GRADUATE STUDIES

I N M AT H E M AT I C S 182

Nonlinear PDEs
A Dynamical
Systems Approach
Guido Schneider
Hannes Uecker

American Mathematical Society

Nonlinear PDEs
A Dynamical
Systems Approach

Guido Schneider
Hannes Uecker
GRADUATE STUDIES
I N M AT H E M AT I C S 182

Nonlinear PDEs
A Dynamical
Systems Approach

Guido Schneider
Hannes Uecker

American Mathematical Society

Providence, Rhode Island
 EDITORIAL COMMITTEE
Dan Abramovich
Daniel S. Freed (Chair)
Gigliola Staffilani
Jeff A. Viaclovsky

2010 Mathematics Subject Classification. Primary 35-01, 35Bxx, 35Qxx, 37Kxx, 37Lxx.

For additional information and updates on this book, visit

www.ams.org/bookpages/gsm-182

Library of Congress Cataloging-in-Publication Data

Names: Schneider, Guido, 1964– | Uecker, Hannes, 1970–
Title: Nonlinear PDEs : a dynamical systems approach / Guido Schneider, Hannes Uecker.
Other titles: Nonlinear partial differential equations
Description: Providence, Rhode Island : American Mathematical Society, [2017] | Series: Gradu-
ate studies in mathematics ; volume 182 | Includes bibliographical references and index.
Identifiers: LCCN 2017012328 | ISBN 9781470436131 (alk. paper)
Subjects: LCSH: Differential equations, Partial. | Differential equations, Nonlinear. | AMS: Par-
tial differential equations – Instructional exposition (textbooks, tutorial papers, etc.). msc | Partial
differential equations – Qualitative properties of solutions – Qualitative properties of solutions.
msc | Partial differential equations – Equations of mathematical physics and other areas of applica-
tion – Equations of mathematical physics and other areas of application. msc | Dynamical systems
and ergodic theory – Infinite-dimensional Hamiltonian systems – Infinite-dimensional Hamiltonian
systems. msc | Dynamical systems and ergodic theory – Infinite-dimensional dissipative dynamical
systems – Infinite-dimensional dissipative dynamical systems. msc
Classification: LCC QA377 .S35584 2017 | DDC 515/.353–dc23 LC record available at https://
lccn.loc.gov/2017012328

Copying and reprinting. Individual readers of this publication, and nonprofit libraries
acting for them, are permitted to make fair use of the material, such as to copy select pages for
use in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Permissions to reuse
portions of AMS publication content are handled by Copyright Clearance Center’s RightsLink
service. For more information, please visit: http://www.ams.org/rightslink.
Send requests for translation rights and licensed reprints to [email protected].
Excluded from these provisions is material for which the author holds copyright. In such cases,
requests for permission to reuse or reprint material should be addressed directly to the author(s).
Copyright ownership is indicated on the copyright page, or on the lower right-hand corner of the
first page of each article within proceedings volumes.
2017
c by Guido Schneider and Hannes Uecker. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www.ams.org/
10 9 8 7 6 5 4 3 2 1 22 21 20 19 18 17
 To Daniela, Max and Jonas
and Anja, Franka and Henrike
Contents

Preface xi

Chapter 1. Introduction 1
§1.1. The three classical linear PDEs 1
§1.2. Nonlinear PDEs 4
§1.3. Our choice of equations and the idea of modulation equations 6
§1.4. Overview 11

Part I Nonlinear dynamics in Rd

Chapter 2. Basic ODE dynamics 15

§2.1. Linear systems 17
§2.2. Local existence and uniqueness for nonlinear systems 34
§2.3. Special solutions 38
§2.4. ω-limit sets and attractors 49
§2.5. Chaotic dynamics 58
§2.6. Examples 64

Chapter 3. Dissipative dynamics 75

§3.1. Bifurcations 76
§3.2. Center manifold theory 85
§3.3. The Hopf bifurcation 91
§3.4. Routes to chaos 98

Chapter 4. Hamiltonian dynamics 109

vii
viii Contents

§4.1. Basic properties 109

§4.2. Some celestial mechanics 116
§4.3. Completely integrable systems 121
§4.4. Perturbations of completely integrable systems 123
§4.5. Homoclinic chaos 128

Part II Nonlinear dynamics in countably many dimensions

Chapter 5. PDEs on an interval 133
§5.1. From finitely to infinitely many dimensions 134
§5.2. Basic function spaces and Fourier series 151
§5.3. The Chafee-Infante problem 167
Chapter 6. The Navier-Stokes equations 179
§6.1. Introduction 179
§6.2. The equations on a torus 186
§6.3. Other boundary conditions and more general domains 197

Part III PDEs on the infinite line

Chapter 7. Some dissipative PDE models 205
§7.1. The KPP equation 206
§7.2. The Allen-Cahn equation 222
§7.3. Intermezzo: Fourier transform 225
§7.4. The Burgers equation 237
Chapter 8. Three canonical modulation equations 249
§8.1. The NLS equation 250
§8.2. The KdV equation 259
§8.3. The GL equation 275
Chapter 9. Reaction-Diffusion systems 295
§9.1. Modeling, and existence and uniqueness 297
§9.2. Two classical examples 302
§9.3. The Turing instability 307

Part IV Modulation theory and applications

Chapter 10. Dynamics of pattern and the GL equation 315
§10.1. Introduction 316
Contents ix

§10.2. The Swift-Hohenberg equation 319

§10.3. The universality of the GL equation 332
§10.4. An abstract approximation result 337
§10.5. Reaction-Diffusion systems 347
§10.6. Convection problems 354
§10.7. The Couette-Taylor problem 370
§10.8. Attractors for pattern forming systems 378
§10.9. Further remarks 395
Chapter 11. Wave packets and the NLS equation 401
§11.1. Introduction 402
§11.2. Justification in case of cubic nonlinearities 404
§11.3. The universality of the NLS equation 411
§11.4. Quadratic nonlinearities 416
§11.5. Extension of the theory 421
§11.6. Pulse dynamics in photonic crystals 429
§11.7. Nonlinear optics 440
Chapter 12. Long waves and their modulation equations 451
§12.1. An approximation result 452
§12.2. The universality of the KdV equation 456
§12.3. Whitham, Boussinesq, BBM, etc. 465
§12.4. The long wave limit 468
Chapter 13. Center manifold reduction and spatial dynamics 473
§13.1. The center manifold theorem 473
§13.2. Local bifurcation theory on bounded domains 478
§13.3. Spatial dynamics for elliptic problems in a strip 482
§13.4. Applications 484
Chapter 14. Diffusive stability 497
§14.1. Linear and nonlinear diffusive behavior 498
§14.2. Diffusive stability of spatially periodic equilibria 507
§14.3. The critical case 523
§14.4. Phase diffusion equations 529
§14.5. Dispersive dynamics 535
Bibliography 541
x Contents

List of symbols 567

Index 569
Preface

“If you want to build a ship, don’t herd people together to collect wood and
don’t assign them tasks and work, but rather teach them to long for the
endless immensity of the sea”
Antoine de Saint-Exupery

This is an introductory textbook about nonlinear dynamics of partial

differential equations (PDEs), with a focus on problems over unbounded
domains and modulation equations. We explain how dynamical systems
methods can be used to analyze PDEs in order to get more insight into the
real world phenomena behind the equations. Our presentation is example-
oriented and the starting point is very often a real world problem. This
means that new mathematical tools are developed step by step in order to
analyze the equations. They are re-applied and improved in subsequent sec-
tions to handle more and more complicated systems. In the end the reader
should have learned mathematical tools for the analysis of some important
classes of nonlinear PDEs and gained insight into nonlinear dynamics phe-
nomena which may occur in PDEs.
The book is divided into four parts. In order to keep the book as an
introductory text and as self-contained as possible, Part I is an introduction
into finite-dimensional dynamics, defined by ordinary differential equations
(ODEs), including bifurcation theory, attractors, and the basics of Hamil-
tonian dynamics. In Part II we explain the major differences between finitely
and infinitely many dimensions and that in principle a PDE on a bounded
domain is isomorphic to a system of countably many ODEs. We give two
main applications of this point of view. The first one is the characterization
of the attractor for the Allen-Cahn equation on an interval, which is also

xi
xii Preface

known as the Chafee-Infante problem. The second one is a very basic intro-
duction to the Navier-Stokes equations, with a focus on periodic boundary
conditions.
Genuine PDE phenomena such as transport, diffusion, and dispersion
can hardly be understood by the interpretation of PDEs as systems of in-
finitely many ODEs. In Part III we consider PDEs which are posed on
the real line. We start with the linear heat equation, and then turn to
nonlinear problems. For famous model equations such as the Kolmogorov-
Petrovsky-Piskounov or Fisher equation, the Korteweg-de Vries (KdV) equa-
tion, the Nonlinear Schrödinger (NLS) equation, and the Ginzburg-Landau
(GL) equation, we discuss the local existence and uniqueness of solutions,
special solutions as fronts and pulses, their stability and instability, soliton
dynamics, the construction of attractors, and some related results.
The equations from Part III all play an important role in mathematics
and have entire monographs devoted to each. Moreover, they have many
connections to physics and other fields of applications, where they are of-
ten used as simplest possible models for the description of some real world
phenomena. In Part IV we explore these connections from a mathematical
perspective. The scalar equations from Part III occur as asymptotic effective
models, or more specifically as modulation equations, for the more compli-
cated systems from physics considered in Part IV. Examples are pattern
forming systems which can be described by the GL equation, light pulses in
nonlinear optics which can be described by the NLS equation, or long waves
in dispersive systems which can be described by the KdV equation. We dis-
cuss how the dynamics of the reduced model equations transfer to the more
complicated systems. Thus, in Part IV we give a mathematically rigorous
presentation of the formalism of modulation equations in the context of real
world applications. While this last part is close to recent research, it is still
in textbook style, and often we do not prove the sharpest or most general
result possible, but instead refer to the literature for extensions.
All chapters are kept as self-contained as possible, such that the reader
can start to read directly about his or her favorite equation. Having a
good background in linear ODEs, cf. §2.1, a starting point for our goals and
objectives are §2.2-§2.3 about basic nonlinear ODE dynamics combined with
Part III. There are other possible combinations, for instance the sections
about dissipative dynamics or the sections about conservative dynamics.
Nevertheless the reader can also read the book from the beginning to the
end. See the Grasshopper’s Guide on page 12 for detailed proposals. All
chapters contain exercises which we strongly recommend not to skip.
Preface xiii

This book grew out of our manuscripts for the lectures and seminars we
gave about ODEs and PDEs at the universities of Bayreuth, Karlsruhe, Old-
enburg, and Stuttgart. We thank the students who attended our lectures
and seminars and urged us to keep the presentation simple and accessi-
ble. Moreover, we thank all friends and colleagues with whom we coop-
erated over the years, mainly on topics from Part IV, in particular, Dirk
Blömker, Tom Bridges, Kurt Busch, Martina Chirilus-Bruckner, Christo-
pher Chong, Walter Craig, Markus Daub, Hannes de Witt, Arjen Doel-
man, Tomas Dohnal, Wolf-Patrick Düll, Wiktor Eckhaus, Jean-Pierre Eck-
mann, Bernold Fiedler, Thierry Gallay, Dieter Grass, Daniel Grieser, Mark
Groves, Tobias Häcker, Mariana Haragus, Ronald Imbihl, Ralf Kaiser, Tasso
Kaper, Klaus Kirchgässner, Markus Kunze, David Lannes, Vincent Lescar-
ret, Karsten Matthies, Ian Melbourne, Andreas Melcher, Johannes Müller,
Robert Pego, Dmitry Pelinovsky, Jens Rademacher, Björn Sandstede, Arnd
Scheel, Zarif Sobirov, Aart van Harten, C. Eugene Wayne, Daniel Wetzel,
Peter Wittwer, and Dominik Zimmermann. We thank Stefanie Siegert and
the unknown referees for a number of additional proposals to improve the
presentation. Especially we thank Alexander Mielke from whom we learned
about nonlinear dynamics and PDEs.
Finally we would like to thank Ina Mette from the AMS for her never
ending motivation to go on with this book project and many helpful com-
ments to transform our lecture notes into a book.

Guido Schneider and Hannes Uecker

Stuttgart and Oldenburg, February 2017
 Chapter 1

Introduction

Mathematicians want to classify things. However, with partial differential

equations (PDEs) they had to stop on a rather unsatisfactory level. The
reason for this is that almost all rules of theoretical physics and engineering,
and many rules in life sciences and economics, are formulated as ordinary
or partial differential equations (ODEs or PDEs). As different as the ap-
plications of differential equations are, as different is the behavior of their
solutions. Therefore, a mathematical theory which wants to cover all differ-
ential equations can only cover the absolute basics. Hence, the books about
PDEs necessarily differ strongly by the choice of examples and by the choice
of mathematical theory which will be applied to the examples. There are
entire books only covering one special important equation. Before we give
the goals and objectives of this book we start with a short review of three
important examples.

1.1. The three classical linear PDEs

In many courses about PDEs the following three examples, namely the
Laplace equation, the heat equation, and the wave equation, play a ma-
jor, sometimes exclusive, role.
Example 1.1.1. The Laplace equation is an equation for an unknown func-
tion u : Ω → R of two or more variables x = (x1 , . . . , xd ) ∈ Ω ⊂ Rd in terms
of certain of its partial derivatives, namely
(1.1) Δu = 0,
where Δ = ∂x21 +. . .+∂x2d . This PDE plays an important role in mathematics
since the real and imaginary part of an analytic function in the complex plane
satisfy the Laplace equation. It also plays a major role in applications. For

1
2 1. Introduction

instance the potential of an irrotational flow of an incompressible fluid such

as water, or a stationary temperature field, or the potential of a stationary
electric field in the absence of charges in Ω, satisfy this equation.
In order to solve this equation uniquely in a domain Ω, additional condi-
tions are needed. To gain an intuition for the required boundary conditions
we consider the factors which should determine a stationary temperature
field u in a room Ω ⊂ R3 as sketched from the side in Figure 1.1. The
temperature will be determined by the temperature at the walls, the win-
dows, the doors and the heating of the room, mathematically speaking by
the conditions at the boundary ∂Ω of Ω.

Γ1
window
door Γ2
heating

Figure 1.1. Different boundary conditions for the temperature field.

There are mainly two different kinds of boundary conditions. At the

heating unit the temperature has a fixed value, while at a window or wall
heat will go through the window or wall. Mathematically speaking the
boundary ∂Ω = Γ1 ∪ Γ2 is split into two parts where in the first part we
have so called Dirichlet conditions
u|Γ1 = g1 ,
and in the second part we have so called Neumann conditions
∂n u|Γ2 = g2 ,
with given functions g1 : Γ1 → R and g2 : Γ2 → R and n : ∂Ω → R3 the
outer normals.
The Laplace equation is the paradigm of an elliptic PDE. It is of second
order, i.e., the highest derivative is of order two. There is an extensive
theory for elliptic systems, especially for second order elliptic systems. The
equilibrium equation of linear elasticity
!
Lu := μΔu + (λ + μ)∇(∇ · u) = 0,
for the displacement vector u : R3 ⊃ Ω → R3 , with constants λ, μ ∈ R
depending on the material, is also a second order elliptic system. Like
the negative Laplace operator −Δ, the linear operator −L defined in this
equation is an example of a so called elliptic operator. Due to the important
role of elasticity in the construction of cars, bridges, planes, etc., there are
1.1. The three classical linear PDEs 3

well developed numerical schemes such as the finite element method (FEM)
or the boundary element method (BEM), which are available for computing
approximate solutions of such systems. Only in very special cases solutions
can be found analytically.
Example 1.1.2. The heat equation or diffusion equation
(1.2) ∂t u = Δu,
with u = u(x, t) and u : Ω × R+ → R, where t denotes time and x denotes
space, describes the evolution of quantities such as heat, chemical concentra-
tions, or the probability distribution of a particle obeying Brownian motion.
It can be derived as follows. Let V ⊂ Ω be an arbitrary subset with
smooth boundary. The change of the total quantity within V equals the
flux through ∂V , i.e.,
  
d
u dx = −
F, n dS = − ∇ · F dx
dt V ∂V V
with F the flux density,
·, · the scalar product in Rd , n : ∂Ω → Rd again
the outer normal, and where we used the Gauss’ integral theorem. Since
this relation is true for all sets V , we find
∂t u = −∇ · F.
Very often the flux density F is proportional to the gradient ∇u of the
concentration u, i.e., F = −a∇u with a constant a > 0. By rescaling time
we finally come to the diffusion equation (1.2).
In order to solve this equation uniquely in a domain Ω × R+ additional
conditions are needed. As in Example 1.1.1 we need boundary conditions,
but also the temperature field at time t = 0 has to be known, i.e., we need
the initial condition u|t=0 = u0 with u0 : Ω → R. Stationary solutions, i.e.,
time-independent solutions, satisfy the Laplace equation (1.1). The heat
equation is the prototype parabolic PDE. There is an extensive theory for
equations of the form ∂t u = Lu with an elliptic operator −L.
Example 1.1.3. The linear wave equation
(1.3) ∂t2 u = Δu,
u = u(x, t) with u : Ω × R → R, is a simple model for, e.g., oscillations of a
string (Ω ⊂ R) or of a membrane (Ω ⊂ R2 ), or the propagation of light in
vacuum. In order to solve this equation uniquely in a domain Ω × (t0 , t1 ),
t0 < 0 < t1 we again need boundary and initial conditions. Like for scalar
second order ODEs we need two initial conditions, namely u|t=0 = u0 and
∂t u|t=0 = u1 with u0 : Ω → R and u1 : Ω → R. The Dirichlet boundary
condition u|∂Ω corresponds to a membrane which is fixed at the boundary.
In this case, the boundary will reflect the waves.
4 1. Introduction

For the wave equation the eigenmodes play a crucial role. An eigenmode
is a solution u(x, t) = eiωt v(x). This yields the eigenvalue problem
−Δv = ω 2 v.
Such problems play an important role in applications, especially in elasticity
theory, where the evolution equations of linear elasticity
∂t2 u = μΔu + (λ + μ)∇(∇ · u)
yield to the eigenvalue problem
−μΔv + (λ + μ)∇(∇ · v) = ω 2 v.
If Ω is a bounded set then under suitable boundary conditions there are
countably many real eigenvalues λn = ωn2 with 0 ≤ λ1 ≤ λ2 ≤ . . . → ∞.
In the construction of cars, bridges, planes, etc., one has to take care that
these so called resonant modes are not periodically excited. Hence, there is a
big industry using FEM and BEM in order to solve these elliptic eigenvalue
problems. The wave equation is the prototype hyperbolic PDE. There is
an extensive theory for equations of the form ∂t2 u = Lu with an elliptic
operator −L.

For reasons explained below we will focus on other examples than the
three classical ones. The fundamental Examples 1.1.1-1.1.3 cannot be and
will not be avoided. However, they will only occur as subproblems which
will help to understand the nonlinear problems under consideration.

1.2. Nonlinear PDEs

We now start discussing our main objectives for this book, namely an intro-
duction to nonlinear PDEs from a dynamical systems point of view, with a
focus on reduction methods, in particular, the use of amplitude and modu-
lation equations.
Many complications with ODEs or PDEs are due to the fact that the
world is nonlinear. Ultimately, to solve a PDE means to look for solutions
u of an abstract equation F (u) = 0. The problem is called linear if for all
α, β ∈ R we have
F (αu + βv) = αF (u) + βF (v).
As a consequence, for linear problems we have the superposition of solutions.
With u, v being solutions, i.e., F (u) = 0 and F (v) = 0, also αu + βv is a
solution, i.e., F (αu + βv) = 0. Most “real life” problems are nonlinear, i.e.,
in general
F (αu + βv) = αF (u) + βF (v),
1.2. Nonlinear PDEs 5

and therefore a sum of two solutions is no longer a solution of the ODE or

PDE. A simple example of a nonlinear function is F (u) = u2 . As a conse-
quence, the theory of linear algebra is not available, and the set of solutions
in general is more complicated than that for linear problems. In science, for
many decades linear problems played a dominating role. Examples 1.1.1 to
1.1.3 are linear. Next we present two famous examples of nonlinear PDEs.
Example 1.2.1. The Navier-Stokes equations
1
∂t u = Δu − ∇p − (u · ∇)u,
R
0 =∇ · u,
describe the evolution of the velocity field u : Ω × R+ → R3 and the pres-
sure field p : Ω × R+ → R of an incompressible fluid, such as water, in
a domain Ω ⊂ R3 . The Reynolds number R measures the ratio between
inertial and viscous forces, and is in some sense proportional to the com-
plexity of the flow. The global existence and uniqueness of smooth solutions
of the three-dimensional (3D) Navier-Stokes equations is one of the seven
one million dollar or Millennium problems in mathematics presented by the
Clay-Foundation in the year 2000. There are a number of reasons for this
choice. On the one hand, the Navier-Stokes equations describe the motion
of fluids, and the answer to this question would allow us to understand
fluids in a much better way. On the other hand, in mathematics the 3D
Navier-Stokes equations are interesting PDEs, which so far have resisted all
attempts to prove the global existence and uniqueness of solutions. This
will be explained in more detail in Chapter 6.
Example 1.2.2. Maxwell’s equations in a medium, for instance a glass
fiber, are given by
∇ · B =0,
∇ × E + ∂t B =0,
∇ · D =ρ,
∇ × H − ∂t D =J.
Here E : Ω × R → R3 is the electric field, D = ε0 E + P is the displacement
field, with ε0 the electric permeability of vacuum, P : Ω × R → R3 is the
electric polarization of the material, B : Ω × R → R3 is the magnetic field,
H = B/μ0 − M is the magnetizing field, with μ0 the magnetic permeability
of vacuum and M : Ω × R → R3 the magnetic polarization of the material,
ρ is the charge density, and J : Ω → R3 the charge flow density. Since the
first and the third equation above are scalar, while the second and fourth
equation are vector valued, so far we have eight equations for the twelve
unknowns Bj , Ej , Mj and Pj for j = 1, 2, 3. Therefore, these equations
6 1. Introduction

have to be closed with constitutive laws P = P (E, H) and M = M (E, H)

describing the reaction of the material to the electric and magnetic field. In
general, these laws are nonlinear. Moreover, as an additional complication
M and P may depend on the past, cf. §11.7.

The world is instationary, i.e., almost all systems evolve with time.
Typical examples are a vibrating beam, the daily change of the weather, or
the motion of the planets in the solar system. Hence, from the beginning
we will consider nonlinear time-dependent systems.
A mathematical concept which is basic to the analytical understanding
of all ODEs and PDEs is the concept of Dynamical Systems. Until the
beginning of the 1960s, Laplace’s principle that with the knowledge of all
physical rules and the present state of the world, the past and future be-
havior of the world for all times can be computed, was widely accepted as a
relevant philosophical foundation of science. Starting already with the work
of H. Poincaré in the 1890s, cf. [Poi57], this principle was finally observed
to be practically useless at the beginning of the 1960s, for instance by the
work of the meteorologist E. Lorenz in 1963 [Lor63]. He observed with
an analog computer for a three-dimensional model for the weather that the
possible time for predictions goes logarithmically with the precision of the
initial conditions, i.e., that long-time weather-forecasts are practically im-
possible. See Figure 1.2 for an illustration of the so called Lorenz attractor
and of the sensitivity of solutions w.r.t. the initial conditions.
Certain ODEs and PDEs, or, more general, dynamical systems, can be
classified as chaotic. The visualization of chaotic dynamical systems was
in fashion in the 1980s. Famous examples are the Mandelbrot and the Julia
sets. In this book, chaos will not play a central role, but one should keep in
mind its existence already in low-dimensional dynamical systems.

1.3. Our choice of equations and the idea of modulation

equations
PDEs play an important role in modern engineering. With the help of
computer simulations, money can be saved, experiments can be replaced,
and data can be gathered which are not available by classical experiments.
However, a numerical simulation of a PDE requires an analytic
understanding of the PDE. The reason for this is again the wide variety
of different types of PDEs. Therefore, very often the numerical simulation of
a PDE needs an adapted numerical scheme based on an analysis of the PDE.
As the example of the crash of the Sleipner oil platform in 1991 shows, a
misuse of numerical schemes can cost a lot of money. In the concrete example
700 million dollars [JR94].
1.3. Our choice of equations and the idea of modulation equations 7

20

0
xp=0
x =1e−3
p
x =1e−5
p
−20
0 10 20

Figure 1.2. Left: Illustration of the attractor of the Lorenz system

ẋ = σ(y − x), ẏ = ρx − y − xz, ż = −βz + xy by one orbit
in 3D phase space, σ = 10, β = 8/3, ρ = 27. Right: x(t) for
three nearby initial conditions, i.e., x1 (0), x2 (0) = x1 (0) + 10−3 ,
and x3 (0) = x1 (0) + 10−5 , y(0), z(0) always the same. The orbits
behave completely different after a certain time, i.e., the orbits to
x2 (0) and x3 (0) deviate from the unperturbed one after t ≈ 7 and
t ≈ 16, respectively. It can be shown that the prediction time goes
logarithmically with the precision of the initial conditions.

Moreover, computers are fast, but never fast enough. A three-

dimensional body [0, 1]3 discretized with 100 points in each direction leads
to 106 variables. A discretization in 1000 points in each direction yields
109 variables. Therefore, due to practical reasons one has to decide before
what quantities shall be computed. Then the scheme can be adapted to the
computation of these quantities.
We are especially interested in problems which cannot be directly stud-
ied numerically, i.e., where first analysis is needed to reduce the di-
mensionality of the problem. This is for instance the case in so called
spatially extended domains, which means that the wave length of typical
solutions is much smaller than the size of the underlying physical domain.
In this case often the modeling over an unbounded domain is more reason-
able. Then, via a multiple scaling perturbation ansatz simpler models can
be derived to describe the phenomena under consideration. These models,
called modulation equations, belong to the best studied nonlinear PDEs
with a status in some scientific areas similar to the three classical linear
PDEs from above. Besides the study of these basic nonlinear PDEs from
a dynamical systems point of view, one of our main objectives will be the
connections between these models and real world problems by going beyond
the formal derivation of these modulation equations. This will be called the
justification of the reduced models.
8 1. Introduction

Example 1.3.1. The digital transport of information in glass fibers is done

by sending 0s and 1s through the fiber. In most modern technologies the
physical realization of a 1 is an electromagnetic pulsemodulating a carrier
wave with a wave length of a few hundred nanometers. There are a number
of relevant questions related to the transport of information. For instance:
• Which form is optimal for a pulse to travel a long distance?
• Which distance do two pulses initially need in order to stay sepa-
rated during the complete journey through the fiber?
• How many kilometers can a pulse travel without an amplifier?
• How do pulses interact if the carrier waves have different frequen-
cies?
There is dispersion in the fiber and thus in general the energy concentrated
in a pulse will spread. Moreover, the fiber behaves nonlinearly. Hence, the
answers to the above questions are not obvious at all. Numerical simulations,
if possible, are much cheaper than experiments. However, suppose that the
length of the fiber is 100km = 105 m. Then, due to the wave length of light of
approximately 10−7 m, a spatial discretization of Maxwell’s equations in the
fiber gives at least about 1012 points, still neglecting all three-dimensional
effects. This number is too big for a direct numerical simulation.
A modulation equation helps. By perturbation analysis the Nonlinear
Schrödinger (NLS) equation

∂τ A = iν1 ∂ξ2 A + iν2 A|A|2 ,

with A(ξ, τ ) ∈ C, τ ∈ R, ξ ∈ R and coefficients ν1 , ν2 ∈ R, can be derived,

describing the evolution of the envelope A of the pulse alone. On the relevant
time scale the dynamics of the envelope of the pulse and the carrier wave
which behaves linearly can approximately be separated. The properties
of the original system, e.g., the refractive index of the material, and the
underlying wave, condense to the coefficients νj ∈ R. The NLS equation
is a universal modulation equation which describes slow modulations in
time and space of the envelope of a spatially and temporarily oscillating
underlying carrier wave in nonlinear dispersive equations.
The spatial discretization can thus be reduced from 1012 points to appro-
ximately 105 or less points, which is quite manageable for numerical schemes.
Moreover, a number of problems can be solved analytically for the NLS equa-
tion, which is a so called completely integrable system. In particular,
if ν1 ν2 > 0 it has explicit so called soliton solutions. These solitons give
the optimal form of pulses for the transport of information. These questions
will be discussed in detail in Chapter 11.
1.3. Our choice of equations and the idea of modulation equations 9

Example 1.3.2. At the end of the 20th century a new generation of high
speed ferries has caused serious problems, especially those that cross the
Channel between England and France and those operating in the Marl-
borough sound in New Zealand. The waves created by these ferries can
propagate without loss of energy over large distances, and thus retain the
potential to create enormous havoc when they come ashore, and as a conse-
quence of a fatal accident and other damage there are now speed limits for
these ferries [Ham99].
Again a modulation equation gives an idea to understand these phenom-
ena. The Korteweg-de Vries (KdV) equation
∂τ A = ν1 ∂ξ3 A + ν2 A∂ξ A,
with τ ∈ R, ξ ∈ R, A(ξ, τ ) ∈ R and coefficients νj ∈ R can be derived
with the help of a perturbation ansatz. The KdV equation is a universal
modulation equation which describes long waves of small amplitude, where
the original system condenses to the coefficients νj ∈ R.
Like the NLS equation, this famous nonlinear equation possesses soliton
solutions, very robust solitary waves. These waves interact like particles,
i.e., after some nonlinear interaction they reform and look exactly as before
the interaction. This observation, made in the middle of the 1960s, that
solutions of a PDE show simultaneously the behavior of a particle and a
wave, had a big influence on nonlinear science due to the similarity with the
particle-wave dualism in quantum mechanics.
For a long time the KdV equation has also been suggested as a model
for the description of tsunamis, water waves of only a few meters height, but
with a length of up to 100km, i.e., in the ocean they cannot be observed by
eye. In the 5000m deep pacific ocean they move with a very high velocity
of around 700km/h. If they approach land they become slower and steeper,
and cause serious floodings. However, data which is now available from the
tsunami at Christmas 2004 in the Indian Ocean seem to indicate that soliton
dynamics had played at least for this tsunami no role on the open sea. The
validity of the KdV equation will be discussed in Chapter 12.
Example 1.3.3. Since the 1960s, systems near the onset of a finite wave
length instability have been analyzed in detail using modulation equations.
These amplitude modulations describe slow changes in time and space of
the envelope of the finite wave length pattern close to the first instability.
The most famous and generic of such equations is the Ginzburg-Landau
(GL) equation
∂τ A = ν2 ∂ξ2 A + ν0 A + ν3 A|A|2 ,
with τ ≥ 0, ξ ∈ R, A(ξ, τ ) ∈ C and coefficients νj ∈ C. Famous pat-
tern forming systems which can be described with the GL equation are
10 1. Introduction

reaction-diffusion systems such as the Brusselator, and hydrodynamical sta-

bility problems, such as the Couette-Taylor problem, Bénard’s problem, or
Poiseuille flow. A big part of Part IV, namely Chapter 10, is devoted to
the justification of this so called GL approximation for various classes of
systems. We explain and prove that the difference of true solutions of the
pattern forming systems and the associated GL approximations remains
small on the natural time scale of this approximation, and thus prove rigor-
ously that the GL equation makes correct predictions about the dynamics
of the original pattern forming systems.
Instead of ”modulation equation”, in particular the GL equation in the
above context is also called the ”amplitude equation”. Although derived
differently, the GL model also plays a crucial role in superconductivity.
A phenomenological model for pattern formation close to the first insta-
bility of a spatially homogeneous solution is the Swift-Hohenberg equation
[SH77]
(1.4) ∂t u = −(1 + ∂x2 )2 u + αu − u3 ,
with u = u(x, t) ∈ R, x ∈ R, t ≥ 0 and control parameter α ∈ R. This fourth
order scalar PDE is probably the simplest example to apply the “Ginzburg-
Landau formalism”. For small α =: ε2 > 0, plugging the ansatz
(1.5) u(x, t) = εA(εx, ε2 t)eix + c.c.
into (1.4) and sorting w.r.t. powers of ε yields the GL equation
(1.6) 2
∂T A = 4∂X A + A − 3|A|2 A
at order ε3 .

As already said, the mathematical analysis of the approximation by these

three ’generic’ modulation equations, namely the KdV, the NLS, and the
GL equation, will be one of the mathematical objectives of Part IV of this
book. Beside these ’generic’ equations there are many more.
Example 1.3.4. The Burgers equation
∂t u = ∂x2 u − ∂x (u2 ),
with t ≥ 0, x ∈ R, and u(x, t) ∈ R arises for instance as a modulation
equation for small amplitude long waves on the surface of a viscous liquid
running down an inclined plane. It describes this system in case when the
trivial solution, the so called Nusselt solution, which possesses a parabolic
flow profile and a flat top surface, is spectrally stable. This is the case when
the inclination angle θ, which serves as a control parameter in this physical
problem, is below a critical value θc . This model is used for instance for
flood forecasts in rivers.
1.4. Overview 11

If the inclination angle is increased, the Nusselt solution becomes unsta-

ble via a so called sideband instability. Above the threshold of instability
the Kuramoto-Shivashinsky-perturbed KdV equation serves as modulation
equation. After some rescaling it has the form

1
∂t u = −∂x3 u − ∂x (u2 ) − ε(∂x2 + ∂x4 )u,
2
√
with t ≥ 0, x ∈ R, u = u(x, t) ∈ R, and where 0 < ε ≈ θ − θc  1 is a
small parameter. Therefore, complicated dynamics that are present in this
equation occur directly at the first instability of the inclined plane problem.
The dynamics is dominated by traveling pulse trains consisting of unstable
pulses. Time series of the position of the pulses indicate the occurrence of
chaotic dynamics. This situation is relevant for cooling units. Again the 3D
Navier-Stokes equations for the water flowing down the unit is replaced by
a simpler model still containing very complicated dynamics.
Another situation where the Burgers equation arises as a modulation
equations are phase or wave number modulations of stable periodic pattern
in a pattern forming system, while phase (or wave number) modulations of
unstable pattern are generically described by Kuramoto-Shivashinsky type
of equations.

In summary, modulation equations are simpler PDEs, which can be de-

rived by perturbation analysis, and which serve as models for more com-
plicated systems. Hence, modulation equations are a part of mathematical
modeling. In Part IV of this book, the derivation and the approximation
properties of the above equations will be explained. We will analyze the orig-
inal system with the help of the modulation equations. We will explain to
which extent conclusions based on the modulation equations can be proven
to be correct. We will show how mathematics can decide which model of all
possible proposed models is the right one. We will explain that modulation
equations are universal models, i.e., exactly the same modulation equation
describes the same phenomena in completely different physical systems. The
much simpler modulation equations itself will be analyzed in Part III of this
book.

1.4. Overview
In order to keep the book as an introductory text and as self-contained as
possible, in Part I we explain basic dynamical systems concepts for ODEs,
such as phase space, fixed points, periodic solutions, attractors, stability and
instability, bifurcations and amplitude equations.
12 1. Introduction

In Part II we start to transfer the dynamical systems concepts from

finite to infinite dimensions. There are major differences due to the non-
equivalence of norms in infinite-dimensional vector spaces and the loss of
compactness of closed bounded sets. We explain that PDEs over bounded
domains can be considered as dynamical systems with countably many de-
grees of freedom. As applications we discuss the Chafee-Infante problem
and the Navier-Stokes equations.
We have already explained in the previous subsection our choice of equa-
tions for the Parts III and IV. In Part III we consider basic model PDEs
posed on the real line, namely the Kolmogorov-Petrovsky-Piscounov (KPP)
or Fisher equation, the Burgers equation, the Nonlinear Schrödinger (NLS)
equation, the Korteweg-deVries (KdV) equation, and the Ginzburg-Landau
(GL) equation. We explain fundamental PDE phenomena as diffusion, dis-
persion, and transport, discuss local and global existence and uniqueness,
and construct stationary solutions, or traveling front and pulse solutions,
using ODE techniques from Part I. We also give some first results for at-
tractors on unbounded domains and a brief introduction to reaction-diffusion
systems.
Part IV is devoted to the analysis of the more complicated systems with
the help of the scalar model equations from Part III, which now reappear
as modulation equations. Additionally we explain useful concepts such as
diffusive stability and spatial dynamics.
At the end of each chapter we collect a number of exercises. We in
general do not claim any originality for them, and many are taken from the
literature, though in some cases we cannot trace back our source. As usual,
the exercises are a crucial part of this book.

1.4.1. Grasshopper’s Guide. To some extent the four parts of this book
are intended to be independent. Moreover, the chapters are kept as self-
contained as possible, such that the reader may start to read directly about
his or her favorite equation. Therefore, we also give the following guide.
Part I can obviously be read independently of the rest of the book. It is
an example-oriented basic course on finite-dimensional dynamical systems
which together with Chapters 5 and 6 (and possibly Chapter 13) yields a two
semester course about finite- and infinite-dimensional dynamical systems.
Chapters 7 and 8 of Part III can subsequently serve as a basis for a seminar.
An alternative one or two semester course is given by §2.2-§2.3 about
basic nonlinear ODE dynamics combined with (parts or all of) Part III and
some parts of Part IV, for instance the beginning of Chapter 10. Other
chapters of Part IV can then serve as a basis for a seminar.
1.4. Overview 13

There are other possibilities, for instance Chapter 3 about dissipative

ODE dynamics combined with some dissipative PDE dynamics, chosen out
of Chapters 5, 7, §8.3, Chapters 9-10, parts of Chapter 13, and Chapter 14.
Similarly, Chapter 4 about conservative ODE dynamics could be combined
with some conservative PDE dynamics, chosen for instance out of §8.1, §8.2,
and Chapters 11 and 12. If the reader is familiar with the contents of Part I
and Part II and is interested in an introduction to the mathematical theory
of modulation equations, then we recommend to start reading in Part IV
and going back to Part III where needed.
Nevertheless, the reader can also work through the book from the be-
ginning to the end.

1.4.2. Recommended literature. Good classical books about PDEs are

[CH89, Joh91, Eva98, Sal08, Vas15], while [Str92, SVZZ13, Olv14,
Log15a] give more elementary introductions to PDEs. Books which look
at PDEs from a dynamical systems point of view are [Hen81, Tem97,
RR04, Rob01]. These books cover and extend material similar to that
in the first two parts of our book, in particular Part II, while for instance
[SS99b, KP13] discuss in more detail parts of what is treated in our Part
III. For a general background on the functional analytic methods in our book
we recommend [Alt16, Wer00], but the needed material can be found in
most books on functional analysis. For more physically oriented introduc-
tions to PDEs see [Fow97, BK00, TM05, Deb05], for an overview of
developments in the theory of PDEs in the 20th century see [BB98], and
for an encyclopedic work on PDEs see [Tay96]. A “visual approach” to
PDE with many motivating pictures is [Mar07]. For ODEs we refer for
instance to [Chi06, HSD04, Tes12, Log15b]. Beginning in Part II, at
the end of most Chapters we give an outlook and hints for further reading.

