intervals. See also Figure 12.)

Exercise 3.60. By refining the analysis used in the proof, replace the log1/3 1

in Proposition 3.35 with log 1 .
180 3. SEMILINEAR DISPERSIVE EQUATIONS

0
Exercise 3.61. [CKSTT13] Let u Ct,loc Sx (R T2 ) be a classical solution
1 2
to the cubic defocusing NLS it u+ 2 u = |u| u. Using the Fourier ansatz u(t, x) =
P i(kx+ 21 |k|2 t)
k(2Z)2 e ak (t), deduce the infinite system of ODE

(3.71) t a(t) = Nt (a(t), a(t), a(t))

where a = (ak )k(2Z)2 and Nt is the trilinear form
X i 2 2 2 2
Nt (a, b, c)k := ak1 bk2 ck3 e 2 (|k1 | |k2 | +|k3 | |k4 | )t .
k1 ,k2 ,k3 (2Z)2 :k1 k2 +k3 =k

(Compare with (1.56)). Let K 1 be a large number, let 0 < < 1, and let
T c()K 2 log K for some small c() > 0 depending only on . Suppose we have
a system b(t) = (bk (t))k(2Z)2 of functions with b Ct1 lk1 ([0, T ] (2Z)2 ) with
b(0) = a(0) which obeys the approximate equation
t b(t) = Nt (b(t), b(t), b(t)) + e(t)
to (3.71), where e(t) and b(t) obey the l1 bounds
Z t Z t
kbkCt1 l1k ([0,T ](2Z)2 . K 1 ; sup k e(t ) dt kl1k . K 1 .
0tT 0 0

Then if c() is sufficiently small depending on , we have the estimate ka

bkCt1 l1k ([0,T ](2Z)2 . K 1/2 . This lemma allows one to use near-solutions to NLS
in Fourier space to approximate actual solutions to NLS, and is a key ingredient in
establishing a certain weak turbulence result for this equation. See [CKSTT13].

3.8. Illposedness results

All happy families resemble one another; each unhappy family is
unhappy in its own way. (Leo Tolstoy, Anna Karenina)
In the past few sections we have developed a wellposedness theory for several
types of NLS and NLW equations, for various regularities Hxs (or Hxs Hxs1 ).
Despite the wide variety of equations and regularities considered, the wellposed-
ness theory for these equations are remarkably similar to each other, especially for
subcritical regularities. In such cases the time of existence depends only on the
norm of the data, and the solution map not only exists and is unique, but enjoys
very strong continuity properties; indeed, the solution map (from Hxs to Ct0 Hxs )
is typically uniformly continuous, Lipschitz, infinitely differentiable, and even real
analytic (see for instance Exercise 3.25).
However, there are certain equations and certain regularities for which the
Cauchy problem does not agree with this picture, either locally or globally in time,
in which case we say that that particular Cauchy problem is illposed. Unlike the sit-
uation with wellposedness, the type of illposedness exhibited can vary substantially
on the equation and on the regularity. At one extreme, there are very dramatic
examples of illposedness, such as blowup - various norms going to infinity in finite
time - beyond which no reasonably strong notion of solution can be salvaged. At
the other extreme there are very mild examples of illposedness, where it may still
be that the solution map exists and could even be continuous, but that the solution
map is known to be unstable (e.g. non-uniformly-continuous or non-Lipschitz),
 3.8. ILLPOSEDNESS RESULTS 181

non-differentiable, or at least non-analytic. Intermediate between these extremes43

are examples of norm explosion - when data of arbitrarily small norm can lead to
solutions of arbitrarily large norm in arbitrarily small time. This is not quite as
dramatic as blowup, because a solution may still exist for each given initial datum,
but it certainly does prevent any continuous dependence of the solution map on the
initial data.
For each of the types of illposedness discussed above, there are examples of
equations and regularities that exhibit that illposedness. In contrast with the well-
posedness theory, which is largely based around the single technique of Duhamel
iteration, illposedness can be achieved by a surprisingly large number of unrelated
methods. We will not be able to discuss all of them here, but we give a represen-
tative sample. For a recent survey of techniques and results, see [Tzv].
We first discuss methods for generating blowup, by which we mean classical (or
strong) solutions which develop a significant singularity in finite time (e.g. the Hxs
norm goes to infinity in finite time). One way to construct these solutions is via
construction of explicit (or nearly explicit) blowup solutions. We have already seen
two examples of this - the blowup solution (3.15) for the pseudoconformal focusing
NLS and the ODE-based blowup solution (3.6) for the focusing NLW. The latter
solution has no decay in space and thus does not lie in any Hxs Hxs1 spaces,
however this can be rectified by a finite speed of propagation; see Exercise 3.9.
In some cases, one cannot construct a blowup solution explicitly, but can cre-
ate an explicit approximate solution to the equation which blows up in finite time.
One can then hope to use perturbation theory to convert this to an exact blowup
solution. This argument can be made to work, but is extremely delicate, because
perturbation theory requires a great deal of wellposedness and stability on the
equation, which is in obvious conflict with our need to make both the exact and
approximate solution to blow up in finite time. One often needs to carefully renor-
malise the solution (usually via rescaling), and obtain stability control in one set of
norms while obtaining blowup in another. See for instance [Mer], [BW] for some
instances of this approach.
In the case of the NLS, there is another, much more indirect, way to force
blowup of a solution, namely the virial argument of Glassey [Gla2], based on the
nonlinear counterpart to (2.38). For simplicity let us consider a classical solution

u Ct,loc Sx (R Rd ) to an algebraic NLS. Consider the quantity

Z Z
V (t) := |x|2 T00 (t, x) dx = |x|2 |u(t, x)|2 dx.
Rd Rd

43There are several other symptoms of illposedness which we will not have space to discuss
here, including breakdown of uniqueness (either for weak or strong solutions); failure of mass or
energy conservation; loss of regularity; or examples of approximate solutions to the equation which
blowup in finite time. The reader is invited to try to list several such illposedness phenomena and
rank them in approximate order of severity.
182 3. SEMILINEAR DISPERSIVE EQUATIONS

This quantity is clearly non-negative. Applying (2.35) and integration by parts

repeatedly, we obtain the virial identity
Z
tt V (t) = 2 Tjj (t, x) dx
Rd
Z
d(p 1) p+1
(3.72) = 2|u|2 + |u| dx
R d p+1
Z
d(p pL2x )
= 4E[u] + |u(t, x)|p+1 dx
p+1 R d

where E[u] is the conserved energy and pL2x := 1 + 42 is the pseudoconformal power.
If we are in the L2x -critical or L2x -supercritical focusing cases p pL2x , = 1, we
thus conclude the bound
tt V (t) 4E[u].
If the energy happens to be negative (which is possible in the focusing case = 1),
this shows that V is a strictly concave function of t. Since V is also non-negative,
we conclude that the solution can only exist classically for a finite amount of time
(in either direction). This argument thus demonstrates blowup in finite time (and
even gives an upper bound on the time of existence in terms of the datum and
the energy). It can be extended to demonstrate blowup for any Hx1 initial data u0
which has negative energy44 and obeys the decay condition hxiu0 L2x . The decay
condition can be removed, basically by working with spatially truncated versions of
the virial identity; see for instance [Naw]. We remark that while negative energy
is a sufficient condition for blowup, it is hardly a necessary condition; for instance,
the solution (3.15) has zero energy, and the solutions constructed in [BW] have
positive energy. The blowup phenomenon has been analyzed much further for the
L2x -critical equation, in the vicinity of soliton solutions; see [MR], [MR2], [MR3].
For the focusing NLW, one can also exploit some positivity properties of the
fundamental solution to establish successively stronger lower bounds on a solution
which eventually leads to blowup. One particularly striking example of this is a
result of John [Joh], who showed that for the three-dimensional focusing NLW with
initial data u(0, x) = u0 (x), t u(0, x) = 0 for positive
Schwartz u0 and sufficiently

small , one has blowup in finite time for p < 2 and global existence for p > 2.
These results have since been generalised extensively; see for instance [GLS], [Hor].
Once one has one solution blowing up, one can use the symmetries of the
equation to generate further solutions blowing up. When the regularity s is low
enough, one can use the symmetries to create classical initial data of arbitrarily
small Hxs norm which blow up in arbitrarily small time, which is a very strong
demonstration of illposedness in that data class Hxs ; we give some examples in the
exercises.
All the known examples of blowup from classical data are for focusing equations;
for many defocusing equations (e.g. Hx1 -subcritical or Hx1 -critical defocusing NLS
or NLW) we have global existence of classical solutions. The question of whether
blowup occurs from classical data for Hx1 -supercritical defocusing NLS or NLW
equations is a major open problem in the subject (analogous to the Navier-Stokes
global regularity problem) and remains very far from resolution. While blowup

44To put it another way, whenever the nonlinear component of the energy exceeds the linear
component, blowup occurs. Compare this with Principle 3.24.
 3.8. ILLPOSEDNESS RESULTS 183

is not known for these equations, we can in many cases establish weaker forms
of illposedness, which are not as dramatic as blowup but do indicate that many of
the techniques discussed in earlier sections to establish wellposedness (e.g. iteration
methods) must necessarily fail. One of the mildest types of illposedness of this form
is that of analytic illposedness, in which one demonstrates that the solution map (say
from Hxs to Ct0 Hxs ), if it exists at all, is not real analytic. In fact one typically shows
the stronger statement of C k illposedness for some k 1, which asserts that the
solution map, if it exists, is not k-times differentiable. This is basically accomplished
by the method of Taylor expansions (i.e. power series methods). Let us illustrate
this with the three-dimensional cubic defocusing NLS (d = 3, p = 3, = +1)
with initial datum u(0) = u0 for some fixed Schwartz u0 , thus we are considering
solutions u() to the Cauchy problem
1
(3.73) it u() + u() = |u() |2 u() ; u(0) = u0 .
2
The global existence theory of this equation (Exercise 3.38) guarantees that the
solutions u() exist and are smooth for all time. A refinement of this theory also
shows us that u() also depend smoothly on , uniformly on any compact time
interval. In particular, we can obtain a Taylor expansion
u() (t, x) = u1 (t, x) + 2 u2 (t, x) + 3 u3 (t, x) + O(4 )
for some smooth functions u1 , u2 , u3 (there is no zeroth order term since u(0) is
clearly zero), where the error is uniformly smooth in t, x on any compact time
interval. We can expand both sides of (3.73) using this expansion and compare
coefficients. One learns that the first coefficient u1 is just the linear solution:
1
it u1 + u1 = 0; u1 (0) = u0
2
it/2
or in other words u1 (t) = e u0 . The second term u2 solves the equation
1
it u2 + u2 = 0; u2 (0) = 0
2
and is hence zero. The third term u3 solves the equation
1
it u3 + u3 = |u1 |2 u1 ; u3 (0) = 0
2
and is hence given by a Duhamel integral
Z t Z t
i(tt )/2
u3 (t) = i e 2
(|u1 | u1 (t )) dt = i ei(tt )/2 (|eit /2 u0 |2 eit /2 u0 ) dt .
0 0
()
From the Taylor expansion of u we thus have obtained the formula
3 Z t
d ()
u (t)|=0 = 3!i ei(tt )/2 (|eit /2 u0 |2 eit /2 u0 ) dt .
d3 0
Rt
This shows that if the map u0 7 0 ei(tt )/2 (|eit /2 u0 |2 eit /2 u0 ) dt is not a
bounded map from Hxs (R3 ) to Ct0 Hxs ([0, T ] R3 ), then the solution map u0 7 u
3
will not be a Cloc map from Hxs (R3 ) to Ct0 Hxs ([0, T ] R3 ), even for data arbitrarily
close to zero in Hxs norm. This lack of boundedness can often be established by
direct computation; in this case, we can achieve this for supercritical regularities
s < sc = 12 (in contrast to the critical case s = sc and subcritical cases s > sc , in
which one does have analytic wellposedness); see Exercise 3.65.
184 3. SEMILINEAR DISPERSIVE EQUATIONS

The method above is fairly general. Roughly speaking, it shows that if the
Duhamel iteration scheme used to construct solutions leaves a certain space X after
finitely many iterations, then the solution map can only have a finite amount of
differentiability in that space. This is not too surprising since the iteration scheme
is closely akin to a power series expansion of the solution in terms of the initial
datum.
Another approach for establishing illposedness is by constructing families of ex-
act solutions to the equation which are close together at time zero but far apart at
other times. In some cases one can use explicit solutions such as solitons and plane
wave solutions, possibly after various symmetries of the equation have been applied;
in other cases one needs to construct solutions by the methods of nonlinear geomet-
ric optics, or more generally by constructing an approximate solution first and then
using stability theory to perturb the approximate solution into the exact solution.
A typical result obtained by this method would be that a certain solution operator
cannot be uniformly continuous from Hxs to Ct0 Hxs even when the size of the datum
and time of existence are set to be small. We have already seen some examples of
this in Exercise 3.5 and the discussion after (3.20), using explicit solutions. We now
briefly sketch how to achieve a similar effect using the approximate solutions of the
preceding section. For sake of concreteness let us just consider the one-dimensional
defocusing cubic NLS (d = 1, p = 3, = +1). Let be a Schwartz function. From
Proposition 3.35 we have constructed (for small > 0 and 1 a 2) solutions w,a
2
to the small-dispersion equation it w + 2 xx w = |v|2 w on the slab [0, 1] R
which has the approximate form
2
|(x)|2 t
w,a (t, x) = aeia (x) + O ()
for 0 t 1, where the error can be controlled in a suitable Hx1 sense. One can
apply the rescaling u,a (t, x) := w,a (t, x) to obtain a class of solutions u to the
original NLS with the approximate form
2
|(x)|2 t
u,a (t, x) = aeia (x) + O ()
where the error is now controlled in some rescaled Hx1 sense. One can exploit scale
invariance (3.9) and Galilean invariance (3.10) to obtain a wider class of exact
solutions u,a,,v to NLS for > 0 and v R of the form
t|v|2
a2 |((xvt)/)|2 t/2 )
u,a,,v (t, x) = 2/(p1) aei(xv+ 2 ((xvt)/)+O (2/(p1) )
where the error has to be interpreted in a suitable norm. If s < 0 is a negative
regularity, then by making v large, and setting |v|(p1)s/2 , these solutions
can become bounded in Hxs . By a suitable variation of the parameters , a, , v one
can then show that the solution operator to this equation cannot be uniformly con-
tinuous from Hxs to Ct0 Hxs , even for small times and small norm, by exploiting the
phase decoherence effect arising from the a2 |((xvt)/)|2 t/2 term; see [CCT2]
for details. Generally speaking, it is not difficult to create (for any equation and
regularity) large data solutions which exhibit these types of instabilities at large
times; the various symmetries are then used to create small data solutions which
are similarly unstable at small times. In order for this to work, one needs the reg-
ularity s to be supercritical with respect to at least one of the symmetries (scaling,
Galilean, or Lorentz). See [Kuk], [Leb], [BGT], [CCT], [CCT2], [CCT3] for
several examples of this technique and further discussion.
 3.8. ILLPOSEDNESS RESULTS 185

One final type of illposedness is the high-to-low frequency cascade, in which

a solution starts off initially with Fourier transform supported primarily at high
frequencies, but quickly creates a substantial presence at low frequencies. For small
s (e.g. negative s), such solutions typically have small Hxs norm at time zero
but large Hxs norm at later times; this norm explosion is a fairly strong form of
illposedness as it shows that the solution operator, if it exists at all, has a severe
singularity in Hxs near the zero solution. These cascading solutions can often be

constructed using the stability theory arising from a higher regularity Hxs . See
[CCT2], [BT] for examples of this strategy.
Exercise 3.62. Let u be a classical solution to an NLW, and let
Z
d1 2
V (t) := |x|2 T00 (t, x) |u| dx.
Rd 2
Establish the following analogue of the virial identity for this quantity, namely
(d 1)(p pH 1/2 ) Z
tt V (t) = 2E[u] + x
|u(t, x)|p+1 dx
p+1 Rd
4
where pH 1/2 := 1 + d1 is the conformal power. (Note the shifting of the dimension
x
d by one; compare this with Exercise 3.2.) This identity is not as useful as the NLS
virial identity because the quantity V does not have a definite sign in general.
Exercise 3.63. Consider a focusing NLS with p pL2x = 1 + d4 , and let
s < sc . Show that there exists classical data of arbitrarily small Hxs norm such that
the solution to the NLS blows up in arbitrarily small time. (Hint: use the virial
identity to create a classical solution with Schwartz initial data which blows up in
finite time, and then use the scaling symmetry to rescale the blowup time to be
arbitrarily small.) This illustrates the principle that one usually does not have a
wellposedness theory at supercritical regularities for focusing equations.
Exercise 3.64. Consider a focusing NLW and let s < sc . Show that there
exists classical data of arbitrarily small Hxs Hxs1 norm such that the solution to
the NLW blows up in arbitrarily small time. (Hint: use Exercise 3.9. In the case
when sc is negative, you may find it convenient to enforce moment conditions on
the initial data to ensure some vanishing of the Fourier coefficients near the origin.)
A similar result holds for s < sl (using the Lorentz invariance instead of the scaling
invariance) but is a little trickier; see [Sog].
Exercise 3.65.R t Let T > 0 be arbitrary. Use a scaling argument to show that

the map u0 7 0 ei(tt )/2 (|eit /2 u0 |2 eit /2 u0 ) dt is not a bounded map from
Hxs to Ct0 Hxs ([0, T ] R3 ) when s < 1/2. Conversely, use Strichartz estimates to
show that this map is bounded for s 1/2. (For sake of this exercise, you may use
the heuristic (A.15) as if it were rigorous. Alternatively, one may use Littlewood-
Paley decomposition.)
Exercise 3.66. [Kat4] Consider the rather artificial nonlinear wave equation
4
u = |u|p for some Hx1 -subcritical power 1 p < 1 + d2 . Let u be a strong
1 2
Hx Lx solution to this equation whose initial position is supported in the ball
{|x| 1} and whose initial velocity is zero (for simplicity). It is possible to establish
the finite speed of propagation property for such solutions, in particular you may
assume that this solution is supported on the ball {|x| 1 + t} for all later times
186 3. SEMILINEAR DISPERSIVE EQUATIONS