1.4.3. Software. There are many software packages for the numerical so-
lution of ODEs and the graphical presentation of solutions. Matlab, Maple,
and Mathematica have built in facilities, and there are various simple to
use Java applets available. We strongly encourage the reader to do own
experiments with any of these programs.
From the above remarks about the very different types of PDEs it readily
follows that there cannot be a general tool for all types of PDEs. However,
tools for specific types of PDEs, both commercial and free are widely avail-
able. We use some short self-written matlab scripts to illustrate some PDE
dynamics, mostly for model problems. However, we do not discuss any
numerical methods behind these programs and refer to [Uec09] and the
references therein. For the computation of so called bifurcation diagrams
we refer to AUTO [Doe07, Dea16] and pde2path [UWR14].
14 1. Introduction

Exercises
1.1. Classify the following PDEs as linear or nonlinear.

d 
d
a) ∂t u = ∂xi ∂xj (aij u)+ ∂xi (bi u), aij , bi : Rd → R smooth functions.
i,j=1 i=1

b) i∂t u = Δu. c) ∂t V = rV − rS∂S V − 12 σ 2 S 2 ∂S2 V , (r, σ ∈ R).

d) ∂t2 u = −∂x4 u. e) ∂t u = Δ(uγ ), (γ > 0).
f) ∂t u = divF (u), F : R → Rd a smooth function. g) ∂t u = ∂x3 u + u∂x u.
1.2. Constant coefficient second order linear partial differential equations in R2 can
be written as
 
Lu = − aij ∂xi ∂xj u + bi ∂xi u + c = 0.
i,j=1,2 i=1,2

The operator L is called elliptic if the eigenvalues of the symmetric matrix A = (aij )
are strictly positive. It is called hyperbolic if they are nonzero, but have different
signs. It is called parabolic if the associated quadratic form (∂x → ξ, ∂y → η)
defines a parabola. Classify
a) 3∂x2 u + 10∂x ∂y u + 15∂y2 u + 36∂x u + 12∂y u + 17 = 0;
b) 3∂x2 u + 4∂x u + ∂y u + 2 = 0.
1.3. Consider the PDE ∂t u = ∂x u for u = u(x, t).
a) Find the general solution for x ∈ R.
b) Solve the PDE for x ∈ (0, 1) with the initial condition u(x, 0) = 1 for
x ∈ (0, 1) under the boundary condition u(1, t) = cos t.
c) Is it possible to solve the PDE for x ∈ (0, 1) with the initial condition
u(x, 0) = 1 for x ∈ (0, 1) and the boundary condition u(0, t) = cos t?
1.4. Consider a membrane Ω = (0, 1)2 which is fixed at the boundary ∂Ω, i.e.,
u|∂Ω = 0.
a) Make an ansatz u(x, y, t) = v(t) sin(mπx) sin(nπy), (n, m ∈ N) for the
solutions of ∂t2 u = Δu. Which equation is satisfied by v?
b) Solve the equation for v with the initial conditions v(0) = 0 and v̇(0) = 1.
c) Sketch for fixed m, n ∈ N the set of (x, y) ∈ Ω, for which u(x, y, t) = 0 for
all t ∈ R.
 Chapter 2

Basic ODE dynamics

The first part of this book is about nonlinear dynamics in Rd . It consists of

this chapter, Chapter 3 about dissipative dynamics, and Chapter 4 about
Hamiltonian dynamics. In this part we provide some basic concepts of non-
linear dynamics. In order to avoid the various functional analytic difficulties
associated with PDEs we restrict to the finite-dimensional situation, i.e., we
consider ODEs

u̇(t) = f (u(t), t),

with u(t) ∈ Rd , f : Rd × I → Rd a continuous vector field which is locally

Lipschitz-continuous w.r.t. its first argument, where I ⊂ R is an open inter-
val, usually I = R, and where u̇(t) denotes the derivative of the function u
w.r.t. time t. In general it is not possible to obtain explicit solutions, and
so our main goal is to provide tools for the understanding of the qualitative
behavior of solutions.
Some notation. The initial value problem consists in finding a solution
of the ODE to an initial value u0 given at an initial time t0 ∈ I. A solution
of the initial value problem is a function u ∈ C 1 (I0 , Rd ) which fulfills the
ODE, where I0 ⊂ I is again an open interval, t0 ∈ I0 , and u(t0 ) = u0 . This
solution is denoted by u(t, t0 , u0 ). If f on the right-hand side of the ODE
does not depend explicitly on time, i.e., f = f (u), then the ODE is called
autonomous, and we may assume t0 = 0 and write u(t, u0 ) for the solution
of the initial value problem.
Absolutely fundamental for the understanding of nonlinear dynamics is
the understanding of the dynamics of linear systems which we therefore con-
sider first. Then we introduce basic concepts of nonlinear dynamics. These

15
16 2. Basic ODE dynamics

are the local and global existence and uniqueness of solutions, special so-
lutions such as fixed points, periodic solutions, homoclinic and heteroclinic
orbits, and further concepts such as stability and instability, invariant mani-
folds, ω-limit-sets, attractors, and chaotic dynamics. Many of these concepts
will later be transferred to nonlinear PDEs. Moreover, the search for special
solutions, such as front or pulse solutions for the PDEs, in later chapters
very often lead to ODE problems as they are considered here.
The behavior and the analysis of an ODE or of a PDE strongly differ
between dissipative and conservative systems. In Chapter 3 we provide the
strategy and the tools to tackle dissipative systems. Such systems are typi-
cally characterized by the existence of compact absorbing sets, i.e., compact
sets into which all solutions finally enter. In dissipative systems very often
more complicated and eventually chaotic dynamics occur through bifurca-
tions if some external parameter is varied. After introducing a number of
elementary bifurcations for one- and two-dimensional systems we introduce
with the Lyapunov-Schmidt reduction and the center manifold theorem two
reduction methods which allow us to find these elementary bifurcations in
higher dimensional systems, too. Chapter 3 is closed by presenting some
routes of bifurcations to chaotic behavior in dissipative systems.
The systems considered in Chapter 3 change the volume in phase space,
but many systems in nature preserve the volume in phase space, especially
those of classical mechanics. Thus, Chapter 4 is devoted to Hamiltonian
ODE dynamics. We provide some tools for their analysis and explain basic
facts about their behavior, which shows fundamental differences compared
to that of the systems of Chapter 3. For instance, a globally attracting fixed
point cannot exist in conservative or volume-preserving systems. Therefore,
stability and instability proofs or the route to chaotic behavior must be
completely different. The starting point of the bifurcation analysis is not a
globally attracting fixed point, but a so called completely integrable system.
In Chapter 4 we also discuss KAM theory which allows to understand the
behavior of systems which are small perturbations of completely integrable
Hamiltonian systems.
The ideas presented in this first part will reappear in subsequent sections.
For instance, Chapter 3 about dissipative ODE dynamics contains basic tools
which will be used in Chapters 5-7, §8.3-Chapter 10, and Chapter 14 about
dissipative PDE dynamics. Similarly, Chapter 4 about conservative ODE
dynamics contains basic tools which will help to understand §8.1, §8.2, and
Chapters 11-12 about conservative PDE dynamics.
We emphazise that the purpose of Part I is not to give a comprehen-
sive overview about ODEs. Rather we present the basic ideas of nonlinear
dynamics as needed in subsequent parts of this book in the analysis of PDEs.
2.1. Linear systems 17

There are a number of excellent textbooks on nonlinear ODE dynamics,

many of them also reviewing basic linear ODE dynamics. An elementary
and very readable account on ODE dynamics and bifurcations is [HK91],
a very applied and example oriented approach is used in [Str94], and an
excellent modern presentation is given in [Tes12]. Alternatives and com-
plements to these textbooks are for instance [Chi06, Ver96, Rob04a,
Rob04b]. More advanced texts include [GH83, KH97, Wig03, HSD04].
In [SH96, Lyn04] discrete dynamical systems and ODEs are treated from a
numerical point of view, and [Dev89] focusses on discrete chaotical dynam-
ical systems. For the bifurcation aspects of ODEs, and in particular center
manifolds, we again refer to [GH83, Wig03], and to [Kuz04, Erm02] for
invariant manifolds from a numerical point of view. Our favorite books on
Hamiltonian systems and KAM theory are [Arn78, Thi88, MH92], see
also [Way96].

2.1. Linear systems

Fundamental for the understanding of nonlinear dynamics is the understand-
ing of linear dynamics. Linear ODEs occur for instance as linearizations
around fixed points or periodic solutions. The solution of these linear ODEs
and the variation of constant formula, which allow us to solve inhomoge-
neous linear problems, will be the basis for stability proofs for fixed points
and periodic solutions of nonlinear ODEs in §2.3. Moreover, this technique
will be generalized to semi-linear dissipative PDEs and a number of conser-
vative PDEs in Parts II-IV for proving the local existence and uniqueness
of solutions of PDEs.
A linear ODE is an equation
(2.1) u̇(t) = A(t)u(t) + g(t)
for an unknown function u ∈ C 1 (I, Rd ), where I ⊂ R is an interval, A(t) ∈
Rd×d is a d×d-matrix with entries aij (t), and where g(t) ∈ Rd is an inhomo-
geneity. We generally think of t as time, and for simplicity we assume that
A and g are at least continuous w.r.t. t. Together with an initial condition
u|t=t0 = u0 ∈ Rd we have an initial value problem. Equation (2.1) is called
homogeneous, if g(t) ≡ 0, i.e., if
(2.2) u̇(t) = A(t)u(t),
and (2.1), respectively (2.2), are called autonomous if A and g in (2.1),
respectively A in (2.2), do not depend on t.
It is well-known that the initial value problems for (2.1) and (2.2) have
unique solutions, and that the solutions of (2.1) form a d-dimensional affine
space and the solutions of (2.2) a d-dimensional vector space. This will
briefly be recalled in §2.1.2. We restrict ourselves to those parts of the
18 2. Basic ODE dynamics

theory which are needed in subsequent sections. That is, we restrict to

A independent of time, and to A periodic in time, i.e., A(t) = A(t + T )
for a T > 0. Other important classes of linear systems are asymptotically
constant systems, i.e., A(t) → A± for t → ±∞. They appear as linearization
around so called homoclinic or heteroclinic orbits.

2.1.1. Notation. Let X be a real or complex vector space. A map · :

X → R is called norm, if for all u, v ∈ X and λ ∈ R, respectively λ ∈ C,
(i) u ≥ 0 and u = 0 if and only if u = 0,
(ii) λu = |λ| u,
(iii) u + v ≤ u + v.

In Rd , the major examples are the 1 -norm u1 = dj=1 |uj |, the Euclidean

or 2 -norm u2 = (uT u)1/2 = ( dj=1 |uj |2 )1/2 , and the ∞ - or maximum-
norm u∞ = maxj=1,...,d |uj |.
Concepts such as convergence in Rd , or later on stability and instability
for ODEs in Rd , are independent of the chosen norm in Rd . The reason for
this is the equivalence of norms in finite-dimensional vector spaces.
Theorem 2.1.1. All norms in Rd are equivalent, i.e., for two norms  · 
and  · ∗ there exist positive constants C1 , C2 such that for all u ∈ Rd we
have
u ≤ C1 u∗ ≤ C2 u.

Proof. Obviously, it is sufficient to establish the estimates between an ar-

bitrary norm · and the ·∞ -norm. By the triangle inequality we have
  ⎛ ⎞
 d 
  d  d
u = 
 uj ej 
≤ |uj |ej  ≤ ⎝ ej ⎠ u∞ .
 j=1  j=1 j=1

For the second estimate let M = {u ∈ Rd : u∞ = 1}. Then f : M →

R, u → u−1 is a continuous map by definition. Suppose that f is un-
bounded on M , i.e., there exists a sequence (un )n∈N with un  → 0 for
n → ∞. Since the finitely many coordinates unj satisfy |unj | ≤ 1 there exists
a convergent subsequence unk → ξ ∈ M for k → ∞. By the continuity of the
norm we have unk  → ξ = 0 for k → ∞ which implies ξ = 0 contradict-
ing ξ ∈ M . Hence, there exists a C > 0 such that supu∈M f (u) = C < ∞
and so u ≥ C1 for all u ∈ M which finally leads to u ≥ C1 u∞ . 
Remark 2.1.2. The reason why we gave a proof of this well known theorem
is that in infinite dimensions there are infinitely many non-equivalent norms.
This has a number of consequences for the subsequent analysis of PDEs. It
is possible that uniqueness but no global existence of solutions is known
2.1. Linear systems 19

in one space, and global existence but no uniqueness of solutions is known

in another space, but there is no space known where both properties hold
simultaneously. So far this is exactly the state of the art for one of the
Millennium problems of the Clay Foundation, namely the global existence
and uniqueness of smooth solutions of the 3D Navier-Stokes equations, which
will be discussed in Chapter 6.

The d × d-matrices form a normed vector space of dimension d2 . With

the matrix multiplication AB they form an algebra. An important matrix
norm is given by the operator norm


∗ Au
A = sup : u ∈ R \ {0} .
d
u
Obviously Au ≤ A∗ u, and so from
ABu ≤ A∗ Bu ≤ A∗ B∗ u
it follows
AB∗ ≤ A∗ B∗ ,
i.e., the matrices form a Banach algebra w.r.t. matrix multiplication and
operator norm. For arbitrary norms the matrix norm and the vector norm
are called compatible if
Au ≤ A u.
Examples of such compatible norms are

d
· = ·1 , A∗ = sup |ajk |,
k=1,...,d j=1
⎛ ⎞1/2

d
· = ·2 , A∗ = ⎝ |ajk |2 ⎠ ,
i,j=1

d
· = ·∞ , A∗ = sup |ajk |.
j=1,...,d k=1

In the following we always use compatible norms, and Rd will be equipped

with the Euclidean norm, if not indicated otherwise. Finally, we remark
that for all norms  t   t
 
 u(τ ) dτ  ≤ u(τ ) dτ.
 
t0 t0

2.1.2. Local existence and uniqueness. We briefly recall that the initial
value problem for (2.2) has a unique solution, and that the solutions of
(2.2) form a d-dimensional vector space. The following local existence and
uniqueness result and many other results in this book are based on the
contraction mapping principle which is absolutely fundamental in nonlinear
20 2. Basic ODE dynamics

analysis. We recall that a metric space M is called complete, if every Cauchy

sequence in M possesses a limit in M .
Theorem 2.1.3. (Contraction mapping principle or fixed point the-
orem of Banach) Let (M, d) be a complete metric space and F : M → M
a contraction, i.e., there exists a κ ∈ (0, 1) such that
d(F (x), F (y)) ≤ κd(x, y)
for all x, y ∈ M . Then F has a unique fixed point x∗ ∈ M , i.e., x∗ = F (x∗ ).
Proof. We first prove the uniqueness. Suppose that there exist two different
fixed points x∗ and y ∗ . Then
d(x∗ , y ∗ ) = d(F (x∗ ), F (y ∗ )) ≤ κd(x∗ , y ∗ ).
Since κ ∈ (0, 1), it follows that d(x∗ , y ∗ ) = 0 and hence x∗ = y ∗ , in contra-
diction to the assumption.
We define the sequence xn+1 = F (xn ) with x0 ∈ M arbitrary, but fixed.
Then for m ≥ n

m−1 
m−1
κn
d(xm , xn ) ≤ d(xj+1 , xj ) ≤ κj d(x1 , x0 ) ≤ d(x1 , x0 ).
1−κ
j=n j=n
Hence, for all ε > 0 there exists an N > 0, such that for all n, m > N :
κN
d(xm , xn ) ≤ d(x1 , x0 ) ≤ ε,
1−κ
i.e., (xn )n∈N is a Cauchy-sequence. Since M is complete, there exists an
x∗ ∈ M , such that x∗ = limn→∞ xn .
The limit x∗ is a fixed point due to the continuity of F , i.e.,
F (x∗ ) = F ( lim xn ) = lim F (xn ) = lim xn+1 = x∗ . 
n→∞ n→∞ n→∞
Corollary 2.1.4. Let (X,  · ) be a Banach space, M be a closed subset
of X, and F : M → M be a contraction. Then F has a unique fixed point
x∗ ∈ M .
Our first version of the local existence and uniqueness of solutions for
(2.2) is as follows.
Lemma 2.1.5. Consider (2.2) with initial condition u|t=t0 = u0 and con-
tinuous A = A(t). Then there exists a δ > 0 independent of u0 such that
(2.2) has a unique solution u ∈ C 1 ((t0 − δ, t0 + δ), Rd ) satisfying u|t=t0 = u0 .
Proof. The proof is based on the application of the contraction mapping
theorem to the integrated ODE
 t
u(t) = u0 + A(s)u(s) ds =: F (u)(t),
t0
2.1. Linear systems 21

where F : M → M with M = C 0 ([t0 − δ, t0 + δ], Rd ). Fix a T0 > 0 and

define C0 = supt∈[t0 −T0 ,t0 +T0 ] A(t) which is finite due to the continuity of
t → A(t). Then we have F (u) − F (v)M ≤ δC0 u − vM , where uM =
supt∈[t0 −δ,t0 +δ] u(t)Rd , such that F is a contraction for instance for δ =
min{1/(2C0 ), T0 }. From u ∈ M it follows that F (u) ∈ C 1 ((t0 −δ, t0 +δ), Rd ).
Since u is a fixed point we also have u = F (u) ∈ C 1 ((t0 − δ, t0 + δ), Rd ). 
Remark 2.1.6. By the last argument, it is easy to see that the m-times dif-
ferentiability of t → A(t) implies inductively the m+1-times differentiability
of t → u(t).

Lemma 2.1.5 only asserts local existence and uniqueness. The way to
show existence and uniqueness of solutions beyond t0 + δ is to prove bounds
on u(t0 + δ). The key tool is Gronwall’s inequality which will be used in
many proofs below. We first restrict to a simple version [Ver96, Theorem
1.2].
Lemma 2.1.7. (Gronwall’s inequality) For t ∈ (t0 , t0 + a) with a > 0,
and φ and ψ non-negative continuous functions assume that
 t
(2.3) φ(t) ≤ ψ(s)φ(s) ds + δ.
t0

Then for all t ∈ (t0 , t0 + a) we have

t
ψ(s) ds
φ(t) ≤ δe t0 .

Proof. Dividing (2.3) by its right-hand side and multiplication of both sides
with ψ(t) yields after integration that
 t  t
ψ(τ )φ(τ )
τ dτ ≤ ψ(τ ) dτ
t0 t0 ψ(s)φ(s) ds + δ t0
  t
t
which implies that ln t0 ψ(s)φ(s) ds + δ − ln δ ≤ t0 ψ(τ ) dτ and finally
that   
t t
ψ(s)φ(s) ds + δ ≤ δ exp ψ(τ ) dτ .
t0 t0
By assumption φ(t) is smaller than the expression on the left-hand side. 
From the integrated ODE
 t
u(t) = u0 + A(s)u(s) ds
t0
we find the inequality
 t 
 
u(t)Rd ≤ u0 Rd +  A(s) u(s)Rd ds

t0
22 2. Basic ODE dynamics

and so by Gronwall’s inequality

 t 
 
(2.4) u(t)Rd ≤ u0 Rd exp  A(s) ds .
t0
Since continuous functions stay bounded on compact intervals, from (2.4)
it follows that for continuous t → A(t) the solutions t → u(t) exist for all
t ∈ R.
Theorem 2.1.8. Consider (2.2) with initial condition u|t=t0 =u0 and a
continuous A=A(t) for all t ∈ R. Then there exists a unique solution
u ∈ C 1 (R, Rd ) satisfying u|t=t0 = u0 .

Proof. We choose an arbitrary, but fixed T0 > 0. We apply Lemma 2.1.5

with the initial condition u|t=t0 = u0 , which gives a unique solution on
[t0 , t0 + δ) and u|t=t0 +δ/2 = u1 . Inductively we apply Lemma 2.1.5 now
with the initial condition u|t=t0 +nδ/2 = un , which gives a unique solution
on [t0 + nδ/2, t0 + nδ/2 + δ) and u|t=t0 +(n+1)δ/2 = un+1 . Doing this until
(n + 1)δ/2 ≥ T0 gives us the solution for all t ∈ [t0 , t0 + T0 ). Proceeding
similarly for negative t − t0 gives the solution for all t ∈ (t0 − T0 , t0 + T0 ).
Since T0 > 0 was arbitrary we are done. 
Lemma 2.1.9. The solutions of (2.2) form a d-dimensional vector space.

Proof. Since the solutions of (2.2) depend linearly and one-to-one on the
initial conditions u0 we have that the set of solutions of (2.2) is isomorphic
via u → u(t0 ) to the space of initial conditions, i.e., Rd . 
Definition 2.1.10. The matrix-valued function t → φ(t) is called funda-
mental matrix if φ̇(t) = A(t)φ(t) and if φ(t0 ) has full rank for a t0 ∈ R.
Remark 2.1.11. From the local existence and uniqueness Theorem 2.1.8,
it immediately follows that φ(t) has full rank for all t ∈ R. If t → φ(t)
is a fundamental matrix then u(t) = φ(t)φ(t0 )−1 u0 solves the initial value
problem with u|t=t0 = u0 . If t → ψ(t) is another fundamental matrix
then φ(t)φ(t0 )−1 = ψ(t)ψ(t0 )−1 such that there exists an invertible matrix
C = φ(t0 )−1 ψ(t0 ) which is independent of t with ψ(t) = φ(t)C.

2.1.3. The variation of constant formula. Associated to (2.2) we define

the linear solution operator S(t, s) : Rd → Rd through
S(t, s)u0 = u(t, s, u0 ),
where u(t, s, u0 ) is the solution of (2.2) with initial value u|t=s = u0 . For
fixed s, t ∈ R the linear map S(t, s) : Rd → Rd is one-to-one, i.e., an invert-
ible matrix, with S(t, s)−1 = S(s, t). The solution of the inhomogeneous
problem (2.1), i.e.,
(2.5) u̇(t) = A(t)u(t) + g(t)
2.1. Linear systems 23

can be expressed in terms of the inhomogeneity g = g(t) and the solution

operator S(t, s). Differentiation of u(t) = S(t, s)y(t) w.r.t. t and using (2.5)
shows that
u̇(t) = ∂t S(t, s)y(t) + S(t, s)ẏ(t) = A(t)S(t, s)y(t) + g(t).
Since S(t, s) solves the homogeneous problem ∂t S(t, s) = A(t)S(t, s) we
obtain
ẏ(t) = S(t, s)−1 g(t) = S(s, t)g(t).
t t
Integration yields y(t) − y(s) = s ẏ(τ ) dτ = s S(s, τ )g(τ ) dτ and therefore
 t
u(t) =S(t, s)y(t) = S(t, s)y(s) + S(t, s) S(s, τ )g(τ ) dτ
s
 t
=S(t, s)u(s) + S(t, τ )g(τ ) dτ,
s

since S(s, s) = I. This formula is called the variation of constant formula.

It will be used in stability proofs and in proofs of the local existence and
uniqueness of solutions of nonlinear problems. In case S(t, s) = e(t−s)A , see
the subsequent §2.1.4, the variation of constant formula specializes to
 t
tA
(2.6) u(t) = e u(0) + e(t−τ )A g(τ ) dτ.
0

2.1.4. The exponential matrix. In general, (2.2) can only be solved ex-
plicitly in case d = 1. For d ≥ 2, if a solution is known of the d-dimensional
problem, then the dimension of the problem can be reduced by one, i.e.,
after the reduction a linear ODE in d − 1 space dimensions has to be solved,
cf. [Cod61, Page 118]. However, in case
(2.7) u̇ = Au, u|t=0 = u0 ,
with A ∈ Rd×d independent of t, all solutions can be computed explicitly.
The ansatz u(t) = eλt u  yields the eigenvalue problem Au = λ u, and in
case that A has d linearly independent eigenvectors φ1 , . . . , φd ∈ Rd with

eigenvalues λ1 , . . . , λd , the general solution reads u(t) = di=1 ci eλi t φi with
ci , . . . , cd ∈ R. In case that there are complex eigenvalues or Jordan blocks
this formula becomes slightly more complicated. Equation (2.7) appears
as linearization around fixed points of nonlinear systems and hence plays a
crucial role.
Remark 2.1.12. Linear scalar equations of nth order

n−1
(n)
y (t) + aj y (j) (t) = 0,
j=0
24 2. Basic ODE dynamics

with aj ∈ R, can always be written as a linear first order system. For the
construction of the explicit solution
 this is not necessary. The r different ze-
roes λk with multiplicity kr , i.e. rk=1 kr = n, of the characteristic equation

λn + n−1 j
j=0 aj λ = 0 allow to construct the general solution, namely

 r −1
r k
y(t) = ck,j eλk t tj
k=1 j=0

with n coefficients ck,j . See [Log15b, Chapter 2].

From a theoretical point of view, the following representation formula

turns out to be useful. The solution of (2.7) is given by
∞
 (tA)n
(2.8) u(t) = etA u0 = u0 .
n!
n=0

We have absolute and uniform convergence w.r.t. t on every compact interval

due to
∞  
∞
 (tA)n  An tn
 u  ≤ u0 Rd ≤ eAt u0 Rd .
 n! 0  d n!
n=0 R n=0

Obviously this also holds for the series differentiated w.r.t. t. Similarly, we
obtain the estimate
etA u0 Rd ≤ etA u0 Rd .

Moreover, etA u0 solves (2.7) due to

∞  ∞ ∞
d  (tA)n  A(tA)n−1  (tA)n
u0 = u0 = A u0 ,
dt n! (n − 1)! n!
n=0 n=1 n=0

where the time derivative and the infinite sums can be interchanged due
to the uniform convergence w.r.t. t on compact intervals. Moreover, A and
the infinite sum can be interchanged due to the boundedness and hence
continuity of A.
The solution operator etA can be expressed in terms of the Jordan normal
form J of A. The change of coordinates u = Sy in u̇ = Au yields ẏ =
S −1 ASy = Jy. We have

(2.9) etA u0 = SetJ y0 = SetJ S −1 u0 ,

2.1. Linear systems 25

or equivalently
 
−1 tA −1 t2 A2
S e S =S 1 + tA + + ... S
2
 
−1 t2 S −1 ASS −1 AS
= 1 + tS AS + + ...
2
 
t2 J 2
= 1 + tJ + + . . . = etJ .
2

Hence, it is sufficient to consider etJ for J a matrix in Jordan normal form,

i.e.,
⎛ ⎞ ⎛ ⎞
J1 0 λj 1 0
⎜ ⎟ ⎜ .. .. ⎟
⎜ J2 ⎟ ⎜ . . ⎟
J =⎜ ⎟ ⎜
with Jj = ⎜ ⎟.
⎝ . .. ⎠ . ⎟
⎝ . . 1⎠
0 Jr 0 λj

Since
⎛⎛ ⎞ ⎞
⎛ tJ ⎞
J1 0 e 1 0
⎜⎜ .. ⎟ ⎟ ⎜ .. ⎟
exp(tJ) = exp ⎝⎝ . ⎠ t⎠ = ⎝ . ⎠
0 Jr 0 etJr

it is sufficient to consider

etJj = et(λj I+Nk ) = eλj tI etNk

due to INk = Nk I with the k × k-matrix

⎛ ⎞
0 1 0
⎜ .. .. ⎟
⎜ . . ⎟
Nk = ⎜⎜ ..
⎟.
⎟
⎝ . 1⎠
0 0

Hence, it remains to compute

∞ ν

tNk t
e = Nkν .
ν!
ν=0

We have
Nkμ = (δi,j−μ ), for μ = 0, ..., k − 1,
Nkμ = 0, for μ = k, k + 1, ...
26 2. Basic ODE dynamics

and so finally
⎛ tk−1
⎞
1 t. . . . . . (k−1)!
⎜ ⎟
⎜0 1 . . . tk−2 ⎟
⎜ (k−2)! ⎟
⎜ .. ⎟
etNk = ⎜ .. . . . . . . . . .
. ⎟
.
⎜ . ⎟
⎜ .. ⎟
⎝ . 1 t ⎠
0 ... 0 1
This representation formula immediately yields a statement about the sta-
bility of the fixed point u = 0 of the ODE u̇ = Au.
Definition 2.1.13. a) The fixed point u = 0 is called asymptotically
stable for u̇ = Au if for all u0 ∈ Rd we have u(t) = etA u0 → 0 for
t → ∞.
b) The fixed point u = 0 is called stable for u̇ = Au if for all u0 ∈ Rd
we have that u(t) = etA u0 stays bounded for all t ≥ 0.
c) In all other cases the origin u = 0 is called unstable.
Theorem 2.1.14. a) If all eigenvalues of A have strictly negative real
parts, then u = 0 is asymptotically stable.
b) If A possesses no eigenvalue with positive real part and if all eigen-
values with real part zero possess the same algebraic and geometric
multiplicity, i.e., no non-trivial Jordan block, then the origin u = 0
is stable.
c) In all other cases, i.e., if A possesses at least one eigenvalue with
strictly positive real part or at least one eigenvalue with real part
zero with algebraic multiplicity bigger than the geometric multiplic-
ity, then the origin u = 0 is unstable.

2.1.5. Linear planar systems. The behavior of the solutions of two-

dimensional autonomous systems u̇ = f (u) can be visualized with the help
of so called phase portraits. As a first step we discuss and visualize the
behavior of linear two-dimensional autonomous systems
(2.10) u̇ = Au, u(t) ∈ R2 , A ∈ R2×2 .
In order to visualize the behavior of (2.10) we have a number of possibilities,
which we shall later also apply to nonlinear systems. In doing so we also
classify the different kinds of fixed points.
1) We plot the vector field f : R2 → R2 . This turns out to be not that
helpful due to the in general strongly varying length of f .
2) Therefore, we plot in most cases the direction field αf /f R2 :
R2 → R2 for a fixed α > 0, i.e., in every point u ∈ (hZ)2 of a grid
with width h we plot a vector of R2 of fixed length α.
2.1. Linear systems 27

3) We plot the flow, i.e., a number of chosen orbits. These are curves
which are defined by the solutions t → u(t) ∈ R2 .
A combination of 2) and 3) is called a phase portrait. The choice of the size
of h and of the orbits depends on the problem, and also is a matter of taste.
Remark 2.1.15. The vector field and the direction field are both tangent
vectors of the solution t → u(t). Hence, the differential equations u̇ = f (u)
and u̇ = f (u)/f (u)R2 have the same orbits, i.e., solution curves in the
phase plane, although their dynamics are very different.

One more option is to plot the nullclines. These are the sets
Nj := {(u1 , u2 ) ∈ R2 : fj (u1 , u2 ) = 0},
for j = 1, 2, where the vector field is vertical, respectively horizontal. The
intersection points of N1 and N2 give the fixed points u∗ of the ODE, i.e.,
points with f (u∗ ) = 0. If the solution starts in a fixed point, the solution
stays in that fixed point, i.e., u(t) = u∗ for all t. Often, nullclines at least
partially coincide with coordinate axis, and, moreover, for (non-degenerate)
linear systems we have N1 ∩ N2 = {(0, 0)} as the only fixed point.
Due to §2.1.4 it is sufficient to consider (2.2) with A ∈ R2×2 in Jordan
normal form. There are the following cases.
a) The eigenvalues have the same algebraic and geometric multiplicity,
i.e.,  
λ1 0
A = with a1) 0 = λj ∈ R or a2) 0 = λ1 = λ2 . In
0 λ2
case a1) we distinguish three subcases i) λ1 = λ2 , ii) λ1 > λ2 > 0
and iii) λ1 > 0 > λ2 . All other cases are obtained from i)-iii) by a
reversal of time t → −t.
b) The eigenvalue has geometric
  multiplicity one and algebraic multi-
λ 1
plicity two, i.e., A = is a Jordan block with λ ∈ R.
0 λ
c) The degenerate case of at least one eigenvalue λ = 0. Besides the
trivial case A = 0 there are the cases c1) 0 = λ1 < λ2 and c2)
λ = 0 with geometric multiplicity one and algebraic multiplicity
two.
In the following we consider a number of examples to visualize these cases,
see Figure 2.1.5.
 
1 0
a1 i): Let A = , i.e., u̇1 = u1 , u̇2 = u2 . For the solutions
0 1
we find u1 (t) = et u1 (0), u2 (t) = et u2 (0). The orbits are straight
lines since u1 (t)/u2 (t) = u1 (0)/u2 (0) = const.. The nullclines are
28 2. Basic ODE dynamics

N1 = {(0, y) : y ∈ R} and N2 = {(x, 0) : x ∈ R}. The fixed point

u∗ = (0, 0) is called a source or unstable node. If time is reversed,
−1 0
i.e., if A = , then u∗ = (0, 0) is called a sink or stable
0 −1
node.
 
2 0
a1 ii): Let A = , i.e., u̇1 = 2u1 , u̇2 = u2 . For the solutions
0 1
we find u1 (t) = e2t u1 (0), u2 (t) = et u2 (0). The orbits are parabolas
since u1 (t)/(u2 (t))2 = u1 (0)/(u2 (0))2 = const.. The phase portrait
is robust w.r.t. small perturbations, i.e., A = diag(1.9, 1.1) has a
similar phase portrait. Again the fixed point u∗ = (0, 0) is called a
source.
 
1 0
a1 iii): Let A = , i.e., u̇1 = u1 , u̇2 = −u2 . For the solutions
0 −1
we find u1 (t) = et u1 (0), u2 (t) = e−t u2 (0). The orbits are hyper-
bolas since u1 (t)u2 (t) = u1 (0)u2 (0) = const.. The phase portrait
is robust w.r.t. small perturbations, i.e., A = diag(1.1, −0.9) gives
a similar phase portrait. The fixed point u∗ = (0, 0) is called a
saddle.
 
0 1
a2 i): Let A = , i.e., u̇1 = u2 , u̇2 = −u1 . Then u  = u1 + iu2
−1 0
solves u ˙ = −iu. The solution u (t) = e−it u(0) leaves the circles
u(t)| = u1 (t) + u2 (t) = |
| 2 2 2 2
u(0)| = const. invariant. Hence, the
orbits are circles, and u∗ = (0, 0) is called a center. The phase por-
trait is not robust w.r.t. small perturbations. In general after the
perturbation we obtain the phase portrait from a2 ii). However,
in applications often additional effects such as a conserved quan-
tity enforce the robustness of centers w.r.t. the class of possible
perturbations. See for instance Chapter 4.
 
1 1
a2 ii): Let A = . For u  = u1 + iu2 we obtain the equation
−1 1
˙ = (1 − i)
u u which is solved by u (t) = et e−it u
(0). In polar coor-
dinates u (t) = r(t)e iφ(t) with r(t) ∈ R and φ(t) ∈ S 1 = R/(2πZ)
we obtain ṙ = r and φ̇ = −1 with solution r(t) = et r(0) and
φ(t) = (φ(0) − t) mod 2π. The orbits are spirals. The phase por-
trait is robust w.r.t. small perturbations. Here u∗ = (0, 0) is called
an unstable vortex or spiral.
 
1 1
b): Let A = , i.e., u̇1 = u1 + u2 , u˙2 = u2 . For the second
0 1
equation we obtain u2 (t) = et u2 (0). The variation of constant
2.1. Linear systems 29

formula applied to the first equation yields

 t
t
u1 (t) = e u1 (0) + et−s es u2 (0) ds = et u1 (0) + et tu2 (0).
0
Here u∗
= (0, 0) is called a degenerate node. The phase portrait is
not generic, since a Jordan block only occurs with probability zero
under all matrices. However, it can make sense to keep the Jordan
block as starting point of the analysis.
 
1 0
c1): Let A = . The general solution is u1 (t) = et u1 (0),
1 0
u2 (t) = u2 (0) + et u1 (0), and there is the line of fixed points u1 = 0.
 
0 1
c2): Let A = . The general solution is u1 (t) = u1 (0) + u2 (0)t,
0 0
u2 (t) = u2 (0), i.e., the flow is parallel to the line of fixed points
u2 = 0.

1 1 1 1

0 0 0 0

−1 −1 −1 −1

−1 0 1 −1 0 1 −1 0 1 −1 0 1

1 1 1 1

0 0 0 0

−1 −1 −1 −1

−1 0 1 −1 0 1 −1 0 1 −1 0 1

Figure 2.1. Phase portraits for a1i), a1ii), a1iii), and b) in the first
row, for a2i), a2ii), c1) and c2) in the second row.

2.1.6. Linear systems with periodic coefficients. Equations

(2.11) u̇(t) = A(t), u(t) with A(t) = A(t + T ),
for a fixed T > 0, appear for instance as linearizations around time-periodic
solutions of nonlinear systems. Hence, they will play an important role in
the following. In contrast to the case of t-independent matrices A, where an
arbitrary shift of time still gives the same system, in case of time-periodic
A = A(t) only integer multiples of the basic period T can be allowed.
Lemma 2.1.16. With u(t, t0 , u0 ) also u(t + nT, t0 + nT, u0 ) with n ∈ N
solves (2.11).
30 2. Basic ODE dynamics

The following theorem is fundamental.

Theorem 2.1.17. (Floquet)Each fundamental matrix φ(t) can be written
as a product
φ(t) = P (t)etB
of two d × d-matrices, with P (t) = P (t + T ) and B a constant d × d-matrix.

Proof. Since
φ̇(t + T ) = A(t + T )φ(t + T ) = A(t)φ(t + T )
with φ(t) also φ(t + T ) is a fundamental matrix. Hence, there exists an
invertible d × d-matrix C, such that
φ(t + T ) = φ(t)C.
Each invertible d × d-matrix C can be written as C = eT B with B a non-
unique d ×d-matrix. As example consider C = diag(λ1 , . . . , λd ) with λj > 0.
Then a logarithm is given by B = diag(ln λ1 , . . . , ln λd ). For the general case
use −1 = eiπ and the expansion of ln(1 + λ) in case of Jordan blocks. See
Exercise 2.5. For P (t) = φ(t)e−tB we obtain
P (t + T ) = φ(t + T )e−(t+T )B = φ(t)Ce−T B e−tB = φ(t)e−tB = P(t). 
Definition 2.1.18. The matrix C = eT B is called monodromy matrix. The
eigenvalues of C are called Floquet multipliers, and the eigenvalues of B are
called Floquet exponents.

Floquet exponents from a different matrix B differ only by adding integer

multiples of 2πi/T . The Floquet multipliers are unique. Suppose that two
fundamental matrices φ = φ(t) and ψ = ψ(t) are given. By Remark 2.1.11
then ψ −1 (t)φ(t) = S is independent of time and hence
Cφ = φ(t)−1 φ(t + T ) = S −1 ψ(t)−1 ψ(t + T )S = S −1 Cψ S.
As a consequence the matrices Cφ and Cψ have the same eigenvalues. The
T -periodic transformation u(t) = P (t)y(t) gives
P (t)ẏ(t) + Ṗ (t)y(t) = u̇(t) = A(t)u(t) = A(t)P (t)y(t)
and thus
ẏ(t) =P (t)−1 (A(t)P (t) − Ṗ (t))y(t)
=P (t)−1 (A(t)P (t) − φ̇(t)e−tB − φ(t)(−B)e−tB )y(t)
=P (t)−1 (A(t)P (t) − A(t)φ(t)e−tB + φ(t)e−tB B)y(t)
=P (t)−1 (A(t)P (t) − A(t)P (t) + P (t)B)y(t) = By(t).
For the stability of u = 0, it is therefore sufficient to consider B. If all
eigenvalues λ of B satisfy Re λ < 0, we have the asymptotic stability of
u = 0, see Theorem 2.1.19.
2.1. Linear systems 31

Alternatively, by Lemma 2.1.16 for all n ∈ N and τ ∈ [0, T ), we have

u(t, 0, u0 ) =u(nT + τ, 0, u0 ) = u(nT + τ, nT, u(nT, 0, u0 ))
=u(nT + τ, nT, u(nT, (n − 1)T, u((n − 1)T, 0, u0 )))
=u(τ, 0, u(T, 0, u(T, 0, . . . , u(T, 0, u0 )) . . .)))
=Φτ ◦ ΦT ◦ . . . ◦ ΦT (u0 ),
where Φt u0 = u(t, 0, u0 ). Since τ ∈ [0, T ) for the long-time dynamics only
the iteration of the time T -map
  −1
ΦT = φ(T )φ(0)−1 = P (T )eT B P (0)e0B = P (0)Cφ P (0)−1 ,
is of interest, where we used P (T ) = P (0) and e0B = I. The proof of the
following theorem is an easy exercise.
Theorem 2.1.19. In a discrete dynamical system un+1 = Cun we have:
a) If all eigenvalues μ of C satisfy the condition |μ| < 1, then u = 0
is asymptotically stable, i.e., limn→∞ u(n, u0 ) = 0.
b) If C has an eigenvalue μ with |μ| > 1, or a non-trivial Jordan block
to an eigenvalue with |μ| = 1, then u = 0 is unstable.

Im
Im
1
1

Re Re

asymptotically stable unstable

Figure 2.2. The eigenvalues of C in cases a) and b) of Theorem 2.1.19.

 
2 −1
Example 2.1.20. Consider the iteration xn+1 = Cxn , with C = .
0 1
The solution
 can be computed explicitly by the transformation
  x =Sy, with
1 1 −1 2 0
S= . We find xn+1 = SB n+1 S x0 , with B = .
0 1 0 1
Example 2.1.21. Consider the 1-periodic ODE u̇(t) = cos2 (2πt)u(t) for
u(t) ∈ R. Using cos2 (2πt) = 1/2 + cos(4πt)/2 the solution with initial
condition u(0) = u0 is given by
 
t sin 4πt
u(t, 0, u0 ) = u0 exp +
2 8π
32 2. Basic ODE dynamics

and therefore
   
sin 4πt t
P (t) = exp and etB = exp .
8π 2
We find the Floquet multiplier e1/2 and the Floquet exponent 1/2. The
time-one-map is given through Φ1 u0 = e1/2 u0 .

The following example [MY60] shows that in the periodic case the eigen-
values of the matrix A(t) have no significance for the stability of u = 0.

Example 2.1.22. Consider u̇ = A(t)u with

 
−1 + 32 cos2 t 1− 32 sin t cos t
A(t) = .
−1− 32 sin t cos t −1 + 32 sin2 t
The characteristic polynomial is given by
     
3 3 3 3
1 − cos2 t+λ 1 − sin2 t+λ − 1− sin t cos t 1+ sin t cos t
2 2 2 2
3   3  
=λ2 + 2λ− cos2 t + sin2 t λ + 1− cos2 t + sin2 t
2 2
9 2 2 9 2 2
+ cos t sin t + 1− cos t sin t
4 4
1 1
=λ2 + λ + ,
2 2

√ the eigenvalues are independent of t and are given by λ1,2 = (−1 ±

i.e.,
i 7)/4. Therefore, we expect u = 0 to be stable, but there is the solution
 
− cos t t/2
e ,
sin t
which is unbounded for t → ∞. With the help of this solution the 2 × 2-
system of ODEs can be reduced to a scalar equation which can be solved
with the method of separation of variables. Hence, the Floquet exponents
can be computed explicitly. They are given by λ1 = 12 and λ2 = −1.

2.1.7. An outlook on amplitude equations. We close this review of lin-

ear ODE theory with a first glimpse at what will be one of the main subjects
of this book, namely reduction methods and amplitude (and modulation)
equations. Consider the weakly damped linear oscillator
(2.12) ü + 2εu̇ + u = 0, u(0) = a, u̇(0) = 0,
with u(t) ∈ R and 0 < ε  1. The explicit solution is
εa 
u(t) = e−εt (a cos(ωt) + sin(ωt)), where ω = 1 − ε2 ,
ω
2.1. Linear systems 33

cf. Remark 2.1.12. However, we might also try an expansion w.r.t. ε, i.e.,
u(t) = u0 (t) + εu1 (t) + O(ε2 ). Plugging this ansatz into (2.12) and sorting
w.r.t. powers in ε yields
O(ε0 ) : ü0 (0) + u0 = 0, u0 (0) = a, u̇0 (0) = 0
⇒ u0 (t) = a cos t,
O(ε1 ) : ü1 + u1 = 2a sin t, u1 (0) = 0, u̇1 (0) = 0
⇒ u1 (t) = −at cos t + a sin t,
and hence uapp1 (t) = a cos t − εta cos t + εa sin t + O(ε2 ). Comparing with
u shows that the expansion only makes sense for t = O(1), and becomes
completely useless after that.
With some physical insight, we may however directly see from (2.12)
that 2ε∂t u corresponds to a weak damping, and hence we suspect that there
are two time scales involved in (2.12). Thus we may try a multi-scale ansatz
of the form
(2.13) u(t) = A(εt)eiω0 t + c.c.,
with ω0 ∈ R an a priori unknown (fast) frequency, and where A = A(τ ) ∈ C
d
is a slowly varying (complex valued) amplitude. Then, e.g, dt u = (iω0 +
d iω t
ε dτ )Ae 0 + c.c., and plugging into (2.12) we obtain
O(ε0 ) : − ω02 + 1 = 0, A(0) = a/2 ⇒ ω0 = 1,
d
O(ε1 ) : 0 = −2i( A + A)eit + c.c..
dτ
This yields A(τ ) = e−τ A(0), and thus
uapp2 (t) = A(τ )eit + c.c. + O(ε) = ae−εt cos(t) + O(ε),
which at least is a much better approximation of the true solution than
uapp1 , see Fig. 2.3.
The equation dτ d
A = −A is called the amplitude equation for the ansatz
(2.13) for the system (2.12), and here can be solved explicitly, like the orig-
inal system. However, already in simple nonlinear ODEs in general neither
the original equation nor the amplitude equation can be solved explicitly.
Moreover, although the amplitude equation is usually a bit “simpler”, this
is not the essential characteristic. The main points are that the amplitude
equation often falls into some universality class, and that it describes the
system on long scales. Thus, if one has to use numerical methods, then the
numerical costs are greatly reduced. For instance, in the present example
we would reduce the numerical costs by a factor 1/ε, e.g., by factor 10 if
ε = 0.1. More drastic cost reductions may occur for PDEs, see Part IV of
this book.
34 2. Basic ODE dynamics

u(t)
u_app1(t)
u_app2(t)
2

-1

-2
0 5 10 15 20 25 30

Figure 2.3. Exact solution and the two approximations for (2.12); ε =
0.1, a = 1.