R
t > 0 for which the solution exists. Show that if the integral Rd u(0, x) dx is
sufficiently large and positive depending on d and p, then the solution u can only
exist for a finite amount of time in the forward direction (i.e. u cannot be a
2
Rstrong solution on [0, +)). If p < 1 + d , show that one only needs the integral
Rd u(0, x) dx to be strictly positive to achieve the same result (i.e. no largeness
hypothesis
R is required). Hints: obtain an integral inequality for the quantity m(t) :=
R d u(t, x) dx, using Holders inequality and finite speed of propagation. First show
that m(t) is convex and monotone increasing, and then obtain even better lower
bounds on this quantity. You may find the comparison principle, Exercise 1.7, to
be useful. For further variations on this theme, see [KTao3].
Exercise 3.67. [LSog], [Sog] Consider a focusing NLW, and let s be such
that 0 < s < sl := d+1 1
4 p1 (so s is supercritical with respect to the Lorentz
invariance). Start with the explicit blowup solution (3.6) with t0 = 0 and apply a
Lorentz transform to it, to create a solution which blows up at the point (0, 0) but is
smooth in the backwards light cone {(t, x) : |x| < t}. Now work on the time slice
t = 1 and localise the initial data to a neighbourhood of the ball {|x| < 1} using
finite speed of propagation, to create smooth initial data (u(1), t u(1)) whose
Hxs (Rd ) Hxs1 (Rd ) norm is arbitrarily small, but which develops a singularity
at time 0. Rescaling this, we can construct data of arbitrarily small Hxs (Rd )
Hxs1 (Rd ) norm with a solution that blows up in arbitrarily small time, which
defeats any hope of a reasonable wellposedness theory at this regularity.

3.9. Almost conservation laws

The Law of conservation of energy tells us we cant get something
for nothing, but we refuse to believe it. (Isaac Asimov)
We have seen how the laws of conservation of mass and energy can be used
to obtain global wellposedness results at the L2x and Hx1 regularities respectively;
generally speaking, they assert that these two norms stay bounded for all time. One
may then ask what happens to the other Hxs norms; after all, the linear Schrodinger
and wave flows preserve the Hxs norm and Hxs Hxs1 norms respectively. In
particular, once one has global wellposedness for one norm, say Hx1 , one can ask
whether the same wellposedness occurs for other regularities also.
Generally speaking, global wellposedness in lower norms implies global well-
posedness in higher norms45, due to persistence of regularity; we have already seen
several examples of this phenomenon in this chapter. However, while the lower
Sobolev norms such as Hx1 may remain bounded uniformly in time, the bounds one
obtains on higher norms such as Hx2 may grow faster than this; see for instance the
exponential bounds in Exercise 3.28. This reflects the fact that the persistence of
regularity arguments do not prohibit a low-to-high frequency cascade scenario, in
which the energy starts off concentrated in low frequencies but moves increasingly
to higher frequencies as time progresses; it is easy to envisage a scenario of this
form where the Hx1 norm stays bounded, but higher norms such as Hx2 go to infin-
ity. Numerical simulations have confirmed this type of weak turbulence behaviour
for the periodic analogues of NLS and NLW, but for the non-periodic defocusing
setting it appears that such phenomena, if they exist at all, do not occur with any
45Assuming of course that the nonlinearity itself is smooth enough to support solutions at
this level of regularity.
 3.9. ALMOST CONSERVATION LAWS 187

great strength generically. (Indeed, the solition resolution conjecture mentioned

earlier is probably not consistent with weakly turbulent behaviour.) It would be of
interest to obtain more theoretical results regarding this issue.
Somewhat dual to this is the problem of starting with a global wellposedness
result, say at Hx1 , and trying to lower the regularity needed for global existence,
say to Hxs for some 0 < s < 1. This is reasonable in the Hx1 -subcritical case with
s subcritical, since in such cases one already knows that the local wellposedness
theory can extend below Hx1 . Indeed, the Hxs local wellposedness theory asserts
in this case that the only way an Hxs -solution can cease to exist is if the Hxs norm
blows up in finite time. Thus the difficulty is to establish some upper bounds on the
growth of the Hxs norm in time; by limiting arguments one can restrict attention
to the global Hx1 solutions, so long as the final bound on the Hxs norm growth
depends only on the Hxs norm of the initial datum rather than on the energy. Here,
the major difficulty is caused by the high-to-low frequency cascade scenario, in
which one starts initially with a very large amount of energy at high frequencies
(which may have small Hxs norm), but a significant fraction of this energy somehow
makes its way to low frequencies, thus causing the Hxs norm to grow substantially.
To summarise, in order to establish good global existence results either for
s above or below the energy regularity Hx1 one needs to control the flow of en-
ergy either from low frequencies to high frequencies or vice versa. In recent years,
two methods have been developed to achieve such a control, namely the Fourier
truncation method of Bourgain, and the subsequent method of almost conserved46
quantities or I-method of Colliander, Keel, Staffilani, Takaoka, and Tao. The two
methods are similar (indeed, the former inspired the latter) but not identical. They
both proceed by selecting a large frequency cutoff N , and declaring frequencies less
than N to be low and greater than N to be high. If the solution has regularity
Hxs for some s < 1, then the low frequency components will have bounded energy
(but with a bound depending on N ), but the high frequency components will have
unbounded or infinite energy. The strategy is then to somehow suppress the un-
bounded energy high frequency component in order that the energy conservation
law can be usefully applied. The Fourier truncation method achieves this by view-
ing the original equation as a weakly coupled system of the high and low frequency
components. Then one attempts to omit the nonlinear effects of the high frequen-
cies, so that one believes the high frequencies to evolve approximately linearly, and
the low frequencies to evolve approximately via the original equation. In particular
one expects the low frequencies to (approximately) conserve their energy (as op-
posed to exporting or importing energy with the high frequencies). On short time
intervals, one can justify this approximation using stability theory; the strategy is
then to iterate this control on short time intervals to control on long-time inter-
vals. Thus turns out to be possible by choosing N to be large, provided that the
initial regularity Hxs is sufficiently close to Hx1 , and provided that the nonlinearity
has a certain smoothing property (roughly speaking, one wants the effect of the
nonlinearity to be bounded in Hx1 even when the solution is only as regular as Hxs ).

46We shall use the term almost conserved quantity rather loosely; for us, it shall mean
a quantity whose time derivative is unexpectedly small or low order in some sense. The
monotone quantities appearing in monotonicity formulae could also be viewed as a type of almost
conserved quantity.
188 3. SEMILINEAR DISPERSIVE EQUATIONS

(large but u (T) (infinite energy, but

Adjusted u (T)
finite energy) lo hi high frequency)
(relatively small energy)
v(T)

selfinteraction
u (T) [smoothing effect]
lo
nonlinear
linear evolution
evolution
[conserves crossterms
Hamiltonian]

Low and medium High frequency

frequency components components
u (0) u (0)
lo hi
Initial datum
(large but finite energy) (infinite energy, but
u(0) high frequency)

Figure 13. The first step in the Fourier truncation method. For
a short time t, one evolves the high frequencies linearly and the
low frequencies nonlinearly (thus preserving the Hamiltonian of
the low frequencies). The error term v arises both from high-low
frequency interactions and high-high frequency interactions; if the
equation has enough smoothing properties, this error will be small
in energy norm and can be safely absorbed into the low frequency
component. One then iterates this scheme for as long as one has
good control on all components.

The Fourier truncation method is surveyed in [Bou9] and will not be detailed
here (but see Figure 13 and Table 2). The I-method proceeds slightly differently;
rather than omit the high frequencies completely, it merely damps them using
a Fourier multiplier I (hence the name I-method). This damping operator is
essentially the mildest operator that makes the high frequencies bounded in energy;
the low frequencies remain undamped by this operator. One then tries to control
the energy E[Iu(t)] of the damped solution Iu to the equation, which consists of
the unadulterated low frequencies and the damped high frequencies. This quantity
turns out to enjoy an almost conservation law, in that the quantity E[Iu(t)] does
not vary very quickly in t. (Note that if I is the identity then E[Iu(t)] would be
constant; thus this almost conservation nature reflects the mild nature of the
operator I.) One can then use this almost conserved quantity to generate long-
time control of the solution in much the same way that a genuine conservation law
can be used to ensure global wellposedness. If all goes well, the time upon which
one ultimately gets a useful control on the solution will be a positive power of N ;
letting N go to infinity will then yield the desired global wellposedness in Hxs . This
 3.9. ALMOST CONSERVATION LAWS 189

method can handle slightly more general nonlinearities and regularities than the
Fourier truncation method, because no smoothing effect is required (this is due to a
certain cancellation arising from a commutator term in the almost conservation
law, which has no counterpart in the Fourier truncation approach), but provides
slightly less information on the solution.

Table 2. An oversimplified comparison between the ways the

Fourier restriction method and I-method treat different frequency
interactions in a nonlinear equation. In both cases the low-low
frequency interactions are considered large, but they do not alter
the Hamiltonian, while the high-high frequency interactions are
treated as error terms. The main difference lies in how the high-
low interactions are treated, with the I-method taking advantage
of commutator cancellations to show that these interactions ap-
proxiately conserve the damped Hamiltonian E(Iu). Also, the
Fourier restriction method takes advantage of smoothing effects
and obtains better (energy-class) control of error terms.

Interaction Fourier restriction method I-method

Low-low Conserves E(ulo ) Conserves E(Iu)
High-low Small error in Hx1 Approximately conserves E(Iu)
High-high Small error in Hx1 Small error in IHx1

Let us illustrate the method with the one-dimensional quintic defocusing NLS
1
(3.74) it u + xx u = |u|4 u,
2
which one can proceed by a relatively simple energy method implementation
of the approach47. Indeed for this equation one can obtain global Hx1 -wellposed
and classical solutions without any difficulty. (On the other hand, this equation
is L2x -critical, and global wellposednes of L2x solutions is unknown.) Now let u
be a classical solution and 0 < s < 1; we are interested in the behaviour of the
s
H
R x 1norm 2of u(t) as t . We alreadyRhave conservation of the energy E[u] :=
1 6
R 2 | x u| + 3 |u| dx and mass M [u] := R |u|2 dx, but we will be reluctant to use
E[u] directly as it will not be controlled purely by the Hxs norm. To create some
almost conserved quantities at the regularity Hxs , let us introduce a large frequency
cutoff N 1 and a spatial Fourier multiplier I defined by
c
Iu() := mN () = m( ),
N
where mN () = m(/N ) and m is a smooth function which equals 1 for || 1 and
is equal to ||s1 for || 2. Thus I is the I dentity operator on low frequencies ||
N , and is essentially an I ntegration operator N 1s ||s1 on high frequencies ||
47This is similar in spirit to the energy cancellation methods for establishing local existence
for various nonlinear equations without performing an iteration scheme, and which can exploit
certain structural cancellations arising from the nonlinearity; see for instance the high-regularity
arguments in Section 4.1, Section 4.4, or Section 6.1. Most applications of the I-method, however,
also require a modified local wellposedness statement which is obtained by standard iterative
means, in order to exploit various local smoothing effects that can only be captured by spacetime
norms.
190 3. SEMILINEAR DISPERSIVE EQUATIONS

N ; this explains why

R this operator is denoted I. We now show that the modified
energy E[Iu(t)] = R 21 |x Iu(t)|2 + 13 |Iu(t)|6 dx obeys an almost conservation law.
Proposition 3.38 (Almost conservation law). Let s > 1/2. Suppose t is a
time such that E[Iu(t)] . 1. Then |t E[Iu(t)]| .s N 1/2 .
The exponent 1/2 here might not be best possible. An improvement of the
exponent here will lead to a better global wellposedness result for a conclusion, as
will be clear from the remainder of this argument.
Proof. For a general classical field v, we have the identity
Z
1
t E[v(t)] = 2Re t v(it v + xx v |v|4 v) dx
R 2
which can be easily verified by integration by parts; note that this reproves the
conservation of energy for (3.74). We now set v := Iu; by applying I to (3.74) we
see that v solves the equation
1
it v + xx v = I(|u|4 u)
2
and hence we have
Z
t E[v(t)] = 2Re It u(I(|u|4 u) |Iu|4 Iu) dx.
R
Thus it will suffice to establish the bound
Z
| It u(I(|u|4 u) |Iu|4 Iu) dx| . N 1/2 .
R
1
Splitting t u = 2 xx u i|u|4 u, we can split this further into
Z
(3.75) | Ixx u(I(|u|4 u) |Iu|4 Iu) dx| . N 1/2
R
and
Z
(3.76) | I(|u|4 u)(I(|u|4 u) |Iu|4 Iu) dx| . N 1/2 .
R
We shall just prove the top order estimate (3.75) and leave the lower order estimate
(3.76) as an exercise. To avoid technicalities we shall cheat somewhat by assuming
heuristics such as the fractional Leibnitz rule (A.14) as if they were rigorous; one
can justify all the cheats performed here by Littlewood-Paley theory and other
tools of Fourier analysis but we shall not do so here. We integrate one of the
partial derivatives by parts, and observe from the hypothesis E[Iu(t)] . 1 that
kIx ukL2x . 1. Thus it suffices to show the commutator estimate
kx [I(|u|4 u) |Iu|4 Iu]kL2x . N 1/2 /
We split u = uhi + ulo , where uhi := P>N/100 u and ulo := PN/100 u. We can
then expand |u|4 u and |Iu|4 Iu into a large number of terms involving five factors
from uhi and ulo . There are three types of terms to consider. First consider the
low-low terms that only involve ulo :
kx [I(|ulo |4 ulo ) |Iulo |4 Iulo ]kL2x .
Because I is the I dentity on low frequencies, we see that both I(|ulo |4 ulo ) and
|Iulo |4 Iulo are equal to |ulo |4 ulo and so the net contribution of these terms is zero.
 3.9. ALMOST CONSERVATION LAWS 191

Next, consider any high-high term that involves two or more factors of uhi . Here
is a typical one (we do not attempt to exploit cancellation here):
kx I(uhi 2 u3lo )kL2x .
The operator x I is a pseudodifferential operator of positive order. Applying the
fractional Leibnitz rule, we can distribute this operator and end up considering
terms, of which the following is typical:
kO(|x Iuhi ||uhi ||ulo |3 kL2x .
Now from the Gagliardo-Nirenberg inequality and the hypothesis E[Iu(t)] . 1 we
have kIukL x
. 1, and in particular kulo kL x
. 1. Also we have already remarked
that kx Iuhi kL2x . 1, which by an easy Fourier analytic argument (exploiting the
high frequency nature of uhi and the hypothesis s > 1/2) implies that kuhi kL x
.s
N 1/2 ; see Exercise 3.68. The desired bound then follows from Holders inequality.
Finally, we must consider high-low terms involving only one factor of uhi .
Here we must use48 the cancellation present in (3.75). A typical term to consider
is
kx [I(|ulo |4 uhi ) |ulo |4 Iuhi ]kL2x ,
where we have used the fact that Iulo = ulo . The expression in brackets is the
commutator of I and |ulo |4 , applied to uhi . Let us write w := |ulo |4 . The Fourier
transform of x (I(wuhi ) wIuhi ) at can be computed to be
Z

i [m( ) m( N )]w()uhi ( ) d.
R N
The integrand vanishes unless || . N and || & N . In such a case, an application
of the mean-value theorem gives the bound (m( N ) m( N )) = O(||m((
)/N )). On the other hand, the expression
Z
||m(( )/N )w()uhi ( ) d
R
is essentially the Fourier transform of (||w)Iuhi . Thus we morally have
x [I(|ulo |4 uhi ) |ulo |4 Iuhi ] . (||w)Iuhi
in some Fourier sense. Assuming this to be rigorous, we are reduced to establishing
that
k(||w)Iuhi kL2x . N 1/2 .
Now we already know that kIuhi kL x
. N 1/2 . Also by distributing the derivative
|| using the fractional Leibnitz rule, we can (morally) replace ||w by an expres-
sion such as O(|||ulo ||ulo |3 ). Since we already know that kulo kLx
. kIukL x
.1
and k||ulo kL2x . kIukL2x . 1, the claim now follows from Holders inequality. 
From the above proposition and the continuity method, we conclude that if
E[Iu(0)] = O(1), then in fact E[Iu(t)] = O(1) for all |t| s N 1/2 . Thus the
quantity E[Iu(t)] is stable for long periods of time. One can now apply scaling
arguments and some Fourier analysis to conclude
48An alternative would be to try to average in time and exploit bilinear refinements to
Strichartz inequality here; this is related to the extra smoothing effect alluded to earlier. However,
the approach given in the text demonstrates that one can use the commutator cancellation in the
I-method as a substitute for such smoothing effects.
192 3. SEMILINEAR DISPERSIVE EQUATIONS

Proposition 3.39. If ku(0)kHxs . 1, then ku(t)kHxs .s N 1s for all |t| s

1
N 2 2(1s)
.
We leave the derivation of this proposition to the exercises. If s > 3/4, the
exponent of N in the bound on |t| is positive, and so by letting N we can
conclude a growth bound on the Hxs norm. In fact we obtain the polynomial bound
1
ku(t)kHxs .s hti(1s)/( 2 2(1s)) .
This, combined with the local Hxs wellposedness theory, easily gives global well-
posedness for this equation in Hxs for all 3/4 < s < 1. (Wellposedness for s 1
already follows from energy conservation and persistence of regularity.)
The above strategy is rather flexible and can be adapted to a variety of sub-
critical equations; see for instance [CKSTT], [CKSTT3], [CKSTT4], [Mat],
[Pec], [Pec2], [Pec3], [Car]. It also combines well with scattering theory (see
[CKSTT7], [CKSTT10]; also see [Bou6] for an application of the Fourier re-
striction method to the scattering problem), to the growth of higher Sobolev norms
(see [CDKS], [Sta], [CKSTT8], [Bou10]) and to the stability theory of solitons
(see [CKSTT8], [CKSTT9]). In many cases it is not practical to obtain a point-
wise bound on the time derivative t E[Iu(t)] as in Proposition 3.38, but all one
Rt
really needs anyway is a bound on the integral t01 t E[Iu(t)] dt of this time deriv-
ative. This additional time averaging allows one to use additional spacetime norms
such as Strichartz norms, which can lead to better estimates. In such cases, one
needs an additional ingredient in the argument, namely a modified local existence
theorem that asserts that whenever E[Iu(t)] is bounded, then certain spacetime
norms of Iu (such as Strichartz norms) are bounded on a time interval centred at
t. This however can be achieved by a routine modification of the local existence
theory; see Figure 14 for a summary of this scheme. One can also exploit other
conservation laws (e.g. mass conservation) to try to improve the powers of N which
appear in the above argument. However, the most powerful methods for improving
the exponents here has proceeded by modifying either the Hamiltonian E[u] or the
almost conserved quantity E[Iu(t)] with additional correction terms to damp out
some nonresonant fluctuations; see Section 4.2. For instance, the quintic NLS
discussed above is in fact known to be globally wellposed in Hxs for all s > 4/9
using this technique; see [Tzi].
Exercise 3.68. Prove that kuhi kL x
.s kx Iuhi kL2x whenever uhi is a Schwartz
function supported on frequencies > N/100. (Hint: use frequency decomposition
and either Bernsteins inequality (A.6) or Sobolev embedding.)
Exercise 3.69. Prove (3.76). (Here one will have to make some use of the
potential energy component of E[Iu], which gives a useful bound on kIukL6x . This
can be combined with the bound one already has on kIukL x
, after decomposing
into high and low frequencies.)
Exercise 3.70. Prove Proposition 3.39. (Hint: choose a 1 such that the
1
rescaled solution u (t, x) := 1/2 u( t2 , x ) obeys E[Iu (0)] . 1 (you may find taking
the Fourier transform to be helpful). Then apply the almost conservation of E[Iu ]
for a long period of time, and then undo the scaling. One can use mass conservation
to control the lower order component of the Hxs norm.)
 3.9. ALMOST CONSERVATION LAWS 193

s s
u(t0 ) in H u(t 2) in H

[rescale] [undo
rescaling]

1 modified 1 modified 1
Iu(t 0) in H local theory Iu(t 1) in H local theory Iu(t 2) in H
on [t 0, t 1] on [t , t 2]
1

E(I(u(t 0)) bounded E(I(u(t 0)) bounded E(I(u(t 0)) bounded

[almost conservation [almost conservation
law] law]

Figure 14. The general scheme of the I-method; compare with

Figure 7. Of course, one would usually iterate the method for more
than the two time steps indicated here. Apart from the rescaling
and the presence of the I operator, one new feature is that the
(modified) local theory plays a quantitative role rather than merely
a qualitative one, as this theory is necessary to control the error
terms in the almost conservation law. However, it is important
that the local theory does not impact the main term in that law,
otherwise the Hx1 norm of Iu(t) could increase exponentially with
each time step.
194 3. SEMILINEAR DISPERSIVE EQUATIONS

log || PMu || 1
H
upper bound on u

upper bound on Iu

log M
N

Figure 15. A log-log plot of the energy of the Littlewood-Paley

pieces PM u(0) of u(0) as a function of M , when u(0) is assumed
to only lie in Hxs for some s < 1. Note the infinite energy at
high frequencies. The operator I smooths out the energy at high
frequencies, giving Iu(0) a large but finite energy. A rescaling is
then needed to make the energy bounded by 1.
 3.9. ALMOST CONSERVATION LAWS 195

log || PMu || 1
H
upper bound on u

upper bound on Iu

log M
N

Figure 16. A log-log plot of the energy of the Littlewood-Paley

pieces PM u(t) of u(t) as a function of M , for some late time t (a
power of N ). The almost conservation law keeps the energy of Iu
bounded but large (because of the scaling), but we do not exclude
the possibility that energy has been moved around in the frequency
ranges M . N . Thus at each time t, the high frequencies still
evolve in a somewhat linear (non-interacting) fashion, but the low
and medium frequencies may share their energy with each other.
As time progresses, more and more frequencies could mix their
energy, potentially leading to a polynomial growth in the Hxs norm.
 CHAPTER A

Appendix: tools from harmonic analysis

Every action of our lives touches on some chord that will vibrate
in eternity. (Sean OCasey)
The nonlinear evolution equations studied here can be profitably analyzed by
viewing these equations as describing the oscillation and interaction between low,
medium, and high frequencies. To make this type of analysis rigorous, we of course
need the notation and tools of harmonic analysis, and in particular the Fourier
transform and Littlewood-Paley theory; the purpose of this appendix is to review
that material. This is only an outline of the material; for a more thorough intro-
duction to these tools from a PDE-oriented perspective, see [Tay], [Tay2].
It is convenient to work in the Schwartz class Sx (Rd ). One particularly im-
portant operation on Schwartz functions (and hence on their dual) is the (spatial)
Fourier transform f 7 f, defined for f Sx (Rd ) by the formula1
Z
f() := f (x)eix dx.
Rd

As is well known, the Fourier transform f 7 f is a Frechet space automorphism

on the Schwartz space Sx (Rd ), with the inversion formula
Z
1
f (x) = f()eix d.
(2)d Rd
Thus every Schwartz function can be formally viewed as a superposition of plane
waves eix . We also have the fundamental Plancherel identity
Z Z
1
|f (x)|2 dx = |f()|2 d
Rd (2)d Rd
as well as the closely related Parseval identity 2
Z Z
1
(A.1) f (x)g(x) dx = f()g() d.
Rd (2)d Rd
The Fourier transform enjoys many symmetry properties, several of which we list
in Table 1. Of particular importance to PDE is the relation lies in the fact that it

1It is customary to omit the factor of 2 from the Fourier exponent in PDE, in order to
simplify the Fourier multipliers associated to any given PDE; of course, this factor then surfaces
in the Fourier inverion formula. In any event, the factors of 2 make only a negligible impact on
the theory, so much so that some authors choose to abuse notation slightly, and simply omit all
factors of 2 in their arguments.
2Some authors reverse the attribution of these two identities, which are easily shown to
be equivalent. Strictly speaking, Parsevals original identity was for Fourier series, whereas
Plancherels theorem concerned Fourier integrals.

329
330 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

diagonalises any constant coefficient operator P ():

P\
()f () = P (i)f().
Thus differential operators amplify high frequencies and attenuate low frequencies;
integration operators of course do the reverse. Note that if P () is skew-adjoint,
then P (i) is automatically skew-adjoint; this can be shown directly, and also fol-
lows from (A.1). Indeed in this case we have P () = ih(/i) for some real-valued
polynomial h : Rd R.

Table 1. Some operations on functions f (x), and their Fourier

transform. Here x0 , 0 R, f, g SRx (Rd ), R\{0}, P : Rd
C is a polynomial, and f g(x) := Rd f (y)g(x y) dy.

f (x) f()
f (x x0 ) eix0 f()
eix0 f (x) f( 0 )
f (x) f()
f (x/) ||d f()
f g(x) f()g()
f (x)g(x) 1
d f g()
(2)
P ()f P (i)f

The Fourier transform can be extended to Lebesgue spaces such as L2x (Rd )
using Plancherels theorem (where it essentially becomes an isometry), and also to
the space of tempered distributions Sx (Rd ) .
An important concept for us shall be that of a Fourier multiplier. If we are
given a locally integrable function m : Rd C of at most polynomial growth, we
can define the associated multiplier m(/i) : Sx (Rd ) Sx (Rd ) via the Fourier
transform by the formula
\ () := m()f()
m(/i)f
or equivalently
Z
1
m(/i)f (x) := m()f()e2ix d.
(2)d Rd

This notation is consistent with that of constant-coefficient differential operators

h(/i). We also have (formally at least) the multiplier calculus
m(/i) = m(/i);
m1 (/i) + m2 (/i) = (m1 + m2 )(/i);
m1 (/i)m2 (/i) = (m1 m2 )(/i).

In particular, Fourier multipliers all (formally) commute with each other. The
function m() is known as the symbol of the operator m(/i). Important examples
of Fourier multipliers include the constant coefficient differential operators h(/i),
the propagators eith(/i) discussed in Section 2.1, and the fractional differentiation
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 331

and integration operators || and hi defined3 for all R, with symbols ||

and hi respectively. This in turn leads to the Sobolev spaces Wxs,p (Rd ) and the
homogeneous Sobolev spaces Wxs,p (Rd ), defined for s R and 1 < p < as the
closure of the Schwartz functions under their respective norms

kf kWxs,p (Rd ) := khis f kLpx (Rd )

and
kf kWxs,p (Rd ) := k||s f kLpx(Rd ) .
Thus these spaces generalise the Lebesgue spaces, which correspond to the cases
s = 0. In the special case p = 2, we write Hxs and Hxs for Wxs,2 and Wxs,2 respectively.
From Plancherels theorem we observe that
1
kf kHxs (Rd ) = khis fkL2 (Rd )
(2)d/2
and similarly
1
kf kH s (Rd ) = k||s fkL2 (Rd ) .
x (2)d/2
Using Calderon-Zygmund theory (see e.g. [Stei]), one can show the identities

kf kWxs,p (Rd ) s,p,d kf kWxs1,p (Rd ) + kf kWxs1,p (Rd ) ;

kf kWxs,p (Rd ) s,p,d kf kWxs1,p (Rd )

for any 1 < p < and s R. Iterating the above inequalities, we obtain that
these Sobolev norms are equivalent (up to constants) to their classical counterparts,
thus
k
X
kf kWxk,p (Rd ) k,p,d kj f kLpx(Rd )
j=0

and
kf kWxk,p (Rd ) k,p,d kk f kLpx (Rd ) .
We will not define Sobolev spaces at p = 1 or p = to avoid the technicalities
associated with endpoint Calderon-Zygmund theory.
Another important class of Fourier multipliers are the Littlewood-Paley multi-
pliers. Let us fix a real-valued radially symmetric bump function () adapted to
the ball { Rd : || 2} which equals 1 on the ball { Rd : || 1}; the
exact choice of bump function turns out in practice to not be important4. Define
a dyadic number to be any number N 2Z of the form N = 2j where j Z is
an integer; any sum over the dummy variable N or M is understood to be over

3For d, the operator || is only defined for Schwartz functions which obey enough
moment conditions that their Fourier transform vanishes to high order at the origin. As we shall
never use integration operators of such low order, we shall ignore this technicality.
4In the classical Littlewood-Paley theory (see e.g. [Stei]), one uses the harmonic extension or
heat extension, which would correspond to the (non-compactly-supported) choices () := e||
2
or () := e|| respectively. However in the modern theory it has turned out to be more
convenient to use compactly supported bump functions (but see Section 6.4).
332 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

dyadic numbers unless otherwise specified. For each dyadic number N , we define
the Fourier multipliers
P\
N f () := (/N )f ()

P\
>N f () := (1 (/N ))f ()

Pd
N f () := ((/N ) (2/N ))f ().

We similarly define P<N and PN . Thus PN , PN , P>N are smoothed out projec-
tions to the regions || N , || 2N , || > N respectively. Note in particular the
telescoping identities
X X X
PN f = PM f ; P>N f = PM f ; f = PM f
MN M>N M

for all Schwartz f , where M ranges over dyadic numbers. We also define
X
PM<N := PN PM = PN
M<N N

whenever M N are dyadic numbers. Similarly define PMN , etc.

Littlewood-Paley projections are extremely handy in the rigorous study of PDE,
because they separate (in a quantitative manner) the rough (high-frequency, oscil-
lating, low regularity) components of a solution from the smooth (low-frequency,
slowly varying, high regularity) components, thus clarifying the nature of various
components of the equation, such as derivatives and various nonlinear interactions
of the solution with itself. The following heuristics are quite useful (see Figure 1).
Principle A.1 (Uncertainty principle). Let N be a dyadic number, and let f
be a function on Rd .
(Low frequencies) If f has Fourier transform supported on frequencies of
magnitude || . N (e.g. if f = PN g for some g), then f should be ap-
proximately constant on spatial balls of radius c/N for small c, and s f
should be dominated by N s f for any s 0. (Thus localisation at fre-
quency scales N forces a spatial uncertainty of 1/N ; this is a manifestation
of the Heisenberg uncertainty principle |x | & 1.)
(High frequencies) If f has Fourier transform supported on frequencies
of magnitude || & N (e.g. if f = PN g for some g), then f should
have approximate mean zero5 on balls of radius C/N for large C, and
s f should be dominated by N s f for any s 0. (Thus exclusion of
frequencies at scales N and below forces spatial oscillation at scale 1/N .)
(Medium frequencies) If f has Fourier transform supported on frequencies
of magnitude || N (e.g. if f = PN g for some g), then both of the above
heuristics should apply, and s f should be comparable to N s f for both
positive and negative s.
We now present some concrete estimates that make the above intuition rigorous.
One easily verifies that PN is a convolution operator, in fact
Z
1 y
PN f (x) = d
(y)f (x + ) dy.
(2) Rd N
5In fact, we expect higher moments to vanish as well, so that f should be approximately
orthogonal to any bounded degree polynomials on these balls.
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 333

1/N

P f
<N
x

P g
N x

P>N h

Figure 1. The uncertainty principle. The low-frequency function

P<N f has frequencies less than N and is thus essentially constant
at spatial scales 1/N . The high-frequency function P>N h has
frequencies greater than N and thus oscillates (with mean essen-
tially zero) at spatial scales 1/N . The medium-frequency func-
tion PN g behaves in both fashions simultaneously.