2.2. Local existence and uniqueness for nonlinear systems

In this section we prove the local existence and uniqueness of solutions for
nonlinear ODEs. We consider
(2.14) u̇(t) = f (u(t), t),
for an unknown function u ∈ C 1 (I, Rd ), where I ⊂ R is an interval and
f : Rd × R → Rd is called the vector field, which is called autonomous if it
does not depend explicitly on time t. An initial value problem consists in
finding solutions of (2.14) to the initial condition u0 ∈ Rd at some time t0 ,
i.e., u|t=t0 = u0 . If d > 1, then (2.14) is sometimes called a system of ODEs.
Sometimes f is not defined for all t ∈ R or for all u ∈ Rd ; the latter is for
instance always the case if we consider a planar ODE in polar coordinates
u=(r, φ) with hence r≥0. However, for a given initial condition u0 at a time
t0 it is clear that for a local solution it is sufficient that f is only defined in
a neighborhood of u0 and t0 . The modifications needed in the theory below
are obvious and thus for notational simplicity we generally assume that f is
defined for all u ∈ Rd and all t ∈ R.
For f locally Lipschitz-continuous w.r.t. u and continuous w.r.t. t we
have the local existence and uniqueness of solutions. A function f : Rd ×I →
Rd with I ⊂ R an open interval is called locally Lipschitz-continuous w.r.t. to
the first variable if for all C1 there exists a C2 such that
max{uRd , vRd } ≤ C1 ⇒ sup f (u, t) − f (v, t)Rd ≤ C2 u − vRd .
t∈I

Theorem 2.2.1. (Picard-Lindelöf ) Consider (2.14) with initial condition

u|t=t0 = u0 and let f : Rd × I → Rd be continuous w.r.t. t and locally
Lipschitz-continuous w.r.t. u. For C1 > 0 define M0 = {u ∈ Rd : u −
2.2. Local existence and uniqueness for nonlinear systems 35

u0 Rd ≤ C1 }, C3 := sup(u,t)∈M0 ×I f (u, t)Rd , and denote the Lipschitz-

constant in M0 by C2 . Moreover, assume that there exists a δ > 0 such that
I ⊃ [t0 − δ, t0 + δ]. Then (2.14) has a unique solution u ∈ C 1 ([t0 − T0 , t0 +
T0 ], Rd ) satisfying u|t=t0 = u0 , where T0 = min(δ, 1/(2C2 ), C1 /C3 ).

Proof. Similar to the proof of Theorem 2.1.5 we apply the contraction map-
ping Theorem 2.1.3 to the integrated ODE
 t
(2.15) u(t) = u0 + f (u(s), s) ds =: F (u)(t),
t0

where F : M → M with
M = C 0 ([t0 − T0 , t0 + T0 ], {u ∈ Rd : u − u0 Rd ≤ C1 })
which is equipped with the metric
d(u, v) = sup u(t) − v(t)Rd =: u − vM .
t∈[t0 −T0 ,t0 +T0 ]

We have
 t
F (u) − u0 M ≤ sup  f (u(s), s) dsRd
t∈[t0 −T0 ,t0 +T0 ] t0
 t
≤ sup | f (u(s), s)Rd ds| ≤ T0 C3 ≤ C1
t∈[t0 −T0 ,t0 +T0 ] t0

for T0 = min(δ, C1 /C3 ) such that F maps M into M . Moreover,

 t
F (u) − F (v)M ≤ sup  f (u(s), s) − f (v(s), s) dsRd
t∈[t0 −T0 ,t0 +T0 ] t0
 t
≤ sup | f (u(s), s) − f (v(s), s)Rd ds|
t∈[t0 −T0 ,t0 +T0 ] t0
 t
≤ sup | C2 u(s) − v(s)Rd ds|
t∈[t0 −T0 ,t0 +T0 ] t0
≤T0 C2 u − vM ,
such that F is a contraction for T0 = min(δ, 1/(2C2 )). From u ∈ M it follows
that F (u) ∈ C 1 ((t0 − T0 , t0 + T0 ), Rd ). Since u is a fixed point we also have
u = F (u) ∈ C 1 ((t0 − T0 , t0 + T0 ), Rd ). 

Remark 2.2.2. The last argument shows that f ∈ C m (Rd × I, Rd ) implies

u ∈ C m+1 (I, Rd ).

For f : Rd × R → Rd locally Lipschitz-continuous the solutions can only

stop to exist if u(t)Rd becomes infinitely large.
36 2. Basic ODE dynamics

Theorem 2.2.3. For locally Lipschitz-continuous f the solution u with

u|t=t0 = u0 ∈ Rd exists for all t ∈ (T− , T+ ), where
T− = inf{t ∈ R : u(t)Rd < ∞} and T+ = sup{t ∈ R : u(t)Rd < ∞}.

Proof. If u(t)Rd is finite, then the local existence and uniqueness Theorem
2.2.1 applies and also u(t − T0 )Rd and u(t + T0 )Rd are finite for some
T0 > 0. 
The following two examples show that f being only continuous is not
sufficient for uniqueness, and that solutions in general do not exist globally.

Example 2.2.4. Consider the one-dimensional ODE u̇ = |u| with initial
value u|t=0 = 0. The right-hand side is not Lipschitz-continuous at u = 0.
This initial value problem has the solution u = 0, but also infinitely many
other solutions, namely

0, for 0 ≤ t ≤ τ
uτ (t) =
(t − τ )2 /4, for τ ≤ t,
solves the ODE for each τ > 0, i.e., there is no uniqueness of solutions.
Example 2.2.5. Consider the one-dimensional ODE u̇ = 1 + u2 with the
initial value u|t=0 = 0. This initial value problem has the solution u(t) =
tan t, i.e., the solution explodes, or blows up, in finite time, it becomes ∞
for t = π/2. Hence, the solution does not exist for all t ∈ [0, ∞), i.e., there
is no global existence of solutions.

Our next goal is to prove the continuity of the solutions w.r.t. the initial
conditions. In order to do so we use Gronwall’s inequality, cf. Lemma 2.1.7.
Lemma 2.2.6. Let f : Rd × R → Rd be continuous and locally Lipschitz-
continuous w.r.t. the first variable. Then each solution u(t, t0 , u0 ) is
Lipschitz-continuous w.r.t. u0 in the following sense: For every T0 > 0
there exist δ > 0 and L > 0 such that for all u1 ∈ Rd with u0 − u1 Rd ≤ δ
we have, for all t ∈ [t0 − T0 , t0 + T0 ],
(2.16) u(t, t0 , u0 ) − u(t, t0 , u1 )Rd ≤ Lu0 − u1 Rd .

Proof. We have
u(t, t0 , u0 ) − u(t, t0 , u1 )Rd
 t
≤ u0 − u1 Rd + f (u(τ, t0 , u0 ), τ ) − f (u(τ, t0 , u1 ), τ )Rd dτ
t0
 t
≤ u0 − u1 Rd + |C2 u(τ, t0 , u0 ) − u(τ, t0 , u1 )Rd dτ |.
t0
Hence, by Gronwall’s inequality we find
u(t, t0 , u0 ) − u(t, t0 , u1 )Rd ≤ u0 − u1 Rd eC2 |t−t0 | . 
2.2. Local existence and uniqueness for nonlinear systems 37

Remark 2.2.7. For continuously differentiable f we have differentiability

of u(t, t0 , u0 ) w.r.t. the data t0 , u0 , see, e.g., [HSD04, Page 402].

Another application of Gronwall’s inequality is the proof of bounds on

the size of the solutions. We state it in a differential form and then present
a simple, but fundamental, example.
Lemma 2.2.8. (Gronwall) Let I ⊂ R be an interval, α, β ∈ R, and φ ∈
C 1 (I, R) a non-negative function with
φ̇(t) ≤ α + βφ(t)
for all t ∈ I. Then, for all t0 , t ∈ I, t ≥ t0 we have
α β(t−t0 ) 
φ(t) ≤ φ(t0 )eβ(t−t0 ) + e −1 .
β

Proof. We introduce ψ(t) = φ(t)eβt which satisfies ψ̇(t) ≤ αe−βt . Integra-

tion yields ψ(t) ≤ ψ(t0 ) + αβ (e−βt0 −e−βt ). Undoing the transformation gives
the result. 
Example 2.2.9. Consider u̇ = u − u3 . For φ(t) = u2 (t) we obtain
φ̇(t) = 2u(t)u̇(t) = 2u2 (t) − 2u4 (t) ≤ 2 − 2u2 (t) = 2 − 2φ(t).
Therefore, from Lemma 2.2.8
u2 (t) ≤ u2 (0)e−2t + 2(1 − e−2t )/2  1 as t → ∞.
Thus, every solution exists globally (in forward time) since it stays bounded
and enters for instance the interval [−2, 2].

As long as they exist, solutions of ODEs (2.14) have the trivial, but
fundamental property
(2.17) u(t + s, t0 , u0 ) = u(t, s, u(s, t0 , u0 )), u(t0 , t0 , u0 ) = u0 .
For autonomous systems we have
u(t, t0 , u0 ) = u(t − t0 , 0, u0 ) =: u(t − t0 , u0 ),
i.e., w.l.o.g. we can always choose the initial time t0 = 0. Then (2.17)
transfers into
(2.18) u(t + s, u0 ) = u(t, u(s, u0 )), u(0, u0 ) = u0 .
Thus, it makes no difference whether we solve the ODE until the time t + s,
or if we solve the ODE until the time s, start again, and solve until the
time t + s. A similar structure occurs for iterations un+1 = f (un ). In the
following we focus on the autonomous case.
38 2. Basic ODE dynamics

Definition 2.2.10. A map u : I × M → M which satisfies (2.18), where

I = R+ , I = R, I = N or I = Z, and where M is a set, is called a dynamical
system or flow.
If I = R+ or I = R then the dynamical system is called continuous.
If I = N or I = Z it is called discrete.
The set M is called the phase space.
The set γ+ (u0 )={u(t, u0 ) : t ≥ 0} is called the forward orbit through u0 ,
the set γ− (u0 ) = {u(t, u0 ) : t ≤ 0} is called the backward orbit through u0 ,
and γ(u0 ) = γ+ (u0 ) ∪ γ− (u0 ) is called the orbit through u0 .

Since with this strict definition, only ODEs (2.14) with solutions existing
globally forward in time in case I = R+ , respectively, forward and backward
in time in case I = R define continuous dynamical systems in the phase
space Rd , we shall not be that strict in the following and call any map u
which fulfills (2.18) a dynamical system.
For ODEs (and PDEs) the dynamical systems property (2.18) expressed
in terms of the family of (nonlinear) solution operators (St )t∈I is given by
(2.19) St+s = St Ss , S0 = I,
where St is defined by St u0 := u(t, u0 ), and where I is here the identity
on M . Due to (2.19), the family of solution operators (St )t∈I is called a
semigroup in case I = R+ .
Remark 2.2.11. Except for the points which are mapped to infinity the
map u0 → u(t, u0 ) is bijective due to u(t, u(−t, u0 ))=u(t − t, u0 )=u0 in case
I = R or I = Z. Since additionally u(t, u0 ) depends continuously on u0
dynamical systems can be interpreted as a flow of homeomorphisms, i.e., as
flow of bijective bi-continuous maps from Rd into Rd . If f is C k then also
u(t, u0 ) is C k w.r.t. u0 , i.e., then the dynamical system can be interpreted
as a flow of C k -diffeomorphisms.

2.3. Special solutions

In this section we introduce special solutions such as fixed points, periodic
solutions, and homoclinic and heteroclinic orbits and basic concepts such as
stability and instability and invariant manifolds.

2.3.1. Fixed Points. Until further notice we consider the autonomous

case. In order to explore the dynamics in phase space we start from the
most simple dynamical objects, namely fixed points.
Definition 2.3.1. A point u∗ ∈ Rd is called fixed point for the ODE (2.14)
if f (u∗ ) = 0.
2.3. Special solutions 39

Example 2.3.2. We consider u̇ = u − u3 with u = u(t) ∈ R. From f (u) =

u − u3 = 0 we obtain the fixed points u∗1 = 0 and u∗2,3 = ±1. See Figure 2.4.

Figure 2.4. The phase portrait of u̇ = u−u3 drawn in the 1D phase

space. In one space dimension, i.e., u = u(t) ∈ R, the complete
qualitative behavior of the dynamics of an autonomous ODE is
known with the knowledge of the fixed points due to topological
reasons. In between two fixed points the “vector field” f (u) ∈ R,
which is a scalar function, cannot change sign. Therefore, the real
line is divided by the fixed points which are connected by so called
heteroclinic solutions. In case of an interval of fixed points the
statement remains true with obvious modifications. Hence, the
dynamics of autonomous one-dimensional ODEs is trivial.

In order to explore the dynamics near a fixed point u∗ we write u = u∗ +v

and make a Taylor expansion of the vector field f around u∗ , i.e.,
d ∗
v̇ = (u + v) = f (u∗ + v) = 0 + ∂u f (u∗ )v + O(v2 ),
dt
using f (u∗ ) = 0. The only approximate system which in general can be
solved explicitly is the linearization at the fixed point u∗ , namely
v̇ = ∂u f (u∗ )v.
In Example 2.3.2 we have ∂u f (u∗j ) = 1 − 3(u∗j )2 , and hence
v̇ = v for u∗1 = 0 and v̇ = −2v for u∗2,3 = ±1.
Therefore, from the linearization we expect that solutions which start close
to u∗2,3 converge towards u∗2,3 for t → ∞, while solutions which start close
to u∗1 will leave any small neighborhood of u∗1 .
Definition 2.3.3. A fixed point u∗ is called stable for the ODE (2.14) if for
any ε > 0 there is a δ > 0 such that u0 − u∗ Rd < δ implies u(t, u0 ) −
u∗ Rd < ε for all t ≥ 0. Otherwise, it is called unstable. A stable fixed
point is called asymptotically stable if u0 − u∗ Rd < δ additionally implies
limt→∞ u(t, u0 ) = u∗ .

For linear systems the statements of Theorem 2.1.14 remain true with
the more general Definition 2.3.3. The following theorem guarantees in
many situations that stability or instability in the linearized system implies
stability or instability for the full system (2.14).
40 2. Basic ODE dynamics

Theorem 2.3.4. Let u∗ be a fixed point for (2.14). Let A = ∂u f (u∗ ) ∈ Rd×d
be the linearization of f in u∗ .
a) If all eigenvalues λj of A satisfy Re λj < 0, then u∗ is asymptoti-
cally stable.
b) If A has an eigenvalue λ with Re λ > 0, then u∗ is unstable.

Proof. W.l.o.g. let u∗ = 0. The proof is based in both situations on the

fact that the nonlinear terms in a neighborhood of u∗ = 0 are much smaller
than the linear terms.
a) In order to use this fact in a) we use the variation of constant formula.
Let etA be the solution operator of the linear system u̇ = Au. Since the
eigenvalues of A have strictly negative real part there are positive constants
μ0 and C0 (necessary due to possible Jordan blocks) with
etA Rd →Rd ≤ SetJ S −1  ≤ SS −1 etJ  ≤ C0 e−μ0 t
for all t ≥ 0 where we used the notation of (2.9). The estimate follows by
using SS −1  < ∞ and by estimating etJ  in the  · ∗ matrix norm
associated to the  · 2 vector norm from page 19. We remark that −μ0
has to be larger than the largest real part of the eigenvalues of A in case of
Jordan blocks. The closer −μ0 gets to the largest real part, the larger the
constant C0 becomes.
For the nonlinear terms g(u) = f (u) − Au = O(u2Rd ) the following
holds: For all b > 0 there exists a δ0 > 0 such that uRd ≤ δ0 implies
g(u)Rd ≤ buRd .
The variation of constant formula, cf. (2.6),
 t
u(t) = etA u0 + e(t−s)A g(u(s)) ds
0
then implies
 t
u(t)Rd ≤ e Rd →Rd u0 Rd +
tA
e(t−s)A Rd →Rd g(u(s)Rd ds
0
 t
≤ C0 e−μ0 t u0 Rd + C0 e−μ0 (t−s) bu(s)Rd ds,
0
and as a consequence
 t
e μ0 t
u(t)Rd ≤ C0 u0 Rd + C0 eμ0 s bu(s)Rd ds.
0

Gronwall’s inequality, cf. Lemma 2.2.8, applied to eμ0 t u(t)Rd finally im-
plies eμ0 t u(t)Rd ≤ C0 u0 Rd eC0 bt , respectively
u(t)Rd ≤ C0 u0 Rd e(C0 b−μ0 )t .
2.3. Special solutions 41

Choosing b = μ0 /(2C0 ) defines δ0 = δ0 (b). Therefore, with μ = μ0 /2, we

find u(t)Rd ≤ C0 u0 Rd e−μt → 0 for t → ∞. Since this additionally
implies that for any given ε > 0 we can choose δ = 12 min{C0−1 ε, δ0 } such
that u0 Rd < δ implies u(t, u0 )Rd ≤ C0 δ < ε, the asymptotic stability of
u∗ = 0 follows.
b) In order to prove b) we show that in the direction of the unstable
subspaces there is a sector with radius ε which is entered by the solutions
along the sides through the origin and which is left by the solutions along
the side opposite to the origin. See Figure 2.5.

Figure 2.5. The phase portrait and the sector in the unstable case
in a typical situation.

We start with a linear change ofcoordinates

 such that after the transform
A1 0
the linear part is of the form A = , where A1 ∈ Rk×k belongs to
0 A2
the part of the spectrum of A with positive real part and A2 ∈ R(d−k)×(d−k)
to the part of the spectrum of A with non-positive real part. Hence, there
exists a σ > 0, such that for all eigenvalues λj of A1 we have Re λj > σ.
Moreover, the change of coordinates is made in such a way that the norm of
the off-diagonal elements of the transformed matrix A is less than γ. In order
 
to do so, we assume ( kj=1 km=1,m=j |ajm |2 )1/2 ≤ γ for which we further
assume γ ≤ σ/20. From linear algebra it is known that this can always be
achieved by using modified Jordan blocks. By changing the length of the

λ 1
vectors of the basis for instance the Jordan block can be changed
0 λ
 
λ r
into for every fixed r > 0. Like in a) we find for the nonlinear
0 λ
terms that for all b > 0 there exists a δ0 > 0 such that uRd ≤ δ0 implies
g(u)Rd ≤ buRd .
k d
Next we define R2 = j=1 |uj | and ρ = j=k+1 |uj | and assume
2 2 2
∗
that u = 0 is stable. Then for all ε > 0 there exists a δ > 0, such that
ρ(0) + R(0) < δ implies ρ(t) + R(t) < ε for all t ≥ 0. For the transformed
42 2. Basic ODE dynamics

system we find

d   k k
d d
2R R = (R2 ) = |uj |2 = 2Re uj ∂t uj
dt dt dt
j=1 j=1


k 
=2Re uj (λj uj + ajm um + gj ).
j=1 m=j

Using

k 
k
2Re uj λj uj ≥ 2σ uj uj = 2σR2
j=1 j=1
and

k  
k 
k
|2Re uj ajm um | ≤ 2( |ajm |2 )1/2 R2 ≤ 2γR2 ,
j=1 m=j j=1 m=1,m=j

together with

k 
|2Re uj gj | ≤ 2Rb ρ2 + R2 ≤ 2Rb(ρ + R)
j=1

yields
d
R ≥ 2σR2 − 2γR2 − 2bR(ρ + R).
2R
dt
Choosing b = σ/10 yields
d
R ≥ σR/2 − bρ.
dt
Similarly, we find
d
ρ ≤ σρ/20 + b(ρ + R).
dt

where we used 2Re dj=k+1 uj λj uj ≤ 0. Since
σR/2 − bρ − σρ/20 − b(ρ + R) ≥ σ(R − ρ)/4
we finally obtain
d
(R − ρ) ≥ σ(R − ρ)/4
dt
and as consequence
R(t) − ρ(t) ≥ (R(0) − ρ(0))eσt/4 .
For solutions with R(0) = 2ρ(0) it follows that R(t) ≥ ρ(0)eσt/4 . However,
this contradicts the assumption of stability, since R(t)+ρ(t) ≤ ε for all t ≥ 0
is not possible, independent of how small ρ(0) > 0 or δ > 0 has been chosen.

2.3. Special solutions 43

Example 2.3.5. We consider again the situation from Example 2.3.2. As

a consequence of Theorem 2.3.4, the linear stability analysis is sufficient to
determine the stability of the fixed points in the nonlinear system. Hence,
the fixed point u∗1 = 0 is unstable, since the linearization A = 1 ∈ R1×1
has the eigenvalue 1, whereas the fixed points u∗2,3 = ±1 are asymptotically
stable, since A = −2 ∈ R1×1 has the eigenvalue −2.

Example 2.3.6. An example of an ODE with a fixed point in the origin

which is stable in the linearized system, but unstable in the full system is
given by u̇ = u3 .

The following theorem states that near a hyperbolic fixed point the flow
can be completely linearized by a change of coordinates h. For a proof we
refer to [Tes12, Theorem 9.9].

Definition 2.3.7. A fixed point u∗ is called hyperbolic, if the linearization

A = ∂u f (u∗ ) has no eigenvalues with Re λ = 0.

Theorem 2.3.8. (Hartman-Grobman) Let u∗ be a fixed point for (2.14),

let St be the flow of (2.14), let A = ∂u f (u∗ ) ∈ Rd×d be the linearization of f
in u∗ , and assume that A has no eigenvalues with zero real part. Then there
exists an homeomorphism h from a neighborhood U of u∗ to a neighborhood
V of u∗ such that for all u0 ∈ U there exists an open interval I0 ⊂ R, 0 ∈ I0
such that for all t ∈ I0 we have
h ◦ St u0 = etA h(u0 ).
Thus, h maps trajectories of (2.14) near u∗ to trajectories of the lineariza-
tion ẏ = Ay.

Remark 2.3.9. There is also a discrete version of the Hartman-Grobman

theorem: Consider the nonlinear map un+1 = f (un ) and assume that f (u∗ ) =
u∗ and that all eigenvalues λ of linearization A = ∂u f (u∗ ) around u∗ satisfy
|λ| = 1. Then there exists a homeomorphism h in a neighborhood U of u∗
such that h(f (u)) = Ah(u) for all u ∈ U .

It is somewhat surprising that even for analytic f the map h is in general

not differentiable; see Exercise 2.9.

2.3.2. Periodic solutions. The first non-trivial dynamical object is a pe-

riodic solution.

Definition 2.3.10. A solution u = u(t) of the ODE (2.14) is called periodic

if u(t + T ) = u(t) for a T > 0 and all t ∈ R. If moreover, u(t) = u(t + τ )
for all 0 < τ < T , then T is called the minimal period.
44 2. Basic ODE dynamics

Example 2.3.11. We consider the two-dimensional ODE

u̇1 = − u2 + u1 (1 − u21 − u22 ),
u̇2 =u1 + u2 (1 − u21 − u22 ).
By introducing polar coordinates u1 = r cos(φ), u2 = r sin(φ) we obtain
ṙ = r − r3 and φ̇ = 1.
In order to understand the dynamics of this ODE we visualize its flow in
the phase plane.

−1

−1 0 1
Figure 2.6. Flow for u̇1 = −u2 + u1 (1 − u21 − u22 ) and u̇2 = u1 +
u2 (1 − u21 − u22 ).

From the phase portrait we find that all solutions converge towards the
circle r = 1 which is a periodic solution with the minimal period T = 2π.
Moreover, the origin r = 0 is an unstable fixed point in the r equation.
As an exercise, we may consider the linearization around (u1 , u2 ) = 0. We
obtain      
u̇1 u1 1 −1
=A with A = ,
u̇2 u2 1 1
with eigenvalues λ1,2 = 1 ± i. Since Re λ1,2 = 1 > 0 we also have with
Theorem 2.3.4 the instability of the origin. On the other hand, from the
phase portrait the periodic solution r = 1 seems to be asymptotically stable.
However, as we see in a moment we have to be more precise when we talk
about stability of periodic solutions.

The stability or instability of non-trivial periodic solutions is a non-

trivial task due to the fact that the derivative u̇per of the periodic orbit
uper solves the linearization v̇ = Df (uper )v around the periodic orbit uper .
Hence, the linearization possesses a Floquet exponent with real part zero,
and so even a generalization of Theorem 2.3.4 to non-autonomous systems
would not be applicable for proving stability in the nonlinear system. In
order to study stability of periodic solutions we proceed as follows. We
2.3. Special solutions 45

introduce a so called Poincaré section, a hyperplane which intersects the

periodic orbit transversally. For our example we choose for instance
S = {(x, y) : u2 = 0, u1 ∈ (1/2, 3/2)}.
Transversality means that in the intersection point u∗ = (u1 , u2 ) = (1, 0) the
Poincaré section S and the vector field f (1, 0) = (0, 1) span the complete
phase space R2 . Then we define the so called Poincaré map Π : S → S as
follows: for u0 ∈ S we let Π(u0 ) be the first intersection point of t → u(t, u0 )
and S for t > 0, i.e., in the example Π(u0 ) = u(2π, u0 ). As Figure 2.7
illustrates, Poincaré maps to different Poincaré sections are conjugated to
each other in the following sense. Let ΠS1 ,S2 be the map from section S1 to
S2 . Then we have ΠS1 ,S1 = ΠS1 ,S2 ◦ ΠS2 ,S2 ◦ ΠS2 ,S1 .

Figure 2.7. Two Poincaré maps to different Poincaré sections are

conjugated to each other.

This fact and the fact that the intersection point u∗ of the periodic
solution is a fixed point of the Poincaré map Π, i.e., Π(u∗ ) = u∗ , lead to the
following definition.

Definition 2.3.12. a) A fixed point u∗ is called stable for the iteration

un+1 = Π(un ) with Π : Rd → Rd if for any ε > 0 there is a δ > 0 such that
u0 − u∗ Rd < δ implies Πn (u0 ) − u∗ Rd < ε for all n ∈ N. Otherwise,
it is called unstable. A stable fixed point is called asymptotically stable if
additionally limn→∞ Πn (u0 ) = u∗ holds.
b) A periodic solution for the ODE (2.14) is called stable, unstable, or
asymptotically stable if the fixed point of the associated Poincaré map is
stable, unstable, or asymptotically stable.
46 2. Basic ODE dynamics

The eigenvalues of the linearization DΠ of the Poincaré map now play

an analogous role as the eigenvalues of the linearization A around a fixed
point, cf. Theorem 2.3.4.
Definition 2.3.13. The eigenvalues of the linearization DΠ of the Poincaré
map are called Floquet multipliers.
Theorem 2.3.14. Let u∗ be a fixed point of the map Π : Rd−1 → Rd−1 and
let A = DΠ ∈ Rd−1×d−1 be the linearization of Π in u∗ .
a) If all eigenvalues λj of DΠ satisfy |λj | < 1, then u∗ is asymptoti-
cally stable.
b) If A has an eigenvalue λ with |λ| > 1, then u∗ is unstable.

Proof. The proof goes along the lines of the proof of Theorem 2.3.4. 
Example 2.3.15. In order to prove the stability of the periodic solution
r = 1 in Example 2.3.11 it remains to compute the Floquet multipliers.
Since r = r − 1 satisfies r˙ = −2
r + O(
r2 ) and since Π(u0 ) = u(2π, u0 ) for
u0 ∈ S due to φ solving φ̇ = 1, we find
DΠ((1, 0)) = e−2·2π ∈ R1×1 .
Thus we have one Floquet multiplier with |e−4π | < 1, which implies the
stability of the periodic solution r = 1.

2.3.3. Homoclinic and heteroclinic solutions. Homoclinic and hetero-

clinic solutions connect fixed points with themselves or other fixed points.
Pulse and front solutions in PDEs correspond to homoclinic and heteroclinic
solutions in associated ODEs.
Definition 2.3.16. A solution u = u(t) of the ODE (2.14) is called hetero-
clinic, if u+ = u− , or homoclinic, if u+ = u− , connection between the fixed
points u− and u+ if limt→−∞ u(t) = u− and limt→∞ u(t) = u+ .

Homoclinic and heteroclinic solutions converge to the fixed points along

special sets, namely to the fixed point u− along the unstable manifold of u−
for t → −∞, and to the fixed point u+ along the stable manifold of u+ for
t → ∞.
Definition 2.3.17. Let u∗ be a fixed point of the ODE (2.14). The set
Ws = {us ∈ Rd : ∃ β > 0 : lim u(t, us ) − u∗ Rd eβt = 0}
t→∞
is called the stable manifold of u∗ . The set
Wu = {uu ∈ Rd : ∃ β > 0 : lim u(t, uu ) − u∗ Rd eβ|t| = 0}
t→−∞
is called the unstable manifold of u∗ .
2.3. Special solutions 47

Example 2.3.18. For u̇1 = −u1 , u̇2 = u2 we have Ws = {(u1 , 0) : u1 ∈

R} and Wu = {(0, u2 ) : u2 ∈ R}.

The following theorem guarantees that the sets Ws and Wu from Defini-
tion 2.3.17 are smooth manifolds and that they are invariant under the flow
of the ODE.

Theorem 2.3.19. (Invariant manifolds) Let u∗ be a fixed point of the

ODE (2.14), f ∈ C k (Rd , Rd ), and let A = ∂u f (u∗ ) ∈ Rd×d be the lineariza-
tion of f in u∗ . Let
Es = span{ϕ : ϕ eigenvector of A to eigenvalues λ with Re λ < 0}
be the so called stable subspace and let
Eu = span{ϕ : ϕ eigenvector of A to eigenvalues λ with Re λ > 0}
be the so called unstable subspace, where eigenvectors include here in all cases
generalized eigenvectors. Then there exists a unique C k -manifold Ws =
Ws (u∗ ) tangential to the stable subspace Es , which coincides with the stable
manifold from Definition 2.3.17, and a unique C k -manifold Wu = Wu (u∗ )
tangential to the unstable subspace Eu , which coincides with the stable man-
ifold from Definition 2.3.17. Moreover, there exists a (non-unique) C k−1 -
center manifold tangential to the center subspace
Ec = span{ϕ : ϕ eigenvector of A to eigenvalues λ with Re λ = 0}.
If f ∈ C ∞ , then Ws , Wu ∈ C ∞ . The center manifold Wc can be chosen to
be in C r for all r < ∞. All these manifolds are invariant under the flow of
the ODE (2.14). A set M is called invariant if u0 ∈ M implies u(t, u0 ) ∈ M
for all t ∈ R.

Proof. The lengthy proof of this theorem is well documented in [Van89].

A sketch of the proof of the existence of the center manifold can be found
in §13.1. 

Example 2.3.20. Consider the equations for the mathematical pendulum

without friction, namely
u̇1 = u2 ,
(2.20)
u̇2 = − sin(u1 ),
where u1 is the angle between the pendulum and the vertical axis. We find
the fixed points u∗ = (kπ, 0) for k ∈ Z. The linearization at (kπ, 0) is given
by
    
d v1 0 1 v1
= ,
dt v2 − cos(kπ) 0 v2
48 2. Basic ODE dynamics

{−1, 1} for k ∈ 2Z + 1
which yields the eigenvalues λ ∈ implying sad-
{−i, i} for k ∈ 2Z
dles for odd
 k and centers
 for even k. The eigenvectors at the saddles are
1 1
ϕ1 = , ϕ2 = , in the stable direction, and unstable direction,
1 −1
respectively. Physical intuition lets us suspect the existence of, for instance,
heteroclinic orbits

γ+ = Ws ((π, 0)) ∩ Wu ((−π, 0)) and γ− = Wu ((π, 0)) ∩ Ws ((−π, 0)),

corresponding to one complete rotation of the pendulum from the unsta-

ble upper rest state to itself in infinite time. Since the points (kπ, 0) and
((k + 2)π, 0) can be identified, γ− , γ+ can also be called homoclinic. These
two orbits separate R2 into two domains, a bounded domain inside and an
unbounded domain outside the two orbits. We further suspect the domain
inside to be filled with periodic orbits corresponding to oscillations of the
pendulum with amplitude < π. See Figure 2.8.

4
f1=0
3 f =0
2
2

−1

−2

−3

−4
−4 −2 0 2 4

Figure 2.8. Phase portrait for the undamped pendulum.

The whole interior of the “eye” is filled with periodic solutions. There-
fore, each of them is stable in the sense of Definition 2.3.12 with Floquet
multiplier 1.

Remark 2.3.21. Instead of “physical intuition” we should rather use the

fact that (2.20) is a Hamiltonian system, see also Chapter 4. Here, (2.20) is
the first order system belonging to the second order equation

(2.21) ü = − sin(u) = f (u)

corresponding to Newton’s law, namely that the change of momentum equals

the acting force. The qualitative behavior of equations such as (2.20) can
be obtained by the following procedure independently of the concrete form
2.4. ω-limit sets and attractors 49

of f . Multiplying (2.21) by u̇ shows that

 
d 1 2
(2.22) u̇ − F (u) = 0, where F = f.
dt 2
Here 12 (u̇)2 and −F (u) are called the kinetic and potential energy, respec-
tively, and (2.22) shows that the total energy E = 12 (u̇)2 −F (u) is conserved.
Hence, orbits of (2.20) lie on level sets of E. It turns out that all equations
from classical mechanics without friction can be written as Hamiltonian
systems in the form
   
q 0 I
(2.23) ∂t = J∇H(q, p), where J =
p −I 0
is skew symmetric, where q ∈ Rd and p ∈ Rd are the position and the
momentum coordinates, and where H : R2d → R is called the Hamiltonian,
see Chapter 4. Also PDEs can have a Hamiltonian structure, see for instance
§8.1 and §8.2.

Stable, center, and unstable manifolds can also be generalized from fixed
points to more complicated objects, for instance to periodic solutions. They
exist for discrete dynamical systems, too.
Remark 2.3.22. Let u∗ be a fixed point of the iteration un+1 = Πun . The
set
Ws = {us ∈ Rd : ∃ β > 0 : lim Πn (us ) − u∗ Rd eβn = 0}
n→∞
is called the stable manifold of u∗ . The set
Wu = {uu ∈ R : ∃ β > 0 :
d
lim Πn (uu ) − u∗ Rd eβ|n| = 0}
n→−∞

is called the unstable manifold of u∗ . They exist as smooth invariant mani-

folds with similar properties as the ones explained in Theorem 2.3.19.

2.4. ω-limit sets and attractors

In this section we are interested in objects which describe the dynamics for
t → ∞. These are so called ω-limit sets and attractors. We characterize
them for two-dimensional autonomous systems and gradient systems. The
concepts of this section will later on be applied in Part II of this book to
PDEs on spatially bounded domains, which very often can be written as
countably infinite-dimensional dynamical systems. Therefore, throughout
this subsection, we consider a general dynamical system X  u0 → St u0 ,
where X is some possibly infinite-dimensional Banach space. The theory has
to be modified in Part III and Part IV where PDEs on unbounded domains
will be handled. For simplicity, the reader may think of St as being defined
50 2. Basic ODE dynamics

by the solutions of some ODE u̇ = f (u) in the phase space X = Rd . This

section follows rather closely [Rob01, §10], including a number of examples.

2.4.1. ω-limit sets. Given some initial condition u0 for a dynamical sys-
tem St u0 = u(t, u0 ) on X, the behavior of the solution for t → ∞ is described
by the ω-limit set, defined by
(2.24) ω(u0 ) = {v ∈ X : ∃ (tn )n∈N with tn → ∞ and lim u(tn , u0 ) = v}.
n→∞
Thus, the ω-limit set of u0 consists of all limit points of the forward orbit
through u0 . Hence, an equivalent characterization is
(2.25) ω(u0 ) = ∩t≥0 ∪s≥t u(s, u0 ).
If γ+ = γ+ (u0 ) is the (forward) orbit through u0 and v ∈ γ+ then ω(v) =
ω(u0 ) such that we also write ω(γ+ ) := ω(u0 ).
Theorem 2.4.1. The set ω(γ+ ) is closed and invariant. If X = Rd and γ+
is bounded, then ω(γ+ ) is compact, connected and non-empty. In general,
if ∪t≥t0 St u0 is compact for some t0 ≥ 0, then ω(γ+ ) is compact, connected
and non-empty.

Proof. We first consider the case X = Rd .

a) ω(γ+ ) as set of limit points is closed.
b) Next, we prove the invariance. Let p ∈ ω(γ+ ). Then there exists
a sequence tn → ∞ such that limtn →∞ u(tn ) = p. We have to prove that
u(t, p) ∈ ω(γ+ ). Since u(t + tn , x0 ) = u(t, u(tn , x0 )) it follows for n → ∞
that
u(t + tn , x0 ) → u(t, p),
which implies that γ(p) ⊂ ω(γ+ ).
c) With γ+ bounded, ω(γ+ ) is bounded. Since ω(γ+ ) is closed by a),
compactness follows.
d) Suppose that γ+ consists of more than one point, i.e., γ+ is not a
fixed point. Then γ+ consists of infinitely many points. Hence, there exists
at least one limit point p of the bounded set γ+ . Suppose that γ+ consists
only of a fixed point. Then γ+ = ω(γ+ ) is also non-empty.
e) Suppose that ω(γ+ ) is not connected, i.e., there exist closed sets A1
and A2 satisfying ω(γ+ ) = A1 ∪ A2 and A1 ∩ A2 = ∅. Since γ+ is bounded,
there exists a R > 0, such that γ+ ⊂ BR (0) = {x ∈ Rd : x ≤ R}. Let
δ > 0 be the distance between A1 and A2 . Define
A3 = {u ∈ BR (0) : δ/4 ≤ dist(u, ω(γ+ ))}
where dist(u, A) = inf a∈A u − a. Obviously the solution must pass A3
infinitely many times, and hence there must be a limit point in A3 which
contradicts the assumption ω(γ) = A1 ∪ A2 .
2.4. ω-limit sets and attractors 51

In the general case, i.e., with X some Banach space, the proof works
the same way, with ω(γ+ ) compact as a closed subset of the compact set
∪t≥t0 St u0 . Note that this compactness argument is used in d) and e). 

2.4.2. Attractors. Attractors are compact sets describing the asymptotic

dynamics of the system in the limit t → ∞. They exist for so called dissi-
pative systems.
Definition 2.4.2. The flow St : X → X is called dissipative if there exists
a compact set B such that for any bounded set M ⊂ X there exists a t0 =
t0 (M ) such that St M ⊂ B for all t ≥ t0 . The set B is then called absorbing.

The goal is to define the so called global attractor A which contains as

much information as possible about the asymptotic behavior for t → ∞. If
the system is dissipative and thus has a compact absorbing set B, a first idea
would be to take ∪u∈B ω(u). However, this set in general does not contain
homo- or heteroclinic connections, which we already know to be relevant for
the asymptotic dynamics. Therefore, a better choice turns out to be
A = ω(B) = {v ∈ X : ∃ tn → ∞, ∃ (un )n∈N ⊂ B,
(2.26)
with limn→∞ u(tn , un ) = v}.
Definition 2.4.3. A set A ⊂ X is called an attractor of the dynamical
system St : X → X if
(i) A is compact and invariant;
(ii) there is a neighborhood U of A such that A attracts U .
The basin of attraction of A is defined as
B(A) = {u ∈ X : dist(St u, A) → 0 as t → ∞},
where dist(A, B) = supa∈A inf b∈B a − b. A set A ⊂ X is called the global
attractor for St if additionally
(iii) A attracts all points in X, i.e., we have B(A) = X.

Theorem 2.4.4. If the flow St is dissipative and if B is a compact absorbing

set, then A = ω(B) is the unique global attractor. Moreover, A is connected,
the maximal compact invariant set, and the minimal set that attracts all
bounded sets.

Proof. The fact that ω(B) is nonempty, compact, invariant and connected
follows as in Theorem 2.4.1. To show that A is the maximal compact in-
variant set, let Y be a compact and invariant set. Then St Y = Y and,
since B is absorbing, St Y ⊂ B for t ≥ t0 , hence Y ⊂ B and therefore
ω(Y ) = Y ⊂ A = ω(B). This shows that A is the maximal compact invari-
ant set.
52 2. Basic ODE dynamics

Next we show that A attracts all bounded sets. Suppose that this is not
the case, then there is a bounded set Y , a δ > 0, and a sequence tn → ∞
with
dist(Stn Y, A) ≥ δ.
Thus, there are un ∈ Y with dist(Stn un , A) ≥ δ/2. Since Y is bounded and
B absorbing we have Stn un ∈ B for n large enough, and as B is compact
there is a subsequence with

(2.27) Stnj unj → v ∈ B and dist(v, A) ≥ δ/2.

However, with vj = St0 unj ∈ B we have

v = lim Stnj unj = lim Stnj −t0 St0 unj

j→∞ j→∞

and hence v ∈ A, which contradicts (2.27). Obviously, A is also the minimal

set that attracts all bounded sets since St A = A. 

Example 2.4.5. Consider u̇ = u−u3 . The attractor is given by A = [−1, 1]

and contains the heteroclinic connections between 0 and ±1. Every set
[−1 − δ, 1 + δ] with δ > 0 is an absorbing set.

2.4.3. Shadowing and upper-semicontinuity of attractors. Clearly,

an important issue is the relation of the flow St in X to that on the attractor.
The following theorem roughly says that given an initial condition in X,
there exists a τ > 0 such that after the time τ the flow can be approximated
by a flow on the attractor for some finite time.

Theorem 2.4.6. Let A be the global attractor for the flow St and let u0 ∈ X.
For all ε > 0 and T > 0 there exists a τ = τ (ε, T ) > 0 and a point v0 ∈ A
such that
Sτ +t u0 − St v0 X ≤ ε for all 0 ≤ t ≤ T.

Proof. From the continuous dependence on the initial conditions for given
ε, T > 0 there exists a δ = δ(ε, T ) such that

u1 − v0  ≤ δ ⇒ St u1 − St v0  ≤ ε for t ∈ [0, T ].

Since A is the global attractor, for any u0 ∈ X and every δ > 0 there exists
a time τ and a v0 ∈ A such that u1 − v0  ≤ δ where u1 = S(τ )u0 . 
An approximation of a solution (St u0 )t≥τ by a single solution in the
attractor cannot be expected in general. However, solutions can be approx-
imated (or shadowed) by so called pseudo-orbits in the attractor. Moreover,
due to the attractivity property of A the approximation becomes better and
better for larger and larger times.
2.4. ω-limit sets and attractors 53

Corollary 2.4.7. (Shadowing) For all u0 ∈ X there exists a sequence

(εn )n∈N of errors εn > 0 with εn → 0, an increasing sequence (tn )n∈N of
times with tn+1 − tn → ∞ for n → ∞, and a sequence (vn )n∈N of points
vn ∈ A such that
St u0 − St−tn vn X ≤ εn for all tn ≤ t ≤ tn+1 ,
and vn+1 − Stn+1 −tn vn X → 0 for n → ∞.

However, in general a flow St u0 cannot be approximated by a single flow

on the attractor as t → ∞.
Example 2.4.8. For (x, y, z) ∈ R3 consider
(2.28) ẋ = z(x + y) + x − xr2 , ẏ = z(−x + y) + y − yr2 , ż = −z|z|,
where r = (x2 + y 2 )1/2 or equivalently in polar coordinates
ṙ = r − r3 , φ̇ = −z, ż = −|z|z.
The global attractor is given by A = {(x, y, z) ∈ R3 : z = 0, x2 + y 2 ≤ 1},
and the dynamics on A is given by ṙ = r − r3 and φ̇ = 0. Hence, the
attractor consists of the origin which is a fixed point, the circle of fixed
points S 1 = {x2 + y 2 = 1}, and the radial heteroclinic connections between
the origin and the points on S 1 .
However, given z0 =0 we obtain z(t)=z0 /(1+|z0 |t) and hence φ(t)=φ0 −
sgn(z0 ) ln(1 + |z0 |t). Thus, the solution converges (algebraically slow) to
A but it does not converge to some particular solution on A. It can only
be approximated by a sequence of solutions (i.e., fixed points on S), with
smaller and smaller errors on longer and longer time intervals.

2
6

0 4

1 2
−2 0
0
0 1 0 500 1000 1500 2000

Figure 2.9. Left: Two orbits for (2.28) approaching the circle S :=
{(x, y, z) ∈ R3 : x2 + y 2 = 1} of fixed points from above and below,
respectively. Right: illustration of the notion of pseudo orbits for
(2.28), here consisting of fixed points on S.

Another important question is the robustness of attractors under per-

turbations of the dynamical system. As the following example shows, in
general we can only expect upper semicontinuity.
54 2. Basic ODE dynamics

Example 2.4.9. For 0 ≤ ε < 1 consider u̇ = f (u, ε) where

⎧
⎪
⎪ −(u + 1) for u < −2,
⎪
⎪
⎨ 1 − (1 − ε)(u + 2) for − 2 ≤ u < −1,
f (u, ε) = −εu for |u| ≤ 1,
⎪
⎪
⎪
⎪ −1 − (1 − ε)(u − 2) for 1 < u ≤ 2,
⎩
1−u for 2 < u,
cf. Figure 2.10. For all ε > 0 the global attractor is given by Aε = {0}.
However, for ε = 0 we have A0 = [−1, 1].

-1

-2
-2 -1 0 1 2

Figure 2.10. The “vector field” for Example 2.4.9 for ε > 0.

Theorem 2.4.10. (Attractor upper semicontinuity) Assume that for

μ ∈ [0, μ0 ) each of the flows (Stμ )t≥0 has a global attractor Aμ such that
∪0≤μ<μ0 Aμ ⊂ Q for some bounded set Q, and that for each t > 0 the flows
Stμ converge to St0 uniformly on bounded subsets M , i.e.,
sup Stμ u0 − St0 u0 X → 0 as μ → 0.
u0 ∈M

Then
dist(Aμ , A0 ) → 0 as μ → 0.

Proof. Let ε > 0. Since A0 attracts Q there exists a t > 0 such that St0 Q is
a subset of the ε/2-neighborhood N (A0 , ε/2) of A0 , i.e., St0 Q ⊂ N (A0 , ε/2).
Next, for μ > 0 sufficiently small, we have
sup Stμ u − St0 u ≤ ε/2
u∈Q

for all u ∈ Q. Since Aμ ⊂ Q we have Aμ = Stμ Aμ ⊂ Stμ Q ⊂ N (A0 , ε). 

Only with a number of additional assumptions, cf. [Rob01, Theorem
10.17], lower semicontinuity and hence continuity can be obtained, too. In
general, as the previous example has shown, lower semicontinuity is wrong.
2.4. ω-limit sets and attractors 55

2.4.4. Planar systems. For autonomous ODEs in two space dimensions

the possible ω-limit sets and attractors are relatively easy.
Theorem 2.4.11. (Poincaré-Bendixson) Consider the ODE (2.14) in
R2 and assume that the positive semiorbit γ+ (u0 ) through u0 is bounded. If
ω(u0 ) contains no fixed point, then ω(u0 ) is a periodic solution. If ω(u0 )
contains a fixed point, but only finitely many, then ω(u0 ) is either a single
fixed point or it consists of fixed points with the homoclinic and heteroclinic
connections between the fixed points.

Idea of the proof. We refrain from giving a complete proof of Theorem

2.4.11, cf. [Tes12, §7.3], since the ideas of the proof will not be used any
more in the following. The proof is based on the Jordan curve theorem saying
that a closed, non-self-intersecting curve separates R2 in an interior and an
exterior part. Since this is only true in R2 this assumption is essential. As
a consequence, Poincaré maps have to be monotonic. See Figure 2.11.

Figure 2.11. Monotonicity of the Poincaré maps.

This yields that the ω-limit set of an orbit γ intersects every Poincaré
section in only one point. If there is no fixed point in ω(γ), then ω(γ) is a
periodic orbit. 
Example 2.4.12. Using the Poincaré-Bendixon theorem allows us to prove
the existence of a periodic solution for
(2.29) ẋ = y and ẏ = −x + y(1 − x2 − 2y 2 ).
A direct consequence of the Poincaré-Bendixon theorem is that a posi-
tively invariant bounded set for ẋ=f (x), x∈R2 , which does not contain
a fixed point, must contain a periodic orbit. The set A = {(x, y) ∈ R2 :
1/4< x2 + y 2 <1} is positively invariant for (2.29). This follows by looking
at the sign of
d
(x(t)2 + y(t)2 ) = 2y 2 (1 − x2 − 2y 2 )
dt
56 2. Basic ODE dynamics

at the boundaries of A. We have 2y 2 (1 − x2 − 2y 2 )|x2 +y2 =1/4 ≥ 0 and

2y 2 (1 − x2 − 2y 2 )|x2 +y2 =1 ≤ 0.

1.5

0.5

−0.5

−1

−1.5

−2
−2 −1 0 1 2

Figure 2.12. Phase portrait of (2.29) with the invariant region.

2.4.5. Gradient systems and Lyapunov functions. Another class of

systems with restricted asymptotic behavior is given by gradient systems.
They are of the form
(2.30) u̇ = −∇V (u),
with the potential V ∈ C 1 (X, R), where X = Rd could be some general
Banach space. For simplicity let X = Rd . Obviously, fixed points of (2.30)
satisfy ∇V (u) = 0, i.e., they are critical points of V . Moreover, V decays
along solutions of (2.30), see Figure 2.13 for a one-dimensional sketch.
Theorem 2.4.13. The function t → V (u(t)) is strictly decaying for solu-
tions u = u(t) of (2.30) except in case that u is a fixed point. In particular,
there are no non-trivial periodic solutions in gradient systems.

Proof. We have
 
d d
(2.31) V (u(t)) = (∇V (u(t))) T
u(t) = −∇V (u(t))2Rd < 0
dt dt
except in case that u is a fixed point. Suppose that there exists a non-trivial
periodic solution with u(t) = u(t + T ) with minimal period T > 0. Then by
(2.31) we have V (u(t)) > V (u(t + T )) = V (u(t)), which is a contradiction.

Remark 2.4.14. The linearization A ∈ Rd×d at a fixed point in a gradient
system is a symmetric matrix. Therefore, all eigenvalues are real.
2.4. ω-limit sets and attractors 57

Saddle

Minimum
Minimum

Figure 2.13. The solutions decay along the gradient of the potential.

It turns out that with only a few additional assumptions the ω-limit sets
and the global attractor of gradient systems can be completely described.

Theorem 2.4.15. Suppose that V (u) → ∞ for uRd → ∞, that the set E
of fixed points is finite, and that the fixed points are all hyperbolic. Then,
for all u0 ∈ Rd we have ω(u0 ) = u∗ for some fixed point u∗ . Moreover, for
(2.30) there exists a compact absorbing set B, and the attractor A = ω(B)
consists of finitely many fixed points and the connecting orbits between the
fixed points.

Proof. See [Rob01, §10.6.1]. 

Most properties of gradient systems are also true in case that a Lyapunov
function exists for the dynamical system.

Definition 2.4.16. A Lyapunov function for a dynamical system St : X →

X is a continuous function Φ : B → R on a positively invariant set B ⊂ X
such that
(i) given u0 ∈ B the function t → Φ(St u0 ) is non-increasing,
(ii) if Φ(St u0 ) = Φ(u0 ) for some t > 0 then u0 is a fixed point.

Obviously, for gradient systems u̇ = −∇V (u) the potential V is a Lya-

punov function on Rd . Conversely, suppose that some dynamical system
St has a Lyapunov function Φ defined on X = Rd with the properties that
Φ(u) → ∞ as uRd → ∞ and dt d
Φ(u(t)) < 0 outside some bounded set B.
For such systems we have global existence of solutions and since B is a com-
pact absorbing set the system is dissipative. Suppose further that the set
E of fixed points is discrete. Then the assertions of Theorem 2.4.15 remain
true. Therefore, such systems are often called gradient-like. Lyapunov func-
tions are very often used to prove stability and instability of fixed points,
cf. [HK91, §9.3-§9.4]. See also Example 2.6.3. The concepts of gradient
systems and Lyapunov functions are used in the analysis of PDEs, too, cf.
§5.3.
58 2. Basic ODE dynamics

2.5. Chaotic dynamics

In contrast to the relatively simple dynamics which can be found for au-
tonomous ODEs in one and two dimensions, for ODEs in dimensions three
and higher very complicated behavior can occur. It is reasonable to expect
that complicated dynamics occurs in very high-dimensional systems such as
for instance the one describing positions and velocities of 1Mol ≈ 6.022×1023
particles of an ideal gas in some container. Statistical mechanics, which
was initiated in 1870 by Ludwig Boltzmann, is based on the insight that
a description of individual particles does not make any sense for such sys-
tems, and that a statistical description is more appropriate. However, it is
surprising that very complicated dynamical behavior already occurs in low-
dimensional systems. This was already observed by Henri Poincaré around
1890, when he studied the N -body problem, cf. Chapter 4. However, this
fact only came apparent to a wider audience with the first computer simu-
lations made in the early 1960s. The meteorologist Edward Lorenz [Lor63]
found that already systems in R3 show a behavior which was later on called
chaotic. As a consequence of the interesting pictures which were produced
in the following years there was a big boom about chaos lasting for almost
30 years, cf. [Gle88, Man91]. According to [Kel93], chaos theory studies
the behavior of dynamical systems that are highly sensitive to initial condi-
tions, an effect which is popularly referred to as the butterfly effect. Small
differences in initial conditions (such as those due to rounding errors in nu-
merical computation) yield widely diverging outcomes for such dynamical
systems, rendering long-term prediction impossible in general.
We will call a dynamical system chaotic if there is a subset of the phase
space such that the flow restricted to this subset is conjugated to shift dy-
namics, which is a prototype of a chaotic dynamical system. The occurrence
of shift dynamics in a dynamical system is proved with the help of an in-
termediate step which is called Smale’s horseshoe. This is a geometric con-
struction of a chaotic dynamical system which is easier to detect in a given
dynamical system. Using this idea we will present with Silnikov chaos an ex-
ample of a three-dimensional ODE which exhibits chaotic behavior. Routes
to chaos in dissipative systems by sequences of local bifurcations will briefly
be described subsequently in §3.4. The occurrence of chaotic behavior in
Hamiltonian systems is discussed in §4.5. In the present book, chaos will
not play a central role, but one should keep in mind its existence already
in low-dimensional dynamical systems. Our presentation of this subject is
based on [GH83].
2.5. Chaotic dynamics 59

2.5.1. Shift dynamics. We will use shift dynamics as a prototype of a

chaotic dynamical system. On the set
Σ2 = {a : Z → {0, 1} : a = (ai )i∈Z }
which is equipped with the distance

d(a, b) = 2−|j| |aj − bj |.
j∈Z

we define the shift σ : Σ2 → Σ2 by (σ(a))i = ai+1 . Similarly, we define

Σ+2 where the index set Z is replaced by N. The subsequent theory will be
formulated for the shift in Σ2 , but can also be formulated in Σ+
2.

Theorem 2.5.1. We have that σ ∈ C(Σ2 , Σ2 ) has the following properties:

(i) There exist non-trivial periodic solutions to every minimal period;
(ii) There exists a dense orbit;
(iii) The sensitivity w.r.t. the initial conditions holds, i.e., for every
a ∈ Σ2 and every δ > 0 there exist b ∈ Σ2 and j ≥ 0 such that
d(σ j (a), σ j (b)) ≥ 1, although d(a, b) ≤ δ.

Proof. (i) The 1-periodic solutions are a = . . . 00000 . . . and a = . . . 11111 . . ..

The 2-periodic solutions are generated by 00, 01, 10 and 11, and the 3-
periodic ones by 000, 001, 010, 100, 110, 101, 011 and 111, etc.
(ii) Consider the orbit to the initial condition a, consisting of all finite se-
quences that generate the periodic solutions, i.e.,
a = . . . 0000|0100011011000001010100110101011111000000010010 . . . .
Right of | we have the position j = 0 and left of | the sequence is filled up
with zeroes. For a given ε > 0 and b ∈ Σ2 we have to find n ∈ N such
that d(b, σ n (a)) ≤ ε. For c ∈ Σ2 we have d(b, c) ≤ ε, if at least bj = cj for
|j| ≤ j0 (ε). The other cj for |j| ≥ j0 (ε) can be arbitrary. Since a contains
all finite sequences the claim follows by shifting a until the finite sequence
(bj )j=−j0 ,...,j0 occurs at the positions between −j0 and j0 .
(iii) Let a ∈ Σ2 and δ > 0, and set bj = aj for j ≤ j0 (δ) and aj0 (δ)+1 =
bj0 (δ)+1 . Then d(a, b) ≤ δ, but d(σ j0 (δ)+1 (a), σ j0 (δ)+1 (b)) ≥ 1. 
A general dynamical system is called chaotic if the flow is conjugated to
shift dynamics on a subset of its phase space, i.e.,
Definition 2.5.2. A discrete dynamical system Π : Rd → Rd is called
chaotic, if there is a set Λ ⊂ Rd and a homeomorphism h : Λ → Σ2 such
that on Λ the flows are conjugated, i.e., σ ◦ h|Λ = h ◦ Π|Λ .
Remark 2.5.3. This is a very strict definition of a chaotic dynamical sys-
tem. Chaos in the sense of [Dev89] for a map f ∈ C(M, M ), with M some
metric space, is defined by
60 2. Basic ODE dynamics

(1) the sensitive dependence on initial conditions;

(2) periodic points are dense in M ;
(3) topological transitivity, i.e., for all open subsets U, V of M , there
exists k ∈ N such that f k (U ) ∩ V = ∅.

Remark 2.5.4. Another definition of chaotic behavior in a system where

all forward orbits are bounded is the occurrence of a positive Lyapunov
exponent [Rue89]. Lyapunov exponents describe how the distance of nearby
solutions evolves in time. They are defined through
 
1
λ(u0 , ϕ) = lim sup ln Du(t, u0 )ϕ
t→∞ t

for ϕ, u0 ∈ Rd . For each initial condition u0 there are d such Lyapunov

exponents. Very often the Lyapunov exponents do not depend on u0 . A
positive Lyapunov exponent implies a sensitive dependence on the initial
conditions.

See Exercise 2.18 for an example of a map where conjugacy to the shift
on Σ+2 can be shown explicitly. A related and famous 1D iteration for which
chaotic behaviour can be shown for certain parameters is the logistic map,
see, e.g., [Dev89], and §3.4.1.

2.5.2. Smale’s horseshoe. The occurrence of shift dynamics in a general

dynamical system is very often proved with the help of an intermediate
step. There is a geometric construction of a chaotic dynamical system,
called Smale’s horseshoe, which is easier to detect in a given dynamical
system than shift dynamics. The construction is as follows. Starting with
S = [0, 1] × [0, 1] ⊂ R2 we define via Figure 2.14 a map f : S → R2 such
that f (S) ∩ S consists of two components, namely the two vertical strips V0
and V1 . There exist two horizontal strips H0 and H1 with f (Hj ) = Vj .

V0 V1

H1

H0

Figure 2.14. Smale’s horseshoe.

2.5. Chaotic dynamics 61

Under iteration of f most of the points of S leave S. The points which

stay in S under all iterations of f define a set
Λ = {x : f i (x) ∈ S, −∞ < i < ∞}.
Remark 2.5.5. The invariant set Λ ⊂ Rd is called hyperbolic since there is
a continuous invariant splitting of the tangent spaces TΛ Rd = EΛu ⊕ EΛs with
the following property: There exist constants C > 0 and λ ∈ (0, 1) with
a) Df −n (x)vRd ≤ Cλn vRd if v ∈ EΛu (x),
b) Df n (x)vRd ≤ Cλn vRd , if v ∈ EΛs (x).
Hyperbolic dynamics, cf. [KH97, Part 4], is one branch in the description
of chaos.

The set Λ has a complicated topological structure.

Lemma 2.5.6. The set Λ is a Cantor set, i.e., an uncountable, compact,
totally disconnected and nowhere dense set which consists entirely of limit
points.

Proof. Each horizontal strip Hi is mapped through f into the vertical strip
Vi = f (Hi ). We consider Vi ∩ Hj which is the image of some thin horizontal
strips Hij . We obtain vertical strips Vij = f 2 (Hij ) by two iterations of f .
See Figure 2.15.

V00 V10 V 11 V 01

H 10
H 11
f
f

H
01
H
00

Figure 2.15. Iteration of the horseshoe map f .

By more forward and backward iterations of f and the intersection of all

horizontal and vertical strips which are obtained in this way we get a closed,
non-empty, completely disconnected set Λ. Each point in Λ is a limit point
of Λ. Thus, Λ is a Cantor set. 

To each point x ∈ Λ we associate an infinite sequence a : Z → {0, 1} via

φ(x) = (ai )∞
i=−∞ if f (x) ∈ Hai .
i
62 2. Basic ODE dynamics

Theorem 2.5.7. There is a one-to-one map φ between Λ and Σ2 such that

the sequence b = φ(f (x)) can be computed from a = φ(x) by shifting the
indices bi = ai+1 . The map φ is a homeomorphism between the metric
spaces (Σ2 , d) and (Λ, ·Rd ).

Proof. We consider φ(f (x)) = (bi )∞

i=−∞ with f
i+1 (x) ∈ H . Hence, f i (x) ∈
bi
Hbi−1 = Hai and therefore bi = ai+1 which implies φ ◦ f = σ ◦ φ.
It remains to prove the continuity of φ : Λ → Σ2 and φ−1 : Σ2 → Λ.
For x ∈ Λ and all ε > 0 we have to show the existence of a δ > 0 such
that d(φ(x), φ(y)) < ε if x − y < δ. To a given ε > 0 there exists
a j0 = j0 (ε) such that d(a, b) ≤ ε is equivalent to aj = bj at least for
|j| ≤ j0 . aj and bj can be arbitrary for |j| > j0 . Therefore, the condition
d(φ(x), φ(y)) < ε uniquely defines two sequences a+ = (ai )jj=0 0
and a− =
−1
(ai )j=−j0 −1 . Associated with these sequences there are strips Va− and Ha+ .
If we choose δ > 0 so small that y ∈ Va+ ∩ Ha− if y − xRd ≤ δ we are done.
The continuity of φ−1 follows in the same way. 

As a consequence it follows that if we find a Smale’s horseshoe in a

dynamical system, then chaotic behavior in the sense of Definition 2.5.2 is
present in this system.

2.5.3. Silnikov chaos. Using the idea of Smale’s horseshoe we present a

first example of a chaotic dynamical system coming from an ODE, namely
Silnikov chaos. The presentation is based on [GH83, §6.5.1]. However, we
skip most analytic arguments and argue mostly by pictures. We consider
an autonomous three-dimensional ODE with a homoclinic orbit γ at the
origin which is assumed to be a hyperbolic fixed point with eigenvalues
λ ∈ R, ω, ω ∈ C, where Im ω = 0. See Figure 2.16(a). Silnikov [Sil65]
proved in 1965 the following result.

Theorem 2.5.8. If | Re ω| < λ, then the flow St can be perturbed in such a

way that the perturbed flow St has a homoclinic orbit γ
 close to γ and that
there exists a subset of R on which the first return map for the perturbed
3

flow St is conjugated to Smale’s horseshoe.

Idea of the proof. By the Hartman-Grobman theorem 2.3.8 we may

assume that the vector field is linear in a neighborhood of the origin, i.e.,
with α := Re ω,
⎛ ⎞ ⎛ ⎞⎛ ⎞
ẋ α −β 0 x
(2.32) ⎝ẏ ⎠ = ⎝β α 0 ⎠ ⎝y ⎠ , ω = α + iβ.
ż 0 0 λ z
2.5. Chaotic dynamics 63

a) b) c)

Σ1
q q

Σ0
R

p p

Figure 2.16. a) The homoclinic orbit in the example of Silnikov,

b) the inner map ψint : Σ0 → Σ1 , and c) the outer map ψout :
Σ1 → Σ0 .

The solutions of (2.32) are given by

⎛ ⎞ ⎛ αt ⎞
x e ((cos βt)x(0) − (sin βt)y(0))
(2.33) ⎝y ⎠ (t) = ⎝eαt ((sin βt)x(0) + (cos βt)y(0))⎠ .
z eλt z(0)

We define two sets

Σ0 ={(x, y, z) : x2 + y 2 = r02 and 0 < z < z1 },

Σ1 ={(x, y, z) : x2 + y 2 < r02 and z = z1 > 0}

and assume that these sets are contained in the previous neighborhood.
The solutions go from Σ0 to Σ1 according to Figure 2.16 b). The inner map
ψint : Σ0 → Σ1 , which maps a point a ∈ Σ0 into the first intersection point
of the associated solution with Σ1 , maps vertical vertical lines from Σ0 into
a logarithmic spiral in Σ1 . The outer map ψout transports a neighborhood
of q through the homoclinic solution into
 0 = {(x, y, z) : x2 + y 2 = r2 , |z| < z1 }.
Σ 0

See Figure 2.16 c). The map ψ is defined by ψ = ψout ◦ ψint for all points
X ∈ Σ0 with ψ(X) ∈ Σ0 and has the same asymptotic behavior as ψint for
z → 0 since for z → 0 the time needed by the solution to come from Σ0
to Σ1 becomes infinite, whereas the time needed by the solution to come
from Σ1 to Σ0 stays finite. Hence, a rectangular set R around the entrance
point p of the homoclinic orbit is mapped into a spiral like structure. See
Figure 2.17. The assumption | Re ω| < λ is necessary that this picture really
occurs, for more details see [GH83, §6.5.1]. Therefore, graphically we have
found a Smale’s horseshoe for ψ. 
64 2. Basic ODE dynamics

ψ(W)
z
W

p θ

Figure 2.17. Smale’s horseshoe in the Silnikov example.

2.6. Examples
The following series of examples is intended to give some familiarity with
the notions and ideas introduced so far.
Example 2.6.1. a) For the potential V (x, y) = (x2 − 1)2 + y 2 we find
−∇V (x, y) = −(4x(x2 − 1), 2y), leading to the fixed point (x, y) = (0, 0),
which is a saddle point of V , and to the fixed points (±1, 0), which are
minima of V . For every r > 1 the set {(x, y) : x2 + y 2 ≤ r2 } is absorbing.
The attractor A is given by [−1, 1] × {0}, consisting of the three fixed points
and the heteroclinic connections between the unstable fixed point (0, 0) and
the stable fixed points (±1, 0).
b) For the potential V (x, y) = (x2 − 1)2 + (y 2 − 1)2 we find that any
neighborhood of the unit square Q = [−1, 1] × [−1, 1] is an absorbing set.
The global attractor is Q = W u ((0, 0)).

(a) (b)
1

−1

−1 0 1

Figure 2.18. Phase portraits for Example 2.6.1 a) and b).

Example 2.6.2. We consider

ẋ = y and ẏ = −cy + x − x3 .
2.6. Examples 65

There is a simple mechanical interpretation of the orbits of this system as the

orbits of a particle moving in the double-well potential F (x) = − 12 x2 + 14 x4
with friction −cy. In partiular, for c = 0 we have a similar situation as in
Example 2.3.20, i.e., orbits are level lines of the energy E(x, ẋ) = 12 ẋ2 +F (x),
see also Example 4.1.2 below for a general discussion.
The system possesses the fixed points (x, y) = (±1, 0) and (x, y) = (0, 0).
The linearization
 at the fixed point (x, y) = (0, 0) yields the eigenvalues
λ1,2 = −c/2 ± c2 /4 + 1. Hence, (0, 0) is a saddle for all values of c. The
unstable eigenspace is spanned by φ1 = (1, 1), the stable one by φ2 = (1, −1).
The linearization
 at the fixed point (x, y) = (1, 0) yields the eigenvalues
λ1,2 = −c/2 ± c2 /4 − 1. Thus, (1, 0) is a center for c = 0 (Figure 2.19(a)),
a stable vortex for 0 < c < 2 (Figure 2.19(b)), a stable node for c ≥ 2
(Figure 2.19(c)), an unstable vortex for −2 < c < 0, and an unstable node
for c ≤ −2. The same classification holds for the fixed point (−1, 0). The
mechanical interpretation is that for c > 2 the friction is so large that the
particle approaches the minima x = ±1 of the energy monotonically.
For c > 0 the system is dissipative, and the stable manifold Ws ((0, 0))
separates the stable manifolds Ws ((1, 0)) and Ws ((−1, 0)), i.e., the domains
of attraction of (1, 0) and (−1, 0). The global attractor consists of the three
fixed points and the unstable manifold of (0, 0).

(a) c = 0 (b) c = 0.5 (c) c = 3

1 1 1

0 0 0

−1 −1 −1

−1 0 1 −1 0 1 −1 0 1

Figure 2.19. Phase portraits for Example 2.6.2; invariant mani-

folds in light grey, and nullclines as dashed lines.

For c = 0 we have homoclinic solutions and so Wu ((0, 0)) = Ws ((0, 0)),

and the center manifolds of (±1, 0) can be defined as small disks around
(±1, 0). To show the existence of the homoclinic solutions for c = 0, instead
of the energy argument that E(x, ẋ) =const= 0 we may also use the symme-
try (reversibility) (t, x, y) → (−t, x, −y) under which the system is invariant.
Hence, with t → (x(t), y(t)) also t → (x(−t), −y(−t)) is a solution. The un-
stable manifold of the origin intersects the fixed space Σ = {(x, 0) : x ∈ R}
of reversibility transversally. W.l.o.g. taking this intersection at t = 0, the
orbit can be extended to t > 0 by reflection at Σ. See Figure 2.20.
66 2. Basic ODE dynamics

Figure 2.20. Persistence of homoclinic connections in planar systems.

In non-reversible systems an additional parameter is needed for the in-
tersection of the stable and unstable manifold. In reversible systems the
fact that the stable manifold intersects the fixed space {y = 0} of the
reversibility operator transversally can be used for the persistence.

This homoclinic orbit persists under small perturbations respecting the

reversibility of the vector field due to the transversal intersection. Such
transversibility and symmetry arguments often also work when other argu-
ments, such as the above energy argument, fail, see, e.g., Remark 13.3.1.
For general perturbations (c > 0) the homoclinic orbit breaks up in accor-
dance with the fact that the probability that two one-dimensional manifolds
intersect in a two-dimensional phase space is zero.

The following examples are applications from mathematical biology; ex-

amples of this type, combined with diffusion lead to the important class of
reaction diffusion systems, see Chapter 9.
Example 2.6.3. (Lotka-Volterra)We consider the predator-prey system
(2.34) ẋ = x(a − y), ẏ = y(x − c), a, c > 0
with x, y > 0 (x = prey, y = predator). In case of no predators, i.e., y = 0,
the prey will grow with some exponential rate according to ẋ = ax. In case
of no prey, i.e., x = 0, the predators will die with some exponential rate
according to ẏ = −cy. In case of predators, i.e., y > 0, the prey will be
killed by the predator via the term −xy with a rate proportional to the
number of predators. On the other hand the term xy gives an exponential
growth of the number of predators with a rate proportional to the number
of preys.
The unique non-trivial fixed point is (x0 , y0 ) = (c, a). Its linearization
   
a − y −x  0 −c
A= = ,
y x − c (x,y)=(c,a) γ 0
√
possesses the eigenvalues λ = ±i ca such that (x0 , y0 ) is a centre for the
linearization. Thus, no stability result can be concluded from Theorem 2.3.4
a) for the nonlinear system.
2.6. Examples 67

However, φ(x, y) = x + y − c ln(x) − a ln(y) is conserved for (2.34), i.e.,

d
dt φ(x, y)= 0. Since (ẋ, ẏ) = (0, 0) for (x, y) = (x0 , y0 ) the solutions move
on the level lines of φ. Calculus yields that φ has a unique critical point,
namely a minimum in (x0 , y0 ). Thus, (x0 , y0 ) is nonlinearly stable and all
other solutions move on periodic orbits around (x0 , y0 ). This behavior agrees
with observed data, for instance of canadian lynx and snowshoe hare pelt-
trading records of the Hudson Bay Company between 1845 and 1935, cf.
[Mur89].
The Lotka–Volterra model (2.34) as the oldest predator-prey model was
partially motivated by the observation that during and shortly after world
war I the fraction of predator fish caught in the mediterranean sea increased,
when the total fishing decreased. Let (x, y)(t) be a periodic solution of (2.34)
with period T . Then
 
1 T 1 T
x= x(t) dt = c, y= y(t) dt = a.
T 0 T 0
T T
This holds due to T1 0 ẋx dt = T1 0 a − y dt = T1 (ln x(T ) − ln x(0)) = 0,
hence y = a, and similarly for x = c. Fishing can be modeled by a simple
modification of (2.34), namely to replace a, c by a − p and c + p, where
0 ≤ p ≤ a denotes the “fishing pressure”. Depending on p, the fraction of
predators hence is
y(p) a−p
= ,
x(p) + y(p) a+c
which is a decreasing function of p.
See also Exercise 2.19 for some modification of (2.34), for which the
conserved quantity φ becomes a genuine Lyapunov function.
Example 2.6.4. (SI and SIS diseases) We are interested in the dynamics
of some disease which proceeds on a time scale much shorter than the lifespan
of its hosts. Thus, we assume that the size of the population is unchanged
and that in the population of size N a fraction S of individuals is healthy,
but susceptible to this disease, while a fraction I is infected. A general ODE
model describing the evolution of the fractions S and I reads
Ṡ = −f (S, I), I˙ = f (S, I),
where f (S, I) is the rate of infections. The simplest model is f (S, I) = βIS,
which in chemistry would be called the law of mass action, see also §9.1.
Here β is called the pairwise infectious contact rate. Using S + I = N yields
I˙ = βI(N − I).
It is easy to see that in this model with β > 0 the whole population becomes
infected.
68 2. Basic ODE dynamics

A slight extension is given by SIS models, where infected recover and

become susceptible again with a rate γ. Thus
Ṡ = −f (S, I) + γI, I˙ = f (S, I) − γI.
Introducing dimensionless variables u = S/N, v = I/N, τ = γt, and using
u + v = 1 we obtain
v = (R0 (1 − v) − 1)v,
where R0 = βN/γ is called the reproductive ratio of the disease and where
denotes the derivative w.r.t. the new time variable τ . For R0 < 1 the disease
dies out, but for R0 > 1 it becomes endemic, i.e., it reaches the steady state
1 − 1/R0 as t → ∞. Vaccination reduces the number of susceptible and
hence R0 . Note that for desease control it is not necessary to vaccinate all,
but sufficiently many to decrease R0 below 1.
Example 2.6.5. (Mathematical ecology) The Kolmogorov form of the
equations for 2-species interaction in mathematical ecology is
(2.35) u̇ = uM (u, v), v̇ = vN (u, v),
where (u, v) = (u, v)(x, t) are population densities and their respective
growth rates M and N are smooth functions from R2+ to R. The mod-
els (2.35) are further classified as
predator-prey (PP) ∂v M < 0 and ∂u N > 0 for u, v > 0,
competition (C) ∂v M < 0 and ∂u N < 0 for u, v > 0,
symbiosis (S) ∂v M > 0 and ∂u N > 0 for u, v > 0.
Usually, further conditions are imposed, namely
(PP1) ∃k0 > 0 such that M (u, 0) < 0 for u > k0 ,
(PP2) ∃ a function l such that N (u, v) < 0 for u > 0 and v > l(u).

(C1) ∃k0 > 0 such that M (u, 0) < 0 for u > k0 ,

(C2) ∃l0 > 0 such that N (u, v) < 0 for v > l0 .

(S1) ∃ a function k such that M (u, 0) < 0 for u > k(v),

(S2) ∃ a function l such that N (u, v) < 0 for v > l(u),
(S3) k(v) = o(v)v→∞ and l(u) = o(u)u→∞ .
Biologically, for instance (PP1) and (C1) essentially mean that even if there
are no predators (v = 0), then the growth of the prey population saturates
at k0 . By (PP2) and (C2), the predators saturate at l(u). Finally, symbio-
sis means that each species thrives with the other, but (S3) ensures limits
to this symbiotic growth. The condition (S3) should be complemented by
demanding that there exists at least one non-trivial fixed point.
2.6. Examples 69

As examples consider
(1) u̇ = u(1 − u − v), v̇ = v(u − v),
(2) u̇ = u(1 − u − v), v̇ = v(1 − u − αv),
(3) u̇ = u(2 arctan(2v) − u), v̇ = v(3 arctan(2u) − v).
In (2), α > 0 is some parameter. Clearly, (1)=(PP), (2)=(C), (3)=(S). After
determining the unique fixed points (u, v)∗ with uv > 0 for (1) and (3), the
phase portraits can be conveniently sketched by considering the signs of the
growth rates M, N . For (1) we may additionally use the fact that, e.g.,
[1/4, 3/4] × [1/4, 3/4] is positively invariant.
For (2) we note that for α = 1 we have M = N , and thus a line {u + v =
1} of fixed points. For α = 1 we again have a unique non-trivial fixed point.
In particular, for α > 1 (α < 1) the v species (the u species) dies out. For
α > 1 the biological interpretation is that for u = v the growth rate N of v
is smaller than that of u, due to higher damping (faster saturation) of the
growth of v by itself, hence u “wins”.

M<0 f1=0 f1=0

N<0
1 1
f =0 f =0
2 2
0.8 0.8
M<0
0.6 M>0 N>0 0.6
N<0
0.4 0.4 M=N<0

0.2 M>0 0.2

N>0 M=N>0
0 0
0 0.5 1 0 0.5 1
f =0
1
6
f =0 M<0 f =0
M<0 1 2
1 N<0 f =0 5 M>0 N<0
2 N>0
0.8 4
M>0
0.6 N<0 3

0.4 2 M>0
N>0
0.2 M>0 1 M<0
N>0 N>0
0 0
0 0.5 1 0 1 2 3 4 5

Figure 2.21. ODE phase portraits for (PP), (C) (α = 1 and α = 2)

and (S). For α = 1 in (C) we have a line {u+v=1} of fixed points. For
equations of the form (2.35) a convenient way to start the phase portrait
is to consider the regions defined by the nullclines, i.e., M = 0 and
N = 0.
70 2. Basic ODE dynamics

Example 2.6.6. The van der Pol equation [VdP26] is given by

(2.36) ü + ε(u2 − α)u̇ + u = 0, u(t) ∈ R,
where α > 0 and 0 ≤ ε  1 are some parameters. As initial conditions
we take u(0) = a and u̇(0) = 0. This describes some oscillator with small
amplitude-dependent damping. It is known and might be expected from the
form of the equation, that for every fixed α > 0 and small ε > 0 there is a
unique periodic solution. For ε = 0 we have solutions u(t) = Aeit + c.c. with
A ∈ C arbitrary, and thus for ε > 0 we try a two-scale ansatz of the form
(2.37) u(t) = A(εt)eiωt + c.c..
2
Using u2 = A2 e2iωt + 2|A|2 + A e−2iωt this yields
O(ε0 ) : − ω 2 + 1 = 0, ⇒ ω = 1,
d
O(ε1 ) : 0 = i(−2 A + αA − A|A|2 )eit − iA3 e3it + c.c.,
dτ
and thus equating the coefficient of ε1 eit to zero yields
d 1
(2.38) A = (αA − A|A|2 ),
dτ 2
which is called the Landau equation for this problem. Introducing polar
coordinates A(τ ) = ρ(τ )eiφ(τ ) gives ρ = 12 ρ(α − ρ2 ), φ = 0, with the
abbreviation = dτ
d
. From this, or directly from (2.38), we can see that |A|
√
converges to α, which predicts that u approaches the circle with radius
√
2 α up to O(ε) terms. Incidentically, although nonlinear, (2.38) can be
explicitly solved. Via r = ρ2 and r = r(α − r), and via v = 1/r and
v = −αv + 1, we find the solution r(t) = αr0 /(r0 + (α − r0 )e−τ ), and hence
 1/2
α
(2.39) ρ(τ ) = ρ0 ,
(α − ρ20 )e−ατ + ρ20
with ρ(0) = ρ0 = a/2, and φ(τ ) = φ0 = 0. Figure 2.22 compares some
numerical solutions to (2.36) with approximations via (2.37) and illustrates
the distortion of the limit cycles of (2.36) from the circles described by (2.37)
as ε becomes larger. Approximation results for this special problem can be
found in [Ver96]. For PDEs on unbounded domains such a perturbation
analysis is one of the most powerful tools. In Part IV of this book and such
PDEs we prove error estimates for such formal approximations.

Exercises
Exercises 2.11 and 2.12 should be done with some software for ODE phase
portraits, e.g., xppaut or pplane, and we also recommend to use such software for
illustration after the analysis for the other planar ODEs, e.g., in Exercises 2.10,
2.15, 2.19, and 2.20.
2.6. Examples 71

2 3

1.5
2
1
1
0.5

0 0

−0.5
−1
−1
eps=0.1
−2
−1.5 eps=0.5
eps=1
−2 −3
0 10 20 30 −3 −2 −1 0 1 2 3

Figure 2.22. Left: numerical solution of (2.36) and approximation via

√
(2.37), α = 1, ε = 0.2. Right: Distortion of circle ρ = 2 α by higher-
order terms.

2.1. Find the general solution of ẋ(t) = Ax(t) with

⎛ ⎞
    7 0 0
3 1 2 1
a) A = , b) A = , c) A = ⎝0 6 1⎠ .
0 4 0 2
0 0 4
2.2. Find the general solutions and the solutions of the initial value problems with
x(0) = ẋ(0) = 1 for: a) ẍ(t)+3ẋ(t)+2x(t) = 0, b) ẍ(t)+5ẋ(t)+4x(t) = cos(3t).
2.3. Solve the initial value problems
d d
a) y = xy, y(0) = 1, b) y = (cos x)y, y(0) = 1.
dx dx
2.4. Prove that eA+B = eA eB for d × d-matrices A, B, if AB = BA.
2.5. Compute real-valued logarithms of the matrices
⎛ ⎞
⎛ ⎞ ⎛ ⎞ a a 1 0
1 1 0   1 1 0
−2 0 ⎜−a a 0 1⎟
⎝0 1 1⎠ , , ⎝0 2 1⎠ , ⎜ ⎟,
0 −2 ⎝ 0 0 a a⎠
0 0 1 0 0 3
0 0 −a a
with a > 0. Are the solutions unique?
 
2.6. Consider
ÿ + 2d ẏ + k(t) 2
+ d2
y = 0, with k(t + T ) = k(t), T = r + π2 ,
0 for t ∈ [0, r),
and k(t) = with r > 0. Compute the evolution operator
1 for t ∈ [r, T ),
U (T, 0) = U (T, r)U (r, 0). Show that the Floquet-multipliers are given by
⎛  ⎞
2
r r
ρ1,2 = e−d(r+ 2 ) ⎝− ± − 1⎠ .
π

2 4

Find the domain of stability in which |ρ1,2 | < 1.