Since is rapidly decreasing and has unit mass, one thus can think of PN as an
averaging operator that blurs f by a spatial scale of O(1/N ), and localises f in
frequency to the ball of radius O(N ), which is consistent with Principle A.1. From
this identity one can easily verify (using Youngs inequality, and the commutativity
of all Fourier multipliers) that the above Littlewood-Paley operators are bounded
(uniformly in N or M ) on every Lebesgue space Lpx (Rd ) with 1 p , as well
as every Sobolev space Wxs,p (Rd ), Wxs,p (Rd ) for s R and 1 < p < . Further-
more, they obey the following easily verified (see Exercise A.1) and extremely useful
Bernstein inequalities for Rd with s 0 and 1 p q :
(A.2) kPN f kLpx (Rd ) .p,s,d N s k||s PN f kLpx (Rd )
(A.3) kPN ||s f kLpx (Rd ) .p,s,d N s kPN f kLpx(Rd )
(A.4) kPN ||s f kLpx (Rd ) p,s,d N s kPN f kLpx(Rd )
d d
(A.5) kPN f kLqx (Rd ) .p,q,d N p q kPN f kLpx(Rd )
d d
(A.6) kPN f kLqx (Rd ) .p,q,d N p q kPN f kLpx(Rd ) .
Thus when the frequency is localised, one can upgrade low Lebesgue integrability to
high Lebesgue integrability, at the cost of some powers of N ; when the frequency
334 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

N is very low, this cost is in fact a gain, and it becomes quite desirable to use
Bernsteins inequality whenever the opportunity arises. These estimates can be
verified by computation of the distributional kernel of PN and PN , and their
derivatives, followed by Youngs inequality. A deeper estimate, requiring some
Calderon-Zygmund theory, is the Littlewood-Paley inequality
X
(A.7) kf kLpx(Rd ) p,d k( |PN f |2 )1/2 kLpx (Rd ) ;
N

see for instance [Stei2]. In a similar spirit, from Plancherels theorem we have the
estimate
X
(A.8) kf kH s (Rd ) s,d ( N 2s kPN f k2L2 (Rd ) )1/2
x x
N
and
X
(A.9) kf kHxs (Rd ) s,d kP1 f kL2x (Rd ) + ( N 2s kPN f k2L2x (Rd ) )1/2
N >1

As a sample application of these estimates, let us present

Lemma A.2 (Hardys inequality). If 0 s < d/2 then
k|x|s f kL2x (Rd ) .s,d kf kH s (Rd ) .
x

Proof. The case s = 0 is trivial, so suppose 0 < s < d/2. Using (A.8) it
suffices to show that
Z X
|f (x)|2
2s
dx .s,d N 2s kPN f k2L2x (Rd ) .
Rd |x| N
We subdivide the left-hand side into dyadic shells and estimate
Z X Z
|f (x)|2 2s
2s
dx . s,d R |f (x)|2 dx
R d |x| |x|R
R
where R ranges over dyadic numbers. Using Littlewood-Paley decomposition and
the triangle inequality, we have
Z XZ
( |f (x)|2 dx)1/2 ( |PN f (x)|2 dx)1/2 .
|x|R N |x|R

On the one hand we have the trivial estimate

Z
( |PN f (x)|2 dx)1/2 kPN f kL2x (Rd )
|x|R

while on the other hand by Bernstein (A.6) and Holder we have

Z
( |PN f (x)|2 dx)1/2 .d Rd/2 kPN f kL d .d (N R)
x (R )
d/2
kPN f kL2x (Rd ) .
|x|R

Combining all these estimates together, we reduce to establishing that

X X X
R2s ( min(1, (N R)d/2 )kPN f kL2x(Rd ) )2 .s,d N 2s kPN f k2L2x (Rd ) .
R N N
s
Writing cN := N kPN f kL2x (Rd ) , this becomes
X
k min((N R)s , (N R)d/2s )cN kl2R (2Z ) .s,d kcN kl2N (2Z )
N
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 335

where 2Z is the space of dyadic numbers. But since 0 < s < d/2, the kernel
min(M s , M d/2s ) is absolutely convergent over dyadic numbers. The claim now
follows from Youngs inequality (or Minkowskis inequality, or Schurs test). 
In a similar spirit we have
Proposition A.3 (Gagliardo-Nirenberg inequality). Let 1 < p < q and
s > 0 be such that
1 1 s
=
q p d
for some 0 < < 1. Then for any u Wxs,p (Rd ) we have
kukLqx(Rd ) .d,p,q,s kuk1
Lp

d kukW s,p (Rd ) .
x (R ) x

In the special case q = , we conclude in particular (by the usual approximation

by Schwartz function argument) that u is in fact continuous (so it lies in Cx0 (Rd )).
Proof. We may of course assume that u is non-zero. The inequality is invari-
ant under homogeneity u(x) 7 u(x) and scaling u(x) 7 u(x/) for any > 0.
Using these invariances we may normalise kukLpx(Rd ) = kukWxs,p(Rd ) = 1.
P
The next step is the Littlewood-Paley decomposition u = N PN u, where N
ranges over dyadic numbers. From the triangle inequality followed by Bernsteins
inequality we have
X
kukLqx(Rd ) kPN ukLqx(Rd )
N
X d d
.d,p,q N p q kPN ukLpx(Rd )
N
X
= N s kPN ukLpx(Rd )
N

On the other hand, from (A.4) and the boundedness of PN have

kPN ukLpx (Rd ) .d,p kukLpx(Rd ) = 1; kPN ukLpx(Rd ) .d,p,s N s k||s ukLpx(Rd ) = N s .
Inserting this into the previous estimate we obtain
X
kukLqx(Rd ) .d,p,s N s min(1, N s ) .,s 1
N
and the claim follows (note that is determined by d, p, q, s). 
Closely related to the above two inequalities is the Hardy-Littlewood-Sobolev
theorem of fractional integration, which asserts that
1
(A.10) kf kLqx (Rd ) .p,q,d kf kLpx (Rd )
|x|
1 1 d
whenever 1 < p < q < and 0 < < d obey the scaling condition p = q + d .
This implies the homogeneous Sobolev embedding
(A.11) kf kLqx(Rd ) .p,q,d kf kWxs,p(Rd )
1 1
whenever 1 < p < q < and s > 0 obey the scaling condition p = q + ds , which in
turn implies the inhomogeneous Sobolev embedding
(A.12) kf kLqx(Rd ) .p,q,s,d kf kWxs,p (Rd )
336 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

whenever 1 < p < q < and s > 0 is such that 1p q1 + ds . We leave the proofs
as exercises. Note that the non-endpoint case p1 < q1 + ds of (A.12) already follows
from Proposition A.3, and we also have an extension to the q = case, namely

kf kCx0 (Rd ) = kf kL d .p,s,d kf kW s,p (Rd )

x (R ) x

1
whenever 1 < p < and s > 0 is such that p < ds . In particular we have

(A.13) kf kCx0 (Rd ) = kf kL d .s,d kf kH s (Rd )

x (R ) x

when s > n/2.

The Sobolev embedding theorem (A.11) is sharp in the following sense: if f
is a rescaled bump function, say f = N (N x) for some Sx (Rd ) and some
N > 0 and R, then one can verify that kf kLqx(Rd ) ,q,d N d/q N and
kf kWxs,p(Rd ) ,s,p,d N s N d/p N , and so from the scaling condition p1 = q1 + ds
we see that both sides of (A.11) are comparable. A useful fact is that these bump
functions are in some sense the only way in which both sides of Sobolev embedding
estimate can be close to comparable. Indeed, we have

Proposition A.4 (Inverse Sobolev theorem). Let 1 < p < q < , s > 0, and
0 < 1.
If p1 = 1q + ds and f is such that kf kWxs,p (Rd ) . 1 and kf kLqx(Rd ) & ,
then there exists a dyadic number N and a position x0 Rd such that
|PN f (x0 )| p,q,d, N d/q , and furthermore
Z
d d
( |PN f (x)|r dx)1/r p,q,d, N q r
|xx0 |C/N

for all 1 r and some large constant C = C(p, q, d, ) > 0.

If p1 < q1 + ds and f is such that kf kWxs,p (Rd ) . 1 and kf kLqx (Rd ) & , then
there exists a dyadic number N p,q,s,d, 1 and a position x0 Rd such
that |PN f (x0 )| p,q,s,d, 1, and furthermore
Z
( |PN f (x)|r dx)1/r p,q,s,d, 1
|xx0 |C

for all 1 r and some large constant C = C(p, q, s, d, ) > 0.

More informally, in order for (A.11) to be close to sharp, f must contain a

large normalised bump function at some position x0 and some frequency N (and
wavelength 1/N ); in order for (A.12) to be sharp, we have a similar conclusion
but with the additional information that the frequency N is comparable to 1. To
put it another way, in order to come within a constant to saturating the Sobolev
embedding theorem, the function must concentrate a significant portion of its Wxs,p
energy in a ball. (See also Lemma B.4, which essentially asserts that if one comes
within an epsilon of the best constant in a Sobolev embedding type theorem, then
one must concentrate nearly all of ones energy in a ball.) The implicit constants
here can be made more explicit, for instance the dependence on is polynomial,
but we will not need such quantitative bounds here. See [BG] for an application
of these types of theorems to nonlinear wave equations.
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 337

Proof. We will just prove the (easier) second half of the theorem here, and
leave the first to Exercise A.7. We have
X
. kf kLqx(Rd ) . kPN f kLqx(Rd ) .
N

Now from (A.6), (A.4), and the hypothesis kf kWxs,p(Rd ) . 1 we have

d d d d
kPN f kLqx (Rd ) .p,q,d,s N p q kPN f kLpx (Rd ) .p,q,d,s N p q min(1, N s ).
P d d
The hypotheses on p, q, s ensure that N N p q min(1, N s ) is geometrically de-
creasing as N 0 or N and is thus convergent. We conclude that there exists
N p,q,s,d, 1 such that kPN f kLqx(Rd ) &p,q,s,d, 1. Since kPN f kLpx (Rd ) .p,q,s,d, 1,
we conclude from Holders inequality that kPN f kL d
x (R )
&p,q,s,d, 1. Thus there
exists x0 Rd such that |PN f (x0 )| &p,q,s,d, 1. Writing PN f = PN/4<<4N PN f ,
we can express PN f (x0 ) as the inner product of PN f with a rapidly decreasing ap-
proximation to the identity centred at x0 . Since we also have kPN f kLpx (Rd ) .d,p,s 1,
an easy application of Holders inequality then gives
Z
|PN f (x)| dx &p,q,s,d, 1
|xx0 |C

for some large C = C(p, q, s, d, ). On the other hand, from Bernsteins inequality
we have kPN f kL d
x (R )
.d,p,s 1. The claim then follows from Holders inequality
again. 
We have seen how Littlewood-Paley technology is useful for understanding lin-
ear operations such as fractional integration. It is also invaluable in understanding
nonlinear operations, such as multiplication (f, g) 7 f g or composition u 7 F (u)
for various explicit functions F . Both of these operations arise constantly in non-
linear PDE, and there are two very useful heuristics that can be used to understand
them:
Principle A.5 (Fractional Leibnitz rule). Let f, g be functions on Rd , and
let D be some sort of differential or pseudodifferential operator of positive order
> 0.
(High-low interactions) If f has significantly higher frequency than g (e.g.
if f = PN F and g = P<N/8 G for some F, G), or is rougher than g (e.g.
f = u and g = u for some u) then f g will have comparable frequency
to f , and we expect D (f g) (D f )g. In a similar spirit we expect
PN (f g) (PN f )g.
(Low-high interactions) If g has significantly higher frequency or is rougher
than f , then we expect f g to have comparable frequency to g. We also
expect D (f g) f (D g), and PN (f g) f (PN g).
(High-high interactions) If f and g have comparable frequency (e.g. f =
PN F and g = PN G for some F, G) then f g should have frequency com-
parable or lower than f , and we expect D (f g) / (D f )g f (D g).
(Full Leibnitz rule) With no frequency assumptions on f and g, we expect
(A.14) D (f g) f (D g) + (D f )g.
Principle A.6 (Fractional chain rule). Let u be a function on Rd , and let
F : R R be a reasonably smooth function (e.g. F (u) = |u|p1 u). Then we
338 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

have the fractional chain rule

(A.15) D (F (u)) F (u)D u
for any differential operator D of positive order > 0, as well as the Littlewood-
Paley variants
P<N (F (u)) F (P<N u)
and
(A.16) PN (F (u)) F (P<N u)PN u.
If F is complex instead of real, we have to replace F (u)D u by Fz (u)D u+Fz D u,
and similarly for (A.16).
Observe that when D is a differential operator of order k, then the heuristics
(A.14), (A.15) are accurate to top order in k (i.e. ignoring any terms which only
differentiate f, g, u k 1 or fewer times). Indeed, the above two principles are
instances of a more general principle:
Principle A.7 (Top order terms dominate). When distributing derivatives, the
dominant terms are usually6 the terms in which all the derivatives fall on a single
factor; if the factors have unequal degrees of smoothness, the dominant term will
be the one in which all the derivatives fall on the roughest (or highest frequency)
factor.
A complete and rigorous treatment of these heuristics (sometimes called parad-
ifferential calculus) is beyond the scope of this text, and we refer the reader to
[Tay2]. We will however give some representative instances of these heuristics in
action.
Lemma A.8 (Product lemma). If s 0, then we have the estimate
(A.17) kf gkHxs (Rd ) .s,d kf kHxs (Rd ) kgkL d + kf kL (Rd ) kgkH s (Rd )
x (R ) x x

for all f, g Hxs (Rd ) L d

x (R ). In particular, if s > d/2, we see from the Sobolev
embedding (A.13) that we have the algebra property
(A.18) kf gkHxs (Rd ) .s,d kf kHxs (Rd ) kgkHxs (Rd ) .
Observe that (A.17) heuristically follows from (A.14), since that latter heuristic
suggests that
his (f g) (his f )g + f (his g)
and the claim then (non-rigorously) follows by taking L2x norms of both sides and
then using the triangle and Holder inequalities.
Proof. The basic strategy with these multilinear estimates is to decompose
using the Littlewood-Paley decomposition, eliminate any terms that are obviously
zero (because of impossible frequency interactions), estimate each remaining com-
ponent using the Bernstein and Holder inequalities, and then sum. One should
always try to apply Bernstein on the lowest frequency possible, as this gives the
most efficient estimates. In some cases one needs to apply Cauchy-Schwarz to
conclude the summation.
6In some cases, there is a special cancellation which allows one to treat the dominant terms
directly. In such cases one often then has to look at the next term in the Taylor expansion in
which all but one derivative falls on one term, and the remaining derivative falls on another. This
phenomenon underlies a number of commutator estimates, such as those discussed in Section 3.9.
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 339

The claim is trivial for s = 0, so assume s > 0. From (A.9) we have

X
(A.19) kf gkHxs (Rd ) .s,d kP1 (f g)kL2x (Rd ) + ( N 2s kPN (f g)k2L2x (Rd ) )1/2 .
N >1

We shall just bound the latter term, and leave the former term to the exercises.
We split7
X
kPN (f g)kL2x (Rd ) . kPN ((P<N/8 f )g)kL2x (Rd ) + kPN ((PM f )gkL2x (Rd ) .
M>N/8

For the first term, observe from Fourier analysis that we may freely replace g by
PN/8<<8N g, and so by Holders inequality
kPN ((P<N/8 f )g)kL2x(Rd ) .d k(P<N/8 f )PN/8<<8N gkL2x (Rd )
X
.d kf kL d
x (R )
kPM gkL2x (Rd )
MN

and so the total contribution of this term to (A.19) is Os,d (kf kL d kgkH s (Rd ) ).
x (R ) x
For the second term, we simply bound
X X
kPN ((PM f )g)kL2x (Rd ) .d k(PM f )gkL2x(Rd )
M>N/8 M>N/8
X
.d kgkL d
x (R )
kPM f kL2x (Rd )
M&N
X
.d,s kgkL d
x (R )
M s kPM f kHxs (Rd )
M&N

and so by Cauchy-Schwarz
X X
(N s kPN ((PM f )g)kL2x (Rd ) )2 .d,s kgk2L d
x (R )
N s M s kPM f k2Hxs (Rd ) .
M>N/8 M&N

Summing this in N (and using the hypothesis s > 0) we see that the total contri-
bution of this term is Os,d (kf kHxs (Rd ) kgkL d ), and we are done.
x (R )

Lemma A.9 (Schauder estimate). Let V be a finite-dimensional normed vector
space, let f Hxs (Rd V ) L d
x (R V ) for some s 0. Let k be the first
k
integer greater than s, and let F Cloc (V V ) be such that F (0) = 0. Then
s d
F (f ) Hx (R V ) as well, with a bound of the form
kF (f )kHxs (Rd ) .F,kf kL (Rd ) ,V,s,d kf kHxs (Rd ) .
x

Note that when F is real analytic, one can deduce this from Lemma A.8; but
the argument below is rather robust and extends to rougher types of function F .
For instance, when s 1 the argument in fact only requires Lipschitz control on F .
The reader should heuristically verify that Lemma A.9 follows immediately from
Principle A.6 in much the same way that Lemma A.8 follows from Principle A.5.

7This is a basic example of a paraproduct decomposition, in which a genuine product such

as f g is split as the sum of paraproducts (combinations of products of Littlewood-Paley pieces).
Paraproducts are usually easier to estimate analytically, especially if derivatives are involved,
because they specifically identify which of the factors is high frequency and which is low frequency,
allowing one to use the flexible estimates (A.2)-(A.6) in a manner adapted to the paraproduct at
hand, instead of relying only on one-size-fits-all tools such as Sobolev embedding.
340 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

The reader may also wish to verify the s = 1 case of this estimate by hand (with
F Lipschitz) in order to get some sense of why this type of estimate should hold.
Proof. The strategy to prove nonlinear is related, though not quite the same
as, that used to prove multilinear estimates. Basically, one should try to split
F (f ) using Taylor expansion into a rough error, which one estimates crudely, and
a smooth main term, which one estimates using information about its derivatives.
Again, one uses tools such as Holder, Bernstein, and Cauchy-Schwarz to estimate
the terms that appear.
k
Let us write A := kf kL d .
x (R )
Since F is Cloc and F (0) = 0, we see that
|F (f )| .F,A,V |f |. This already establishes the claim when s = 0. Applying (A.9),
it thus suffices to show that
X
( N 2s kPN F (f )k2L2x (Rd ) )1/2 .F,A,V,s kf kHxs (Rd ) .
N >1

for all s > 0.