 
3 4
2.7. i) Solve xn+1 = Bxn with B = . ii) Illustrate selected orbits of of
4 3
xn+1 = Bxn for
      1   
1 0 2 0 −1 0 2 0 0 1
a) , b) , c) , d) , e) .
0 1 0 12 0 1 0 − 12 −1 0
72 2. Basic ODE dynamics

2.8. Consider the one-dimensional ODEs

0, if u = 0,
a) u̇ = u − u2 , b) u̇ = −u + 4u3 − u5 , c) u̇ =
−u3 sin(1/u), if u = 0.
Find the fixed points and compute their linearization. Which fixed points are stable
and which fixed points are unstable? Sketch the phase portraits.
2.9. Consider f : R3 → R3 , f (u) = Au + g(u) with
Au = (au1 , acu2 , cu3 ), g(u) = (0, acu1 u3 , 0),
−1
where a > 1 > c > a > 0. According to the discrete Hartman-Grobman theorem,
cf. Remark 2.3.9, there exists a homeomorphism h such that h−1 ◦ f ◦ h = A. Show
that h cannot be Lipschitz-continuous.
Hint: Clearly h−1 ◦ f n ◦ h = An . Show that this implies
c−n h2 (u1 , 0, cn u3 ) − an h2 (a−n u1 , 0, u3 ) = nh1 (u1 , 0, cn u3 )h3 (a−n u1 , 0, u3 ).
Next show that h2 (u1 , 0, 0) = 0 and h2 (0, 0, u3 ) = 0, if h is Lipschitz-continuous.
Then obtain a contradiction for n → ∞.
2.10. Consider ẋ = y, ẏ = −cy − x + x3 , with c ∈ R a parameter. Find the fixed
points and compute their linearization. Which fixed points are stable and which
fixed points are unstable? Sketch the phase portrait for different values of c. Find
the stable, the unstable and the center manifolds for the fixed points.
2.11. Find the possible ω-limit sets for ẋ = y, ẏ = x + εy − x3 + 0.1x2 y, with
ε ∈ [−0.09, −0.07]. Compute values of ε ∈ [−0.09, −0.07] where a qualitative
change of the periodic orbits occurs. Hint: Unstable objects can be found by
t → −t.
2.12. Consider
      2 
ẋ 1 1 x (x + αy 2 )x
= − .
ẏ −1 1 y (αx2 + y 2 )y
Plot the phase portrait for α = 1, 5, 10. Find the fixed points and the periodic
solutions. Which of them are stable? Find the maximal α∗ > 1, such that there
exists a non-trivial stable periodic solution for all α ∈ [1, α∗ ]. (Hint: Consider the
phase portrait for α ∈ (10, 12) by computing the ω-limit set for the initial condition
(x, y) = (0.1, 0.1).) Let α = 12. Find the fixed points and the associated stable
and unstable manifolds.
2.13. Consider u̇ = f (u) with f ∈ C 1 (R2 , R2 ). Let Ω ⊂ R2 be open and simply
connected. Assume the existence of a b ∈ C 1 (R2 , R) with div(bf ) > 0 in Ω. Use
the integral law of Gauss to show the non-existence of a periodic orbit in Ω.
2.14. Use the idea from Example 2.4.12 to prove that ẋ = x−y −x3 , ẏ = x+y −y 3 ;
has a periodic solution.
2.15. Discuss the stability of the fixed point (x, y) = (0, 0) and sketch the phase
portraits for the following systems; compare with ẋ = y, ẏ = −x and explain the
qualitative differences.
a) ẋ = y, ẏ = −x3 . b) ẋ = y 1999 , ẏ = −x1999 .
Hint for a) Consider V (x, y) = αx4 + y 2 with suitable α.
2.6. Examples 73

2.16. Consider ẍ + δ(x)ẋ + 25x = 0 with δ(x) = 8 for |x| > 1 and δ(x) = −6 for
|x| ≤ 1. In order to show the existence of a periodic orbit consider the Poincaré
map G1 : S1 → S2 and G2 : S2 → S3 , where S1 = {(x, ẋ) : x = −1, ẋ ≥ 0},
S2 = {(x, ẋ) : x = 1, ẋ ≥ 0} and S3 = {(x, ẋ) : x = 1, ẋ ≤ 0}. Use then the
symmetry of the problem.
2.17. Let Σ be the set of all 0 − 1 sequences (sj )j∈N with the following property.
If sj = 0, then sj+1 = 1, i.e., Σ consists of all sequences without two succeeding
zeroes. Prove that:
a) the shift σ maps Σ into itself;
b) there exists a dense orbit in Σ ;
c) the set of periodic orbits is dense in Σ .

2 is conjugated to the tent map f : [0, 1) → [0, 1)

2.18. Prove that the shift σ on Σ+
defined by

2x, for x ∈ [0, 1/2),

f (x) =
2x − 1, for x ∈ [1/2, 1).
∞
Hint: show that φ ◦ f = σ ◦ φ where φ(x) = (xj )j∈N for x = j=1 xj 2−j .
2.19. The dynamics of the prey predator system (2.34) is not robust under pertur-
bations. Consider (2.34) with the modification ẋ = x(a(x) − y) with a(x) = ae−βx
for a β > 0.
a) Give a biological interpretation of the modification.
b) Show the asymptotic stability of the fixed point (x, y) = (c, a(c)).
c) Use the Lyapunov function φ(x, y) = x + y − c ln x − a(c) ln y to prove that all
solutions starting with x(0) > 0 and y(0) > 0 converge towards this fixed point.
d) Sketch the phase portrait.
2.20. Consider the 2-species interaction systems
  
u̇ = u 1+v2
2 − u , u̇ = u v − u2 − 1 , u̇ = u − u2 + 1+v
u
,
(1) (2) (3)
v̇ = v (u − v) , v̇ = v (u − v + 3) , v̇ = 2v − v + 1+u
2 v
,
all on u, v > 0. For each system, compute the nontrivial fixed point and its lin-
earized stability, and sketch the phase portrait. Classify the systems according to
Example 2.6.5.
 Chapter 3

Dissipative dynamics

In this chapter we provide the strategy and the tools to tackle dissipative
systems, which are characterized by the existence of a compact absorbing
set. In such systems very often through so called bifurcations complicated
and eventually chaotic dynamics occur if some external parameter is varied.
In applications such an external parameter can be for instance an external
heating or the concentration of a chemical substance. A typical scenario is as
follows. For small values of this parameter all solutions are attracted to some
asymptotically stable fixed point. If the value of the parameter is increased
the fixed point becomes unstable. Then more complicated dynamics can
be expected in a neighborhood of the unstable fixed point, for instance
new fixed points or time-periodic solutions may bifurcate, i.e., appear in a
neighborhood of the first unstable fixed point. A further increase of the
external parameter leads to instabilities of the bifurcating solutions. Then
quasi-periodic solutions can occur. The next bifurcation may already lead
to chaotic dynamics.

Figure 3.1. Complicated dynamics occurs in dissipative systems

as a parameter is increased.

75
76 3. Dissipative dynamics

After introducing a number of elementary bifurcations for one- and two-

dimensional systems we introduce with the Lyapunov-Schmidt reduction and
the center manifold theorem two reduction methods which allow to find these
elementary bifurcations in higher dimensional systems, too. Center manifold
theory turns out to be a very powerful tool. Besides the construction of the
bifurcating solutions it also yields information on their stability. If a fixed
point changes from stable to unstable, then all nearby solutions are attracted
with some exponential rate towards the center manifold, i.e., the interesting
non-trivial dynamics happens on the center manifold of the fixed point. So
called normal form transformations allow to analyze the dynamics on the
center manifold. We will present this method in the context of the proof
of the Hopf bifurcation theorem, i.e., we use it to prove the bifurcation of
time-periodic solutions. The chapter is closed by sketching some routes of
bifurcations to chaotic behavior in dissipative systems.

3.1. Bifurcations
We present a number of elementary bifurcations and explain how the implicit
function theorem and the Lyapunov-Schmidt reduction can be used to prove
their occurrence in more complicated systems.

3.1.1. Examples of elementary bifurcations. We start with a globally

attracting fixed point which becomes unstable when an external parameter
is changed. The following examples are the simplest ones which however
turn out to be the ’generic’ (see Remark 3.1.10) bifurcations occurring at a
fixed point.

Example 3.1.1. (Pitchfork bifurcation of fixed points) Consider

ẋ = f (x, μ) = μx − x3 ,

with x = x(t) ∈ R and μ ∈ R. The linear stability analysis of x = x∗1 = 0

gives: x∗1 = 0 is asymptotically stable for μ < 0 and unstable for μ > 0.
At μ = 0 a real eigenvalue crosses the imaginary axis and two further fixed
√
points x∗2,3 = ± μ bifurcate from x∗1 = 0. There is an exchange of stability:
√
for μ < 0, x∗1 = 0 is stable; for μ > 0, x∗1 = 0 is unstable and x∗2,3 = ± μ
are stable, since for μ > 0 the linearization A = (μ − 3x2 )|x=x∗2,3 = −2μ has
the negative eigenvalue −2μ. Since the fixed points only exist for μ > 0,
where x∗1 = 0 is unstable, this bifurcation is called a supercritical pitchfork
bifurcation. In case f (x, μ) = μx + x3 = 0 we can explicitly compute the
√
bifurcating unstable branches x∗2,3 = ± −μ for μ < 0. These exist where the
primary solution x∗1 = 0 is stable, and the bifurcation is called a subcritical
pitchfork bifurcation. See Figure 3.2.
3.1. Bifurcations 77

x
x

μ
μ

Figure 3.2. Super- and subcritical pitchfork bifurcation of fixed

points.

There are two other elementary bifurcations of fixed points, namely the
transcritical bifurcation and the saddle-node bifurcation.
Example 3.1.2. (Transcritical bifurcation of fixed points) Consider
ẋ = μx − x2 ,
with x = x(t) ∈ R and μ ∈ R. The trivial fixed point x = x∗1 = 0 is
asymptotically stable for μ < 0 and unstable for μ > 0. For μ = 0 a
real eigenvalue crosses the imaginary axis. There exists another fixed point
x∗2 = μ, which coincides with the trivial solution x∗1 = 0 for μ = 0. Since
in general we know a priori only the trivial solution, we say that the fixed
point x∗2 = μ bifurcates from the trivial solution x∗1 = 0. For the transcritical
bifurcation an exchange of stability takes place: for μ < 0, x∗1 = 0 is stable
and x∗2 = μ is unstable; for μ > 0, x∗1 = 0 is unstable and x∗2 = μ is stable.
See Figure 3.3.

x
Im Im

Re μ
Re
μ <μ 0 μ>μ0

Figure 3.3. A real eigenvalue crosses the imaginary axis leading

(here) to a transcritical bifurcation.

Example 3.1.3. (Saddle-node or flip bifurcation of fixed points)

Consider
ẋ = μ − x2 ,
√
with x = x(t) ∈ R and μ ∈ R. Two fixed points x∗1,2 = ± μ appear at
√
μ = 0. The linearization around x∗1,2 gives ∓2 μ. Thus, x∗1 is stable and x∗2
is unstable. See Figure 3.4. The origin of the name saddle-node bifurcation
can be seen in ẋ = μ − x2 , ẏ = −y: for this system, (x∗2 , 0) is a saddle and
(x∗1 , 0) is a stable node. See Figure 3.4.
78 3. Dissipative dynamics

Figure 3.4. Saddle-node-bifurcation.

There is another elementary bifurcation which may occur when a glob-

ally attracting fixed point becomes unstable, namely the bifurcation of pe-
riodic solutions from a fixed point.
Example 3.1.4. (Hopf bifurcation) Consider
(3.1) ẋ1 = μx1 + x2 − x1 (x21 + x22 ) ẋ2 = −x1 + μx2 − x2 (x21 + x22 ),
and
 
μ 1
with xj (t) ∈ R and μ ∈ R. The linearization A = around x = 0
−1 μ
possesses the eigenvalues λ1,2 = μ±i, i.e., two complex conjugate eigenvalues
cross the imaginary axis for μ = 0. See Figure 3.5. Introducing polar
coordinates x1 = r sin φ and x2 = r cos φ with r ≥ 0 and φ ∈ R/(2πZ) gives
ṙ = μr − r3 and φ̇ = 1.
Hence, a family of periodic solutions
√ √
{x = xper (t, μ, φ0 ) : x1 = μ sin(t + φ0 ), x2 = μ cos(t + φ0 )}
bifurcates from the trivial solution x = 0 at μ = 0. This is called a super-
critical Hopf bifurcation. For fixed μ > 0 the family attracts every solution
with an exponential rate O(exp(−2μt)), see Figure 3.5.

Im Im 1

Re −1
Re

−1 0 1
Figure 3.5. Two complex conjugate eigenvalues cross the imagi-
nary axis and the phase portrait for (3.1) for μ > 0.

In §3.3 we shall see that this bifurcation occurs generically when a fixed
point loses stability due to two complex conjugate eigenvalues crossing the
imaginary axis.
3.1. Bifurcations 79

Example 3.1.5. (Subcritical Hopf bifurcation with turning point)

Consider
      
ẋ ε −1 x x
= + (r − r )
2 4
,
ẏ 1 ε y y
with r2 = x2 + y 2 . The occurrence of x2 + y 2 on the right-hand side suggests
the use of polar coordinates which yields ṙ = εr + r3 − r5 , φ̇ = 1. Thus,
forε < 0, with small |ε|, there exists an unstable periodic solution r ≡ r0 =
O( |ε|). The second (stable) fixed point r1 = O(1) of ṙ = εr +r3 −r5 yields
a stable periodic solution, see Figure 3.6. The bifurcation is subcritical. The
small amplitude non-trivial branch exists in the parameter regime (ε < 0)
where the trivial solution is stable. Due to the turning point and the O(1)
amplitude stable periodic orbits in the subcritical regime, this is also called
a hard bifurcation since in systems described by such a model the solution
may suddenly “jump” to the O(1) amplitude stable periodic orbit, because
in applications noise will push the solution beyond the unstable periodic
orbit. In contrast, supercritical bifurcations (stable non-trivial solutions
only start to exist after the trivial solution becomes unstable) are soft, since
√
then the bifurcating stable periodic orbits have O( ε) amplitude.

|| r ||

Figure 3.6. Subcritial Hopf bifurcation with turning point

3.1.2. Bifurcations of fixed points. It is the purpose of this section

to prove the occurrence of transcritical and pitchfork bifurcations of fixed
points from a fixed point in case when the branch of the bifurcating solutions
cannot be computed explicitly. The following analysis is based on scaling
arguments and the implicit function theorem. The detection of fixed points
for the ODE ẋ = f (x, μ), where μ ∈ R, leads to the algebraic equation
(3.2) f (x, μ) = 0.
Throughout this section we restrict to analytic f : R2 → R. Assume that
a solution (x0 , μ0 ) of (3.2) is known, i.e., f (x0 , μ0 ) = 0, and assume that
∂x f (x0 , μ0 ) = 0. Then, by the implicit function theorem there exists a
unique smooth solution x = x(μ) of (3.2), i.e., f (x(μ), μ) = 0, in a neigh-
borhood of (x0 , μ0 ). This solution can be extended outside the neighborhood
80 3. Dissipative dynamics

of (x0 , μ0 ) with the same argument until the assumption ∂x f (x(μ), μ)) = 0
is no longer satisfied. In such a point (x0 , μ0 ) a new branch of solutions can
bifurcate from this family of solutions η → (x, μ)(η). This so called bifur-
cation point can be analyzed with the Newton polygon which is explained
subsequently. It turns out that generically only two situations for the bifur-
cations of fixed points can occur, namely the transcritical bifurcation from
Example 3.1.2 or the pitchfork bifurcation from Example 3.1.1.
Scaling arguments. One way to establish the existence of bifurcating
solutions in the general case are scaling arguments and the implicit function
theorem.
Example 3.1.6. Let f (x, μ) = μx + x2 + sin x. Then f (0, μ) = 0 for all
μ ∈ R. Hence, x = x∗1 = 0 is the trivial solution for all μ ∈ R. For all values
of μ ∈ R we have ∂μ f |x=0 = 0. Hence, it is sufficient to consider
∂x f |x=0 = (μ − 2x + cos x)|x=0 = μ + 1.
Thus, a bifurcation can only take place when μ + 1 = 0. Therefore, we
introduce the small bifurcation parameter α = μ + 1. In order to find
Example 3.1.2 in the present problem we rescale x = αy and introduce
F (y, α) = α−2 f (αy, 1 + α) = y + y 2 + O(α).
Thus, for α = 0 we have the simple equation F (y, 0) = y + y 2 having the
solutions y1∗ = 0 and y2∗ = −1. According to
∂y F |(y,α)=(yj∗ ,0) = (1 + 2y)|(y,α)=(yj∗,0) = 0

we can apply the implicit function theorem to solve F = 0 in a neighborhood

of (y, α) = (yj∗ , 0) for y and and obtain y1∗ = 0 + O(α) and y2∗ = −1 + O(α).
Hence, beside the trivial solution x∗1 = 0 we also found the bifurcating
solution x∗2 = α + O(α2 ).
Example 3.1.7. Consider f (x, μ) = μx + sin x. Again x = x∗1 = 0 becomes
unstable at μ = −1. Let α2 = μ + 1 and x = αy. The rescaled problem
1
F (y, α) = α−3 f (αy, 1 + α2 ) = y − y 3 + O(α2 ) = 0
6
can be explicitly solved for α = 0. Using √ the implicit function theorem
we obtain y1∗ = 0 + O(α2 ) and y2,3∗ = ± 6 + O(α2 ), hence x∗ = 0 and
√ 1
x∗2,3 = ± 6α + O(α3 ) for α > 0.

Also more general situations can be handled by scaling arguments.

Example 3.1.8. Consider
(3.3) f (x, ε) = x2 + xε + ε3 = 0.
3.1. Bifurcations 81

We have the trivial solution (0, 0) for which the assumptions of the implicit
function theorem are not satisfied, i.e., ∂x f (0, 0) = ∂ε f (0, 0) = 0. Again we
are interested in non-trivial solutions x = x(ε) near the origin.
We make the ansatz x(ε) = εv(ε) and obtain
F (v, ε) = ε−2 f (εv, ε) = v 2 + v + ε = 0.
For ε = 0 we find the non-trivial solution v1∗ = −1. We additionally have
∂v F (−1, 0) = −1 = 0 such that we can apply the implicit function theorem
and obtain a smooth solution v = v1∗ (ε) = −1 + O(ε). Hence, we find
a non-trivial solution x∗1 = −ε + O(ε2 ) for f = 0. However, the ansatz
x(ε) = ε2 v(ε) yields
F (v, ε) = ε−3 f (ε2 v, ε) = O(ε) + v + 1 = 0.
For ε = 0, we find the non-trivial solution v2∗ = −1 and ∂v F (−1, 0) = 1 = 0.
Hence, we can apply the implicit function theorem and obtain a smooth
solution v = v2∗ (ε) = −1 + O(ε). Therefore, we found a second curve of non-
trivial solutions x∗2 = −ε2 +O(ε3 ) for f = 0. The expansions correspond to
the solutions
ε ε2
x1,2 (ε) = − ± − ε3
2 4
of (3.3), which only can be computed explicitly since (3.3) is a second order
polynomial w.r.t. x.

The Newton-polygon. In the last example there exist at least two curves
of non-trivial solutions. Since we have a polynomial in the example we can
be sure that we found all solutions. For non-polynomial problems the scaling
argument can be made rigorous with the help of the Weierstrass preparation
theorem which allows to bring analytic f into a polynomial form w.r.t. one
of the variables, cf. [CH82, §2.8]. With this preparation it is then clear
that the solutions which we will find with the scaling arguments are the only
non-trivial solutions near (x, ε) = (0, 0).
A systematic approach to find the scalings is as follows. Assume that
f can be expanded
∞ in some convergent power series near the origin, i.e.,
mn ∈ R is nonzero,
f (x, ε) = a x m εn . Whenever the coefficient a
m,n mn
make a dot at (m, n) ∈ N0 × N0 . Then take the lower convex hull of all dots
in the N0 × N0 -plane. This hull is the so called Newton-polygon with finitely
many line segments with endpoints (mi , ni ) and (mi+1 , ni+1 ) and slopes −αi .
Associated with each of these lines there are pi solutions x∗i (ε) = εαi vi∗ (ε)
of f (x, ε) = 0, where pi = mi − mi−1 .
Example 3.1.9. The Newton polygon for
f (x, ε) = x3 + 3x2 ε + 2xε2 + ε5 = 0
82 3. Dissipative dynamics

3 5

2
3

2
1

0 0
0 1 2 0 1 2 3

Figure 3.7. Newton polygons for f (x, ε) = x2 +xε+ε3 and for f (x, ε) =
x3 + 3x2 ε + 2xε2 + ε5 .

yields α1 = 3, p1 = 1 and α2 = 1, p2 = 2. With the ansatz x(ε) = εα v(ε)

we obtain
ε3α v 3 + 3ε1+2α v 2 + 2ε2+α v + ε5 = 0.
The first three terms are of the same leading order for α = 1. The third and
the fourth term are of the same leading order for α = 3.
For α = 1 we obtain F (v, ε) = v 3 + 3v 2 + 2v + O(ε2 ) = 0. For ε = 0
we find the non-trivial solutions v1∗ = −1 and v2∗ = −2. Since ∂v F (−1, 0) =
(3v 2 +6v+2)|v=−1 = −1 = 0 and ∂v F (−2, 0) = 2 = 0 by the implicit function
theorem we find the non-trivial solutions v1∗ (ε) = −1 + O(ε2 ) and v2∗ (ε) =
−2 + O(ε2 ) or equivalently x∗1 (ε) = −ε + O(ε3 ) and x∗2 (ε) = −2ε + O(ε3 ).
For α = 3, we obtain F (v, ε) = O(ε2 )+2v +1 = 0. For ε = 0, we find the
non-trivial solution v3∗ = −1/2. Since ∂v F (−1/2, 0) = 2 = 0 by the implicit
function theorem we find the non-trivial solution v3∗ (ε)=−1/2 + O(ε2 ) or
equivalently x∗3 (ε)=−ε3 /2 + O(ε5 ).
Remark 3.1.10. (The genericity of transcritical and pitchfork bifur-
cations) If w.l.o.g. we assume that the trivial solution is given by x = x∗ =
0, then there exists a smooth function g : R2 → R with f (x, μ) = xg(x, μ).
If we further assume that the bifurcation point is given by (x, μ) = (0, 0),
then ∂x f (0, 0) = ∂μ f (0, 0) = 0. This gives the condition g(0, 0) = 0 such
that
g(x, μ) = g10 x + g01 μ + g20 x2 + g11 xμ + g02 μ2 + O(|x|3 + |μ|3 ),
with coefficients gij ∈ R. Generically we have g10 = 0 and g01 = 0 such that
−1
we find a bifurcating branch with x = −g10 g01 μ + O(μ2 ), i.e., a transcritical
bifurcation. However, symmetries such as f (x, μ) = −f (−x, μ) can force
g10 = 0. Solving the equation g(x, μ) = 0 w.r.t. μ and using the Newton
polygon we find by the ansatz μ(x) = x2 s(x) that g01 μ and g20 x2 are of the
same order. We obtain
G(x, s) = x−2 g(x, x2 s) = g01 s + g20 + O(x) = 0
3.1. Bifurcations 83

−1 −1
and hence s = −g01 g20 such that μ = −g01 g20 x2 +O(x3 ), i.e., depending on
−1
the sign of g01 g20 a sub- or a supercritical pitchfork bifurcation occurs. More
coefficients can vanish, but this is a degenerated situation, which requires
more symmetries.
By a small perturbation the vanishing coefficients can be made non-
zero. This is called unfolding of the bifurcation, cf. [GS85, Chapter III].
Such unfoldings are robust w.r.t. other small perturbations, i.e., additional
parameters different from the unfolding parameters will not change the so-
lution set qualitatively.

3.1.3. The Lyapunov-Schmidt reduction. We consider now an ODE

ẋ = f (x, μ)
with x(t) ∈ Rd in case d > 1 under the assumption that at μ = μ0 one
simple eigenvalue crosses the imaginary axis and that all other eigenvalues
have negative real part. See Figure 3.8. We remark that for the subsequent
analysis it is sufficient that all eigenvalues except of one are bounded away
from the imaginary axis, i.e., eigenvalues with positive real part are allowed,
too.

Figure 3.8. Spectral situation in case of a fixed point becoming unsta-

ble via a pitchfork or transcritical bifurcation.

In order to establish a pitchfork or a transcritical bifurcation we use

the so called Lyapunov-Schmidt method, which allows to reduce the d-
dimensional problem f (x, μ) = 0 to a one-dimensional one. So far we have
restricted ourselves to problems f : R × R → R. In principle, the assertions
from §3.1.2 remain valid also in the general case f : B × R → B with B
a Banach space. This means that under the previous spectral assumption
generically only transcritical and pitchfork bifurcations occur. In contrast
to the examples above, x and μ are no longer equivalent. We distinguish
between the variable x and the parameter μ.
Only to avoid a number of technicalities we restrict ourselves in the
following to B = Rd . By the implicit function theorem we can compute a
solution x = x(μ) for growing μ until M = ∂x f (x(μ), μ) ∈ Rd×d is no longer
invertible. We denote this point by (x∗0 , μ0 ). Under the previous spectral
84 3. Dissipative dynamics

assumption non-invertibility is equivalent to the fact that exactly one of

the d eigenvalues of M is zero. W.l.o.g. we assume that the associated
eigenvector is given by e1 . In order to apply the implicit function theorem
in this situation, we split the system into two parts, namely into a part
where the implicit function theorem can be applied and into a part where
it cannot be applied. Denote by P1 a projection on span{e1 } = (1, 0, . . . , 0)
and let P2 = I − P1 . Moreover, we set xj = Pj x, denote by Q2 a projection
on the range of M , and let Q1 = I − Q2 . Then, we consider
Q1 f ((x1 , x2 ), μ) = 0, and Q2 f ((x1 , x2 ), μ) = 0.
We find that ∂x2 Q2 f (x0 , μ0 ) ∈ Rd−1×d−1 is invertible since it possesses the
(d − 1) non-zero eigenvalues of M . Hence, the second equation can be
solved locally w.r.t. x2 , i.e., there exists a solution x2 = x2 (x1 , μ). Inserting
this solution into the first equation gives the so called reduced bifurcation
problem
f(x1 , μ) = Q1 f ((x1 , x2 (x1 , μ)), μ) = 0.
After this so called Lyapunov-Schmidt reduction we are in the same situation
as in the previous section. We have to find the zeroes of a smooth function
f : R2 → R. If symmetries are present in the problem, then the projections
can be chosen in such a way that the symmetries are preserved by the
reduction [GS85, Chapter VII.3].
Example 3.1.11. Consider
f1 (x, y, ε) =εx − yx − x3 = 0,
f2 (x, y, ε) =y + 2x2 + y 2 ε2 = 0.
The origin (x, y) = (0, 0) is a solution for all ε ∈ R, and we are interested in
non-trivial solutions close to it. The linearization
 
ε 0
∂(x,y) (f1 , f2 )|(x,y)=(0,0) =
0 1
has the eigenvalues ε and 1. Hence, a bifurcation is only possible for ε = 0.
The kernel is given by {(x, 0) : x ∈ R} and the range by {(0, y) : y ∈ R}.
Thus, the above system is already in the form needed for the Lyapunov-
Schmidt reduction, and the second equation can be solved w.r.t. y. In order
to obtain an approximate solution we consider an iteration of the second
equation, namely
 2  
y = −2x2 − y 2 ε2 = −2x2 − −2x2 − y 2 ε2 ε2 = −2x2 + O ε2 x4 .
Inserting the solution y = −2x2 + O(ε2 x4 ) into the first equation gives the
bifurcation equation
    
f1 (x, y(x, ε), ε) = εx − −2x2 + O ε2 x4 x − x3 = εx + x3 + O ε2 x5 = 0.
3.2. Center manifold theory 85

 
Dividing this equation by x gives ε+x2 +O ε2 x4 = 0 which can be analyzed
by the Newton polygon. We find √ a subcritical pitchfork bifurcation, i.e.,
non-trivial solutions x∗ (ε) = ± −ε + O(ε) and y ∗ (ε) = O(ε) for ε < 0.

Consequence. When a fixed point becomes unstable by a simple eigen-

value crossing the imaginary axis, generically a transcritical or a pitchfork
bifurcation occurs. Hence, even in higher-dimensional phase spaces for this
spectral situation no new bifurcations can occur.
Remark 3.1.12. The Lyapunov-Schmidt reduction has certain disadvan-
tages. It does not provide information about the stability of bifurcating solu-
tions. Treating Hopf bifurcations via Lyapunov-Schmidt reduction leads to
an infinite-dimensional problem. In order to find 2π/ω-time-periodic solu-
tions of the ODE
 ẋ(t) iωjt
= f (x, μ) the problem is transferred by using Fourier
series x(t) = j∈Z x j e to an infinite-dimensional stationary problem for
the Fourier coefficients x j . With the help of the Lyapunov-Schmidt reduc-
tion the problem can be reduced to a two-dimensional one, cf. [CH82,
§1.4]. The construction of homoclinic and heteroclinic solutions with this
method again leads to an infinite-dimensional problem and requires special
properties of the underlying ODE, cf. [PS07].

3.2. Center manifold theory

Center manifold theory is an alternative way to find the elementary bifur-
cations from above also in higher space dimensions. Additionally, it often
yields information on the stability of the bifurcating solutions. Moreover, in
contrast to the Lyapunov-Schmidt reduction with this method Hopf bifur-
cations and the occurrence of small amplitude homoclinic and heteroclinic
solutions can be handled as finite-dimensional problems.
If a fixed point becomes unstable, all solutions are attracted with some
exponential rate towards the center manifold, i.e., the interesting non-trivial
dynamics happens on the center manifold of the unstable fixed point. In
general, only polynomial approximations of the vector field on the center
manifold are known. If the center manifold has two and more space dimen-
sions, so called normal form transformations help us to analyze the dynamics
on the center manifold. In the next section we use center manifold theory
and normal forms to prove a general Hopf bifurcation theorem.
We already formulated in Theorem 2.3.19 the invariant manifold theo-
rem. The part about the center manifold is now made more precise. We
consider
u̇c =Bc uc + gc (uc , us , uu ),
(3.4) u̇s =Bs us + gs (uc , us , uu ),
u̇u =Bu uu + gu (uc , us , uu ),
86 3. Dissipative dynamics

(a) (b)

Figure 3.9. When a fixed point becomes unstable the bifurcating so-
lutions can be found on the exponentially attracting center manifold.
(a),(b) before and after the bifurcation of a stable periodic orbit on Mc .

with uc ∈ Ec = Rdc , us ∈ Es = Rds , uu ∈ Eu = Rdu some finite-dimensional

vectors, Bc a matrix with eigenvalues on the imaginary axis, Bs a matrix
with eigenvalues with negative real part, Bu a matrix with eigenvalues with
positive real part, and gc : Rd → Ec , gs : Rd → Es and gu : Rd → Eu ,
d = dc + ds + du , are C r+1 -maps without constant and linear terms.
Theorem 3.2.1. (Center manifold theorem) There exists a neighbor-
hood U ⊂ Ec of uc = 0 and a C r -map h : U  uc → h(uc ) such that the
manifold
Wc = {u = uc ⊕ h(uc ) : uc ∈ U, (us , uu ) = h(uc )}
is invariant under the flow of (3.4). Wc is called the center manifold. The
reduced flow is determined by
(3.5) u̇c = Bc uc + gc (uc , hs (uc ), hu (uc )).
The function h contains no constant and no linear terms w.r.t. uc such that
the center manifold Wc is tangential to the central subspace Ec associated to
the eigenvalues with vanishing real part. In general the center manifold is
not unique.

Some parts of the theorem, namely the invariance, the existence, and
Lipschitz-continuity instead of r-times differentiability of the center manifold
will be proven in §13.1. Here we will discuss some of the assertions of the
theorem and concentrate on its application by giving a number of examples.
With the first example we explain how center manifold theory can be used
to handle bifurcation problems although the central eigenvalues are only on
the imaginary axis for one particular value of the bifurcation parameter
Example 3.2.2. For μ close to zero consider the trivial decoupled system
(3.6) ẋ = μx − x3 , ẏ = −y.
3.2. Center manifold theory 87

For μ < 0, the origin (x, y) = (0, 0) is stable. For μ = 0, we find the one-
dimensional center manifold Wc = {(x, y) ∈ R2 : y = 0}. In order to handle
non-zero values of μ with the center manifold theorem the above system is
extended to
ẋ = μx − x3 , ẏ = −y, μ̇ = 0.
For this extended system we find the two-dimensional center manifold Wc =
{(μ, x, y) ∈ R3 : y = 0}. Note that after introducing μ̇ = 0 the term μx is
no longer a linear, but a nonlinear term. Since μ̇ = 0 implies that μ is a
constant, the two-dimensional center manifold is foliated by one-dimensional
invariant manifolds. See Figure 3.10. Hence, the additional equation μ̇ = 0
can be canceled again and on the two-dimensional center manifold μ can be
considered again as a parameter. Therefore, by applying the center manifold
theorem in a sloppy way the two-dimensional bifurcation problem (3.6) can
be reduced to
ẋ = μx − x3
in the one-dimensional center manifold Wc = {(x, y) ∈ R2 : y = 0}. Obvi-
ously, the reduction is trivial in this case, i.e., h = 0. In summary, bifurcation
problems can be handled with the help of the center manifold theorem by
introducing the equation μ̇ = 0.

y
μ

x x

Figure 3.10. Reduction of the system ẋ = μx − x3 , ẏ = −y, μ̇ = 0

to a two-dimensional center manifold which is foliated by invariant one-
dimensional manifolds.

The next example shows how to compute approximations of the reduc-

tion function h and of the reduced system on the center manifold.
Example 3.2.3. For μ close to zero consider
ẋ = μx + x3 − xy, ẏ = −y + 2x2 .
Like above we extend the system by the equation μ̇ = 0. The linearized
system is given by
ẋ = 0, ẏ = −y, μ̇ = 0,
88 3. Dissipative dynamics

and hence Ec = {(μ, x, y) ∈ R3 : y = 0}. Therefore, we make the ansatz

y = h(x, μ) = ax2 + bμx + cμ2 + O(|μ|3 + |x|3 ),
and from ẏ = −y + 2x2 we obtain 2axẋ + μẋ + . . . = −(ax2 + bμx + cμ2 +
. . .) + 2x2 . Since ẋ = μx + . . ., by comparing the coefficients it follows
x2 : 0 = −a + 2, xμ : 0 = −b, μ2 : 0 = −c, ...
As a general rule, no powers μn without x can occur in h, and therefore
Wc = {(μ, x, y) ∈ R3 : y = 2x2 + O(|μ|x2 + |x|3 )}.
Moreover, the function h cannot contribute to the quadratic terms of the
reduced system which here is given by
ẋ = μx + x3 − x(2x2 ) + h.o.t = μx − x3 + h.o.t.,
i.e., the fixed point (x, y) = (0, 0) is stable also for μ = 0. We explain below
that stability on the center manifold implies stability in the full system in
such a situation. At μ = 0 a supercritical pitchfork bifurcation occurs.

The following two examples are about the non-uniqueness and the non-
smoothness of center manifolds.
Example 3.2.4. In order to illustrate the non-uniqueness of the center
manifold we consider
(3.7) ẋ = x2 , ẏ = −y.
Obviously, the central subspace Ec is given by the x-axis. The solutions of
the ODEs are given by x(t) = 1−tx x0
0
and y(t) = y0 e−t . Elimination of time
t yields y(x) = (y0 e−1/x0 )e1/x . For x < 0 we have limx→0,x<0 y (n) (x) = 0,
i.e., every solution approaches the origin in a flat way. For x > 0 we find
that y = 0 is the only solution which approaches the origin. Thus, we find
infinitely many different C ∞ -center manifolds which are tangential to Ec at
the origin by glueing together the obits in the left half plane with the positive
real axis. This shows that center manifolds are non-unique in general. The
only analytic center manifold, i.e., with a convergent power series, is the
x-axis. See Figure 3.11 for the phase portrait of (3.7).

Example 3.2.5. In order to illustrate the non-smoothness (and non-unique-

ness) of the center manifold we consider
ẋ = −μx, ẏ = −y, μ̇ = 0
with 0 < μ < 1. The vector field is C ∞ and Ec is given by the (x, μ)-plane.
Obviously Wc = Ec is a center manifold. However, the trajectories satisfy
d y 1/μ . If r < 1/μ < r + 1 with
dx y = μ x and are given through y(x) = C|x|
r ∈ N, then the trajectories for fixed μ are in C r , but not in C r+1 . Each of
3.2. Center manifold theory 89

0.5

−0.5

−1

−1 −0.5 0 0.5 1

Figure 3.11. The phase portrait of ẋ = x2 , ẏ = −y.

these curves is tangent to y = 0 and so the whole of these trajectories forms

another center manifold. In a ball of radius μ < 1/r this center manifold is
C r , i.e., the larger r is chosen, the smaller is the center manifold.

Remark 3.2.6. In case of no eigenvalues with positive real part there is a

neighborhood in which all solutions are attracted to the associated center
manifold Wc with some exponential rate O(e−βt ) for a β > 0. More precisely,
in [Van89, Theorem 5.17] it is shown that in this case there are strictly
positive constants C and β, such that for all x0 in a neighborhood of the
center manifold there is a t0 ∈ R and a xc ∈ Wc such that
x(t, x0 ) − x(t − t0 , xc ) ≤ Ce−βt .
As a consequence the stability of bifurcating solutions is solely determined
by the reduced ODE on the center manifold.

Similarly, center manifolds can be defined for discrete dynamical sys-

tems.
Example 3.2.7. We consider the discrete dynamical system
xn+1 = xn + xn yn yn+1 = λyn − x2n
with 0 < λ < 1. We find Ec = {(x, y) ∈ R2 : y = 0}. In order to compute
the center manifold we make the ansatz
y = h(x) = ax2 + bx3 + O(x4 )
and find
a(x + x(ax2 + . . .))2 + b(x + x(ax2 + . . .))3 + . . . = λ(ax2 + bx3 + . . .) − x2
which yields a = − 1−λ
1
and b = 0. Hence, we have
x2
y = h(x) = − + O(x4 )
1−λ
90 3. Dissipative dynamics

and find for the reduced equation

 
x3n x2n
xn+1 = xn − + O(xn ) = xn 1 −
5
+ O(xn ) .
4
1−λ 1−λ

Therefore, x = 0 is asymptotically stable in the reduced equation which

implies the asymptotic stability of the origin (x, y) = (0, 0) in the full system,
similarly to the previous Remark 3.2.6.

Example 3.2.8. (Saddle-Node bifurcation on center manifold) We

consider

(3.8) ẋ = ε + x2 + y 2 and ẏ = −y + x2

with small ε. For ε = 0 we have the fixed point (x, y) = (0, 0) with eigenval-
ues 0, −1 with neutral direction (1, 0). Thus we expand the center manifold
as y = h(x, ε) = ax2 + bxε + cε2 + . . ., which yields a = 1, b = −2, c = 2
and hence ẋ = ε + x2 + O(x4 ) as reduced
√ equation. We
√ have a saddle-node
bifurcation with two fixed points − −ε (stable) and −ε (saddle) for ε < 0
and no fixed point for ε > 0. Thus we do not actually need h. See Figure
3.12 for the phase portrait.

0.4
y=x2+0.2x+0.01

0.3

0.2

0.1

−0.1
−0.5 0 0.5

Figure 3.12. Saddle-node bifurcation on center manifold; ε= − 0.1. .

As already said, some parts of the center manifold theorem, cf. Theorem
3.2.1, namely the invariance, the existence, and the Lipschitz-continuity of
the center manifold will be proven in §13.1. There we will explain that center
manifold theory is not restricted at all to the finite-dimensional situation.
In Part IV it is used for the construction of bifurcating spatially periodic
solutions of pattern forming systems, but also in the construction of traveling
wave solutions in unbounded cylindrical domains.
3.3. The Hopf bifurcation 91

3.3. The Hopf bifurcation

In case that two complex conjugate eigenvalues cross the imaginary axis, the
analysis of the system can be reduced to the analysis of the associated two-
dimensional center manifold. For the ODE on the center manifold, however,
a large number of coefficients have to be computed, namely six coefficients
for the quadratic terms and eight coefficients for the cubic terms. Hence, at
a first glance a big zoo of possible dynamics can be expected. However, this
is not true. By normal form transformations the problem can be reduced in
polar coordinates to
(3.9) ṙ = ν1 r + ν2 r3 + . . . and φ̇ = 1 + . . . ,
with ν1 , ν2 ∈ R, i.e., to the computation of two efficient coefficients. Ignoring
the higher order terms this system has already been discussed in Example
3.1.4. Therefore, in case that two complex conjugate eigenvalues cross the
imaginary axis, under some non-degeneracy condition, always time-periodic
solutions occur, either as sub- or supercritical bifurcation.
Theorem 3.3.1. Consider ẋ = Aμ x + g(x) with x(t) ∈ Rd and g(x) =
O(x2 ) for x → 0. Assume that for μ = μ0 the matrix Aμ possesses two
eigenvalues λ± = ±iω with ω = 0 and that all other eigenvalues possess
strictly negative real part. Furthermore assume that d Re λ±
dμ |μ=μ0 = 0. If
ν2 = 0 in (3.9), or more precisely γr = 0 in (3.13) below, then a one
parametric family of periodic solutions bifurcates from x = 0 at μ = μ0 . The
period of the bifurcating solutions is 2π/ω + O(|μ − μ0 |) and their amplitude
is of order O(|μ − μ0 |1/2 ).

Proof. For the somewhat lengthy proof we introduce the new bifurcation
parameter ε = μ − μ0 and extend the ODE system with ε̇ = 0. Then we
apply the center manifold theorem and reduce the full system to a system
on the three-dimensional center manifold associated to the eigenvalues λ±
and the variable ε.
On the center manifold Mc for arbitrary coordinates (y, z) ∈ R2 the
reduced system can be written as
ẏ = a11 y+a12 z+a120 y 2 +a111 yz+a102 z 2 +a130 y 3 +a121 y 2 z
+a112 yz 2 +a003 z 3 +O(ε2 (|x|+|y|)+|y|4 +|z|4 ),
(3.10) ż = a21 y+a22 z+a220 y 2 +a211 yz+a202 z 2 +a230 y 3 +a221 y 2 z
+a212 yz 2 +a203 z 3 +O(ε2 (|x|+|y|)+|y|4 +|z|4 ),
ε̇ = 0,
where the values of the real-valued coefficients a· = a· (ε) depends  on our

a11 a12
choice of basis. The only restriction so far is that the matrix A =
a21 a22
92 3. Dissipative dynamics

possesses the eigenvalues

λ± (μ) = ±iω + O(ε).

At a first view, all kinds of dynamics seem to be possible. However, the

system can be simplified heavily with a so called normal form transform. In
order to do so we diagonalize (3.10) and obtain
1 εc + α1 εc
ċ1 = iωc1 + α110 1 101 −1
1 c2 + α 1 c c
+α020 1 2
1 011 1 −1 + α002 c−1 + . . . ,
(3.11) ċ−1 = −iωc−1 + α110 εc1 + α101 εc−1
2 2
2 c2 + α 2 c c
+α020 2 2
1 011 1 −1 + α002 c−1 + . . . ,
ε̇ = 0,
with c1 = c−1 and coefficients α· ∈ C. The idea of the normal transform is
to eliminate all terms which are not in resonance with the linear ones. As
an example consider c2−1 in the equation for c1 . It oscillates as e−2iωt and is
therefore not in resonance with c1 , which oscillates as eiωt . Therefore, this
c2−1 term be eliminated. With this heuristic argument the only terms which
remain in the equation for c1 are those of the form εm cn1 cn−1−1 , and in the
m n−1 n
equation for c−1 those of the form ε c1 c−1 . This heuristic argument can
be made rigorous by making a number of near identity changes of variables.