We first throw away a rough portion of F (f ) in PN F (f ). Fix N, s, and split
f = P<N f + PN f . Note that f and P<N f are both bounded by OV,d (A). Now F
k
is Cloc , hence Lipschitz on the ball of radius OV,d (A), hence we have
F (f ) = F (P<N f ) + OF,A,V,d(|PN f |),
and thus
kPN F (f )kL2x (Rd ) .F,A,V,d kPN F (P<N f )kL2x (Rd ) + kPN f kL2x (Rd ) .
To control the latter term, observe from the triangle inequality and Cauchy-Schwarz
that X
N 2s kPN f k2L2 (Rd ) .s (N )s N s kPN f k2L2 (Rd )
x x
N N
and summing this in N and using (A.9) we see that this term is acceptable. Thus
it remains to show that
X
( N 2s kPN F (P<N f )k2L2x (Rd ) )1/2 .F,A,V,s kf kHxs (Rd ) .
N >1

We will exploit the smoothness of P<N f and F by using (A.4) to estimate

(A.20) kPN F (P<N f )kL2x (Rd ) .d,k N k kk F (P<N f )kL2x (Rd ) .
Applying the chain rule repeatedly, and noting that all derivatives of F are bounded
on the ball of radius OV,d (A), we obtain the pointwise estimate
|k F (P<N f )| .F,A,V,d,k sup |k1 (P<N f )| . . . |kr (P<N f )|
k1 +...+kr =k

where r ranges over 1, . . . , k and k1 , . . . , kr range over non-negative integers that

add up to k. We split this up further using Littlewood-Paley decomposition as
X
|k F (P<N f )| .F,A,V,d,k sup |k1 (PN1 f )| . . . |kr (PNr f )|
k1 +...+kr =k 1N ,...,N <N
1 r

where we adopt the convention that PN := PN when N > 1 and P1 := P1 . By

giving up a factor of r! = Ok (1) we may take N1 N2 . . . Nr . where k1 , . . . , kr
range over all positive integers that add up to k. Now from (A.4) we have
ki
kki (PNi f )kL d .d,k N
x (R ) i kf kL d .d,k,A N
x (R )
ki
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 341

for i = 1, . . . , r 1, and similarly

kkr (PNr f )kL kr
d .d,k Nr kPNr f kL2 (Rd )
x (R ) x

and hence we have

kk F (P<N f )kL2x (Rd ) .F,A,V,d,k
X
sup N1k1 . . . Nrkr kPNr f kL2x(Rd ) .
k1 +...+kr =k 1N ...N <N
1 r

Performing the sum in N1 , then N2 , then finally Nr1 , and rewriting N := Nr , we

obtain X
kk F (P<N f )kL2x (Rd ) .F,A,V,d,k (N )k kPN f kL2x(Rd ) .
1N <N
By Cauchy-Schwarz and (A.20) we conclude
X
kPN F (P<N f )k2L2x (Rd ) .F,A,V,d,k (N )k N k kPN f k2L2x (Rd ) .
1N <N

Summing this in N and using (A.9) we see that this term is acceptable (note that
k depends only on s, so the dependence on k is not a concern). 

R In computations
R involving
x
momentum, one often encounters expressions such
as Rd uv dx or Rd u(x)( |x| v). The following lemma is useful for controlling
these quantities.
Lemma A.10 (Momentum estimate). Let u, v S(Rd ) for some d 3, and let
K be a kernel on Rd which is smooth away from the origin, and obeys the estimates
|K(x)| .d 1; |K(x)| .d |x|1
xj
away from the origin. (For instance, we could have K(x) 1, or K(x) |x| for
some j = 1, . . . , d.) Then we have
Z
| u(x)K(x)v(x) dx| .d kukH s (Rd ) kvkHx1s (Rd )
x
Rd
for all 0 s 1 (in particular, the estimate is true for s = 1/2).
Intuitively speaking, the justification for this lemma is that we can integrate
by parts 1/2 times to move half of the derivative from v onto u, ignoring the
mild symbol K in between. By standard limiting arguments we may now extend
R 1/2
the bilinear form (u, v) 7 Rd u(x)K(x)v(x) dx to all u, v Hx (Rd ), dropping
the hypothesis that u, v is Schwartz.
Proof. A standard regularisation argument (replacing K by K for some
approximation to the identity , and then letting 0, taking advantage of the
hypothesis that u, v are Schwartz) allows us to assume that K is smooth on all of
Rd (including the origin), provided of course that our estimates are uniform in K.
By real or complex interpolation it will suffice to establish the estimates
Z
| u(x)K(x)v(x) dx| .d kukL2x (Rd ) kvkH 1 (Rd )
x
Rd
and Z
| u(x)K(x)v(x) dx| .d kukH 1 (Rd ) kvkL2x (Rd ) .
x
Rd
342 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

The first estimate is immediate from Holders inequality (estimating u, v in L2x and
K in L x ). For the second estimate, we integrate by parts (again taking advantage
of the hypothesis that u, v are Schwartz) and use the triangle inequality to estimate
Z Z Z
| u(x)K(x)v(x) dx| | (u)(x)K(x)v(x) dx|+| u(x)(K)(x)v(x) dx|.
Rd Rd Rd
The first term can be estimated by Holders inequality as before. The second term
can be estimated by Cauchy-Schwarz (placing v in L2x ) followed by Lemma A.2
(with s = 1 and d 3), and we are done. 
Exercise A.1. Prove (A.2)-(A.6). (Hint: for each of the estimates, use Fourier
analysis to write the expression in the left-hand norm as the convolution of the
expression in the right-hand norm with some explicit kernel, and then use Youngs
inequality.) Discuss why these estimates are consistent with Principle A.1.
Exercise A.2. Deduce (A.13) directly from the Fourier inversion formula and
Cauchy-Schwarz, and show that it fails at the endpoint s = d/2.
Exercise A.3 (Lorentz characterisation of Lpx ). [KTao] p
P Let f Lx (R ) for
d

some 1 < p < . Show that one can decompose f = k ck k , where k ranges
over the integers, k is a function bounded in magnitude by 1 and supported on a
set
P of kmeasure at most 2k , and ck are a sequence of non-negative reals such that
p 1/p
( k 2 |ck | ) p kf kLpx(Rd ) . (Hint: let f (x) := inf{ : |{|f | > }| < x} be the
(left-continuous) nondecreasing rearrangement of |f |. Set ck to equal f (2k1 ), and
ck k be the portion of f where f (2k ) < |f (x)| f (2k1 ).)
Exercise A.4 (Dual Lorentz characterisation of Lqx ). Let f Lqx (Rd ) for some
1 < q < . Show that
X Z
kf kLqx(Rd ) q sup( 2k(1q) | f (x) dx|q )1/q
Ek Ek
k

where for each k, Ek ranges over all bounded open sets of measure 2k . (Hint: use the
R R 2k
nondecreasing rearrangement again. Show that supEk | Ek f (x) dx| q 0 f (t) dt,
and then decompose the interval [0, 2k ] dyadically.)
Exercise A.5. Use Exercises A.3, A.4 to prove (A.10). (Hint: first establish
the estimate Z
1 d d
| k | .d, min(2 d k 2k , 2 d k 2k )
Ek |x|

for all k, k , where Ek and k are as in the preceding exercises.) Deduce (A.11)
and (A.12) as a consequence.
Exercise A.6 (Lorentz refinement of Sobolev embedding). For 1 < p, q < ,
define the Lorentz norm
X Z
p
k( p)
kf kLq,p d := sup(
x (R )
2 q | f (x) dx|p )1/p
Ek Ek
k
where Ek is as in Exercise A.4. By repeating the proof of Exercise A.5, refine the
estimate (A.11) to
kf kLqx(Rd ) .p,q,d kf kLq,p d .p,q,d kf k
x (R ) Wxs,p (Rd )
under the same hypotheses on p, q, d, s.
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 343

Exercise A.7.R Prove the first half of Proposition A.4. (Hint: first use Exercise
A.6 to show that | Ek f (x) dx| p,q,d, 2k(11/q) for some k Z and some set Ek
of measure 2k . Then perform a Littlewood-Paley decomposition of f to conclude
that kPN f kL d &p,q,d, 2
x (R )
k(11/q)
for some N p,q,d, 2k/d .)
Exercise A.8 (Relationship between Sobolev and isoperimetric inequalities).
Let d > 1. Prove the endpoint Sobolev estimate
d/(d1)
k|f |kL1 (Rd )
|{|f (x)| }| .d x

d/(d1)
for any > 0 and f Sx (R ). (Hint: estimate |f | pointwise by |f | |x|1d1 . Let
d

E := {|f (x)| } and obtain a pointwise bound for 1E |x|1d1 . Alternatively,

use induction on d to establish the slightly stronger estimate kf kLd/(d1)(Rd ) .d
k|f |kL1x(Rd ) .) If Rd is a bounded domain with smooth boundary, deduce the
isoperimetric inequality
|| &d ||(d1)/d
where || is the surface area of . (Hint: set f to be a smoothed out version
of 1 .) It is well known that among all domains with fixed volume, the ball has
the smallest surface area; comment on how this is compatible with the heuristics
supporting Proposition A.4.
Exercise A.9. Give a heuristic justification of Principle A.5 using the Fourier
transform and the elementary estimate h + is .s his + his for all , Rd .
Exercise A.10. Complete the proof of Lemma A.8.
Exercise A.11. Generalise Lemma A.8 by replacing Hxs with Wxs,p for some
1 < p < , and replacing the condition s > d/2 with s > d/p. (Hint: you will need
the Littlewood-Paley estimate (A.7).)
Exercise A.12. Let the assumptions and notation be as in Lemma A.9, but
k+1 k
suppose that F lies in Cloc rather than just in Cloc . Establish the Lipschitz
estimate
kF (f ) F (g)kHxs (Rd ) .F,kf kL (Rd ) ,kgkL (Rd ) ,V,s,d kf gkHxs (Rd ) .
x x

(Hint: One could repeat the proof of Lemma A.9, but a slicker proof is to use the
R1
fundamental theorem of calculus to write F (f ) F (g) = 0 DF ((1 )f + g) (f
g) d, where DF is the differential of F , and then apply Lemma A.9 and Lemma
A.8.)
Exercise A.13 (Fractional chain rule). [CWein] Let p > 1, and let F
1
Cloc (C C) be a function of pth power type, in the sense that F (z) = O(|z|p ) and
F (z) = Op (|z|p1 ). Let 0 s < 1 and 1 < q < r < obey the scaling condition
d dp
q = r (p 1)s. Show that

kPN F (f )kWxs,q (Rd ) .d,p,q,r,s kf kpWxs,r (Rd )

for all f Wxs,r (Rd ) and N > 0. (Note that this is rather easy to justify heuris-
2
tically from Principle A.6.) If furthermore p > 2, and F is Cloc with 2 F (z) =
p2
Op (|z| ), establish the stronger estimate
kPN (F (f )F (g))kWxs,q (Rd ) .d,p,q,r,s (kf kWxs,r (Rd ) +kgkWxs,r (Rd ) )p1 kf gkWxs,r (Rd )
344 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

for all f, g Wxs,r (Rd ) and N > 0. It is possible to remove the PN projections,
and obtain the stronger estimates
kF (f )kWxs,q (Rd ) .d,p,q,r,s kf kpWxs,r (Rd )
and
kF (f ) F (g)kWxs,q (Rd ) .d,p,q,r,s (kf kWxs,r (Rd ) + kgkWxs,r (Rd ) )p1 kf gkWxs,r (Rd )
but this is quite challenging (requiring tools such as the Littlewood-Paley square
function inequality and the Fefferman-Stein vector-valued maximal inequality) and
beyond the scope of this text.
Exercise A.14. If I is an interval in R and 2 q, r < , establish the
inequality
X
kukLqt Lrx (IRd ) .d,q,r ( kPN uk2Lqt Lrx (IRd ) )1/2 .
N
(Hint: use (A.7). To interchange the norm and square function, first consider the
extreme cases q, r = 2, and then interpolate, for instance using the complex
method.)
Exercise A.15. If I is a bounded interval in R, and u, t u L2t (I), use
elementary arguments to obtain a localised Gagliardo-Nirenberg inequality
1/2 1/2
kukL
t (I)
. kukL2(I) kt ukL2 (I)
t t

and the Poincare inequality

Z
1
ku ukL2t (I) . |I|kt ukL2t (I) .
I I

Exercise A.16 (Hardy inequality revisited). Let u Sx (Rd ). Use integration

by parts to establish the identity
Z Z
2 (d + )2
|x| |x u(x)| dx = |x| |u(x)|2 dx
Rd 4 Rd
Z
d+
+ |x| |x u(x) u(x)|2 dx
R d 2
for any > d, and use this to establish another proof of Hardys inequality in the
case s = 1.
Exercise A.17. Give another proof of Lemma A.10 which does not use inter-
polation, but relies instead on Littlewood-Paley decomposition. (Hint: you may
need to decompose K smoothly into dyadic pieces also and use arguments similar
to those used to prove Lemma A.2.) The techniques of interpolation and of dyadic
decomposition are closely related; the latter tends to be messier but more flexible.
Exercise A.18 (Localisation of Hxs functions). Let u Hxs (Rd ) for some s 0,
and let Sx (Rd ). Show that for any R 1 we have
x
ku(x)( )kHxs (Rd ) .s,d, kukHxs (Rd ) .
R
This very useful fact allows one to smoothly localise functions in Hxs to large balls,
uniformly in the size of the ball. (Hint: prove this for s a positive integer by
induction first, and then use interpolation. You may find the Hardy or Sobolev
 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS 345

inequalities to be useful. An alternate approach is to perform a Fourier decompo-

sition of and work entirely in frequency space.) Similarly, if u Hx1 (Rd ) and
d 3, establish the bound
x
ku(x)( )kH 1 (Rd ) .,d kukH 1 (Rd ) .
R x x

Exercise A.19 (Radial Sobolev inequality). Let d 3, and let u be a Schwartz

function on Rd . Establish the inequality
d
k|x| 2 1 |u|kL d .d kuk 1
x (R ) H (Rd )
x

d
for all x R , as well as the variant
k|x|s |u|kL d .d,s kukH 1 (Rd )
x (R ) x

for all d2 1 s d1
2 . (Hint: if |x| = R, truncate u smoothly to the region
|x| R using Exercise A.18, use polar coordinates, and use the Gagliardo-Nirenberg
inequality.)
Exercise A.20. If f is spherically symmetric, show that one can take x0 =
Os,p,q,d, (1/N ) in Proposition A.4; thus Sobolev embedding is only sharp near the
origin (using the natural length scale associated to the frequency). (Hint: if x0 is
too far away from the origin, use the symmetry to find a large number of disjoint
balls, each of which absorb a significant portion of energy.)
Exercise A.21 (Littlewood-Paley characterisation of Holder regularity). Let
0 < < 1 and 1 p . If f Sx (Rd ), we define the Holder norm kf kp (Rd )
by the formula
kf h f kLpx (Rd )
kf kp (Rd ) := kf kLpx(Rd ) + sup
hRd :0<|h|1 |h|

where f h (x) = f (x + h) is the translate of f by h. Show that

kf kp (Rd ) p,,d kf kLpx(Rd ) + sup N kPN f kLpx(Rd ) .
N 1

(Hint: to control the latter by the former, express PN f as an average of functions

of the form f h f . To control the former by the latter, obtain two bounds for
the Lpx norm of PN f h PN f , using the triangle inequality in the high frequency
case N & |h|1 and the fundamental theorem of calculus in the low frequency case
N . |h|1 .) The latter expression is essentially an example of a Besov norm, which
often functions as a substitute for the Sobolev norm which is a little more tech-
nically convenient in several PDE applications, particularly those in which one is
concerned about controlling interactions between high and low frequencies. Con-
clude in particular that
kf kWx,p (Rd ) .p,,d, kf kp (Rd ) .p,,d kf kWx,p(Rd )
for any > 0; thus Holder norms are within an epsilon of their Sobolev counter-
parts.
d
Exercise A.22 (Morrey-Sobolev inequality). If 0 < < 1 and d < p 1 ,
show that
kf k d .p,,d kf k
(R ) Wx1,p (Rd )
346 A. APPENDIX: TOOLS FROM HARMONIC ANALYSIS

for all f Sx (Rd ), where was defined in the previous exercise. This reflects a
general principle, that if there is some surplus regularity in the Sobolev embed-
ding theorem that causes one to go past L x , this additional regularity will then
manifest itself as Holder continuity, and one can again recover endpoint estimates.
Exercise A.23 (Hodge decomposition). Let : Hxs (Rd Rd ) be a vector
field. Show that one has a unique decomposition = cf +df into a curl-free vector
field cf Hxs (Rd Rd ) and a divergence-free vector field df Hxs (Rd Rd ),
thus curlcf = cf = 0 and divdf = df = 0 in the sense of distributions.
Verify the identities cf = 1 ( ) and df = 1 ( ). If s = 0, show
that cf and df are orthogonal. (You may either use the Fourier transform, or
take divergences and curls of the decomposition = cf + df to solve for cf and
df .)
Exercise A.24 (Div-curl lemma). Let , : LR2x (Rd Rd ) be vector fields
such that div = 0 and curl = 0. Show that Rd = 0, and also that
d/2
Hx (Rd ); this is a simple example of a div-curl lemma, that exploits
a certain high-high frequency cancellation between divergence-free and curl-free
vector fields, and forms a key component of the theory of compensated compactness;
see for instance [CLMS]. (Note that Holders inequality would place in L1x ,
d/2
which is not enough for Sobolev embedding to place into Hx 3. To prove the
lemma, use Hodge theory to write and as the curl and gradient of a Hx1 2-form
and scalar field respectively, then use Littlewood-Paley decomposition.)
Exercise A.25 (Sobolev trace lemma). Let f Sx (Rd ) for d 2, and view
d1
R Rd1 {0} as a subset of Rd in the usual manner. Show that
kf kH s (Rd1 ) .d,s kf kH s+1/2 (Rd )
x x

and
kf kHxs (Rd1 ) .d,s kf kH s+1/2 (Rd )
x

for all s > 0. (This can be done from the Fourier transform; it is also worthwhile to
try to prove this from Littlewood-Paley theory, following the heuristics in Principle
A.1.) Show that these estimates fail when s = 0, and also that the loss of 1/2 a
derivative cannot be reduced. One can of course generalise this lemma to other
subsets of Rd of various codimension, and other Sobolev spaces, but we shall not
do so here.
Exercise A.26 (Agmon division lemma). Let f Sx (R) be such that f (0) = 0.
Show that kf (x)/xkHxs1 (R) .s kf kHxs (R) for all s > 1/2. (Hint: write f (x)/x =
R1
0
f (tx) dt, then take Fourier transforms.)
 Bibliography

[AF] M. Ablowitz, A. Fokas, The inverse scattering transform for the Benjamin-Ono equation, a
pivot for multidimensional problems, Stud. Appl. Math. 68 (1983), 110.
[Ale] A. D. Alexandrov, Uniqueness theorems for surfaces in the large, V. Vestnik Leningrad
Univ. Mat. Mekh. Astronom. 13 (1958), 58; Amer. Math. Soc. Transl. Ser 2. 21 (1962),
412416.
[Arn] V.I. Arnold, Mathematical methods in classical mechanics (3rd edn). Nauka, Moscow, 1974.
English translation, 2nd edn, Springer-Verlag, Berlin, 1989.
[BZS] J. Baez, Z. Zhou, I.E. Segal, The global Goursat problem and scattering for nonlinear wave
equations, J. Func. Anal. 93 (1990), 239296.
[BG] H. Bahouri, P. Gerard, High frequency approximation of solutions to critical nonlinear wave
equations, Amer. J. Math. 121 (1999), 131175.
[BS] H. Bahouri, J. Shatah, Decay estimates for the critical semilinear wave equation, Ann. Inst.
H. Poincare Anal. Non Lineaire 15 (1998),783789.
[Bea] M. Beals, Self-Spreading and strength of Singularities for solutions to semilinear wave
equations, Annals of Math 118 (1983), 187214.
[BB] M. Beals, M. Bezard, Low regularity local solutions for field equations, Comm. Partial Dif-
ferential Equations 21 (1996), 79124.
[BT] I. Bejenaru, T. Tao, Sharp well-posedness and ill-posedness results for a quadratic non-linear
Schrodinger equation, to appear, J. Func. Anal.
[Ben] T. Benjamin, Internal waves of permanent form in fluids of great depth, J. Fluid Mech. 29
(1967), 559562.
[BKPSV] B. Birnir, C. Kenig, G. Ponce, N. Svanstedt, L. Vega, On the ill-posedness of the IVP
for the generalized Korteweg-de Vries and nonlinear Schrodinger equations, J. London Math.
Soc. 53 (1996), 551559.
[BPS] B. Birnir, G. Ponce, N. Svanstedt, The local ill-posedness of the modified KdV equation,
Ann. Inst. H. Poincare Anal. Non Lineaire 13 (1996), 529535.
[BC] P. Blue, J. Colliander, Global well-posedness in Sobolev space implies global existence for
weighted L2 initial data for L2 -critical NLS, preprint.
[BSmi] J. Bona, R. Smith, The initial-value problem for the Korteweg-de Vries equation, Philos.
Trans. Royal Soc. London Series A 278 (1975), 555601.
[Bou] J. Bourgain, Fourier restriction phenomena for certain lattice subsets and applications to
nonlinear evolution equations, Parts I, II, Geometric and Funct. Anal. 3 (1993), 107156,
209262.
[Bou2] J. Bourgain, Approximation of solutions of the cubic nonlinear Schrodinger equations by
finite-dimensional equations and nonsqueezing properties, Int. Math. Res. Notices, 1994, no.
2, (1994), 7990.
[Bou3] J. Bourgain, Aspects of longtime behaviour of solutions of nonlinear Hamiltonian evolution
equations, Geom. Func. Anal. 5 (1995), 105140.
[Bou4] J. Bourgain, Periodic nonlinear Schrodinger equations and invariant measures, Comm.
Math. Phys. 166 (1995), 126.
[Bou5] J. Bourgain, Periodic Korteweg de Vries equation with measures as initial data, Selecta
Math. (N.S.) 3 (1997), 115159.
[Bou6] J. Bourgain. Scattering in the energy space and below for 3D NLS, Jour. DAnal. Math.,
75 (1998), 267297.
[Bou7] J. Bourgain, Global well-posedness of defocusing 3D critical NLS in the radial case, J.
Amer. Math. Soc. 12 (1999), 145-171.

363
364 BIBLIOGRAPHY

[Bou8] J. Bourgain, Refinements of Strichartz Inequality and Applications to 2d-NLS With Crit-
ical Nonlinearity, Inter. Math. Res. Not., (1998), p. 253284.
[Bou9] J. Bourgain, New global well-posedness results for non-linear Schrodinger equations, AMS
Publications, 1999.
[Bou10] J. Bourgain, A remark on normal forms and the I-Method for periodic NLS, J. Anal.
Math. 94 (2004), 125157.
[BK] J. Bourgain, V. Kaloshin, On diffusion in high-dimensional Hamiltonian systems, J. Func.
Anal. 229 (2005), 161.
[BW] J. Bourgain, W. Wang, Construction of blowup solutions for the nonlinear Schrodinger
equation with critical non-linearity, Dedicated to Ennio De Giorgi. Ann. Scuola Norm. Sup.
Pisa Cl. Sci. (4) 25 (1997), no. 1-2, 197-215.
[BC] J. P. Boyd, G. Chen, Weakly nonlinear wavepackets in the Korteweg-de Vries equation: The
KdV/NLS connection, Math. Comput. Simulation 56 (2001), 317328.
[BGT] N. Burq, P. Gerard, N. Tzvetkov, An instability property of the nonlinear Schrodinger
equation on S d , Math. Res. Lett. 9 (2002), 323335.
[BP] N. Burq, F. Planchon, On well-posedness for the Benjamin-Ono equation, preprint.
[Car] X. Carvajal, Sharp global well-posedness for a higher order Schrodinger equation, preprint.
[Caz] T. Cazenave, An introduction to nonlinear Schrodinger equations, Textos de Metodos
Matematicos 26, Instituto de Matematica UFRJ, 1996.
[Caz2] T. Cazenave, Semilinear Schrodinger equations, Courant Lecture Notes in Mathematics,
10. New York University, Courant Institute of Mathematical Sciences, AMS, 2003.
[CL] T. Cazenave, Orbital stability of standing waves for some nonlinear Schrodinger equations,
Comm. Math. Phys. 68 (1979), 209243.
[CSS] T. Cazenave, J. Shatah, A.S. Tahvildar-Zadeh, Harmonic maps of the hyperbolic space and
development of singularities in wave maps and Yang-Mills fields, Ann. Inst. H. Poincare Phys.
Theor.68 (1998), 315-349.
[CWeis] T. Cazenave, F.B. Weissler, Some remarks on the nonlinear Schrodinger equation in the
critical case, Nonlinear semigroups, Partial Differential Equations and Attractors, Lecture
Notes in Math. 1394 (1989), 1829.
[CWeis2] T. Cazenave, F.B. Weissler, Critical nonlinear Schrodinger Equation, Non. Anal. TMA,
14 (1990), 807836.
[CWY] S.Y.A. Chang, L. Wang, P. Yang, Regularity of Harmonic Maps, Comm. Pure. Appl.
Math. 52 (1999), 10991111.
[Chi] H. Chihara, Local existence for semilinar Schrodinger equations, Math. Japon. 42 (1995),
3552.
[CCT] M. Christ, J. Colliander, T. Tao, Asymptotics, frequency modulation, and low-regularity
illposedness of canonical defocussing equations, Amer. J. Math. 125 (2003), 12351293.
[CCT2] M. Christ, J. Colliander, T. Tao, Ill-posedness for nonlinear Schrodinger and wave equa-
tions, to appear, Annales IHP.
[CCT3] M. Christ, J. Colliander, T. Tao, Instability of the periodic nonlinear Schrodinger equa-
tion, preprint.
[CKis] M. Christ, A. Kiselev, Maximal operators associated to filtrations, J. Funct. Anal. 179
(2001).
[CWein] M. Christ, M. Weinstein, Dispersion of small amplitude solutions of the generalized
Korteweg-de Vries equation, J. Funct. Anal. 100 (1991), 87109.
[Chr] D. Christodoulou, Global solutions of nonlinear hyperbolic equations for small initial data,
Comm. Pure Appl. Math, 39 (1986), 267282.
[CKla] D. Christodoulou, S. Klainerman, Asymptotic properties of linear field equations in
Minkowski space, Comm. Pure Appl. Math. (1990), 137199.
[CT] D. Christodoulou, A. Tahvildar-Zadeh, On the regularity of spherically symmetric wave
maps, Comm. Pure Appl. Math, 46 (1993), 10411091.
[CT2] D. Christodoulou, A. Tahvildar-Zadeh, On the asymptotic behavior of a spherically sym-
metric wave map, Duke Math. J. 71 (1993), 3169.
[CLMS] R. Coifman, P.L. Lions, Y. Meyer, S. Semmes, Compensated compactness and Hardy
spaces, J. Math. Pure Appl. 72 (1993), 247286.
[Col] J.D. Cole, On a quasi-linear parabolic equation occurring in aerodynamics, Quart. Appl.
Math. 9 (1951), 225236.
 BIBLIOGRAPHY 365

[CDKS] J. Colliander, J. Delort, C. Kenig, G. Staffilani, Bilinear Estimates and applications to

2D NLS, Trans. Amer. Math. Soc. 353 (2001), 33073325.
[CKSTT] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Global well-posedness for
KdV in Sobolev spaces of negative index, Electron. J. Diff. Eq. 2001 (2001) No 26, 17.
[CKSTT2] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Sharp global well-posedness
for KdV and modified KdV on R and T, J. Amer. Math. Soc. 16 (2003), 705749.
[CKSTT3] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Multilinear estimates for
periodic KdV equations, and applications, preprint 2001, J. Funct. Anal. 211 (2004), 173218.
[CKSTT4] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Global well-posedness for
Schrodinger equations with derivative, Siam J. Math. 33 (2001), 649669.
[CKSTT5] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, A refined global well-
posedness for the Schrodinger equations with derivative, SIAM J. Math. Anal. 34 (2002),
6486.
[CKSTT6] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Almost conservation laws
and global rough solutions to a nonlinear Schrodinger equation, Math Res. Lett. 9 (2002),
5-6, 659682.
[CKSTT7] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Existence globale et diffu-
sion pour lequation de Schrodinger nonlineaire repulsive cubique sur R3 en dessous lespace
denergie , Journees Equations aux Derivees Partielles (Forges-les-Eaux, 2002), Exp. No.
X, 14, 2002.
[CKSTT8] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Polynomial upper bounds
for the orbital instability of the 1D cubic NLS below the energy norm, Discrete Contin. Dyn.
Syst. 9 (2003), 3154.
[CKSTT9] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Polynomial growth and
orbital instability bounds for L2 -subcritical NLS below the energy norm, Commun. Pure Appl.
2 (2003), 3350.
[CKSTT10] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Scattering for the 3D cubic
NLS below the energy norm, Comm. Pure Appl. Math. 57 (2004), 9871014.
[CKSTT11] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Global well-posedness and
scattering in the energy space for the critical nonlinear Schrodinger equation in R3 , to appear,
Annals Math.
[CKSTT12] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Symplectic nonsqueezing of
the KdV flow , to appear, Acta Math.
[CKSTT13] J. Colliander, M. Keel, G. Staffilani, H. Takaoka, T. Tao, Growth of higher Sobolev
norms for a periodic nonlinear Schrodinger equation, preprint.
[CST] J. Colliander, G. Staffilani, H. Takaoka, Global well-posedness of the KdV equation below
L2 , Math Res. Letters 6 (1999), 755-778.
[Cof] C.V. Coffman, Uniqueness of the ground state solution for uu+u3 = 0 and a variational
characterization of other solutions, Arch. Rat. Mech. Anal. 46 (1972), 8195.
[CS] P. Constantin, J.C. Saut, Effets regularisants locaux pour des equations disperives generales,
C. R. Acad. Sci. Paris. Ser. I. Math. 304 (1987), 407410.
[DG] P. DAncona, V. Georgiev, On the continuity of the solution operator of the wave maps
system, Comm. Pure Appl. Math. 57 (2004), 357383.
[EL] J. Eells, L. Lemaire, A report on harmonic maps, Bull. London Math. Soc. 10 (1978), no.
1, 168.
[EL2] J. Eells, L. Lemaire, Another report on harmonic maps, Bull. London Math. Soc. 20 (1988),
no. 5, 385524.
[ES] J. Eells, H. Sampson, Harmonic mappings of Riemannian manifolds, Amer. J. Math. 86
(1964), 109160.
[Eva] C. Evans, Partial Differential Equations, Graduate Studies in Mathematics 19, American
Mathematical Society, 1998.
[FW] D. Fang, C. Wang, Some Remarks on Strichartz Estimates for Homogeneous Wave Equa-
tion, preprint.
[Fos] D. Foschi, Inhomogeneous Strichartz estimates, Jour. Hyperbolic Diff. Eqs 2, No. 1 (2005)
124.
[FK] D. Foschi, S. Klainerman, Bilinear space-time estimates for homogeneous wave equations,
Ann. Sci. de lENS, 33 no.2 (2000).
[Fri] A. Friere, Wave maps to compact Lie groups, preprint.
366 BIBLIOGRAPHY

[FPT] G. Furioli, F. Planchon, E. Terraneo, Unconditional well-posedness for semi-linear