3.3.1. Normal form transforms. We consider the autonomous system

ẋ = Ax + f (x)

for x(t) ∈ Rd , with A ∈ Rd×d , and

f (x) = f2 (x) + f3 (x) + f4 (x) + . . . ,

with fm (kx) = k m fm (x) for all k ≥ 0, i.e., fm is a vector in Rd whose entries

are homogeneous polynomials of degree m in the variables x1 , . . . , xd . Hence,
⎛ ⎞
fm1
⎜ ⎟
fm = ⎝ ... ⎠
fmd
is an element of the vector space
⎧ ⎛ m1 md ⎞ ⎫
⎪
⎨ m 1 +...+m d =m αm
1
1 ...m d
x 1 · . . . · x d ⎪
⎬
⎜ . ⎟
⎠ : αm1 ...md ∈ R ,
j
Vm = u = ⎝ ..
⎪
⎩  m1 md
⎪
⎭
m1 +...+md =m αm1 ...md x1 · . . . · xd
d

the space of vector valued homogeneous polynomials of degree m in the

variables x1 , . . . , xd .
3.3. The Hopf bifurcation 93

We look for a near identity change of variables which allows us to elim-

inate as many terms as possible in order to make the system as simple as
possible. Therefore, we make the ansatz
x = y + h(y),
where
h(y) = h2 (y) + h3 (y) + h4 (y) + . . .
with hm ∈ Vm . We obtain
∂h
ẋ = ẏ + ẏ = A(y + h(y)) + f (y + h(y)),
∂y
and therefore
 
∂h −1
ẏ = 1 + [A(y + h(y)) + f (y + h(y))]
∂y
∂h2
=Ay − Ay + Ah2 (y) + f2 (y) + O(y3 ).
∂y
In order to eliminate all quadratic terms in f2 we have to find an h2 such
that
∂h2
− Ay + Ah2 (y) + f2 (y) = 0.
∂y
With the above interpretation of h2 as an element of the vector space V2 ,
∂h2
(LA h2 )(y) = − Ay + Ah2 (y)
∂y
m
is a linear map of V2 into itself which acts linearly on the coefficients αm .
1 ...md

Similarly, in order to eliminate terms of order m we have to solve the

linear system
∂hm
− Ay + Ahm (y) + fm (y) = 0,
∂y
where fm represents the nonlinear terms of degree m after the application
of the transformations h2 to hm−1 .
For our purposes it is sufficient to restrict ourselves to the case of a
diagonal A, i.e., A = diag(λ1 , . . . , λd ). Then in the space Vm the linear map
LA possesses the eigenvectors y1m1 · . . . · ydmd ej , where ej is the j th unit vector

of Rd . The associated eigenvalues are given by μ = dk=1 mk λk − λj . In
order to see this, we consider the j th component

d
∂hmj
λk yk − λj hmj = μhmj
∂yk
k=1
of the eigenvalue equation LA hm = μhm . Inserting the above eigenvectors
shows the statement. Therefore, LA hm = gm can be solved w.r.t. hm in all
eigenspaces with eigenvalue μ = 0. We found
94 3. Dissipative dynamics

term non-resonance condition can be eliminated

α20 c21 −2λ1 − 0λ2 + λ1 = −2i + i = 0 yes
α11 c1 c−1 −0λ1 − 2λ2 + λ1 = −2(−i) + i = 0 yes
α02 c2−1 −0λ1 − 2λ2 + λ1 = −2(−i) + i = 0 yes
α30 c31 −3λ1 − 0λ2 + λ1 = −3i + i = 0 yes
α21 c1 c−1 −2λ1 − 1λ2 + λ1 = −2i − (−i) + i = 0
2 no
α12 c1 c2−1 −1λ1 − 2λ2 + λ1 = −i − 2(−i) + i = 0 yes
α03 c3−1 −0λ1 − 3λ2 + λ1 = −3(−i) + i = 0 yes
Table 1. Non-resonance conditions for the terms in the first equation of (3.11).

Lemma 3.3.2. Assume A = diag(λ1 , . . . , λd ). Then in order to eliminate

the term y1m1 ·. . .·ydmd in the j th equation we need the non-resonance condition

d
(3.12) mk λk − λj = 0.
k=1

3.3.2. Continuation of the proof of Theorem 3.3.1. We now make a

normal form transform for (3.11). In Table 1 we list the various terms in
the first equation of (3.11) and their non-resonance conditions. Hence, after
the transformation we obtain for the new variables bj = cj + O(|ε|(|c1 | +
|c−1 |) + |c1 |2 + |c−1 |2 ) that
ḃ1 =iωb1 + β1 εb1 + γ1 b21 b−1 + O(|ε2 |(|b1 | + |b−1 |) + |b1 |4 + |b−1 |4 ),
ḃ−1 = − iωb−1 + β−1 εb−1 + γ−1 b1 b2−1 + O(|ε2 |(|b1 | + |b−1 |) + |b1 |4 + |b−1 |4 ),
with b1 = b−1 , β1 = β−1 = βr + iβi , and γ1 = γ−1 = γr + iγi , where βr , βi ,
γr , γi ∈ R. Introducing polar coordinates b1 = reiφ gives the system
ṙ = βr εr + γr r3 + O(ε2 r + r4 ),
(3.13)
φ̇ = ω + βi ε + γi r2 + O(ε2 + r3 ).
Hence, we have a system which can be analyzed for small ε. Ignoring the
√
higher order terms we have r = O( ε) for the bifurcating time-periodic
√
solutions. For the scaled variable r defined by r = ε r we find the approx-
imate time-periodic solution r02 = −βr /γr . Depending on the sign of γr we
have a sub- or supercritical bifurcation of time-periodic solutions. In order
to prove the persistence of these solutions under the neglected O(ε2 )-terms
we construct the associated Poincaré map Πε for which the periodic solution
is a fixed point. The fixed point is therefore a zero of
r(0), ε) = ε−1/2 (Πε (
F ( r(0)) − r(0)).
We have F ( r0 (0), 0) = −4πβr ω −1 = 0 such that the
r0 (0), 0) = 0 and ∂1 F (
implicit function theorem can be applied and a fixed point r(0) = rε (0) of
3.3. The Hopf bifurcation 95

Πε exists for ε > 0, too. Associated to this fixed point are periodic solutions
of (3.13), (3.11), and finally of (3.10). Therefore, we are done. 

3.3.3. An example and further remarks. In applications very often

the parameter space has more than one dimension, i.e., the parameters are
given by p ∈ Rm . In principle, this situation can be handled in the same
way as above by varying the parameters individually. Generically, the set
{p ∈ Rm : F (x(p), p) = 0, ∂x F (x(p), p) = 0} of possible bifurcation values
becomes a (m − 1)-dimensional manifold. Here we consider the following
example of a 2-parameter bifurcation/stability diagram.
Example 3.3.3. Bifurcation diagram for a (toy) problem from chem-
istry. The system
(3.14) α̇ = μ − α(κ + β 2 ), β̇ = −β + α(κ + β 2 )
serves as a (drastically reduced) model for so called cubic autocatalysis in
chemistry. Here, α, β are concentrations (hence α, β ≥ 0) of some substances
and the parameters μ, κ > 0 are some reaction rates. Unlike other reagents,
here represented by α, that participate in the chemical reaction, a catalyst,
here represented by β, is not consumed by the reaction itself. The above
model describes a situation where the so called educt α is supplied into the
system at constant rate μ and converted into β with rate κ + β 2 . Thus, the
catalyst β catalyses its own production, hence the name autocatalysis.
The unique fixed point of (3.14) is given by
(α∗ , β ∗ ) = (μ/(κ + μ2 ), μ),
with the associated linearization
 
−(μ2 + κ) −2μ2 /(κ + μ2 )
(3.15) A= .
μ2 + κ (μ2 − κ)/(μ2 + κ)
First, we discuss the eigenvalues of A which are given in terms of p = trace A
and q = det A by
p p2
(3.16) λ1,2 = ± − q,
2 4
This associated bifurcation/stability diagram is plotted in Figure 3.13. The
meaning of this diagram is as follows. Starting, e.g., with p, q in the sn
regime (p < 0 and 0 < q < p2 /4) and crossing, e.g., the line q = p2 /4,
the fixed point changes type from sn to sf, cf. (3.16). Next, depending on
the nonlinearity, we may expect a Hopf bifurcation when crossing the line
p = 0, q > 0, which is therefore called Hopf line. The point (p, q) = (0, 0) is
called Bogdanov-Takens or co-dimension-2 point since two parameters are
needed to describe the possible bifurcations in its neighborhood.
96 3. Dissipative dynamics

q
Hopf line

sn sf uf un

p
saddle points

Figure 3.13. Bifurcation diagram for ẋ = Ax in terms of the trace p

and the determinant q of A. The abbreviations stand for: sn: stable
node, two stable real eigenvalues; sf : stable focus, two stable conju-
gate complex eigenvalues; uf : unstable focus, two unstable conjugate
complex eigenvalues; un: unstable node, two unstable real eigenvalues;
saddle points: two real eigenvalues, one stable, the other unstable.

The transfer of Figure 3.13 to the fixed point (α0 , β0 ) and to the pa-
rameters μ, κ yields to the solution of a number of algebraic equations. For
instance, the Hopf line is given by solving the 4th order equation
trace A = −(μ4 − (1 − 2κ)μ2 + κ/(1 + κ))/(μ2 + κ) = 0,
hence
1 1/2
μ1,2 (κ) = √ (1 − 2κ) ± (1 − 8κ)1/2 .
2
In summary, we obtain the bifurcation diagram plotted in Figure 3.14, while
Figure 3.15 shows two selected phase portraits. We will come back to such
systems in Chapter 9.

Remark 3.3.4. Besides the analytical study of bifurcating branches close

to bifurcation, i.e., the analysis of the reduced equation, there is the big field
of numerical path following (or continuation) and bifurcation analysis. The
basic idea of continuation is as follows. Given a solution (x0 , μ0 ) ∈ R2 of
f (x, μ) = 0 with ∂x f (x0 , μ0 ) = 0 we choose a small δ > 0, let μ = μ0 + δ,
and use the Newton scheme to compute x(μ). In detail, we use the iteration
xn+1 = xn − (∂x f (xn , μ))−1 f (xn , μ)
with starting point x0 . The scheme converges for δ > 0 sufficiently small and
we set x(μ) = limn→∞ xn . Replacing (x0 , μ0 ) by (x(μ0 + δ), μ0 + δ) we can
start again and compute solutions x = x(μ) until ∂x f (x(μ), μ) = 0. In case of
∂μ f (x0 , μ0 ) = 0 we can interchange the role of x and μ and obtain a solution
3.3. The Hopf bifurcation 97

1/4

sn

sf

uf
un
1 2 μ
Hopf line

Figure 3.14. The bifurcation diagram for (3.14) is to be read as follows.

If, e.g., we fix κ = 0.1 then we cross the Hopf line at approximately
μ1 ≈ 0.41 and μ2 ≈ 0.77. At these lines Hopf bifurcations can be
expected. Path following methods may allow us to follow the family of
periodic solutions in the parameter plane.

f1=0 2.5
2 f =0
1
f2=0 f =0
2 2

1.5
1.5

1
1

0.5 0.5

0
0 0 0.5 1 1.5 2 2.5
0 0.5 1 1.5

Figure 3.15. Phase portraits for (μ, κ) = (0.5, 0.2) (left) and (μ, κ) =
(0.5, 0.1) (right)

μ = μ(x), i.e., we have a smooth curve (x, μ) = (x, μ)(s) parameterized with
s until simultaneously ∂x f (x(s), μ(s)) = ∂μ f (x(s), μ(s)) = 0.
There are variants of this idea [Kel77, Kuz04, Doe07, Sey10] which
automatically allow the continuation of branches around folds and beyond
bifurcation points, the detection and localization of bifurcation points, and
branch switching at bifurcation points. One standard method is so called
(pseudo-)arclength continuation, which is implemented in the package Auto,
[Doe07, Dea16], see also xppaut, [Erm02]. Many of these methods can
also be applied to bifurcation problems in PDEs and are important tools
there. A recent package specifically designed for elliptic systems in two
space dimensions is pde2path, [UWR14, DRUW14].
98 3. Dissipative dynamics

3.4. Routes to chaos

The chapter is closed by sketching two routes of bifurcations to chaotic be-
havior in dissipative systems, namely period-doubling, which is based on an
infinite series of local bifurcations, and homoclinic explosion, which is a so
called global bifurcation. There are many other routes to chaotic behav-
ior, but we will only comment on one of them, namely the Ruelle-Takens
scenario.
The theory of turbulence developed by Landau [LL91] in 1944 is based
on the assumption that more and more pairs of complex conjugate eigenval-
ues cross the imaginary axis. This route to chaos is called the Landau-Hopf
scenario. In 1971 Ruelle and Takens [RT71] showed that these infinitely
many unstable eigenvalues are not necessary for the occurrence of chaos.
The scenario starts with a stable fixed point and provides a very short route
to chaotic behavior only using local bifurcations. The first bifurcation is a
Hopf bifurcation leading to time-periodic solutions. Then the time-periodic
solution becomes unstable via a pair of complex-conjugate Floquet multi-
pliers crossing the unit circle leading to quasi-periodic solutions. The next
bifurcation leads to a three-dimensional invariant torus where nearby chaotic
behavior can be found. This route to chaos plays a certain role in hydro-
dynamical applications, but now we focus on the route with infinitely many
period-doublings.

3.4.1. Period-doubling. Period-doubling is realized in nature in a num-

ber of systems, as cardiac diseases, leaking water-taps, laser dynamics, and
various chemical reactions. It occurs if a periodic orbit becomes unstable
and a stable periodic solution with roughly the double period occurs, and
if this repeats under further increase (or decrease) of some parameter, see
Figure 3.16.

Figure 3.16. Sequence of period-doublings for a periodic solution

For the analysis of this phenomenon we consider the associated Poincaré

map Π. The fixed point of Π, which is associated to the periodic orbit,
becomes unstable via a real Floquet multiplier crossing the unit circle at
3.4. Routes to chaos 99

−1. The fixed point becomes also unstable for the second iterate of the
Poincaré map Π2 , but now via a real Floquet multiplier crossing the unit
circle at 1. On the center manifold associated to the Floquet multiplier −1
we have the following situation. If
Πx = −x + αx2 + O(x3 ), then Π2 x = x + βx3 + O(x4 ),
such that for Π2 a pitchfork bifurcation occurs. The two stable bifurcating
fixed points for Π2 corresponds to a 2-periodic solution for Π itself, since for
Π no bifurcation of fixed points occurs. Hence, a new periodic orbit with
twice the period is bifurcating from the old one. Assuming that this new
periodic orbit becomes unstable in the same way and that this procedure
goes on and on we finally come to chaotic dynamics. A famous ODE example
showing this behavior is by Rössler [Rös76].
Example 3.4.1. The Rössler system. Consider the ODE
⎛ ⎞ ⎛ ⎞
x −(y + z)
∂t ⎝y ⎠ = ⎝ x + by ⎠ ,
z b + z(x − a)
where typically b = 0.1 and a ∈ R serves as a bifurcation parameter. Starting
from an asymptotically stable periodic orbit for a=4 we find for increasing a
a period-doubling sequence, cf. Figure 3.17. See, e.g., [PJS92, §3.3] and the
references therein for a more detailed introduction to the Rössler system.

3.4.2. The logistic map. There is a discrete model problem for the period-
doubling route to chaos, namely the logistic map
xn+1 = μxn (1 − xn ) = F (xn )
with μ ≥ 0 and xn ∈ R. We have for the nth iteration F n (x) → −∞ for
n → ∞ if x < 0 or x > 1. For μ ∈ [0, 4] the map F maps the interval [0, 1]
into itself. In the following we restrict ourselves to values μ and x0 in these
sets. More details can be found in [Dev89], including a discussion of chaos
in the strict
√ sense of Definition 2.5.2 in the logistic map for parameter values
μ > 2 + 5.
The condition F (x) = μx(1 − x) = x gives the fixed points x∗1 = 0 and
x∗2 = 1 − 1/μ. At μ0 = 1 a transcritical bifurcation of fixed points occurs.
The linearization around the fixed point x∗ is given by yn+1 = F (x∗ )yn
where F (x∗ ) = μ(1 − 2x∗ ). For x∗ = 1 − 1/μ we obtain
F (x∗ ) = μ(1 − 2(1 − 1/μ)) = 2 − μ.
Hence, this fixed point is stable for μ ∈ (1, 3) and becomes unstable at
μ1 = 3 via some period-doubling. A stable two-periodic solution appears.
See Figure 3.18.
100 3. Dissipative dynamics

(a) a = 4 (b) a = 6

2.5
6
2
1.5 4
z

z
1
0.5 2

5
5 5 10
0 0 5
0 −5 0
−5 −5 −5
y x y x

(c) a = 8.5 (d) a = 9

15
12
10
8 10
z

6
4 5
2

10 10
0 10 0 10
0 0
−10 −10 −10 −10
y x y x

Figure 3.17. Period-doubling in the Rössler system, b = 0.1 fixed.

Figure 3.18. The map F 2 for μ = 2.9, μ = 3 and μ = 3.1. A pitchfork

bifurcation occurs which corresponds to the occurrence of a two-periodic
solution for F .

A further increase of μ yields an instability of the two-periodic solution

at a value μ = μ2 . We find a pitchfork bifurcation for F 4 and so the
occurrence of a four-periodic solution for F . A further increase of μ yields an
instability of the four-periodic solution at a value μ = μ3 . We find a pitchfork
bifurcation of F 8 and so the occurrence of an eight-periodic solution for F .
Interestingly, there is an infinite sequence of such bifurcations and so a
further increase of μ yields an instability of the 2n−1 -periodic solution at a
3.4. Routes to chaos 101

n
value μ = μn . We find a pitchfork bifurcation of F 2 and so the occurrence
of a 2n -periodic solution for F .
Even more interestingly, the period-doublings show some asymptotic
behavior. It can be proved rigorously by computer-assisted proofs that the
limit
μn − μn−1
lim ≈ 4.6692,
n→∞ μn+1 − μn
called the Feigenbaum constant, exists, cf. [CE80]. As a consequence we
have the existence of μ∞ = limn→∞ μn ≈ 3.57, too. A recent overview about
the theoretical background of the occurrence of these limits is [Avi11]. For
most values μ > μ∞ the system exhibits chaotic behavior. In Figure 3.19
the ω-limit set for starting point x0 = 1/2 is plotted as a function over the
bifurcation parameter μ. There are isolated regions on the μ-axis where
no attractive chaotic
√ behavior occurs, the so called windows of stability.
Beginning at 1+ 8 ≈ 3.83 there is for instance a range of parameters μ with
a stable 3-periodic solution. There is a general theory [Dev89, §1.10] that
for maps from R to R solutions of period 3 imply the existence of periodic
solutions of every period m ∈ N, known as the Theorem of Sarkovskii.

Figure 3.19. The ω-limit set for starting point x0 = 1/2 is plotted as a
function over the bifurcation parameter μ. For every fixed μ the iterates
xN , . . . xN +M with N and M sufficiently large are plotted.

For ODEs the instability occurs in a one-dimensional center manifold

where the dynamics can be described via a one-dimensional Poincaré map
and hence it can be expected that the route to chaos via period-doublings
may occur in ODE systems, too. Another example is the so called chemostat,
see Exercise 3.6.

3.4.3. Homoclinic explosion – the Lorenz attractor. The Lorenz at-

tractor [Lor63, Spa82] is a famous example of a three-dimensional ODE
with chaotic dynamics. It was found by the meteorologist E. Lorenz in 1963
102 3. Dissipative dynamics

as a lowest order approximation for convection in fluids and is considered as

a cartoon weather model. It is given by
ẋ = σ(y − x),
(3.17) ẏ = ρx − y − xz,
ż = −βz + xy,
with σ = 10 and β = 8/3 fixed. Numerical simulations of this simple model
for ρ = 27 show a complicated irregular dependence of the solutions on the
initial conditions and the occurrence of a so called ”strange” attractor. A
rigorous proof for the occurrence of chaotic dynamics in the Lorenz model
(similar to shift dynamics (Σ2 , σ) defined in §2.5.1) has been given [Tuc02].
We now explore the route to chaos in a bit more detail.
For any σ, β, ρ > 0 any large enough sphere around (0, 0, ρ + σ) is ab-
sorbing. This can be shown with the Lyapunov function
V (x, y, z) = x2 + y 2 + (z − ρ − σ)2 .
With α = min{2σ, 2, β} we obtain
d
V = − 2σx2 − 2y 2 − 2βz 2 + 2β(ρ + σ)z
dt
= − 2σx2 − 2y 2 − β(z − ρ − σ)2 − βz 2 + β(ρ + σ)2
≤ − αV + β(ρ + σ)2 .
Hence, for t large enough by Gronwall’s inequality we obtain
2β(ρ + σ)2
V (t) ≤ .
α
By Theorem 2.4.4 there exists the global attractor A = ω(B), for which
numerical simulations show its geometric complexity. The attractor has a
dimension less than three since the divergence of the vector field
∂y1 (σ(y2 − y1 )) + ∂y2 (ρy1 − y2 − y1 y3 ) + ∂y3 (−βy3 + y1 y2 ) = −(σ + 1 + β)
is negative and, therefore, every test volume shrinks to zero for t → ∞. For
ρ = 27 numerical experiments show a non-integer Hausdorff-dimension of A
of approximately 2.04.
The behavior occurs by a global bifurcation which is called homoclinic
explosion. The route to chaos for the Lorenz system is as follows when
ρ is increased from 0 to 27, for σ = 10 and β = 8/3 fixed. The z-axis
is an invariant set, and the origin is a stable fixed point for ρ < 1. The
linearization of (3.17) around 0 is
⎛ ⎞
−σ −σ 0
∇f (x, y, z) = ⎝ρ − z −1 −x⎠ .
y x −β
3.4. Routes to chaos 103

At (x, y, z) = (0, 0, 0) we find the eigenvalues

σ + 1 1
λ1,2 = − ± (σ + 1)2 + 4σ(ρ − 1), λ3 = −β,
2 2
such that at ρ = 1 a bifurcation of fixed points occurs which turns out to
be a supercritical pitchfork bifurcation. For ρ > 1 we have two non-trivial
fixed points X1∗ , X2∗ with

(3.18) z = ρ − 1 and x = y = ± β(ρ − 1).
∗ ) are the roots of
The eigenvalues of ∇f (X1,2
p(λ) = λ3 + (σ + β + 1)λ2 + β(σ + ρ)λ + 2σβ(ρ − 1),
∗ are stable until ρ = ρ
From this we find that the fixed points X1,2 Hopf ≈
24.74, where two complex conjugate eigenvalues cross the imaginary axis.
It turns out that a subcritical Hopf bifurcation occurs. This means that for
< ∗ exist, which shrink to X ∗
ρ ≈ ρHopf unstable periodic solutions near X1,2 1,2
as ρ → ρHopf .
Figure 3.20 gives some numerical illustrations: In general for ρ ∈ (1, 24]
the two parts S1 and S2 of the one-dimensional unstable manifold of the
origin are connected with the three-dimensional stable manifolds of the fixed
points X1∗ and X2∗ . Though it is difficult to see, there is a value ρ = ρglobal ≈

(a) S1,2 for ρ = 13. (b) S1,2 for ρ = 13.91.

15

10
z

5 10

0
−10 −5 0 5 −10
10 y
x

(c) S1,2 for ρ = 24. (d) A visualization of the attractor.

Figure 3.20. The unstable manifolds of the origin in the Lorenz

system for different ρ (a)-(c), and the “Lorenz attractor” for ρ = 27
(d), visualized by one trajectory.
104 3. Dissipative dynamics

13.91 where the two parts S1 and S2 are connected with the two-dimensional
stable manifold of the origin and form two homoclinic connections, see Figure
3.20(b). For ρ < ρglobal the part S1 connects to X1∗ and S2 to X2∗ , see Figure
3.20(a), and vice versa for ρ > ρglob . This behavior is the origin of a so called
homoclinic explosion, cf. [Wig88], and creates the chaotic behavior in the
system. Therefore, chaotic behavior is present in the system already for ρ
close to ρglobal , but becomes only attractive for larger values of ρ. Figures
3.20(c) and (d)) illustrate the different behavior for ρ = 24 < ρHopf and
ρ = 27 (Figure 3.20(d)).
An elementary but more detailed introduction to the Lorenz system
can be found in [Str94, §9], including explanations of simple mechanical
and electronic systems able to simulate the Lorenz system, together with
applications to send encrypted messages.
Exercises
3.1. For the following ODEs ẋ=f (x) determine all fixed points and their stability
in dependence of the parameter μ∈R. What bifurcations occur at what μ?
a) ẋ=μ + 6 + 4x − x2 , b) ẋ = 2 − μ + x(μ−4) + 3x2 − x3 , c) ẋ=μ + x(μ−1) − x2 .
3.2. Compute the non-trivial solutions close to the origin for
 
εx − yx2 − x4
f (x, y, ε) = = 0.
y + x2
3.3. Check the stability of (x, y) = (0, 0) for the following systems by calculating
the center manifold, cf. [Wig03].
(a) ẋ = −xy − x6 , ẏ = −y + x2 , (b) ẋ = x2 y − x5 , ẏ = −y + x2 , and
      
xn+1 0 1 xn 0
(c) = + .
yn+1 − 12 23 yn −yn3
3.4. Which of the following systems has periodic orbits close to z = (x, y) = (0, 0)?
Does Theorem 3.3.1 apply?
 2 
−μ 1
2 z − |z| z,
2
(a) ż =
 −12 −μ 
−μ 1
(b) ż = z − |z|2 z, μ ∈ R, with small |μ|.
 −1 −μ
  2 
0 −1 x − xy
(c) ż = + .
1 0 xy
3.5. This exercise (from [Str94]) brings together a number of concepts treated
above, namely bifurcation, center manifold calculations, and non-trivial gradient
dynamics. The system
ϑ̇1 = f1 (ϑ) := k sin(ϑ1 − ϑ2 ) − sin ϑ1 ,
ϑ̇2 = f2 (ϑ) := k sin(ϑ2 − ϑ1 ) − sin ϑ2 ,
with ϑ = (ϑ1 , ϑ2 ) and parameter k > 0 describes two rotating magnets between
two fixed magnets, in a geometry as follows:
3.4. Routes to chaos 105

a) The system has exactly nine fixed points in [−π, π]2 for k < 1/2. Determine
these and their stability.
b) What bifurcation occurs at k = 1/2? What new fixed points emerge?
c) Find a potential V such that ϑ̇ = −∇V (ϑ).
d) Sketch a phase portrait for 0 < k < 1/2 and a phase portrait for k > 1/2.
Remark. Doing c) first and then the rest is a good idea.
3.6. The Chemostat. The chemostat is an industrially used “predator-prey sys-
tem” to cultivate bacteria. In case of 3 species the system is modeled by
m1 s(t)
ṡ(t) = (1 − s(t)) − x1 (t),
a1 + s(t)
 
m1 s(t) m2 x2 (t)
ẋ1 (t) = x1 (t) −1− ,
a1 + s(t) a2 + x1 (t)
 
m2 x1 (t) m3 x3 (t)
ẋ2 (t) = x2 (t) −1− ,
a2 + x1 (t) a3 + x2 (t)
 
m3 x2 (t)
ẋ3 (t) = x3 (t) −1 .
a3 + x2 (t)
a) Explain the modeling.

b) Let σ(t) = 1 − s(t) − 3k=1 xk (t). Show that σ̇(t) = −σ(t) and use this to prove
that the ω-limit set of any solution (s(t), x1 (t), . . . , x3 (t)) is contained in

3
Ω = {(s, x1 , . . . , x3 ) : s + xk = 1}.
k=1
3
c) Substitute s = 1 − k=1 xk into the equations for ẋk and try to reproduce the
period-doubling shown in Figure 3.21.
3.7. Let Qc (x) = x2 + c. Prove that for all c < 14 there exists a unique μ > 1 such
that Qc is conjugated to Fμ (x) = μx(1 − x) through the map h(x) = αx + β.
3.8. Consider the iteration xn+1 = Tλ (xn ), where
$
2x for x ∈ [0, 1/2],
Tλ (x) = λ
2 − 2x, for x ∈ (1/2, 1].
a) Prove, x∗ = 0 is an asymptotically stable fixed point for λ ≥ 0 sufficiently
small.
b) Compute Tλ2 = Tλ ◦ Tλ and find graphically the 2-periodic solutions,
i.e., solve Tλ2 (x) = x by finding the intersection points of the functions
x → Tλ2 (x) and x → x.
106 3. Dissipative dynamics

(a) (b) (c)

0.5
0.4 0.4 0.4

0.3 0.3

3
0.3

x
x
0.2 0.2
0.2 0.1
0.1

0.6
0.4 0.6
0.4 0.8 0.8
0.3 0.6 0.6 0.4 0.6
0.2 0.4 0.2 0.4 0.2 0.4
0.1 0.2 0.2
0.2 x x x x
2 1 2 1

(d) (e) (f)

0.5 0.5

0.4 0.4

0.3 0.3

0.1 0.15 0.2 0.25 0.1 0.15 0.2 0.25

Figure 3.21. Period-doubling bifurcation in the 3 component

chemostat. (m1 , m2 , m3 ) = (10, 4, 3.5), (a1 , a2 ) = (0.08, 0.23), and
a3 = 0.4, 0.3, 0.225, 0.2 in (a)-(d), respectively. (e,f) Poincaré sec-
tion to x1 = 0.3 for a3 = 0.225 and a3 = 0.2, respectively. For
a3 = 0.2 the time span is t ∈ [0, 200].

c) Find for λ = 4 the structure of the set

Sλ = {x : Tλn (x) ∈ [0, 1] ∀n ∈ N0 }
What kind of dynamics do you expect in Sλ ?
3.9. With z = x + iy und c = a + ib the discrete dynamical system
πa,b (x, y) = (x2 − y 2 + a, 2xy + b)
can be written in the complex form fc (z) = z 2 + c. The Mandelbrot set is sketched
in Figure 3.9, and is the set of all c ∈ C for which the sequence (zn )n∈N , defined
through zn+1 = fc (zn ), z0 = 0, is bounded.
a) Prove that πa,b possesses an asymptotically stable fixed point (x0 , y0 ) if
|fc (z0 )| < 1. Express the fixed points of Dπa,b (x0 , y0 ) in terms of f  .
b) Find the fixed points of πa,b and compute their stability. Show that the
set of all c, for which an asymptotically stable fixed point z0 (c) exists is
given by the interior of the cardioid
c(t) = 1
4 − 14 (1 − eit )2 , t ∈ R.
c) Compute the fixed points of = πa,b ◦ πa,b . Show that the non-trivial
2
πa,b
2-periodic solutions of πa,b are asymptotically stable for |c + 1| < 14 .
3.4. Routes to chaos 107

d) Which periods can be expected in the different parts of the Mandelbrot

set? For this consider the eigenvalues of the linearization in the fixed
point z0 (c) for different values of c at the boundary of the cardioid.

Figure 3.22. The Mandelbrot set, cf. [Man91].

 Chapter 4

Hamiltonian dynamics

In the systems considered in Chapter 3, the evolution changes the volume of

sets in phase space. However, many systems in nature conserve this volume,
especially those of classical mechanics. For these systems we will discuss
their dynamical properties, as stability and instability, and the occurrence
of chaotic behavior. Our starting point of the bifurcation analysis of dissi-
pative systems usually was a system with a globally attracting fixed point.
In conservative systems such things cannot exist, and so we start from a
completely integrable system, i.e., from a system in which all solutions can
in principle be computed explicitly. It is the main purpose of this section to
contrast the behavior of dissipative and conservative systems. Hence, essen-
tial parts in usual courses about Hamiltonian systems will be skipped. For
an overview we refer to the textbooks [Arn78, Thi88, MH92, HZ11], or
the selection of reprints [MM87].

4.1. Basic properties

The basic rule of classical mechanics is that the force f : Rd → Rd acting
on a mass point at a position q ∈ Rd equals the product of mass m and
acceleration q̈, i.e.,

(4.1) mq̈ = f (q).

For simplicity we set m = 1 in the following. In a conservative system, to a

given force f : Rd → Rd there exists a potential U : Rd → R with f = −∂q U .
Introducing the momentum variable p = q̇ ∈ Rd gives the first order system
     1   
q̇ p 2 ∂p (p2 ) ∂p H
(4.2) = = = = J∇H,
ṗ −∂q U −∂q U −∂q H

109
110 4. Hamiltonian dynamics

where H : R2d → R given by H = 12 p2 + U is called the Hamiltonian and

where  
0 I
J= ∈ R2d×2d
−I 0
is a skew symmetric operator with I the identity matrix in Rd×d . All classical
mechanical systems can be written as a Hamiltonian system
(4.3) u̇ = J∇H(u),
with u(t) ∈ R2d , H : R2d→ R the Hamiltonian, and J ∈ R2d×2d the skew
symmetric operator from above. For ∇H locally Lipschitz-continuous, there
exists a unique solution u = u(t, u0 ) of (4.3) with initial condition u|t=0 = u0 ,
cf. Theorem 2.2.1. An important property of (4.3) is the conservation of
energy.
Theorem 4.1.1. The Hamiltonian H is constant along a solution, i.e.,
H(u(t, u0 ))=H(u0 ).

Proof. Let u = u(t) be a solution of the Hamiltonian system (4.3). Then

d
H(u(t)) =(∇H)T u̇(t) = (∇H)T J∇H = 0,
dt
due to the skew symmetry of J. 
In case d = 1 in (4.1) the full phase portrait can be constructed graphi-
cally.
Example 4.1.2. Consider an ODE
(4.4) ẍ = f (x)
with locally Lipschitz-continuous but otherwise arbitrary f : R → R. In
Figure 4.1 we explain how to draw the phase portrait without any formulas
or calculations. Given f , without formulas, we may draw the potential
energy −F (x). In the top left we do thisby first labeling the 4 zeros of
x
f by x1 , . . . , x4 . Setting, e.g., −F (x) = − 0 f (ξ) dξ we may plot −F and
find that x1 , . . . , x4 are stationary points of −F ; here they are extrema since
the roots of f are simple. Note that given −F we obtain E = 12 ẋ2 − F by
simply adding a parabola in ẋ to −F (u) at each x, see the bottom right for
an illustration. Thus, minima of −F are minima of E, while maxima of −F
are saddle points for E with stable direction (0, 1) and unstable direction
(1, 0). To draw orbits we may think of small balls rolling around on the
energy surface. For instance, consider a ball starting at (x, ẋ) = (x5 , 0).
It will slowly start to move to the right, thereby loosing potential energy
−F and taking up speed, hence gaining kinetic energy 12 ẋ2 . At x = x2 it
will have maximum speed, and then has to roll uphill, thus loosing kinetic
energy but gaining potential energy. Thus, it will roll precisely until x6 ,
4.1. Basic properties 111

defined by −F (x6 ) = −F (x5 ), where all kinetic energy has been transformed
to potential energy again. The ball will now roll back, and all together
we obtain a periodic orbit γ1 . Similar periodic orbits are obtained for all
starting positions (x, 0) with x2 < x < x3 , or, equivalently for all (x, ẋ)
inside the region bounded by the homoclinic orbit γ2 to (x3 , 0) and passing
through (x7 , 0). In a similar way all orbits can be constructed graphically.
For instance, the orbit γ3 corresponds to a ball coming from the far left with
some large positive speed and rolling through the potential all the way to
a position x8 on the (far) right where it reaches some maximal potential
energy −F (x8 ) and then rolls back.

−F(x)
f(x)
0

γ3
4

γ1 2

γ2
0

-2
0 1
x1 x2 x3 x4 -2 -1
x7 x5 x6 -4 -4 -3

Figure 4.1. Phase portrait for a general scalar Newtonian system

ẍ = f (x).

The volume of sets in the phase space of Hamiltonian systems is pre-

served.
Theorem 4.1.3. (Liouville) Let
u(t, Ω) = {v ∈ R2d : ∃u0 ∈ Ω : u(t, u0 ) = v}
be the image of a bounded and measurable Ω ⊂ R2d under the map u(t, ·),
and let μ be the Lebesgue measure in phase space, i.e., dμ = 1du. For all
t ∈ R we then have  
dμ(u) = dμ(u).
Ω u(t,Ω)
112 4. Hamiltonian dynamics

Proof. A change of coordinates yields

 
dμ(u) = |Du0 u(t, u0 )| dμ(u0 ).
u(t,Ω) Ω

The Jacobi matrix Y (t) = Du0 u(t, u0 ) satisfies the differential equation
d
Y (t) = ((D(J∇H))(u(t, u0 )))Y (t).
dt
Furthermore, the determinant |Y (t)| satisfies
d
|Y (t)| = trace(D(J∇H))|Y (t)|.
dt
See Exercise 4.1 for a proof of this formula in R2 . Using Y (0) = I, and

2d 
2d 
2d 
2d 
2d
trace(DJ∇H) = δik ∂ uk Jij ∂uj H = Jij ∂ui ∂uj H = 0
i=1 k=1 j=1 i=1 j=1

due to the skew symmetry of J and due to the symmetry of the matrix
(∂ui ∂uj H)i,j , we obtain the assertion. 
The invariant Lebesgue measure in phase space is called Liouville mea-
sure. The theory of measure preserving dynamical systems is the subject
of ergodic theory, see, e.g., [Wal82, Kre85]. Complicated dynamical be-
haviour in Hamiltonian systems is described statistically in this theory.

4.1.1. Dynamics near a fixed point. As a direct consequence of the

invariance of the phase space volume Hamiltonian systems cannot possess
asymptotically stable fixed points. We start with the discussion of the lin-
earization of the system at the fixed point. The Hamiltonian H must be
quadratic in order to obtain a linear differential equation, i.e.,
1 
2d
1
H(u) =
M u, u = mij ui uj .
2 2
i,j=1

Then
1  1
2d 2d
∂ uk H = (mij δik uj + mij ui δjk ) = (mik + mki )ui ,
2 2
i,j=1 i=1

i.e., w.l.o.g. M can be considered as symmetric. Hence, a linear Hamiltonian

system is of the form
  u̇ = JM u
0 I
with J = and M = M T .
−I 0

Lemma 4.1.4. Let λ be an eigenvalue of JM . Then also −λ, λ and −λ

are eigenvalues of JM .
4.1. Basic properties 113

Proof. We have det J = 1, J 2 = −I and J T = −J. For the characteristic

polynomial of JM we obtain
p(λ) = det(JM − λI) = det(J) det(JM − λI) = det(J 2 M − λJ)
= det(−M − λJ) = det(−M − λJ) det(J)
= det(−M J + λI) = det((−M J + λI)T )
= det(−J T M T + λI) = det(JM + λI) = p(−λ).
Hence, with λ also −λ is an eigenvalue. Since JM is a real-valued matrix,
the complex conjugate of an eigenvalue must be an eigenvalue, too. 
A direct consequence is the following lemma, which again implies that
in Hamiltonian systems no asymptotically stable fixed points can exist.
Lemma 4.1.5. A fixed point of a Hamiltonian system can only be stable
if all eigenvalues of the linearization lie on the imaginary axis with same
geometric and algebraic multiplicity.

For general ODEs in case that all eigenvalues lie on the imaginary axis
the nonlinear terms decide about stability. For Hamiltonian systems the
quadratic approximation of the Hamiltonian at the fixed point gives addi-
tional information.
Theorem 4.1.6. Let H(u) = 12 uT Au + O(u3 ) with A strictly positive (or
strictly negative) definite. Then u = 0 is stable.

Proof. Let
ρ0 (r) = min{H(u) : |u| = r}, ρ1 (r) = max{H(u) : |u| = r}.
Then
1 1
ρ0 (r) = λmin r2 + O(r3 ), ρ1 (r) = λmax r2 + O(r3 ),
2 2
where λmin > 0, respectively λmax > 0, is the smallest, respectively the
largest eigenvalue, of the positive definite matrix A. Then there exists an
r0 > 0, such that
1
ρ0 (r) ≥ λmin r2 and ρ1 (r) ≤ λmax r2 .
4
for all r ∈ [0, r0 ]. Given ε > 0 we choose
λmin
0≤δ≤ min(ε, r0 ).
4λmax
Then
4 4
|u(t, u0 )|2 ≤ ρ0 (u(t, u0 )) ≤ H(u(t, u0 ))
λmin λmin
4 4 4λmax 4λmax 2
= H(u0 ) ≤ ρ1 (u0 ) ≤ |u0 |2 ≤ δ ≤ ε2 .
λmin λmin λmin λmin
114 4. Hamiltonian dynamics

Hence, the solution u = u(t, u0 ) cannot leave the ε-neighborhood, if the

initial condition u0 is contained in the δ-neighborhood. 
Example 4.1.7. Let H = q 2 + p2 . Then
q̇ = p, ṗ = −q.
√
The orbits are circles with radius H and u = (q, p) = (0, 0) is stable.

4.1.2. Lyapunov’s subcenter theorem. Our next goal is the existence

of periodic solutions. It turns that in each neighborhood of a fixed point with
imaginary eigenvalues we can always find periodic solutions if the eigenvalues
satisfy some non-resonance condition. Before we prove this result, we need
two preparations.
First we introduce angle and action variables for linear systems. We
consider the harmonic oscillator
q̈ = −ω 2 q,
which we write as Hamiltonian system
q̇ = ωp = ∂p H, ṗ = −ωq = −∂q H,
with Hamiltonian H = 12 ω(p2 + q 2 ). Introducing polar coordinates
√ √
q = 2I cos(φ), p = 2I sin(φ)
shows that the new variables φ and I satisfy the Hamiltonian system
φ̇ = ω = ∂I H, I˙ = 0 = −∂φ H
with Hamiltonian H = ωI. The 2π-periodic variable φ is called angle vari-
able and the variable I which is preserved under the flow is called action
variable. Such variables play a fundamental role in the description of com-
pletely integrable Hamiltonian systems in the following.
Secondly we explain a reduction method for systems which are at least
partly given in action angle variables. Let H = H(φ, q, I, p) be 2π-periodic
w.r.t. φ. Moreover, assume that ∂I H > 0 in some open subset of R2d . Then
H(φ, q, I, p) = h can be solved for I = −K(q, p, φ, h).
From d
dt H = 0 and H(φ, q, −K(q, p, φ, h), p) = h it follows that
∂qi H + ∂I H · (−∂qi K) =0,
∂pi H + ∂I H · (−∂pi K) =0,
d d d −1
and hence using dt φ = ∂I H and dt φ = ( dφ t) yields
(4.5) ∂φ qi =∂t qi ∂φ t = ∂pi H/∂I H = ∂pi K(q, p, φ, h),
(4.6) ∂φ pi = − ∂t pi ∂φ t = ∂qi H/∂I H = −∂qi K(q, p, φ, h).
4.1. Basic properties 115

This is a 2(d−1)-dimensional Hamiltonian system which is 2π-periodic w.r.t.

the new time variable φ.
With these two preparations we are now going to prove
Theorem 4.1.8. (Lyapunovs subcenter theorem) Let u = 0 be a fixed
point of the Hamiltonian system u̇ = J∇H(u). Let ±iω, with ω = 0, be sim-
ple eigenvalues of the linearization JD 2 H(0), and let all other eigenvalues
λj fulfill λj = inω for all n ∈ Z. Then there is a neighborhood U of 0 and a
two-dimensional manifold M ⊂ U which is filled with periodic solutions with
period close to 2π/ω. Moreover, M is tangential to the subspace spanned by
the eigenvectors which are associated to the eigenvalues ±iω.

Proof. We seek small solutions x = ε x, with 0 < ε  1 a small parameter,

and consider the rescaled Hamiltonian
 ε ( 1
H x) = 2 H(εx).
ε
By this rescaling the quadratic part of the Hamiltonian stays independent
of ε, whereas the higher order terms become small. W.l.o.g. assume that
⎛ ⎞
ω 0 0
 ε (0) = D 2 H(0) = ⎝ 0 ω 0 ⎠ ,
D2 H
0 0 A
with A ∈ R2(d−1)×2(d−1). This form can always be achieved by interchanging
the second and the (d + 1)th coordinate. The coordinates associated to the
matrix A are denoted by x √ q1 and p1 we
. For √the first two coordinates
introduce polar coordinates q1 = 2I cos(φ) and p1 = 2I sin(φ). In the
new coordinates we have
1 T
) = ωI + x
Hε (I, φ, x  Ax + O(ε).
2
We look for solutions with Hε (I, φ, x ) = ω. For ω = 0 we have ∂I Hε = 0
in a neighborhood of the origin such that Hε (I, φ, x ) = ω can be solved
w.r.t. I. We write this solution as I = 1 − Kε (φ, x ). Inserting this ansatz
) = ω yields
in Hε (I, φ, x
1 T
) =
Kε (φ, x  A
x x + O(ε).
2ω
By the above reduction we obtain a 2π-periodic Hamiltonian system with
new time variable φ and associated Poincaré map ψε ( x0 ) = x(2π, x
0 ). A
fixed point x ∗
 of the Poincaré map ψε yields a periodic solution x ∗ ) =
(t, x
(t + 2π, x
x ∗
 ). Thus, a periodic solution can be obtained via a zero of the
function F ( x∗ , ε) = ψε (x∗ ) − x∗ . We have F (0, 0) = 0 since K0 (φ, x ) =
2ω 
1 T
x Ax and thus 
x = 0 is a solution of the associated linear autonomous
Hamiltonian system x ˙ = ω1 JA x. For the same reason we have Dx∗ F (0, 0) =
116 4. Hamiltonian dynamics

2π
e ω JA − I. Due to the non-resonance assumption, Dx∗ F (0, 0) has only non-
zero eigenvalues and hence is invertible. Then, by the implicit function
theorem, F ( x∗ , ε) = 0 can be solved for x
∗ = x
∗ (ε). The associated solution
) is non-trivial since I = 0. The period is
(I, φ, x
 T  2π  2π  2π
dφ dφ dφ 2π
dt = = = = + O(ε).
0 0 φ̇ 0 ∂I H 0 ω + O(ε) ω
This family of periodic solutions is tangential to span{I, φ} since 
x =
O(ε). 

4.2. Some celestial mechanics

To give some illustrations of computations in Hamiltonian dynamics we
review some very basic celestial mechanics. A good reference is [Gut94].

4.2.1. The 1-body problem. Let q ∈ R3 be the position of a mass point,

 (q), i.e.,
e.g., earth, that moves in a radially symmetric potential U (q) = U
U (q) only depends on q, e.g. the gravitational potential of the sun. Then
 (q)q/q,
q̈ = −∇U (q) = −U
or, in Hamiltonian form
 
d q 1
= J∇H(q, p), with H(q, p) = U (q) + p2 .
dt p 2
This is a 6-dimensional first order ODE, or, more precisely a Hamilton-
ian system with 3 degrees of freedom. Using the angular momentum, see
Exercise 4.8, ⎛ ⎞
a2 b3 − a3 b2
C = q × q̇, a × b = ⎝a3 b1 − a1 b3 ⎠
a1 b2 − a2 b1
the dimension can be reduced. Let C = 0. Then C is orthogonal to the
orbital plane {αq + β q̇ : α, β ∈ R}. The area swept until t then is by
Leibniz’s sector formula

1 t
F (t) = |q(s) × q̇(s)| ds.
2 0
This yields Kepler’s second law: the line segment joining a planet and
the Sun sweeps out equal areas during equal intervals of time, or in modern
d
formulation: dt F (t) = 12 |q(t) × q̇(t)| = 12 C.
Now assume that the orbital plane is the q1 -q2 plane, let q = (q1 , q2 ) =
(x, y), and introduce polar coordinates, i.e.
   
x cos φ
q= =r .
y sin φ
4.2. Some celestial mechanics 117

Then    
cos φ − sin φ
p := q̇ = ṙ + rφ̇ ,
sin φ cos φ
and C and H in polar coordinates become
1  (r).
(4.7) C = r2 φ̇ = 0, H = (ṙ2 + r2 φ̇2 ) + U
2
From (4.7) we obtain φ̇ = 0 and ṙ2 = 2H − r2 φ̇2 − 2U  (r). This yields a scalar
first order equation as follows. Locally we can assume r = r(φ); then
d d d ṙ
ṙ = r(φ) = ( r)φ̇ ⇒ r := r= .
dt dφ dφ φ̇
In particular
2(H − U ) 2(H − U )r4
(r )2 = − r2 = − r2 ,
φ̇2 C2
which yields the Clairaut ODE

r = ±g(r), with g(r) = 2C −2 (H − U (r))r4 − r2 .
Instead of r we use the inverse radius σ=1/r, which yields the so called
fundamental equation of the 1-body problem
σ = g(1/σ)σ 2 .
For the gravitational potential U (r) = −Ar−1 we obtain

(4.8) σ = − −σ 2 + ασ + β, with α = 2A/C 2 , β = 2H/C 2 .
Lemma 4.2.1. a) We have H ≥ −A2 /(2C 2 ) (lower energy bound).
b) For H = −A2 /(2C 2 ) the orbit is a circle with radius C 2 /A.

Proof. Completing the square we write −(σ )2 = σ 2 − ασ − β = (σ − α2 )2 −

δ/4, where δ = α2 + 4β. Thus δ ≥ 0 and hence H ≥ −A2 /(2C)2 . For
H = −A2 /(2C)2 we have δ = 0 and hence σ = 0 and σ ≡ α/2. 
Henceforth let δ = α2 + 4β > 0. We seek solutions of
(4.9) σ = −g(σ)
√ √
 α− δ α+ δ
with g(σ) = −σ 2
+ ασ + β and σ ∈ I := [ , ] to have a real
2 2
radicand. In particular, an unbounded orbit is only possible for
√
0 ∈ I ⇔ α ≤ δ ⇔ β ≥ 0 ⇔ H ≥ 0,
√
while the orbit is always bounded if H < 0. Moreover, from σ ≤ α+ δ
2 we
obtain a minimal distance of the orbit to the origin, i.e.,
2
r ≥ rmin := √ .
α+ δ
118 4. Hamiltonian dynamics

√ √
Let σ(0) = σ0 with max{0, α−2 δ } < σ0 < α+2 δ . Then we have a local
√
solution as g is locally Lipschitz near σ0 . Substituting u = 12 δx + α2 we
obtain
 σ  σ  σ
du du dx
φ= =  = √
σ0 −g(u) σ0 − −u + αu + β
2 0 − 1 − x
σ
2

 − arccos σ
= arccos σ 0 ,
0 = √2 (σ0 − α2 ) and σ
with σ  = √2 (σ − α2 ). From −1 < σ
0 < 1 we obtain
δ δ
σ0 ) < π and thus
0 < arccos(
2
(4.10) √ (σ − α/2) = σ  = cos(φ + b) with b = arccos(σ0 ).
σ
W.l.o.g. we choose the initial condition σ0 = 1/rmin and obtain
p 
(4.11) r(φ) = with p = C 2 /A, e = 1 + 2HC 2 /A2 .
1 + e cos(φ)
We distinguish three cases.
(1) 0 ≤ e < 1 ⇔ H < 0: Then r(φ) is defined for all φ ∈ R and
2π-periodic in φ. Going back to cartesian coordinates we obtain
     
x q1 cos φ
= =r ,
y q2 sin φ
and cos φ = (p/r − 1)/e yields (p − ex)2 = r2 = x2 + y 2 , and thus
 2
ep y2 p2
x+ + = .
1 − e2 1 − e2 (1 − e2 )2
This is Kepler’s first law: The body moves on an ellipse with focal points
(0, 0) and (−(2ep)/(1 − e2 ), 0), numerical eccentricity e and major semi-axis
a = −A/(2H). The point (rmin , 0) is called peri center (perihel for a planet
in the solar system) and (rmax , π) is called apo center (apohel). Examples for
numerical eccentricities e are e = 0.0167 for Earth, e = 0.2056 for Mercury,
and e = 0.9673 for Halley’s comet. The relatively large eccentricity of
Mercury is of great importance historically since already in the 19th century
it allowed the observation of the perihel precession of Mercury: after each
elliptical orbit Mercury’s perihel is shifted by a few angular seconds. This
contradicts the above (newtonian) calculations, but could be explained by
Einstein’s relativity theory.
(2) e = 1 ⇔ H = 0: the existence interval is −π < φ < π, and
geometrically the orbit is a parabola opening to the left, y 2 = −px + p2 .
(3) e > 1 ⇔ H > 0: the orbit is the hyperbola y 2 = (e2 −1)x2 −2epx+p2 .
Thus we found the orbits for the 1-body problem in implicit form and
without time dependence, determined by parameters H, A and C. Next, the
orbits can be characterized via initial conditions and the time-dependence
4.2. Some celestial mechanics 119

major semi axis F1

sun F1
apohel P2 perihel sun sun

F2 F2

Figure 4.2. Kepler-ellipse, parabola and hyperbola. The areas F1 ,

F2 are meant to illustrate Kepler’s 2nd law.

can be reintroduced. This allows to derive Kepler’s 3rd law: if aγ is the

length of an orbit’s major semi-axis, and Tγ its period, then Tγ2 /a3γ is inde-
T2 4π 2
pendent of the orbit γ. More precisely, in this calculus we have 3 =
a GM
for all orbits in the solar systems, where G is the gravitational constant and
M the mass of the sun, which is a very good approximation to observations.
This is only natural, as Kepler derived his laws from observations.

4.2.2. The restricted 3-body problem. N bodies which move under

the influence of gravity have the Hamiltonian

N
p2i  mi mj
H= + .
2mi |qi − qj |
i=1 i=j

The solution of the associated differential equations, and the associated ques-
tion about the mechanical stability of our solar system, have been considered
as essential for mankind. However, it turned out that only the two body
problem (N = 2), see above and Exercise 4.9, can be solved explicitly and
already the three body problem shows chaotic behavior.
There is one intermediate problem, namely the so called restricted three
body problem. There it is assumed that the third body K3 has a very small
mass compared to the other two bodies K1 and K2 . The restricted three
body problem is obtained by neglecting the forces of K3 on K1 and K2 , such
that their motion is not affected by K3 , i.e., they move on Kepler ellipses
around their center of mass.
In a coordinate system which moves with the two larger bodies of reduced
masses μ = m1m+m1
2
and 1−μ, which lie fixed in −μ and 1−μ, the Hamiltonian
for the motion of the third body is given by
1
H(q, p) = (p21 + p22 + p23 ) + q2 p1 − q1 p2 + U (q),
2
where
q2 q2 1−μ μ
U (q) = − 1 − 2 −  − .
2 2 2 2
(q1 + μ) + q2 + q3 2 (q1 − 1 + μ)2 + q22 + q32
120 4. Hamiltonian dynamics

The second and third term in H and the first term in U come from the
Coriolis force in the rotating coordinate system. There are five equilibria,
called Lagrangian points, shown in Figure 4.3.

q2

P4

P1 −μ P2 1−μ P3 q1

P
5

Figure 4.3. The equilibria in the restricted three body problem.

For the stability of these equilibria we first consider

⎛ ⎞
∂q21 U ∂ q1 ∂ q2 U 0 0 −1 0
⎜ ∂ q1 ∂ q2 U ∂q22 U 0 1 0 0 ⎟
⎜ ⎟
⎜ 0 0 ∂ q3 U 0 0 0 ⎟
2
2
M = D H(u) = ⎜ ⎜ ⎟.
⎜ 0 1 0 1 0 0 ⎟⎟
⎝ −1 0 0 0 1 0 ⎠
0 0 0 0 0 1

For the linearisation JM in these points (especially ∂qj ∂q3 U |Pi = 0 for j =
1, 2 and i = 1, . . . , 5) we thus obtain
⎛ ⎞
0 1 0 1 0 0
⎜ −1 0 0 0 1 0 ⎟
⎜ ⎟
⎜ 0 0 0 0 0 1 ⎟
JM = ⎜ ⎜ ⎟.
⎜ −∂q1 U
2 −∂q1 ∂q2 U 0 0 1 0 ⎟ ⎟
⎝ −∂q1 ∂q2 U −∂q22 U 0 −1 0 0 ⎠
0 0 −∂q23 U 0 0 0

We find that the q3 , p3 -part decouples and leads to the eigenvalue problem

0 = λ2 + ∂q23 U.

Since ∂q23 U > 0 we have λ1,2 ∈ iR. For the remaining eigenvalues we have

0 = λ4 + λ2 (∂q21 U + ∂q22 U + 4) + (∂q21 U )(∂q22 U ) − (∂q1 ∂q2 U )2 .

4.3. Completely integrable systems 121

It turns out that the points P1 , P2 and P3 are saddles and therefore unstable.
In the points P4 and P5 we find
√ √
1 3 2 3 2 9 3 3 1
q1 = − μ, q2 = ± , ∂q1 U = − , ∂q2 U = − , (∂q1 ∂q2 U ) = −
2
( − μ).
2 2 4 4 2 2
For 4( 12 −μ)2 < 1 the eigenvalues are purely imaginary, i.e., the points P4 and
P5 are linearly stable. Unfortunately M is indefinite, such that we cannot
conclude on the nonlinear stability of P4 and P5 with the above theorem.
Nevertheless P4 and P5 are realized in nature and play an important role
for space missions. For instance, Sun and Jupiter can be taken as the big
bodies, and in an angle of 60 degrees before and after Jupiter on his orbit
there are the so the called Greeks and Trojans, some families of asteroids.

4.3. Completely integrable systems

If there are several non-resonant eigenvalues on the imaginary axis, then
there are several families of periodic solutions. In this and the following
section we discuss situations with even more complex structures. We start
with the linear Hamiltonian system
(4.12) ẋ = JM x.
We assume that all eigenvalues iωj of the matrix JM are semi-simple and
on the imaginary axis. Then the system can be transformed into
(4.13) q˙j = ωj pj , p˙j = −ωj qj , j = 1, . . . , d,
i.e., into a Hamiltonian system with Hamiltonian

d
ωj
H= (qj2 + p2j ).
2
j=1

Clearly this system is the direct sum of d Hamiltonian systems with d inde-
ω
pendent Hamiltonians Hj = 2j (qj2 + p2j ). The Ij = 2ω1 j Hj are conserved also
d
for the flow of (4.12), i.e., dt Ij (x(t)) = 0 for solutions x = x(t) of (4.12).
th
For the j system the orbits are circles, i.e.,

qj + ipj = 2Ij eiφj with φj (t) = φj (0) + ωj t mod 2π.
For (4.12) the phase space decomposes into d-dimensional tori
{u ∈ Rd : I1 = c1 , . . . , Id = cd }.
For one or more vanishing Ij s we have dimensions of the tori between 1 and
d. The d-dimensional tori contain so called quasi-periodic solutions
x(t) = g(ω1 t, . . . , ωd t)
122 4. Hamiltonian dynamics

with g : S 1 × . . . × S 1 → R2d . If the non-resonance condition

ω · n = ω1 n1 + . . . + ωd nd = 0 for all (n1 , . . . , nd ) ∈ Z × . . . × Z \ {0, . . . , 0}
holds, then the orbits are dense in the associated d-dimensional torus. If for
instance ω1 = 2ω2 and all other ωi are non-resonant, then the solutions are
dense in d − 1-dimensional tori.
In the following we study whether this situation persists under perturba-
tions or not. We expect that it is more simple to destroy a torus filled with
low-dimensional solutions associated to a set of resonant ωs than a torus
with dense solutions associated to non-resonant ωs. Therefore, we expect
non-resonance conditions to play an important role. Moreover, such tori
are not only important for linear Hamiltonian systems but also for special
nonlinear systems, which are called completely integrable, see below.
A coordinate transform y = T (x) will in general destroy the Hamiltonian
structure of a Hamiltonian system. Only so called symplectic transforma-

tions keep the Hamiltonian structure. For H(y) = H(x) we have

2d
∂xj H(T (x)) = 
∂yk H(y)∂ xj yk ,
k=1
and hence

ẏ = (DT )ẋ = (DT )J∇x H(x) = (DT )J(DT )T ∇y H(y) 
= J∇y H(y),
if T is a so called symplectic transformation.
Definition 4.3.1. Let J −1 = −J = J T . The bilinear form
ω(v1 , v2 ) = v1T Jv2
is called the symplectic structure induced by J. A transformation y = T (x)
is called canonical or symplectic if
((DT )(x))J(DT )T (x) = J ∀x ∈ R2d .
A Hamiltonian system is called completely integrable if it can be transformed
into the form
(4.14) I˙j = −∂φj H = 0, φ˙j = ∂Ij H = ωj , j = 1, . . . , d
by a symplectic transformation. The Hamiltonian H = H(I1 , . . . , Id ) and
the frequencies ωj = ωj (I1 , . . . , Id ) only depend on the conserved quantities
I1 , . . . , Id . If the set
{x ∈ R2d : I1 = const1 , . . . , Id = constd }
is smooth and compact then it is a d-dimensional torus. The Ij and φj are
called action and angle variables, respectively.
Lemma 4.3.2. The map x0 → x(t, x0 ) is symplectic for all t.
4.4. Perturbations of completely integrable systems 123

Proof. Consider R(t, ·) = (Dx(t, ·))J(Dx(t, ·))T . Then R|t=0 = J and R

solves the linear ODE
Ṙ =(D ẋ)J(Dx)T + (Dx)J(D ẋ)T
=JD 2 H(Dx)J(Dx)T + DxJ(JD 2 H(Dx))T
=JD 2 HR + R(JD 2 H)T .

Since (D 2 H)T = D 2 H and J T = −J we have

JD 2 HJ + J(JD 2 H)T = JD 2 HJ + J(D 2 H)T J T = 0.
Hence, R = J is the unique solution. 

Example 4.3.3. Let F = F (I,  φ). Then the map induced by φ = ∂ F and
I
I = ∂φ F is symplectic, cf. [Arn78, §48].

There are various sufficient conditions that ensure that a Hamiltonian

system is completely integrable. Here we will only cite one major criterion.
Definition 4.3.4. Let F : R2d → R and G : R2d → R be in C 1 . Then
{F, G} = (∇F )T J∇G
is called the Poisson bracket of F and G.
Theorem 4.3.5. (Liouville’s theorem) Let H : R2d → R, and let I1 =
H, I2 , . . . , Id be independent integrals in involution, i.e. {Ij , Ik } = 0 for i, j =
1, . . . , d. Then the Hamiltonian system is completely integrable.

Proof. See [Arn78, §49]. 

4.4. Perturbations of completely integrable systems

The question occurs how robust completely integrable systems are under
perturbations. The answer turns out to be rather delicate and has to do with
number theory. The starting situation is as follows. Consider a Hamiltonian
of the form
(4.15) H(φ, I) = H0 (I) + εH1 (φ, I, ε),
with I ∈ Rd , φ ∈ T d , and ε a small parameter. The associated Hamiltonian
system reads
φ̇ = ∂I H0 + ε∂I H1 (I, φ), I˙ = −ε∂φ H1 (I, φ).
Hence, the action variable I only changes slowly in time. The idea is to
find a change of coordinates such that the transformed system is of the
original completely integrable form. According to Example 4.3.3 we obtain
124 4. Hamiltonian dynamics

a symplectic transformation if we take a so called generating function F :

R2d → R and define a relation between the old and new variables through
φ = ∂ F (φ, I)
I
 and 
I = ∂φ F (φ, I).
We look for a transformation near the identity and therefore choose
(4.16)  = φI + εf (φ, I),
F (φ, I) 
which yields
φ =∂IF = φ + ε∂If, I = ∂φ F = I + ε∂φ f,
and therefore
φ =φ − ε∂If + O(ε2 ), I = I + ε∂φf + O(ε2 ).
Plugging this into the Hamiltonian gives
H(  I)
 φ,  I,
 =H0 (I + ε∂ f ) + εH1 (φ,  0) + O(ε2 )
φ

=H0 (I)  I,
 + ε[∂ H0 · ∂ f + H1 (φ,  0)] + O(ε2 ).
I φ
The idea is to eliminate the terms of order O(ε) by finding f such that
 I,
∂IH0 · ∂φf + H1 (φ,  0) = 0.

If we find such an f , then we can go on and find in the next step another
symplectic transformation which then eliminates the O(ε2 ) terms, etc., until
finally all perturbations are eliminated. Before we do so we look at the prob-
lem to eliminate the terms of order O(ε) in more detail. Given H1 (φ,  I,
 0)
 we seek f : Td × Rd → R, such that
and H0 (I)
(4.17) ∂IH0 (I)  I)
 · ∂ f (φ,  I,
 = −H1 (φ,  0).
φ

Since f is periodic in φ we expand, for fixed I,  the problem into Fourier

series, cf. §5.2.2. Thus,
   
 =
f (φ)  =
ak eik·φ , H1 (φ) bk eik·φ , k = (k1 , . . . , kd ).
k∈Zd k∈Zd
Next we set
ω(I)  = (ω1 , . . . , ωd )(I).
 = ∂ H0 (I) 
I
 0 we obtain i(ω · k)ak = −bk . Thus, if ω · k = 0, then
For k =
ibk
ak = ,
ω·k

is determined and the term bk eik·φ can be removed from H1 . If there are no
resonances at all, i.e., if ω · k = 0 for all k ∈ Z \ {0}, then formally all O(ε)
terms can be removed except of b0 (I).  However, this term can be included
 as a correction. Obviously the non-resonance condition for the
into H0 (I)
elimination of the higher order terms is not changed since the left-hand side
4.4. Perturbations of completely integrable systems 125

of (4.17) is not changed. If there are no resonances until the nth step, then
the perturbation up to terms of order O(εn ) can be removed. We have the
following approximation theorem.
Theorem 4.4.1. If the normal form transformations allow to remove all
terms up to order O(εn ), i.e., if ∂φH = O(εn+1 ), then there exist C1 and
ε0 > 0 such that for all ε ∈ (0, ε0 ) we have in the original coordinates
sup I(t) − I(0) ≤ C1 ε.
t∈(−1/εn ,1/εn )

˙

 I) ≤ C2 ε for a C2 > 0, and, since I  ≤
Proof. We have I − I(φ, = ∂φH
C3 ε n+1 for a C3 > 0, we have
 − I(0)
I(t)  ≤ C3 |t|εn+1 ≤ C3 ε
for all |t| ≤ 1/εn . This yields

I(t) − I(0) ≤I(t) − I(t)  − I(0)
+ I(t)  
+ I(0) − I(0)
≤(2C2 + C3 )ε =: C1 ε.

So far we did not consider the convergence of the above Fourier series
in the normal form transforms. This turns out to be complicated due to
so called small divisor problems. This means that for given ω ∈ Rd and
(arbitrary small) δ > 0 there always is a k ∈ Zd such that
|k · ω| < δ.
Hence, the divisors in the series for f become arbitrarily small and the
convergence of the Fourier series is a serious problem. The problem is solved
by restricting the set of possible frequencies.
Definition 4.4.2. A vector ω ∈ Rd is called of type (L, γ) if for all k ∈
Zd \ {(0, . . . , 0)} we have
|k · ω| ≥ L|k|−γ

We remark that for given γ > d and almost all ω ∈ Rd there exists
a L > 0 such that ω is of type (L, γ), cf. [Arn88, page 114]. To study
the analytic properties of the generating function F we use the following
functions spaces.
Definition 4.4.3. For n ∈ N define the spaces

1,n = {a : Zd → C : a1,n = |a0 | + |ak ||k|n < ∞}.
k∈Zd

Remark 4.4.4. We have that a ∈ 1,n implies F −1 a ∈ Cbn , where

Cbn = {f : Td → R : f n times continuously differentiable},
126 4. Hamiltonian dynamics


which is equipped with the norm f Cbn = n|j|=0 ∂φj f C 0 , where f C 0 =
b  b
supφ∈Td |f (φ)|, and where F −1 : a → f is defined by f (φ) = k∈Zd ak eik·φ ,
cf. §5.1.
Lemma 4.4.5. Let ω ∈ Rd be of type (L, γ) and b ∈ 1,n with b0 = 0. Then
ibk
a, defined by ak = k·ω for k = 0, a0 = 0, is in 1,n−γ .

Proof. We have
 
   bk 
a1,n−γ = |ak |k| |=   ≤ L−1 b . 
 k · ω |k|
n−γ n−γ
 1,n
k∈Zd \{0} k∈Zd \{0}

The correspondence of the exponent n in the weight of 1,n in Fourier

space to regularity w.r.t. φ in physical space implies that in each iterative
step we lose regularity of the Hamiltonian H. In order to eliminate the
perturbation completely, infinitely many steps are necessary. Hence, there
will be a loss of infinitely many derivatives. This problem can be solved by
working in a space of analytic functions or by using some artificial smoothing
in the so called hard implicit function theorem, cf. [SR89], when working
with Hamiltonians of finite differentiability. The convergence is based on the
quadratic convergence of the Newton scheme. This approach results in the
famous KAM-theorem named after Kolmogorov, Arnold and Moser. We
choose the analytic version and set
Aσ,ρ (I ∗ ) = {(I, φ) ∈ Rn × Cn : |I − I ∗ | < ρ, |Im(φj )| < σ, j = 1, . . . , d}.
We define the norm of a function f which is analytic w.r.t. φ on Aσ,ρ (I ∗ )
by
f σ,ρ = sup |f (I, φ)|.
(I,φ)∈Aσ,ρ (I ∗ )
In nonlinear Hamiltonian systems in general the frequencies vary in a non-
trivial manner with I, i.e., ∂I ω = ∂I2 H0 does not vanish. Since there is
a dense set of resonant frequencies in Rd it cannot be expected that the
phase space is completely filled with tori after the perturbation. Therefore,
in any neighborhood of a torus Td with a non-resonant ω there is a torus
Td with a resonant ω. This means that next to any torus for which the
transformations can be carried through to arbitrary order there is a torus
in which low order perturbations influence the dynamics in the torus and
may destroy the torus. Nevertheless, almost all tori persist in the following
sense.
Theorem 4.4.6. (KAM) Let ω(I ∗ ) = ω ∗ be of type (L, γ) and let the
Hessian ∂I2 H0 be invertible in I ∗ . Then there exists an ε0 > 0 such that
for f σ,ρ < ε0 the Hamiltonian system has quasi-periodic solutions with
frequencies ω ∗ , i.e., the torus to I = I ∗ persists under the perturbation. Let
V ⊂ Rd be an open set with finite Lebesgue measure, where the Hessian ∂I2 H0
4.4. Perturbations of completely integrable systems 127

is invertible. For all δ > 0 there is an ε0 > 0, such that for all ε ∈ (0, ε0 )
there is a set Pε ⊂ V × T d with the following properties. The Lebesgue
measure μ of (V × T d ) \ Pε is less than δ, and for all (I0 , φ0 ) ∈ Pε the orbit
through (I0 , φ0 ) is quasi-periodic.

Proof. See [Arn78, Way96] or [KP03, §2] for a review. 

Hence, for small perturbations most of the phase space is still filled with
tori. In between the tori chaotic behavior may occur. This is explained
subsequently in §4.5.
In R4 the KAM-theorem yields a stability theorem since the invariant
tori form two-dimensional hypersurfaces in the three-dimensional energy sur-
faces. In higher space dimensions the d-dimensional tori cannot separate do-
mains in the 2d − 1-dimensional energy surface, but we expect that solutions
need a long time to wander around the tori. This is called Arnold-diffusion
and is mathematically formulated in Nekhoroshov’s Theorem below.
An important motivation of these investigations again comes from celes-
tial mechanics, in particular the question of the stability of our solar system.
If we ignore mutual gravitational forces between the planets then we obtain
a completely integrable system. The forces between the planets compared to
that of the sun have a ratio of ε ≈ 1/1000. Hence, the interplanetary forces
can be considered as small perturbations. Quite obviously it is impossible to
say whether our solar system is resonant or non-resonant, i.e., whether the
ratios between different rotation times are rational or irrational. In spite
of the fact that the rotation times of Jupiter and Saturn have a ratio of
about 2/5, our solar system seems to be remarkably stable. As said above,
heuristically, even in the resonant case we expect the solutions to need a
long time to wander around the remaining tori. This can be made precise
for so called steep Hamiltonians, cf. [AKN06, §6.3.4].
Definition 4.4.7. An analytic function is called steep if it has no real ex-
trema and if all complex extrema are isolated.
Theorem 4.4.8. (Nekhoroshov) Let H0 = H0 (I) be a steep function.
Then in the perturbed Hamiltonian system for a sufficiently small perturba-
tion εH1 we have
(4.18) I(t) − I(0)Rd < εb
for 0 ≤ t ≤ 1
ε exp( cε1a ), where a, b, c > 0 only depend on H0 .
Remark 4.4.9. KAM-theory is used as an explanation for the so called
Kirkwood gaps. Figure 4.4 shows the number of asteroids in the main aster-
oid belt as a function of their orbits major semi-axis in astronomical units
(AU), where 1 AU is the length of the major semi-axis of the earth’s orbit.
128 4. Hamiltonian dynamics

At certain values there are gaps in the distribution, and these correspond to
low resonances between the periods of the asteroids and Jupiter.

Figure 4.4. Kirkwood-gaps (ssd.jpl.nasa.gov/images/ast histo.ps)

Courtesy NASA/JPL-Caltech.

4.5. Homoclinic chaos

It is the purpose of this section to explain that in the part of the phase
space which is not filled with invariant tori for small ε > 0 chaotic behavior
can be expected. In between these tori there are periodic solutions and
their stable and unstable manifolds. If there exists a heteroclinic connection
with a transversal intersection of stable and unstable manifolds then a Smale
horseshoe map and hence shift dynamics and chaotic behavior can be found.
If for a map f : Rd → Rd the stable and unstable manifolds Ws and Wu of
a hyperbolic fixed point p intersect transversally in a point q, then due to
the invariance of the manifolds there must be infinitely many intersections.
See Figure 4.5. Hence complicated dynamics can be expected. Recall that
a fixed point p for an iteration f : Rd → Rd is called hyperbolic if the
linearization ∂x f (p) possesses no eigenvalues on the unit circle.
Theorem 4.5.1. (The Smale-Birkhoff homoclinic orbit theorem)
Let f : Rd → Rd be a diffeomorphism such that p is a hyperbolic fixed point,
and let q = p be another point in which there is a transversal intersection
of the stable manifold Ws (p) and the unstable manifold Wu (p). Then there
4.5. Homoclinic chaos 129

Wu

p q

Ws

Figure 4.5. A transversal homoclinic point q implies infinitely

many intersection points of the stable and unstable manifolds due
to the invariance of the manifolds.

exists a (hyperbolic) set Λ on which an iteration of f is homeomorphic to

shift dynamics.

Idea of the proof in R2 : We are done if we find for an iteration of f

a Smale’s horseshoe, cf. Figure 2.14. W.l.o.g. let the saddle p be in the
origin. By the Hartman-Grobman theorem, cf. Theorem 2.3.8, the saddle
(x, y) = (0, 0) has a neighborhood in which after some change of coordinates
the dynamics is given by
xn+1 = λxn and yn+1 = μyn
with |μ| > 1 > |λ|. W.l.o.g. we can assume μ, λ > 0. If this is not the case
we consider the second iteration f 2 . Then we consider
S = {(x, y) ∈ R2 : 0 ≤ x ≤ δ, |y| ≤ δ}
for δ > 0 sufficiently small. The k th iteration of f applied to S for k suffi-
ciently large looks as sketched in Figure 4.6. Hence, we found a horseshoe
in case of a homoclinic transversal point. 

This idea can be applied to 2π-time-periodic systems by considering the

time 2π-map Πε . We derived from a 2d-dimensional Hamiltonian system the
2(d − 1)-dimensional 2π-time-periodic Hamiltonian system (4.5)-(4.6). The
d-dimensional tori break up and periodic solutions occur which are fixed
points for the associated time 2π-map Πε . For a variety of systems numer-
ical experiments indicate a transversal intersection of the associated stable
and unstable manifolds and the occurrence of chaotic behavior between the
persisting invariant tori, cf. [Wig03].
Remark 4.5.2. For time-periodic perturbations of an autonomous system
with a homoclinic orbit the occurrence of a transversal intersection of the
130 4. Hamiltonian dynamics

f k (S)

Figure 4.6. Smale’s horseshoe in case of a homoclinic transversal

point. The light gray rectangle is the set S and the dark gray set
is f k (S).

stable and unstable manifolds can be established by finding single zeroes of

the associated so called Melnikov function, cf. [GH83, §4.5].

Exercises
4.1. Prove that ẏ = (traceM )y for y = detY , where Y (t) ∈ R2×2 satisfies Ẏ = M Y
for M = M (t) ∈ R2×2 .
4.2. The “6 − 12 Lennart-Jones potential” models the forces between two neutral
particles (atoms or molecules), namely an attractive van der Waals force at long
ranges and a repulsive force at short ranges due to overlapping electron orbitals.
In a simple (dimensionless) form it is given by F (u) = au−12 − bu−6 where u is
the distance between the particles and a, b > 0 are suitable constants. Choose
a = 0.001 and b = 1 and discuss the phase portrait of the system ü = −F  (u).
4.3. Consider the pair ẍ + ω 2 x = 0, ÿ + μ2 y = 0 of (uncoupled) harmonic oscilla-
tors. Write this as a Hamiltonian system. Find two integrals in polar coordinates.
Discuss the cases (i) ω/μ rational and (ii) ω/μ irrational.
4.4. Given r, μ > 0, write down explicitly a circular solution of the 1-body problem
q̈ = −μq/q3 , q ∈ R2 , i.e., find initial conditions q0 , q̇0 such that the solution
satisfies q(t) = r for all t ∈ R.
4.5. In dimensionless form, the first “Post-Newtonian” approximation for the orbit
of a planet around the sun is
∂θ2 u + u = α + εu2 ,
where u = 1/r and (r, θ) are the polar coordinates of the planet and α, ε > 0 are
parameters. Discuss the phase portrait of this system.
4.6. Let M ∈ Rn×n be nonsingular and symmetric and F : Rn → R be smooth.
Write the Newtonian equation M ẍ + ∇F (x) = 0 as a Hamiltonian system.
4.7. Write the 4th order ODE u + qu + f (u) = 0 as a Hamiltonian system
for (u, u , u , u ). Hint. Let z=(u, u ) and derive a system T z  +∇V (z)=0 with
non-singular T ∈R2×2 .
4.5. Homoclinic chaos 131

 
d q 1
4.8. Consider the 1-body problem = J∇H(q, p), H(q, p) = U (q) + |p|2 .
dt p 2
Show that the angular momentum C = q × q̇ is constant.
4.9. (The 2-body problem) Consider two mass points with positions qj ∈ R3
and masses mj that move under mutual gravitational attraction. The equations
are
m1 q̈1 = F21 , m2 q̈2 = F12 ,
with
Fij = mi mj g(|qi − qj |)(qi − qj ), g(r) = G/r 3 .
This problem can be completely reduced to the 1 body problem. For this consider
the center of mass qs = (m1 q1 + m2 q2 )/ms , with ms = m1 + m2 . Find the ODE
for qs and express the orbits q1,2 via qs and orbits of the 1 body problem for the
distance q = q2 − q1 .
4.10. Let F, G, H : R2n → R be smooth. Show that (a)
{F, {G, H}} + {G, {H, F }} + {H, {F, G}} = 0.
(b) F is an integral of u̇ = J∇H iff {F, H} = 0. (c) d
dt F (u) = {F, H}.
 φ). Show that the map induced by φ = ∂ F and I = ∂φ F is
4.11. Let F = F (I, I
symplectic.
 Chapter 5

PDEs on an interval

The second part of this book is about nonlinear dynamics in countably many
dimensions. It contains this chapter about PDEs on an interval and Chapter
6 about the Navier-Stokes equations.
We start with ordinary differential equations in R∞ , where R∞ stands
for the spaces RN or RZ , i.e., for the spaces of real or complex (identifying C
with R2 ) sequences (aj )j∈N or (aj )j∈Z . In this book these countably many
ODEs arise from PDEs, for which the spatial variable lives in a bounded
domain. For function spaces on such domains very often a countable basis
exists. By an expansion of the PDE w.r.t. this basis, for instance by an
expansion into Fourier series in case of rectangular domains and suitable
boundary conditions, the PDE can be transformed into an ODE in R∞
Example. Consider the linear heat equation ∂t u = ∂x2 u for x∈[0, π]
with boundary condition u(0, t)=u(π, t)=0. Expanding


u(x, t)= k (t) sin(kx),
u
k∈N

the PDE is formally equivalent to the countably many (uncoupled) ODEs

d
k = − k 2 u
dt u k for the sequence of Fourier coefficients (
uk )k∈N .
There are major differences between finitely and infinitely many dimen-
sions due to the non-equivalence of norms in infinite-dimensional spaces
and due to the loss of compactness of bounded closed sets. As a conse-
quence, in infinite dimensions there can be stability w.r.t. one norm, but
instability w.r.t. another norm. On the other hand there is a large class of
equations, namely dissipative systems with smoothing properties, where the
choice of the phase space does not matter. We however do not aim at a

133
134 5. PDEs on an interval

complete functional analytic treatment of such PDEs posed on bounded do-

mains with dynamical systems concepts. For this we refer to the textbooks
[Hen81, Hal88, Rob01].
Most dynamical system questions addressed in this part will be more
involved when one considers PDEs on unbounded domains, as we do in Parts
III and IV of this book. Hence, one of the main purposes of this section is
to prepare for the additional difficulties as they appear in transferring the
dynamical systems concept to PDEs posed on unbounded domains.
In §5.1 we consider the non-equivalence of norms, the loss and regain of
compactness, and the local existence and uniqueness theory for countable
many linear and nonlinear differential equations in R∞ . In §5.2 we discuss
a number of basic function spaces, in particular those that are isomorphic
via Fourier series to some sequence spaces. We explain local existence and
uniqueness of solutions for some prototype linear and nonlinear PDEs, most
of which will later be considered also over unbounded domains, and, more-
over, explain how to prove global existence results. For these, the main
tools are energy estimates and Gronwall type inequalities. We also give a
characterization of the attractor of the so called Chafee-Infante problem,
the scalar equation ∂t u = ∂x2 u + αu − u3 on an interval (0, π) with Dirichlet
boundary conditions u|x=0,π = 0, where α ∈ R is a parameter.

5.1. From finitely to infinitely many dimensions

We consider systems of countably many linear and nonlinear differential
equations. We discuss continuity of solutions w.r.t. time and some abstract
local existence and uniqueness theory for ODEs in R∞ . Moreover, we explain
how to differentiate and integrate in spaces of infinitely many dimensions
and very briefly recall some basic facts from functional analysis, in particular
compactness, which plays a crucial role for the dynamical systems point of
view for PDEs

5.1.1. Non-equivalent norms. Concepts such as convergence in Rd or

stability and instability for ODEs in Rd are independent of the chosen norm
in Rd . The reason for this is the equivalence of norms in finite-dimensional
vector spaces, cf. Theorem 2.1.1. Setting u = (u1 , . . . , ud ), examples for
norms in Rd have been
⎛ ⎞1/2

d d
u1 = |uj |, u2 = ⎝ |uj |2 ⎠ ,
j=1 j=1
5.1. From finitely to infinitely many dimensions 135


and more generally up := ( dj=1 |uj |p )1/p , 1 ≤ p ≤ ∞, and finally u∞ =
maxj=1,...,d |uj |. We have for instance
u∞ ≤ up ≤ d1/p u∞ .
In infinite dimensions there are infinitely many non-equivalent norms. The
norms which we use in this section are as follows.
Definition 5.1.1. For p ∈ [1, ∞) and θ ∈ R let
 1/p

up,θ (RZ ) = |un |p max(1, |n|)pθ .
n∈Z
For p = ∞ and θ ∈ R let
u∞,θ (RZ ) = sup |un | max(1, |n|)θ .
n∈Z
We set
p,θ (RZ ) = {u : Z → R : up,θ (RZ ) < ∞}.

Similarly, we define ·p,θ (RN ) and p,θ (RN ). We use the abbreviations
·p,θ and p,θ for ·p,θ (R∞ ) and p,θ (R∞ ).

The norms for different p or different θ are not equivalent. As a conse-

quence a sequence can converge in one norm towards 0 while it diverges to
∞ in another norm.
Example 5.1.2. For the sequence (um )m∈N , with um ∈ 1,2 (RN ) for fixed
m defined through umn = δnm /n, we have u 1,0 = 1/m → 0 for m → ∞,
m

while u 1,2 = m → ∞ for m → ∞.

m
Remark 5.1.3. The spaces p,θ are Banach spaces, i.e., complete normed
vector spaces. We recall that a metric space M is called complete, if every
Cauchy sequence in M possesses a limit in M . The spaces 2,θ are Hilbert
spaces, i.e., complete normed vector spaces where the norm is induced by a
scalar product. The space c00 = {u : Z → R : un = 0 for finitely many n}
equipped with the 1 norm is not complete. See Exercise 5.1.