Schrodinger equations in H s , Harmonic analysis at Mount Holyoke, (South Hadley, MA,
2001), 147156.
[FT] G. Furioli, E. Terraneo, Besov spaces and unconditional well-posedness for the nonlinear
Schrodinger equation in H s ,Comm. in Contemp. Math. 5 (2003), 349367.
[GLS] V. Georgiev, H. Lindblad, C. D. Sogge, Weighted Strichartz estimates and global existence
for semilinear wave equations, Amer. J. Math. 119 (1997) 12911319.
[GNN] B. Gidas, N.W. Ni, L. Nirenberg, Symmetry of positive solutions of nonlinear elliptic
equations in Rn , Mathematical analysis and applications, Part A, pp. 369402, Adv. in Math.
Suppl. Stud., 7a, Academic Press, New York-London, 1981.
[Gin] J. Ginibre, Le probleme de Cauchy pour des EDP semi-lineaires periodiques en variables
despace (dapres Bourgain), Seminaire Bourbaki 1994/1995, Asterisque 237 (1996), Exp.
796, 163187.
[GO] J. Ginibre, T. Ozawa, Long range scattering for nonlinear Schrodinger and Hartree equa-
tions in space dimension n >= 2, Comm. Math. Phys. 151 (1993) 619645.
[GSV] J. Ginibre, A. Soffer, G. Velo, The global Cauchy problem for the critical non-linear wave
equation, J. Funct. Anal. 110 (1982), 96130.
[GTV] J. Ginibre, Y. Tsutsumi, G. Velo, Existence and uniqueness of solutions for the generalized
Korteweg de Vries equation, Math. Z. 203 (1990) 936.
[GV] J. Ginibre, G. Velo, Smoothing Properties and Retarded Estimates for Some Dispersive
Evolution Equations, Comm. Math. Phys., 123 (1989) , 535573.
[GV2] J. Ginibre, G. Velo, Generalized Strichartz Inequalities for the Wave Equation, Jour. Func.
Anal., 133 (1995), 5068.
[GV3] J. Ginibre, G. Velo, Scattering theory in the energy space for a class of nonlinear
Schrodinger equations, J. Math. Pure. Appl. 64 (1985), 363401.
[GV4] J. Ginibre, G. Velo, The global Cauchy problem for the nonlinear Schrodinger equation
revisited, Ann. Inst. H. Poincare Anal. Non Lineaire 2 (1985),309327.
[GV5] J. Ginibre, G. Velo, Conformal invariance and time decay for nonlinear wave equations
I., Ann. Inst. H. Poincare, Pyhs. Theor. 47 (1987), 221261.
[GV6] J. Ginibre, G. Velo, Scattering theory in the energy space for a class of non-linear wave
equations, Commun. Math. Phys. 123 (1989), 535573.
[GV7] J. Ginibre, G. Velo, Smoothing properties and existence of solutions for the generalized
Benjamin-Ono equation, J. Diff. Eq. 93 (1991),150232.
[Gla] R.T. Glassey, Asymptotic behavior of solutions to certain nonlinear Schrodinger-Hartree
equations, Commun. Math. Phys. 53 (1977), 918.
[Gla2] R.T. Glassey, On the blowing up of solutions to the Cauchy problem for nonlinear
Schrodinger operators, J. Math. Phys. 8 (1977), 17941797.
[Gli] J. Glimm, Solutions in the large for nonlinear hyperbolic systems of equations, Commun.
Pure Appl. Math. 18, (1965), 697715.
[Gri] M. Grillakis, Regularity and asymptotic behaviour of the wave equation with a critical non-
linearity, Ann. of Math. 132 (1990), 485509.
[Gri2] M. Grillakis, Regularity for the wave equation with a critical nonlinearity, Commun. Pure
Appl. Math., 45 (1992), 749774.
[Gri3] M. Grillakis, A priori estimates and regularity of nonlinear waves, in Proceed. Inter. Con-
gress of Math. 1994, Birkhauser, 1187 - 1194.
[Gri4] M. Grillakis, Classical solutions for the equivariant wave map in 1 + 2 dimensions, to
appear in Indiana Univ. Math. J.
[Gri5] M. Grillakis, On nonlinear Schrodinger equations. , Commun. Partial Differential Equa-
tions 25 (2000), no. 9-10, 18271844.
[Gri6] M. Grillakis, personal communication.
[GSS] M. Grillakis, J. Shatah, W. Strauss, Stability theory of solitary waves in the presence of
symmetry I, J. Funct. Anal. 74 (1987), 160197.
[GSS2] M. Grillakis, J. Shatah, W. Strauss, Stability theory of solitary waves in the presence of
symmetry II, J. Funct. Anal. 94 (1990), 308348.
[Gro] M. Gromov, Pseudoholomorphic curves in symplectic manifolds, Invent. Math. 82 (1985),
307347.
[Gu] C. Gu, On the Cauchy problem for harmonic maps defined on two-dimensional Minkowski
space, Commun. Pure Appl. Math., 33,(1980), 727737.
 BIBLIOGRAPHY 367

[Har] J. Harmse, On Lebesgue space estimates for the wave equation, Indiana Univ. Math. J. 39
(1990), 229248.
[HN] N. Hayashi, P.I. Naumkin, Asymptotics for large time of solutions to the nonlinear
Schrodinger and Hartree equations, Amer. J. Math. 120 (1998), 369389.
[HO] N. Hayashi, T. Ozawa, Remarks on nonlinear Schrodinger equations in one space dimension,
Diff. Integ. Eq. 7 (1994), 453461.
[Hel] F. Helein, Regularite des applications faiblement harmoniques entre une sur surface et un
sphere, C.R. Acad. Sci. Paris Ser. I Math., 311 (1990), 519524.
[Hel2] F. Helein, Regularite des applications faiblement harmoniques entre une sur face et une
varitee Riemannienne, C.R. Acad. Sci. Paris Ser. I Math., 312 (1991), 591596.
[Hid] K. Hidano, Scattering problem for the Nonlinear Wave Equation in the Finite Energy and
Conformal Charge space, J. Func. Anal. 187 (2001), 274307.
[HSW] N. J. Hitchin, G. B. Segal, R.S. Ward, Integrable systems: twistors, loop groups, and
Riemann surfaces, Oxford University Press, 1999.
[HZ] H. Hofer, E. Zehnder, Symplectic invariants and Hamiltonian dynamics, Birkhauser Ad-
vanced Texts, 1994.
[Hop] E. Hopf, The partial differential equation ut + uux = uxx , Commun. Pure Appl. Math. 3
(1950), 201230.
[Hor] L. Hormander, Lectures on nonlinear hyperbolic differential equations, Mathematiques and
Applications, 26. Springer-Verlag, 1997.
[IK] A. Ionescu, C. Kenig, Global well-posedness of the Benjamin-Ono equation in low-regularity
spaces, preprint.
[Ior] R.J. Iorio Jr., On the Cauchy problem for the Benjamin-Ono equation, Commun. Partial
Diff. Eq. 11 (1986), 10311181.
[Joh] F. John, Blow-up of solutions of non-linear wave equations in three dimensions, Manuscript.
Math. 28 (1979), 235268.
[Jor] K. Jorgens, Das Anfangswertproblem im Grossen fur eine Klasse nichtlinearer Wellengle-
ichungen, Math. Z. 77 (1961), 295308.
[Kap] L. Kapitanski, Some Generalizations of the Strichartz-Brenner Inequality, Leningrad Math.
J., 1 (1990), 693676.
[Kap2] L. Kapitanski, Weak and Yet Weaker Solutions of Semilinear Wave Equations, Commun.
Part. Diff. Eq., 19 (1994), 16291676.
[Kap3] L. Kapitanski, Global and unique weak solutions of nonlinear wave equations Math. Res.
Letters, 1 (1994), 211223.
[KTop] T. Kappeler, P. Topalov, Global well-posedness of KdV in H 1 (T, R), preprint.
[Kat] T. Kato, Quasilinear equations of evolution, with applications to partial differential equa-
tions, Springer Lecture Notes in Math. 448 (1975), 2750.
[Kat2] T. Kato, The Cauchy problem for the Korteweg-de Vries equation, Pitman Research Notes
in Math. 53 (1979), 293-307
[Kat3] T. Kato, On the Korteweg-de Vries equation, Manuscripta Mathematica, 28 (1979) 8999.
[Kat4] T. Kato, Blow-up of solutions of some nonlinear hyperbolic equations, Commun. Pure
Appl. Math. 33 (1980), 501505.
[Kat5] T. Kato, On the Cauchy problem for the (generalized) Korteweg-de Vries equation, Adv.
in Math. Suppl. Stud., Stud. in Appl. Math. 8 (1983), 93128.
[Kat6] T. Kato, On nonlinear Schrodinger equations. Ann. Inst. H. Poincare Phys. Theor. 46
(1987), 113129.
[Kat7] T. Kato, Abstract evolution equations, linear and quasilinear, revisited. Functional analysis
and related topics, 1991 (Kyoto), 103125, Lecture Notes in Math., 1540, Springer, Berlin,
1993.
[Kat8] T. Kato, An Lq,r -theory for nonlinear Schrodinger equations, Spectral and scattering
theory and applications, Adv. Stud. Pure Math., vol. 23, Math. Soc. Japan, Tokyo, 1994,
223238.
[Kat9] T. Kato, On nonlinear Schrodinger equations, II. H s -solutions and unconditional well-
posedness, J. dAnalyse. Math. 67, (1995), 281306.
[Keel] M. Keel, Global existence for critical power Yang-Mills-Higgs equations in R3+1 , Princeton
University Thesis, 1996.
[KTao] M. Keel, T. Tao, Endpoint Strichartz estimates, Amer. J. Math., 120 (1998), 955980.
368 BIBLIOGRAPHY

[KTao2] M. Keel, T. Tao, Local and global well-posedness of wave maps on R1+1 for rough data,
Internat. Math. Res. Not. 21 (1998), 11171156.
[KTao3] M. Keel, T. Tao, Small data blowup for semilinear Klein-Gordon equations, Amer. J.
Math. 121 (1999), 629669.
[KK] C. Kenig, K. Koenig, On the local well-posedness of the Benjamin-Ono and modified
Benjamin-Ono equations, preprint.
[KPV] C. Kenig, G. Ponce, L. Vega, Well-posedness of the initial value problem for the Korteweg-
de Vries equation, J. Amer. Math. Soc. 4 (1991), 323347.
[KPV2] C. Kenig, G. Ponce, L. Vega, Well-posedness and scattering results for the generalized
Korteweg-de Vries equation via the contraction principle, Commun. Pure Appl. Math. 46
(1993), 527560.
[KPV3] C. E. Kenig, G. Ponce, L. Vega, The Cauchy problem for the Korteweg-de Vries equation
in Sobolev spaces of negative indices, Duke Math J. 71 (1993), 121.
[KPV4] C. Kenig, G. Ponce, L. Vega, A bilinear estimate with applications to the KdV equation,
J. Amer. Math. Soc. 9 (1996), 573603.
[KPV5] C. Kenig, G. Ponce, L. Vega, On the ill-posedness of some canonical dispersive equations,
Duke Math. J. 106 (2001), 617633.
[Ker] S. Keraani,On the defect of compactness for the Strichartz estimates of the Schrodinger
equations, J. Diff. Eq. 175, (2001), 353392.
[Kla] S. Klainerman, Uniform decay estimates and the Lorentz invariance of the classical wave
equation, Commun. Pure Appl. Math. 38 (1985), 321332.
[Kla2] S. Klainerman, On the regularity of classical field theories in Minkowski space-time R3+1 ,
Prog. in Nonlin. Diff. Eq. and their Applic., 29, (1997), Birkhauser, 113150.
[Kla3] S. Klainerman, A Commuting Vectorfield Approach To Strichartz Type Inequalities And
Applications To Quasilinear Wave Equations, Internat. Math. Res. Notices (2001), no. 5,
221274.
[Kla4] S. Klainerman, PDE as a unified subject, Geom. Func. Anal., special volume (2000), 279
315.
[KM] S. Klainerman, M. Machedon, Space-time estimates for null forms and the local existence
theorem, Commun. Pure Appl. Math. 46 (1993), no. 9, 12211268.
[KM2] S. Klainerman, M. Machedon, Smoothing estimates for null forms and applications, Duke
Math J. 81 (1995), 99103.
[KM3] S. Klainerman, M. Machedon, Remark on Strichartz-type inequalities. With appendices
by Jean Bourgain and Daniel Tataru. Internat. Math. Res. Notices 5 (1996), 201220.
[KM4] S. Klainerman, M. Machedon, On the regularity properties of a model problem related to
wave maps, Duke Math. J. 87 (1997), 553589.
[KR] S. Klainerman, I. Rodnianski, On the global regularity of wave maps in the critical Sobolev
norm, Internat. Math. Res. Not. 13 (2001), 656677.
[KS] S. Klainerman, S. Selberg, Remark on the optimal regularity for equations of wave maps
type, Commun. Partial Diff. Eq., 22 (1997), 901918.
[KTat] S. Klainerman, D. Tataru, On the optimal regularity for Yang-Mills equations in R4+1 ,
J. Amer. Math. Soc. 12 (1999), 93116.
[KTzv] H. Koch, N. Tzvetkov, On the local well-posedness of the Benjamin-Ono equation in
H s (R), Internat. Math. Res. Not. 26 (2003).
[KTzv2] H. Koch, N. Tzvetkov, Nonlinear wave interactions for the Benjamin-Ono equation,
Internat. Math. Res. Not. 30 (2005), 18331847.
[KdV] D.J. Korteweg, G. de Vries, On the change of form of long waves advancing in a rectangular
canal, and on a new type of long stationary waves, Philos. Mag. 539 (1895), 422443.
[Kri] J. Krieger, Global regularity of wave maps from R3+1 to H2 , Comm. Math. Phys. 238
(2003), 333366.
[Kri2] J. Krieger, Global regularity of wave maps from R2+1 to H2 . Small energy, preprint.
[Kri3] J. Krieger, Stability of spherically symmetric wave maps, preprint.
[KF] S. Kruzhkov, A. Faminskii, Generalized solutions of the Korteweg-de Vries equation (Rus-
sian), Mat. Sb. (N.S.) 120 (1983), 296-425; English translation in Math USSR Sbornik 48
(1984), 391421.
[Kuk] S. Kuksin, On squeezing and flow of energy for nonlinear wave equations, Geom. Func.
Anal. 5 (1995), 668701 .
 BIBLIOGRAPHY 369

[Kuk2] S. Kuksin, Infinite-dimensional symplectic capacities and a squeezing theorem for Hamil-
tonian PDEs, Commun. Math. Phys. 167 (1995), 521552.
[Kuk3] S. Kuksin, Analysis of Hamiltonian PDEs. Oxford Lecture Series in Mathematics and its
Applications, 19. Oxford University Press, Oxford, 2000.
[Kwo] M.K. Kwong, Uniqueness of positive solutions of u u + up = 0 in Rn , Arch. Rat. Mech.
Anal. 105 (1989), 243266.
[LShu] O.A. Ladyzhenskaya, V.I. Shubov, Unique solvability of the Cauchy problem for the equa-
tions of the two dimensional chiral fields, taking values in complete Riemann manifolds, J.
Soviet Math., 25 (1984), 855864. (English Trans. of 1981 Article.)
[Leb] G. Lebeau, Nonlinear optics and supercritical wave equation, Bull. Soc. Roy. Sci. Liege, 70
(2002), 267-306. Hommage a Pascal Laubin.
[Lem] L. Lemaire, Applications harmoniques de surfaces riemanniennes, J. Diff. Geom. 13 (1978).
[LStr] J. Lin, W. Strauss, Decay and scattering of solutions of a nonlinear Schrodinger equation,
Journ. Funct. Anal. 30, (1978), 245263.
[Lin] H. Lindblad, A Sharp Counterexample to Local Existence of Low Regularity Solutions to
Nonlinear Wave Equations, Duke Math J., 72, (1993), 503539.
[LSog] H. Lindblad, C. D. Sogge, On Existence and Scattering with Minimal Regularity for Semi-
linear Wave Equations, Jour. Func. Anal., 130 (1995 ), 357426.
[LSog2] H. Lindblad, C. D. Sogge, Restriction theorems and semilinear Klein-Gordon equations
in (1+3) dimensions, Duke Math. J. 85 (1996), no. 1, 227252.
[MNNO] S. Machihara, M. Nakamura, K. Nakanishi, T. Ozawa, Endpoint Strichartz estimates
and global solutions for the nonlinear Dirac equation, J. Func. Anal. 219 (2005), 120.
[MNO] S. Machihara, K. Nakanishi, T. Ozawa, Small global solutions and the nonrelativistic limit
for the nonlinear Dirac equation, Rev. Mat. Iberoamericana. 19 (2003), 179194.
[MNO2] S. Machihara, K. Nakanishi, T. Ozawa, Nonrelativistic limit in the energy space for
nonlinear Klein-Gordon equations, Math. Ann. 322 (2002), 603621.
[MSW] B. Marshall, W. Strauss, S. Wainger, Lp Lq estimates for the Klein-Gordon equation,
J. Math. Pures Appl. (9) 59 (1980), no. 4, 417440.
[MN] N. Masmoudi, K. Nakanishi, From nonlinear Klein-Gordon equation to a system of coupled
nonlinear Schrdinger equations, Math. Ann. 324 (2002), no. 2, 359389.
[Mat] C. Matheus, Global well-posedness for a NLS-KdV system on T, preprint.
[Mcl] K. McLeod, Uniqueness of positive radial solutions of u + f (u) = 0 in Rn . II. Trans.
Amer. Math. Soc. 339 (1993), no. 2, 495505.
[Mer] F. Merle, Construction of solutions with exactly k blow-up points for the Schrodinger equa-
tion with critical non-linearity, Commun. Math. Phys. 149 (1992), 205214.
[Mer2] F. Merle, Determination of blow-up solutions with minimal mass for non-linear
Schrodinger equation with critical power, Duke Math. J. 69 (1993), 427453.
[Mer3] F. Merle, Asymptotics for L2 minimal blowup solutions of critical nonlinear Schrodinger
equation, Ann. Inst. Henri Poincare 13 (1996), 553565.
[MR] F. Merle, P. Raphael, Blowup dynamic and upper bound on the blowup rate for the critical
nonlinear Schrodinger equation, Ann. Math. 161 (2005), 157222.
[MR2] F. Merle, P. Raphael, Sharp upper bound on the blowup rate for the critical nonlinear
Schrodinger equation, GAFA 13 (2003), 591642.
[MR3] F. Merle, P. Raphael, On universality of blowup profile for L2 critical nonlinear
Schrodinger equation, Invent. Math. 156 (2004), 565672.
[Miu] R. Miura, Korteweg-de Vries equation and generalizations. I. A remarkable explicit non-
linear transformation, J. Mathematical. Phys. 9 (1968), 12021204.
[Miu2] R. Miura, The Korteweg-de Vries Equation: A Survey of Results, SIAM Rev. 18(3) (1976),
412459.
[MST] L. Molinet, J.-C. Saut, N. Tzvetkov, Ill-posedness issues for the Benjamin-Ono and related
equations, SIAM J. Math. Anal. 33 (2001), 982988.
[Mon] S. J. Montgomery-Smith, Time Decay for the Bounded Mean Oscillation of Solutions of
the Schrodinger and Wave Equation, Duke Math J. 19 (1998), 393408.
[Mor] C. Morawetz, Time decay for the nonlinear Klein-Gordon equation, Proc. Roy. Soc. A 306
(1968), 291296.
[MS] S. Muller, M. Struwe, Global existence of wave maps in 1 + 2 dimensions with finite energy
data, Topological Methods in Analysis 7 (1996), 245259.
370 BIBLIOGRAPHY

[MT] T. Muramatu, S. Taoka, The initial value problem for the 1-D semilinear Schrodinger
equation in Besov spaces, J. Math. Soc. Japan 56 (2004), 853888.
[NSU] A. Nahmod, A. Stefanov, K. Uhlenbeck, On the well-posedness of the wave map problem
in high dimensions, Commun. Anal. Geom. 11 (2003), 4983.
[Nak] K. Nakanishi, Scattering Theory for Nonlinear Klein-Gordon Equation with Sobolev Critical
Power, Internat. Math. Research Not. 31 (1999), 31-60.
[Nak2] K. Nakanishi, Energy scattering for non-linear Klein-Gordon and Schrodinger equations
in spatial dimensions 1 and 2, J. Func. Anal. 169 (1999), 201225.
[Nak3] K. Nakanishi, Unique global existence and asymptotic behaviour of solutions for wave
equations with non-coercive critical nonlinearity, Commun. Partial Differential Equations. 24
(1999), 185221.
[Nak3] K. Nakanishi, Local well-posedness and Illposedness in the critical Besov spaces for semi-
linear wave equations with quadratic forms, Funk. Ekvac. 42 (1999), 261-279.
[Nak4] K. Nakanishi, Nonrelativistic limit of scattering theory for nonlinear Klein-Gordon equa-
tions, J. Differential Equations 180 (2002), no. 2, 453470.
[NTT] K. Nakanishi, H. Takaoka, Y. Tsutsumi, Counterexamples to bilinear estimates related with
the KdV equation and the nonlinear Schrodinger equation, Methods Appl. Anal. 8 (2001),
569578.
[Naw] H. Nawa, Asymptotic and limiting profiles of blowup solutions of the nonlinear Schrodinger
equation with critical power, Commun. Pure. Appl. Math. 52 (1999), 193270.
[Obe] D. Oberlin, Convolution estimates for some distributions with singularities on the light
cone, Duke Math. J. 59 (1989), 747757.
[Ono] H. Ono, Algebraic solitary waves in stratified fluids, J. Phys. Soc. Japan 39 (1975), 1082
1091.
[Oza] T. Ozawa, Long range scattering for nonlinear Schrodinger equations in one space dimen-
sion, Commun. Math. Phys. 139 (1991), 479493.
[Oza2] T. Ozawa, On nonlinear Schrodinger equations of derivative type, Indiana Univ. Math. J.
45 (1996), 137163.
[Pec] H. Pecher, Global solutions with infinite energy for the 1-dimensional Zakharov system,
preprint.
[Pec2] H. Pecher, The Cauchy problem for a Schroedinger-Korteweg-de Vries system with rough
data, preprint.
[Pec3] H. Pecher, Rough solutions of a Schroedinger-Benjamin-Ono system, preprint.
[Poh] K. Pohlmeyer, Integrable Hamiltonian systems and interaction through quadratic con-
straints, Commun. Math. Phys., 46 (1976), 207221.
[Pon] G. Ponce, On the global well-posedness of the Benjamin-Ono equation, Diff. Int. Eq. 4
(1991), 527542.
[PS] G. Ponce, T. Sideris, Local regularity of nonlinear wave equations in three space dimensions,
Commun. Partial Diff. Eq. 18 (1983), 169177.
[Pri] P. Price, A montonicity formula for Yang-Mills fields, Manuscripta Math. 43 (1983), 131
166.
[Rau] J. Rauch, The u5 Klein-Gordon equation. II. Anomalous singularities for semilinear wave
equations in Nonlinear partial differential equations and their applications. College de France
Seminar, Res. Notes in Math. 51, Pitman Publ.,(1981) 335364.
[RR] J. Rauch, M. Reed, Nonlinear microlocal analysis of semilinear hyperbolic systems in one
space dimension, Duke Math. J., 49 (1982), 397475.
[Ros] R. Rosales, I. Exact solution of some nonlinear evolution equations. II. The similarity
solution for the Korteweg-de Vries equation and the related Painleve transcendent, Ph. D.
thesis, California Institute of Technology, 1977.
[RV] E. Ryckman, M. Visan, Global well-posedness and scattering for the defocusing energy-
critical nonlinear Schrodinger equation in R1+4 , preprint.
[SacU] J. Sacks, K. Uhlenbeck, The existence of minimal immersions of 2-spheres, Annals of
Math. (2) 113 (1981), 124.
[Sad] C. Sadosky, Interpolation of Operators and Singular Integrals, Marcel Dekker Inc., 1976.
[Sau] J. C. Saut, Sur des generalisations de lequation de Korteweg-de Vries, Universite de Paris-
Sud Departement de Mathematique, Orsay, (1978).
[Sau2] J. C. Saut, Quelques generalisations de lequation de Korteweg-de Vries. II, J. Differential
Equations 33 (1979), 320335.
 BIBLIOGRAPHY 371

[Sau3] J.C. Saut, Sur quelques generalizations de lequation de Korteweg-de Vries, J. Math. Pures
Appl. 58 (1979), 21-61.
[Schn] G. Schneider, Approximation of the Korteweg-de Vries equation by the non-linear
Schrodinger equation, J. Differential Equations 147 (1998), 333354.
[Scho] R. Schoen, Analytic aspects of the harmonic map problem, Math. Sci. Res. Inst. Publ. 2
(1984), Springer, Berlin, 321358.
[SchU] R. Schoen, K. Uhlenbeck, A regularity theory for harmonic maps, J. Diff. Geom. 17 (1982),
307335.
[SY] R. Schoen, S.T. Yau, Lectures on harmonic maps, Conference proceedings and Lecture notes
in Geometry and Topology, II. International Press.
[Seg] I.E. Segal, Non-linear semi-groups, Ann. of Math. 78 (1963), 339364.
[Seg2] I.E. Segal, The global Cauchy problem for a relativistic scalar field with power interactions,
Bull. Soc. Math. France 91 (1963), 129135.
[Seg3] I. E. Segal, Space-time Decay for Solutions of Wave Equations, Adv. Math., 22 (1976),
304311.
[Sel] S. Selberg, Multilinear space-time estimates and applications to local existence theory for
non-linear wave equations, Princeton University Thesis.
[Sha] J. Shatah, The Cauchy problem for harmonic maps on Minkowski space, in Proceed. Inter.
Congress of Math. 1994, Birkhauser, 11261132.
[Sha2] J. Shatah, Weak solutions and development of singularities of the SU (2) -model. Com-
mun. Pure Appl. Math., 41 (1988), 459469.
[SStra] J. Shatah, W. Strauss, Breathers as homoclinic geometric wave maps, Physics D 99
(1996), 113133.
[SStru] J. Shatah, M. Struwe, Regularity results for non-linear wave equations, Ann. of Math.
138 (1993) 503518.
[SStru2] J. Shatah, M. Struwe, Well Posedness in the energy space for semilinear wave equations
with critical growth, Inter. Math. Research Not., 7 (1994), 303309.
[SStru3] J. Shatah, M. Struwe, Geometric Wave Equations, Courant Lecture Notes in Mathemat-
ics 2 (1998).
[SStru4] J. Shatah, M. Struwe, The Cauchy problem for wave maps, Internat. Math. Res. Not.
11 (2002) 555571.
[ST] J. Shatah, A. Tavildar-Zadeh, Regularity of harmonic maps from the Minkowski space into
rotationally symmetric manifolds., Commun. Pure Appl. Math. 45 (1992), 947971.
[ST2] J. Shatah, A. Tavildar-Zadeh, On the Cauchy problem for equivariant wave maps, Commun.
Pure Appl. Math., 47 (1994), 719753.
[Sid] T. Sideris, Global existence of harmonic maps in Minkowski space, Commun. Pure Appl.
Math., 42 (1989),113.
[Sjo] P. Sjolin, Regularity of solutions to Schrodinger equations, Duke Math. J.,55 (1987), 699
715.
[SSog] H. Smith, C.D. Sogge, Global Strichartz estimates for nontrapping perturbations of the
Laplacian, Commun. PDE 25 (2000), 21712183.
[Sog] C. D. Sogge, Lectures on Nonlinear Wave Equations, Monographs in Analysis II, Interna-
tional Press, 1995.
[Sta] G. Staffilani, On the growth of high Sobolev norms of solutions for KdV and Schrodinger
equations, Duke Math J. 86 (1997), 109142.
[Stef] A. Stefanov, Strichartz estimates for the Schrodinger equation with radial data, Proc. Amer.
Math. Soc. 129 (2001), no. 5, 13951401.
[Stei] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton
University Press, 1970.
[Stei2] E. M. Stein, Harmonic Analysis, Princeton University Press, 1993.
[Stra] W. A. Strauss, Nonlinear wave equations, CBMS Regional Conference Series in Mathemat-
ics, no. 73, Amer. Math. Soc. Providence, RI, 1989.
[Stri] R. S. Strichartz, Restriction of Fourier Transform to Quadratic Surfaces and Decay of
Solutions of Wave Equations, Duke Math. J., 44 (1977), 70 5774.
[Stri2] R. Strichartz, Asymptotic behavior of waves, J. Funct. Anal. 40 (1981), no. 3, 341357.
[Stru] M. Struwe, On the evolution of harmonic mappings of Riemannian surfaces, Comment.
Math. Helv. 60 (1985), 558581.
372 BIBLIOGRAPHY

[Stru2] M. Struwe, Globally regular solutions to the u5 Klein-Gordon equation, Ann. Scuola Norm.
Sup. Pisa Cl. Sci. 15 (1988), 495513.
[Stru3] M. Struwe, Equivariant wave maps in two dimensions, Comm. Pure Appl. Math. 56
(2003), 815823.
[Stru4] M. Struwe, Radially symmetric wave maps from the (1+2)-dimensional Minkowski space
to a sphere, Math Z. 242 (2002), 407414.
[Stru5] M. Struwe, Radially symmetric wave maps from (1+2)-dimensional Minkowski space to
general targets, Calc. Var. 16 (2003), 431437.
[SSul] C. Sulem, P.-L. Sulem, The nonlinear Schrodinger equation: Self-focusing and wave col-
lapse, Springer-Verlag, New York, (1999), p. 25.
[Tak] H. Takaoka, Well-posedness for the one-dimensional Schrodinger equation with the deriv-
ative nonlinearity, Adv. Diff. Eq. 4 (1999), 561580.
[Tao] T. Tao, Low regularity semi-linear wave equations, Commun. Partial Differential Equations,
24 (1999), no. 3-4, 599629.
[Tao2] T. Tao, Spherically averaged endpoint Strichartz estimates for the two-dimensional
Schrodinger equation, Commun. PDE 25 (2000), 14711485.
[Tao3] T. Tao, Ill-posedness of one-dimensional wave maps at the critical regularity, Amer. J.
Math., 122 (2000), 451463.
[Tao4] T. Tao, Multilinear weighted convolution of L2 functions, and applications to non-linear
dispersive equations, Amer. J. Math. 123 (2001), 839908.
[Tao5] T. Tao, Global regularity of wave maps I. Small critical Sobolev norm in high dimension,
Internat. Math. Res. Not. 7 (2001), 299-328.
[Tao6] T. Tao, Global well-posedness of wave maps II. Small energy in two dimensions, Commun.
Math. Phys. 224 (2001), 443544.
[Tao7] T. Tao, Global well-posedness of the Benjamin-Ono equation in H 1 (R), J. Hyperbolic Diff.
Eq. 1 (2004) 2749.
[Tao8] T. Tao, On the asymptotic behavior of large radial data for a focusing non-linear
Schrodinger equation, Dynamics of PDE 1 (2004), 148.
[Tao9] T. Tao, Global well-posedness and scattering for the higher-dimensional energy-critical
non-linear Schrodinger equation for radial data, New York J. Math. 11 (2005), 5780.
[Tao10] T. Tao, Geometric renormalization of large energy wave maps, Journees EDP, Forges les
Eaux, 7-11 June 2004, XI 132..
[Tao11] T. Tao, Recent progress on the Restriction conjecture, to appear, Park City Proceedings.
[Tao12] T. Tao, Spacetime bounds for the energy-critical nonlinear wave equation in three spatial
dimensions, preprint.
[TV] T. Tao, M. Visan, Stability of energy-critical nonlinear Schrodinger equations in high di-
mensions, Electron. J. Diff. Eq. 2005 (2005), No. 118, 128.
[TVZ] T. Tao, M. Visan, X. Zhang, The nonlinear Schrodinger equation with power type nonlin-
earities, preprint.
[Tat] D. Tataru, Local and global results for wave maps I, Commun. PDE 23 (1998), 17811793.
[Tat2] D. Tataru, On global existence and scattering for the wave maps equation, Amer. J. Math.
123 (2001), 3777.
[Tat3] D. Tataru, On u = |u|2 in 5 + 1 dimensions, Math. Res. Letters 6 (1999), 469-485.
[Tat4] D. Tataru, Rough solutions for the wave maps equation, Amer. J. Math. 127 (2005), 293
377.
[Tat5] D. Tataru, The wave maps equation, preprint.
[Tay] M. Taylor, Partial Differential Equations I. Basic Theory, Applied Mathematical Sciences
115, Springer-Verlag, New York, Berlin, Heidelberg 1997.
[Tay2] M. Taylor, Tools for PDE. Pseudodifferential operators, paradifferential operators, and
layer potentials. Mathematical Surveys and Monographs, 81. American Mathematical Society,
Providence, RI, 2000.
[TU] C.L. Terng, K. Uhlenbeck, 1 + 1 wave maps into symmetric spaces, Comm. Anal. Geom. 12
(2004), 343388.
[Tom] M.M. Tom, Smoothing properties of some weak solutions of the Benjamin-Ono equation,
Diff. Int. Eq. 3 (1990) 683694.
[Tomas] P. Tomas, A restriction theorem for the Fourier transform, Bull. Amer. Math. Soc. 81
(1975), 477478.
 BIBLIOGRAPHY 373

[Tsu] Y. Tsutsumi, L2 solutions for nonlinear Schrodinger equations and nonlinear groups, Funk.
Ekva. 30 (1987), 115125.
[Tsu2] Y. Tsutsumi, The Cauchy problem for the Korteweg-de Vries equation with measures as
initial data, SIAM J. Math. Anal. 20 (1989), 582588.
[Tzi] N. Tzirakis, The Cauchy problem for the semilinear quintic Schrodinger equation in one
dimension 18 (2005).
[Tzv] N. Tzvetkov, Remark on the local ill-posedness for the KdV equation, C. R. Acad. Sci. Paris
329 (1999), 10431047.
[Tzv2] N. Tzvetkov, Ill-posedness issues for nonlinear dispersive equations, lecture notes.
[Tzv3] N. Tzvetkov, On the long time behavior of KdV type equations (after Martel-Merle),
Bourbaki seminar March 2004.
[Uhl] K. Uhlenbeck, Connections with Lp bounds on curvature, Commun. Math. Phys 83, (1982)
3142.
[VV] A. Vargas, L. Vega, Global well-posedness for 1D nonlinear Schrodinger equation for data
with an infinite L2 norm, J. Math. Pures Appl. 80 (2001), 10291044.
[Veg] L. Vega, Schrodinger equations: pointwise convergence to the initial data, Proc. Amer.
Math. Soc., 102 (1988), 874878.
[Vil] M. Vilela, Regularity of solutions to the free Schrodinger equation with radial initial data,
Illinois J. Math. 45 (2001), 361370.
[Vil2] M. Vilela, Inhomogeneous Strichartz estimates for the Schrodinger equation, preprint.
[Vis] M. Visan, The defocusing energy-critical nonlinear Schrodinger equation in dimensions four
and higher, Ph.D. thesis, in preparation.
[Wei] M.I. Weinstein, Nonlinear Schrodinger equations and sharp interpolation estimates, Com-
mun. Math. Phys. 87 (1983), 567576.
[Wei2] M. Weinstein, Modulational stability of ground states of nonlinear Schrodinger equations,
SIAM J. Math. Anal. 16 (1985), 472491.
[Wei3] M. Weinstein, Lyapunov stability of ground states of nonlinear dispersive equations,
Comm. Pure Appl. Math. 39 (1986), 51-68.
[Yaj] K. Yajima, Existence of solutions for Schrodinger evolution equations, Commun. Math.
Phys. 110 (1987), 415426.
[Zho] Y. Zhou, Global weak solutions for the 1 + 2 dimensional wave maps into homogeneous
spaces Ann. Inst. Henri Poincare (C) Analyse nonlineaire 16 (1999), 411422.
[Zho2] Y. Zhou, Uniqueness of weak solutions of 1 + 1 dimensional wave maps, Math. Z. 232
(1999), 707719.
[Zho3] Y. Zhou, Local existence with minimal regularity for nonlinear wave equations, Amer. J.
Math, 119 (1997), 671703.