5.1.2. Linear differential equations in R∞ . For notational simplicity in

the following we work with equations in RN . The results for RZ are exactly
the same. We consider linear differential equations
d d 
u = Au, i.e., uk = akj uj .
dt dt
j∈N

We briefly recall the basic notions of semigroup theory which is the abstract
version of the subsequent analysis.
136 5. PDEs on an interval

We are not interested in such equations in greatest generality and there-

fore restrict ourselves mainly to equations having to do with PDEs, i.e., we
consider A in diagonal form or with Jordan blocks of finite size. In this
situation the equation can be solved explicitly, but the analytic properties
of the solutions still turn out to be rather subtle. In case of A in diagonal
form
d
uk = λk uk
dt
we find the solutions
uk (t) = eλk t uk (0).
Solutions u(t) to linear differential equations in Rd with constant coefficients
are arbitrarily smooth w.r.t. t. In RN this is no longer true. Even for the
boundedness additional conditions are necessary.
Lemma 5.1.4. Let supk∈N Reλk = α < ∞. Then for every θ ∈ R, T0 > 0,
and p ∈ [1, ∞] the curve t → u(t) is bounded in p,θ for t ∈ [0, T0 ].

Proof. We have u(t)p,θ ≤ (supk∈N |eλk t |)u(0)p,θ ≤ eαt u(0)p,θ . 

The next question is the continuity of the curve t → u(t) in the spaces
p,θ . Which conditions do we have to impose on the eigenvalues λk to have
continuity? We put this question into a bigger framework, namely the theory
of semigroups. The solution operator T (t) = diag(eλ1 t , eλ2 t , . . .) defined
through T (t)u(0) = u(t) is an example of a semigroup of bounded linear
operators, here from p,θ to p,θ .
Definition 5.1.5. Let (X,  · ) be a Banach space. A one-parameter family
T (t), 0 ≤ t < ∞, of bounded linear operators from X to X is called semi-
group of bounded linear operators on X, if
(i) T (0) = I,
(ii) T (t + s) = T (t)T (s) for all t, s ≥ 0.
The linear operator A : D(A) → X, defined by
T (t)u − u
D(A) = {u ∈ X : lim exists },
t↓0 t
T (t)u − u
Au = lim , for u ∈ D(A),
t↓0 t
is called the infinitesimal generator of T (t).

According to the semigroup property, for the continuity of the maps

t → T (t) or t → T (t)u the continuity at t = 0 is sufficient, cf. Remark
5.1.11. There are different concepts of continuity for semigroups. The first
one is as follows.
5.1. From finitely to infinitely many dimensions 137

Definition 5.1.6. The semigroup is called operator-continuous if

lim T (t) − I = 0
t↓0

where  ·  denotes the operator norm.

Example 5.1.7. Consider u̇ = Au with u(t) ∈ Rd and A a d × d-matrix

with constant coefficients. Then T (t) = etA defines an operator-continuous
semigroup in Rd since
∞

etA − I =  (At)k /k! ≤ |t|Ae|t|A → 0 for t → 0.
k=1

Theorem 5.1.8. Let supk∈N |λk | = α < ∞. Then for every θ ∈ R and
p ∈ [1, ∞] the associated semigroup is operator-continuous in p,θ for all
t ∈ R.

Proof. We have u(t) − u(0)p,θ ≤ (supk∈N |eλk t − 1|)u(0)p,θ ≤ |eαt −

1|u(0)p,θ → 0 for t → 0. This implies T (t) − Ip,θ →p,θ ≤ |eαt − 1| → 0
for t → 0 and so continuity holds. 
For completeness we remark that a semigroup T (t) of bounded linear
operators on X is operator-continuous if and only if the generator A : X →
X is bounded, cf. [Paz83, §1, Theorem 1.2]. Hence, as seen in the above
example, the solutions of finite-dimensional ODEs always define an operator-
continuous semigroup.
Since linearized operators in PDEs are usually unbounded, the generated
semigroups are in general only strongly continuous.

Definition 5.1.9. A semigroup T (t), 0 ≤ t < ∞, of bounded linear opera-

tors in X is called strongly continuous semigroup, or C0 -semigroup, if

lim T (t)u − u = 0 for each u ∈ X.

t↓0

Theorem 5.1.10. For every θ ∈ R, p ∈ [1, ∞), and u(0) ∈ p,θ , the curve
t → u(t) is continuous in p,θ for t ≥ 0 if and only if supj∈N Reλj = α < ∞.

Proof. Let ε > 0. Using the triangle inequality in p,θ we have that

u(t) − u(0)p,θ

N ∞

=( |(e λn t
− 1)un (0)| |n| )
p pθ 1/p
+( |(eλn t − 1)un (0)|p |n|pθ )1/p
n=1 n=N +1
= s1 + s2
138 5. PDEs on an interval

for a N ∈ N suitably chosen in the following. In order to prove that s1 +s2 <
ε for t > 0 sufficiently small we first estimate s2 by choosing N so big that
∞
s2 ≤ (eαt + 1)( |un (0)|p |n|pθ )1/p < ε/2.
n=N +1

For this N we then find a t0 > 0 such that for all t ∈ (0, t0 ) we can estimate
s1 ≤ ( max |eλn t − 1|)u(0)p,θ < ε/2.
n=1,...,N

Therefore, we are done. 

Remark 5.1.11. Since T (t0 + h)u−T (t0 )u = (T (h)−I)T (t0 )u → 0 and
T (t0 −h)u−T (t0 )u = −T (t0 −h)(T (h)−I)u → 0 for h ↓ 0 the right-continuity
in t0 = 0 implies the continuity in every t0 > 0 if the semigroup is uniformly
bounded on every compact interval, cf. Lemma 5.1.4. In fact the assumption
of the uniform boundedness on every compact interval is satisfied for C0 -
semigroups due to a deep result from functional analysis, namely the uniform
boundedness principle, cf. [Paz83, §1.2, Theorem 2.2].

For u̇ = Au, with A in diagonal form, for p ∈ [1, ∞) every bounded

trajectory is also continuous in t. However, there is no uniformity w.r.t. the
initial conditions u(0). In ∞,θ the assumption about the boundedness of
the eigenvalues λj is also necessary for continuity as the following example
shows.
Example 5.1.12. Let λk = −k 2 . Then there exists an u(0) ∈ ∞,0 , for
instance u(0) = (1, 1, 1, . . .), such that u(t) − u(0)∞,0 = 1 for every t > 0
such that continuity cannot hold.

Solutions of ODEs in Rd are smooth if the data is smooth. In R∞ for

u(t) = (eλn t un (0))n∈N additional conditions about the eigenvalues λn are
necessary.
(m)
• The mth derivative u(m) (t) is given by un (t) = λm
ne
λn t u (0). For
n
u(0) ∈ p,θ we can guarantee the m-times differentiability of t →
u(t) in p,θ if the eigenvalues are in a set
{λ ∈ C : tReλ ≤ a − m ln |Imλ|}
for some constant a ∈ R, cf. [Paz83, §2.4, Theorem 4.8]. As an
example we consider λn = − ln n+in. In order to have the m-times
differentiability we need that u(m) (t) ∈ p,θ which follows if
sup |(− ln n + in)m e−(ln n)t | = sup |(− ln n + in)m n−t | < ∞.
n∈N n∈N

This means that the curve is one time differentiable for t ∈ (1, 2],
two times differentiable for t ∈ (2, 3], etc.
5.1. From finitely to infinitely many dimensions 139

• For u(0) ∈ p,θ we have the analyticity of t → u(t) if the eigenvalues

are in a sector
Sa,b = {λ ∈ C : Reλ ≤ a − b ln |Imλ|}
for some constants a ∈ R and b ≥ 0, cf. [Paz83, §2.5, Theorem
5.2]. As an example we consider λn = −n + (−1)n in. The function
t → eλn t un (0) can be extended analytically into a sector of the
complex plane. For t = tr + iti we find u(tr + iti ) ∈ p,θ if
| ≤ sup |en(−tr +|ti |) | < ∞
n in)(t
sup |e(−n+(−1) r +iti )

n∈N n∈N
which holds if |ti | < tr . Such generators are called sectorial and play
a major role in the analysis of dissipative systems. The associated
semigroup (eλn t )n∈N is called analytic. We come back to this in
§6.3.

Figure 5.1. The picture shows the regions where the spectrum of the
generators must be contained in to have a continuous (left panel), a
differentiable (middle panel), or an analytic (right panel) semigroup.

We refer to the textbook [Paz83] for a thorough introduction to semi-

group theory. Generators of C0 -semigroups are characterized by the theorem
of Hille-Yosida, cf. [Paz83, §1.3, Theorem 3.1] or the Lumer-Phillips theo-
rem, cf. [Paz83, §1.4, Theorem 4.3].

5.1.3. Differentiation and integration in Banach spaces. Before we

proceed with the consideration of nonlinear infinite-dimensional ODEs, we
need some additional functional analytic tools. For the stability of fixed
points in ODEs the linearization, i.e., the derivative A = Df ∈ Rd×d of a
vector field f : Rd → Rd plays a central role. Hence, the concept of deriva-
tives has to be generalized from Rd to general Banach spaces X. The same
is true for integration. In the iteration scheme used in the proof of the local
existence and uniqueness theorem for ODEs, a continuous function on an
interval with values in Rd is integrated. If this iteration scheme is trans-
ferred to PDEs, then a continuous function on an interval with values in
some infinite-dimensional Banach space has to be integrated. Such integra-
tions occur in other iteration schemes used for PDEs, too. Hence, we have
140 5. PDEs on an interval

to define the integral of a continuous function on an interval with values in

some Banach space. It turns out that the usual definition with Riemann
sums is sufficient for our purposes.
Let us start with the derivatives. The Gateaux derivative is a general-
ization of the concept of the directional derivative. In Banach spaces the
derivative is also called Fréchet derivative, cf. [AA11].
Definition 5.1.13. Suppose that X and Y are Banach spaces, that U ⊂ X
is open, and consider F : X → Y . The Gâteaux derivative DF (u)[v] of F
at u ∈ U in the direction v ∈ X is defined as

F (u + τ v) − F (u) d 
DF (u)[v] = lim = F (u + τ v) .
τ →0 τ dτ τ =0
If the limit exists for all v ∈ X, then F is called Gâteaux differentiable at u.
F : X → Y is called differentiable in u ∈ U if there exists a bounded linear
operator A = A(u) : X → Y such that
F (u + h) − F (u) − A(u)hY
lim = 0.
h→0 hX

The function f : R2 → R with

1, if y = x2 , x = 0
f (x) =
0, elsewhere
is a finite-dimensional example of a function for which every directional
derivative exists, but which is not differentiable. In infinite-dimensional
spaces less ’exotic’ examples are possible.
Example 5.1.14. Consider X = Y = L2 (0, 1) and F (u)(x) = sin(u(x)).
We show that F is Gâteaux differentiable, but not differentiable at u = 0.
We have
F (u + τ v) − F (u) sin(τ v(x))
lim = lim = cos(0)v(x) = v(x)
τ →0 τ τ →0 τ
due to the differentiability of sin : R → R. For the Fréchet differentiability
we can vary v not only along lines. Due to the above computed Gâteaux
derivative the only possible candidate for the derivative A(0) is the identity.
We set

nπ, if x ∈ (0, 1/n4 )

vn (x) =
0, elsewhere
and find
F (vn ) − F (0) − A(0)vn Y vn Y
= = 1 → 0
vn X vn X
although vn L2 = π/n → 0 for n → ∞. We remark that with the choice
X = Y = Cb0 ([0, 1]) equipped with the sup-norm the map F would be
analytic.
5.1. From finitely to infinitely many dimensions 141

However, in spaces with additional algebra properties, virtually all re-

sults from complex power series carry over.
Example 5.1.15. Let (X, ·) be a Banach space. Then Y = L(X, X) =
{F : X → X : linear, continuous} is a Banach space equipped with the
operator norm
A = sup{AuX : uX = 1}.
For A, B ∈ Y we have that AB and An are in Y with AB≤ AB,
and An  ≤ An . Therefore, the series F (A) = exp(A) = ∞ 1 k
k=0 k! A is
convergent in Y . We have the continuity and the Fréchet differentiability
of F : Y → Y with DF (A) = F (A). The analyticity of F follows like for
real-valued power series, cf. Exercise 5.3.

Next we come to the integration of continuous functions f : [a, b] → X

with values in a Banach space X. Let P = {x0 , . . . , xn } with a = x0 < x1 <
. . . < xn−1 < xn = b be a partition of the interval. Its fineness is defined by
P  = max{|xj+1 − xj | : j = 0, . . . , n − 1}.
Let ξ = (ξ1 , . . . , ξn ) with ξj ∈ [xj−1 , xj ]. Then define the Riemann sum

n
S(ξ, P ) = f (ξj )(xj − xj−1 ).
j=1

Definition 5.1.16. A function f : [a, b] → X is called Riemann integrable

if the limit
lim S(ξ(n), P (n))
n→∞
exists for every sequence (ξ(n), P (n)) with limn→∞ P (n) = 0. If the limit
exists, then we define the Riemann integral by
 b
f (x) dx = lim S(ξ, P ).
a P →0

Theorem 5.1.17. Continuous functions f : [a, b] → X are Riemann inte-

grable.

Proof. We have to show that for every ε > 0 there exists a δ > 0 such that
for all (ξ 1 , P1 ), (ξ 2 , P2 ) with P1  ≤ δ and P2  ≤ δ we have S(ξ 1 , P1 ) −
S(ξ 2 , P2 )X ≤ ε.
We set P3 = P1 ∪ P2 and choose an arbitrary ξ 3 . Then by the triangle
inequality we have
S(ξ 1 , P1 )−S(ξ 2 , P2 )X ≤S(ξ 1 , P1 )−S(ξ 3 , P3 )X + S(ξ 3 , P3 −S(ξ 2 , P2 )X
≤ε/2 + ε/2 ≤ ε.
142 5. PDEs on an interval

where we used

N1 
N3
S(ξ , P1 ) − S(ξ , P3 )X =
1 3
f (ξj1 )(
xj −x
j−1 ) − f (ξj3 )(xj − xj−1 )X
j=1 j=1


N1 
α(j)−1
 1 
(5.1) ≤ f (ξj ) − f (ξ 3  |xk+1 − xk | ,
k+1 ) X
j=1 k=α(j−1)
%α(j)−1
xj−1 , x
where [ j ] = k=α(j−1) [xk , xk+1 ]. See Figure 5.2.

x0 x x2 x3 x x5 x6 x7
1 4

~
x0 ~
x ~x
1 2

Figure 5.2. The partition P1 is drawn below the line and P3 above the
line. In this example we have α(0) = 0, α(1) = 3, α(2) = 7, . . ..

By uniform continuity of f , which follows from the continuity of f on the

compact interval [a, b], we have that for all ε > 0 there exists a δ > 0 such
that |y − y| < δ implies f (y) − f ( ε
y )X < 2(b−a) . Hence, if P3  ≤ P1  < δ
is chosen sufficiently small, (5.1) can be estimated by

N3
ε ε
≤ |xk − xk−1 | ≤ . 
2(b − a) 2
k=1
Remark 5.1.18. Not only the Riemann integral can be generalized to func-
tions u : R → X, with X some Banach space, but also the Lebesgue integral,
cf. [Alt16, §A1].

5.1.4. Nonlinear differential equations in R∞ . Since in general the

ODEs in R∞ obtained from PDEs have unbounded λk s the right-hand side
is no longer Lipschitz-continuous from p,θ to p,θ . Thus, the Picard-Lindelöf
theorem no longer applies and has to be replaced. The simplest idea to
obtain a contraction as in the proof of the Picard-Lindelöf theorem is the
use of the variation of constant formula, cf. §2.1.3.
For simplicity we first consider
d
(5.2) u = Λu + B(u, u),
dt
where u(t) ∈ p,θ , where Λ is a diagonal matrix with entries λk satisfying
(5.3) sup Reλk = β < ∞,
k
and where B is a bilinear symmetric map from p,θ into p,θ satisfying
(5.4) B(u, v)p,θ ≤ CB up,θ vp,θ .
5.1. From finitely to infinitely many dimensions 143

In order to prove the local existence and uniqueness of solutions of (5.2) on

an interval [0, T0 ] we use the variation of constant formula to rewrite (5.2)
into
 t
(5.5) u(t) = etΛ u(0) + e(t−τ )Λ B(u(τ ), u(τ )) dτ =: F (u)(t).
0

Definition 5.1.19. a) A function u ∈ C 0 ([0, T0 ], p,θ ) which satisfies (5.5)

is called a mild solution of (5.2).
b) A function u ∈ C 1 ([0, T0 ], p,θ ), with Λu ∈ C([0, T0 ], p,θ ), is called a
strong solution of (5.2), if (5.2) holds in p,θ for every t ∈ (0, T0 ).

Clearly, every strong solution is a mild solution. Conversely, a mild

solution which satisfies u ∈ C 1 ([0, T0 ], p,θ ) and Λu ∈ C([0, T0 ], p,θ ) is a
strong solution. In the following until further notice solutions will always
mean mild solutions.

Theorem 5.1.20. Assume (5.3) and (5.4). For all C1 > 0 there exists a
T0 > 0 such that for all w ∈ p,θ with wp,θ ≤ C1 we have a unique solution
u ∈ C([0, T0 ], p,θ ) of (5.2) with initial condition u(0) = w.

Proof. We fix a C2 > 0 and show that for T0 ∈ (0, 1) sufficiently small the
right-hand side of (5.5) is a contraction in the set

M = C([0, T0 ], {u(t) ∈ p,θ : u(t) − etΛ u(0)p,θ ≤ C2 }),

and apply the contraction mapping theorem. M is a complete metric space,

but since the metric is induced by a norm we will use the norm notation in
the following. We use the abbreviation C3 = supu∈M uM ≤ C1 eβ + C2 .
In a first step we prove that F maps M into itself. We have

F (u) − (etΛ u(0))t≥0 M = sup F (u)(t) − eΛt u(0)p,θ

t∈[0,T0 ]
 t
≤ sup  e(t−τ )Λ B(u(τ ), u(τ )) dτ p,θ
t∈[0,T0 ] 0
 t
≤ sup eβ(t−τ ) B(u(τ ), u(τ ))p,θ dτ
t∈[0,T0 ] 0

≤T0 eβT0 CB C32 ≤ C1

for T0 > 0 sufficiently small.

144 5. PDEs on an interval

Secondly, we find that F is a contraction since

F (u) − F (v)M = sup F (u)(t) − F (v)(t)p,θ
t∈[0,T0 ]
 t
≤ sup  e(t−τ )Λ (B(u(τ ), u(τ )) − B(v(τ ), v(τ ))) dτ p,θ
t∈[0,T0 ] 0
 t
≤ sup eβ(t−τ ) B(u(τ ), u(τ )) − B(v(τ ), v(τ ))p,θ dτ
t∈[0,T0 ] 0

≤ T0 eβT0 sup B(u(τ ) + v(τ ), u(τ ) − v(τ ))p,θ

τ ∈[0,T0 ]

≤ 2T0 eβT0 CB C3 u − vM ≤ u − vM /2

for T0 > 0 sufficiently small. 
This procedure of constructing solutions to ODEs in R∞ can be extended
to a wider class of problems. We consider again
d
(5.6) u = Λu + B(u, u),
dt
where u(t) ∈ p,θ , but now with the following assumptions:
• Λ a diagonal matrix satisfying
(5.7) etΛ up,θ ≤ Cθ−r eβt t−α up,r
for an α ∈ [0, 1), a constant Cθ−r , and θ − r ≥ 0.
• B a bilinear symmetric map from p,θ into p,r satisfying
(5.8) B(u, v)p,r ≤ CB up,θ vp,θ .
The property described by equation (5.7) is called smoothing since the evo-
lution operator maps for t > 0 the space p,r into p,θ and since functions
whose Fourier coefficients are in p,θ are smoother than functions whose
Fourier coefficients are only p,r . Many of the subsequent examples will
satisfy estimates like (5.7).
Example 5.1.21. Consider λn = −n2 . We have the decay estimate
(eλn t un )n∈N p,θ ≤ sup |e−n t nθ |up,0 ≤ Ct−θ/2 up,0 ,
2

n∈N

which corresponds to smoothing of functions in physical space, see Example

5.2.19.
Remark 5.1.22. Smoothing is not directly related to regularity w.r.t. time
t, as the following examples show. In case λn = 0 for all n ∈ N all eigenvalues
are identical and contained in a sector. However, the associated semigroup
is the identity which is not smoothing from p,r into p,θ for r < θ although
we have an analytic semigroup (w.r.t. time). In case λn = −n2 + i(−1)n n3
5.1. From finitely to infinitely many dimensions 145

obviously the eigenvalues are not contained in a sector and the semigroup is
not analytic (w.r.t. time t = tr + iti ) since
| = sup |e−n
2 +i(−1)n n3 )(t 2t n+1 n3 t
sup |e(−n r +iti ) r +(−1) i
|=∞
n∈N n∈N
for ti = 0. However, we have the decay estimate
2 ±in3 )t
(eλn t un )n∈N p,θ ≤ sup |e(−n nθ |up,0 ≤ Ct−θ/2 up,0 .
n∈N

In order to prove the local existence and uniqueness of solutions of (5.6)

on an interval [0, T0 ], we again use the variation of constant formula
 t
tΛ
(5.9) u(t) = e u(0) + e(t−τ )Λ B(u(τ ), u(τ )) dτ =: F (u)(t).
0
We show that for T0 ∈ (0, 1) sufficiently small the right-hand side of (5.9) is
a contraction in the set
M = C([0, T0 ], {u(t) ∈ p,θ : u(t) − etΛ u(0)p,θ ≤ C2 }),
where u(0)p,θ ≤ C1 and C2 > 0 is a fixed constant. In a first step we
prove that F maps M into itself. With C3 as above we have
F (u)−(etΛ u(0))t≥0 M = sup F (u)(t) − etΛ u(0)p,θ
t∈[0,T0 ]
 t
≤ sup  e(t−τ )Λ B(u(τ ), u(τ )) dτ p,θ
t∈[0,T0 ] 0
 t
≤ sup Cθ−r (t − τ )−α eβ(t−τ ) B(u(τ ), u(τ ))p,r dτ
t∈[0,T0 ] 0

≤ Cθ−r (1 − α)−1 T01−α eβT0 CB C32 ≤ C1

for T0 > 0 sufficiently small.
Secondly, we find
F (u) − F (v)M = sup F (u)(t) − F (v)(t)p,θ
t∈[0,T0 ]
 t
≤ sup  e(t−τ )Λ (B(u(τ ), u(τ )) − B(v(τ ), v(τ ))) dτ p,θ
t∈[0,T0 ] 0
 t
≤ sup Cθ−r (t − τ )−α eβ(t−τ ) B(u(τ ), u(τ )) − B(v(τ ), v(τ ))p,r dτ
t∈[0,T0 ] 0

≤ Cθ−r (1 − α)−1 T01−α eβT0 sup B(u(τ ) + v(τ ), u(τ ) − v(τ ))p,θ
t∈[0,T0 ]

≤ 2Cθ−r (1 − α)−1 T01−α eβT0 CB C3 u − vM ≤ u − vM /2

for T0 > 0 sufficiently small. Hence, the contraction F possesses a unique
fixed point in M and so we have proved
146 5. PDEs on an interval

Theorem 5.1.23. Assume (5.7) and (5.8). For all C1 > 0 there exists a
T0 > 0 such that for all w ∈ p,θ with wp,θ ≤ C1 we have a unique solution
u ∈ C([0, T0 ], p,θ ) of (5.6) with initial condition u(0) = w.
Remark 5.1.24. Both theorems obviously also hold if B is replaced by a
general locally Lipschitz-continuous map from p,θ into p,θ or p,r respec-
tively, i.e., for instance in the latter case that for all C1 there exists an L
such that max(up,θ , vp,θ ) ≤ C1 implies
N (u) − N (v)p,r ≤ Lu − vp,θ .
Every polynomial nonlinearity is locally Lipschitz-continuous in this sense.

Moreover, Theorem 5.1.20 and Theorem 5.1.23 are prototypes for other
local existence and uniqueness theorems for semi-linear evolutionary PDEs
below.

5.1.5. A first look on Fourier series. PDEs with periodic boundary

conditions for the spatial coordinates can be transferred to ODEs in R∞ with
the help of Fourier series. Here we give the definition and some elementary
properties. For later purposes we consider here the d-dimensional situation.
Definition 5.1.25. A series of the form

(5.10) u(x) = k eik·x ,
u
k∈Zd

is called Fourier series, its partial sums u(x) = |k|≤N u k eik·x , are called
Fourier polynomials of order N , and u k is called the k Fourier coefficient.
th

See §5.2.2 for more details, in particular a number of convergence results

for (5.10). Since our main interest is in nonlinear PDEs we also have to
handle products of functions in physical space with Fourier series. The
point-wise multiplication in physical space correspond in Fourier space to
convolution. That is,
⎛ ⎞⎛ ⎞ ⎛ ⎞
   
u(x)v(x) = ⎝ k eik·x ⎠ ⎝
u vm eim·x ⎠ = ⎝ k−m vm ⎠ eik·x .
u
k∈Zd m∈Zd k∈Zd m∈Zd

This motivates the definition of the convolution


(u ∗ v)k = k−m vm .
u
m∈Zd
For the control of the nonlinear terms in Fourier space we need
Lemma 5.1.26. (Young’s inequality for convolutions) For p ∈ [1, ∞]
we have
u ∗ vp ≤ 
 up 
v  1 .
5.1. From finitely to infinitely many dimensions 147

Proof. For p ∈ [1, ∞] we find

   

u ∗ vp =( | k−m vm |p )1/p ≤ ( (
u |
ul vm |)p )1/p
k∈Z m∈Z l∈Z m∈Z
 
=( |
ul | (
p
|
vm |) )
p 1/p
≤ 
up 
v  1 .
l∈Z m∈Z

The case p = ∞ is obvious. 

Young’s inequality for convolutions allows us to prove that the 2,θ -spaces
are closed under convolution if θ is sufficiently big. In order to do so we prove
the following version of Sobolev’s embedding theorem
Lemma 5.1.27. For m − d/2 > n there exists a C > 0, such that
u1,n (Rd ) ≤ C
 u2,m (Rd ) .

Proof. With ρk = max(1, |k|) the estimate follows from

  (n−m)

u1,n = |
uk |
ρnk = |
uk |
ρm k
k ρ
k∈Zd k∈Zd
⎛ ⎞1/2 ⎛ ⎞1/2
 
≤⎝ ⎠ ⎝ 2(n−m) ⎠
|
uk |2 ρ2m
k ρk ≤ C
u2,m ,
k∈Zd k∈Zd
 2(n−m)
since k∈Zd ρk < ∞, due to m − d/2 > n by assumption. 
We use this embedding to establish
Lemma 5.1.28. a) For all m ≥ 0 there exists a C > 0, such that for all
, v ∈ 1,m we have
u
u ∗ v1,m ≤ C
 u1,m 
v 1,m .
, v ∈ 2,m we
b) For all m > d/2 there exists a C > 0, such that for all u
have
u ∗ v2,m ≤ C
 u2,m 
v 2,m .

k ≤ C(
Proof. a) Since ρm ρm
k−l + ρ m m
l ) with C = 2 for ρ k = max(1, |k|) using
Lemma 5.1.26 it follows that
   

u ∗ v1,m = | k−l vl ρm
u k |≤ |
uk−l vl ρm
k |
k∈Zd l∈Zd k∈Zd l∈Zd
 
≤C uk vl |
(| ρm
k + |
uk vl |
ρm
l )
k∈Zd l∈Zd
u1,0 
≤C( v 1,m + 
u1,m 
v 1,0 ) ≤ 2C
u1,m 
v 1,m .
148 5. PDEs on an interval

b) With Lemma 5.1.26 we have

 

u ∗ v2,m =( | k−l vl ρm
u k | )
2 1/2

k∈Zd l∈Zd
 
≤( ( |
uk−l vl |(
ρm m
k−l + ρ
2 1/2
l )) )
k∈Zd l∈Zd
≤C(|
u|
ρ ∗ |
v |2,0 + |
m
u| ∗ |
v |
ρm 2,0 )
u2,m 
≤C( v 1,0 + 
u1,0 
v 2,m ).

v 1,0 ≤
The final assertion follows from Sobolev’s embedding theorem 
v 2,m for m > d/2.
C 
We now give a number of classical examples of nonlinear PDEs over in-
tervals with periodic boundary conditions. In fact, over unbounded domains
each of these equations will play an important role in this book. For the
modeling and physical background of the equations we refer in particular to
Part III.

Example 5.1.29. Let u(x, t) = u(x+2π, t) ∈ R, A(X, T ) = A(X +2π, T ) ∈

C, and u k (T ) be the associated Fourier coefficients.
k (t), A
a) The Kolmogorov, Petrovsky, Piskounov (KPP) equation ∂t u = ∂x2 u + u −
u2 transforms into

k = −k 2 u
∂t u k + u
k − k−m u
u m .
m∈Z

b) The Allen-Cahn equation ∂t u = ∂x2 u + u − u3 transforms into


k = −k 2 u
∂t u k + u
k − k−m u
u m−l u
l .
m∈Z l∈Z

c) The Burgers equation ∂t u = ∂x2 u + ∂x (u2 ) transforms into


k = −k 2 u
∂t u k + ik uk−m u
m .
m∈Z

d) The Korteweg-deVries (KdV) equation ∂t u = ∂x3 u + ∂x (u2 ) transforms

into

k = −ik 3 u
∂t u k + ik k−m u
u m .
m∈Z

e) Using (F A)j = A −j the Nonlinear Schrödinger (NLS) equation ∂T A =

2 2
i∂X A + i|A| A transforms into

k = −ik 2 A
∂T A k + i k+l−m A
A l A
m .
m∈Z l∈Z
5.1. From finitely to infinitely many dimensions 149

f) The complex Ginzburg-Landau (GL) equation ∂T A = (1+iα)∂X 2 A+RA−

(1 + iβ)|A|2 A, with α, β, R ∈ R, transforms into


k = −(1 + iα)k 2 A
∂T A k + Ak − (1 + iβ) k+l−m A
A l Am .
m∈Z l∈Z

Except for the KdV equation, for all equations from above the local
existence and uniqueness theory can be handled with Theorem 5.1.20 and
Theorem 5.1.23. The linear parts are given by the eigenvalues λk , k ∈ Z,
where
a), b) λk = −k 2 + 1, c) λk = −k 2 ,
d) λk = −ik 3 , e)λk = −ik 2 , f) λk = −(1 + iμ)k 2 + 1.

In a), b), e) and f) we only need that (eλk t )k∈Z : 2,θ → 2,θ is bounded for
fixed t. Since the nonlinear terms in a), b), e) and f) are bi- and trilinear
maps from 2,θ → 2,θ for θ > 1/2 we have the local existence and uniqueness
for these equations in 2,θ for θ > 1/2 according to Theorem 5.1.20. Since
for c) and d) the nonlinear terms are only bilinear maps from p,θ+1 into p,θ
we need an estimate
(5.11) (eλk t )k∈Z p,θ →p,s+1 ≤ C max(1, t−α )
with α ∈ [0, 1) for the semigroup in order to apply our local existence and
uniqueness result from Theorem 5.1.23. According to Example 5.1.21 such
an estimate is true for c) with α = 1/2, but not for d). The KdV equation
is a so called a quasilinear (hyperbolic) equation. There is local existence in
2,θ for θ = 3, for instance. However, the proof is more involved, cf. [Paz83,
§8, Theorem 5.6] or §8.2 for further remarks.

5.1.6. Loss and regain of compactness. We close this section with a

number of comments on compactness, which is a crucial concept to define
attractors in dynamical systems. In metric spaces there are the following
equivalent characterizations of compact sets, cf. [Alt16, §2.5].

Definition 5.1.30. Let (M, d) be a complete metric space.

a) A set A ⊂ M is compact if every covering of A by open sets contains
a finite subcovering.
b) A set A ⊂ M is (sequentially) compact if every sequence in A has a
convergent subsequence with limit in A.
c) A set A ⊂ M is compact if A is closed and pre-compact, where a set
A ⊂ M is said to be pre-compact if for every % ε > 0, there exists a finite
subset {s1 , s2 , . . . , sn } of A such that A ⊂ nk=1 B(sk , ε), where B(sk , ε)
denotes the open ball around sk with radius ε.
150 5. PDEs on an interval

Compactness arguments in the sense of b) were used a number of times

in Part I. Examples are the existence, respectively the non-emptyness, of ω-
limit sets and attractors. For the term ’pre-compact’ also the term ’totally
bounded’ is used in the literature. In Rd compact sets can be characterized
by the theorem of Heine-Borel.
Theorem 5.1.31. In Rd a set is compact if and only if it is closed and
bounded.

In infinite-dimensional spaces compactness is more restrictive due to

the fact that the theorem of Heine-Borel is no longer true as the following
example shows.
Example 5.1.32. Consider the closed unit ball in (RN , ·∞ ). The sequence
(un )n∈N with un = en satisfies un − um ∞ = δnm such that no convergent
subsequence can exist. Variants of this example works in all norms from
above such that the closed unit ball is not compact in any of the norms
from Definition 5.1.1.

The equivalence of compactness to boundedness and closedness is a

precise distinction between finite- and infinite-dimensional Banach spaces
[Alt16, Satz 2.9]. There are famous theorems about the characterization of
pre-compact subsets of function spaces. These are the Arzela-Ascoli theorem
[Alt16, Satz 2.11], the theorem of Riesz [Alt16, Satz 2.15], and Sobolev’s
embedding theorem [Alt16, Satz 8.9].
Compactness in infinite-dimensional spaces will be regained by smooth-
ing properties of the evolution operators. For instance the evolution operator
of Example 5.1.21 maps bounded balls of 2,0 into bounded balls of 2,1 for
every fixed t > 0. Since the subsequent version of Sobolev’s embedding
theorem 5.1.33 guarantees that bounded balls of 2,1 are pre-compact sets
of 2,0 , the evolution operator of Example 5.1.21 maps bounded balls of 2,0
into pre-compact sets of 2,0 . This property will be used for showing that
ω-limit sets and attractors for such systems are non-empty. The following
theorem is also known under the name Rellich’s embedding theorem.
Theorem 5.1.33. The space p,θ can be compactly embedded into the space
p,r for all p ≥ 1 and θ > r.

Proof. For notational simplicity we restrict to the index set N. Compactly

embedded means that every bounded set of p,θ is pre-compact in p,r . Due
to the homogeneity of the spaces it is sufficient to prove that the unit ball of
p,θ can be covered by finitely many balls of p,r with radius ε. In order to
do so we consider the first n0 coordinates. The restriction of the unit ball of
p,θ to these coordinates is a pre-compact set in Rn0 . Hence, for every ε > 0
it can be covered by finitely many balls Bε (zj ) of Rn0 w.r.t. the p,r -norm
5.2. Basic function spaces and Fourier series 151

and with zj ∈ Rn0 for j = 1, . . . , N . We claim that the unit ball of p,θ
is contained in the union of the balls Bε/2 ((zj , 0)) of p,r (RN ) if we choose
(n0 + 1)r−θ ≤ ε/2. This follows since for u = ((uk )k=1,...,n0 , u∞ ) in the unit
ball of p,θ (RN ) we have a j ∈ {1, . . . , N } such that

u−(zj , 0)p,r ≤ ((uk )k=1,...,n0 , 0)−(zj , 0)p,r +(0, u∞ )p,r ≤ ε/2+ε/2 = ε,

where ((uk )k=1,...,n0 , 0) − (zj , 0)p,r ≤ ε/2 due to the construction of the
points zj and where (0, u∞ )p,r ≤ ε/2 due to
∞
 ∞

(0, u∞ )p,r ≤( |un | |n| )
p pr 1/p
≤ sup |n| r−θ
( |un |p |n|pθ )1/p
n0 +1 n=n0 +1,...,∞ n0 +1

≤(n0 + 1) r−θ
(0, u∞ )p,θ ≤ ε/2

for n0 sufficiently large since (0, u∞ )p,θ ≤ 1. 

5.2. Basic function spaces and Fourier series

PDEs posed on spatially bounded domains are very often isomorphic to
ODEs in R∞ . Thus, the abstract set-up from the last section can often be
applied to solve PDEs posed on spatially bounded domains. However, a big
part of PDE theory is concerned with problems coming from the bound-
ary of the considered domains. These play almost no role in this book,
i.e., they are circumvented by considering almost all systems subsequently
with periodic boundary conditions or on the real line. This allows us to
concentrate on phenomena coming from the equations. In other words, a
complete functional analytic treatment of PDEs posed on bounded domains
with dynamical systems concepts is beyond the scope of this book. For this
we refer to the textbooks [Hen81, Hal88, Tem97]. However, in Part IV of
this book some of the methods to handle problems posed on cylindrical do-
mains R × Σ, with Σ ⊂ Rd a bounded domain, are explained. In this section
we concentrate on PDEs where the spatial coordinate lives on a bounded
interval with periodic boundary conditions. Such problems can easily be
related to ODEs in R∞ with the help of Fourier series. These explanations
are embedded in some theoretical background about basic function spaces
and Fourier series.

5.2.1. Basic function spaces. The solution u = u(·, t) of a PDE is for

fixed t in some function space. Here, we introduce some basic function
spaces following [Alt16, Wlo87]. In the following let Ω ⊂ Rd be an open
set, x = (x1 , . . . , xd ) ∈ Ω, n = (n1 , . . . , nd ) a multi-index, |n| = n1 + . . . + nd ,
and ∂xn = ∂xn11 . . . ∂xndd .
152 5. PDEs on an interval

Continuous and differentiable functions. The space of continuous

functions in Ω is
C 0 (Ω, R) = {u : Ω → R : u is continuous},
equipped with the norm
uC 0 = sup |u(x)|.
b
x∈Ω

The space of m-times continuously differentiable functions in Ω is

C m (Ω, R) = {u : Ω → R : ∂xj u is continuous for |j| = 0, . . . , m},
equipped with the norm

uCbm = ∂xj uC 0 .
b
0≤|j|≤m

From the definition it is clear that for u ∈ C m (Ω, R) we have uCbm < ∞,
if Ω is bounded. More generally, we define
Cbm (Ω, R) = {u ∈ C m (Ω, R) : uCbm < ∞}.
For Ω = Ω = R the function u(x) = x is in C 0 , but not in Cb0 . For Ω
bounded, Cbm is dense in Cb0 . For the treatment of unbounded Ω we define
m
Cb,unif (Ω, R) = {u : Ω → R :∂xj u is uniformly continuous for
|j| = 0, . . . , m, uCbm < ∞}.
For Ω = R the function u(x) = sin(x2 ) is in Cb0 , but not in Cb,unif
0 . For
similar reasons Cb (R, R) is not dense in Cb (R, R), but Cb,unif (R, R) in
n 0 n
0
Cb,unif (R, R). All these spaces are Banach spaces.
Hölder spaces. The spaces Cb0 and Cbm are not the optimal choice for
solving linear PDEs. Even for arbitrarily smooth boundary ∂Ω the boundary
value problem
(5.12) Δu = f in Ω, u|∂Ω = 0,
for f ∈ Cb0 in general does not possess a solution u with optimal regularity,
cf. Example 5.2.4 on page 157. Optimal regularity holds for the subsequently
defined Hölder-continuous functions and Sobolev functions, i.e., for instance
for (5.12) from f ∈ C 0,α it follows u ∈ C 2,α . For α ∈ (0, 1] we define
C 0,α (Ω, R) = {u : Ω → R : u is α -Hölder-continuous, uC 0,α < ∞}
equipped with the norm
|u(x) − u(y)|
uC 0,α = uC 0 + sup ,
b
x,y∈Ω,x=y,|x−y|≤1 |x − y|α
5.2. Basic function spaces and Fourier series 153

and, for k ∈ N and α ∈ (0, 1],

C k,α (Ω, R) ={u : Ω → R : ∂xj u ∈ C 0 for |j| = 0, . . . , k,
∂xk u ∈ C 0,α , uC k,α < ∞}
equipped with the norm

uC k,α = uC k−1 + ∂xj uC 0,α .
b
|j|=k

All these function spaces are Banach spaces, cf. Exercise 5.9. C 0,1 (Ω, R) is
the space of Lipschitz-continuous functions.
Lebesgue and Sobolev spaces. Unfortunately, the above spaces are
not equipped with a scalar product and so tools from linear algebra related
to orthogonality are not available. A natural choice of a scalar product for
functions would be

(5.13)
u, vL2 = u(x)v(x) dx.
Ω
However, if the above spaces are equipped with the above scalar product
they are not complete w.r.t. the induced norm. For instance the sequence
(un )n∈N with
⎧
⎨ 1, for |x| ≤ 1 − 1/n,
un (x) = 0, for |x| ≥ 1,
⎩
n(1 − |x|), for |x| ∈ (1 − 1/n, 1),
is a Cauchy sequence w.r.t. the norm induced by the L2 -scalar product.
However, the limit function is not in Cb0 although un ∈ Cb0 for all n ∈ N.
Since the limit of a Cauchy sequence of Riemann integrable functions
is in general no longer Riemann integrable the Riemann integral has to
be replaced by the Lebesgue integral in order to define complete function
spaces [Alt16, §A1]. In order to define the Lebesgue and Sobolev spaces we
introduce
C ∞ (Ω, R) = {u : Ω → R : u is arbitrarily many times differentiable}
and
C0∞ (Ω, R) = {u ∈ C ∞ (Ω, R) : u has compact support in Ω}
where the support of a function is defined by supp(u) = clRd {x ∈ Ω : u(x) =
0}. The Lebesgue spaces are defined by
 1/p
∞
L (Ω, R) = cl·Lp (C0 (Ω, R)), where uLp =
p
|u(x)| dx
p
Ω
for all p ∈ [1, ∞). By construction all these spaces are Banach spaces consist-
ing of equivalence classes of Cauchy sequences, with two Cauchy sequences
154 5. PDEs on an interval

in the same class if their difference converges to zero. The Lp -spaces con-
structed in this way coincide with the spaces known from measure theory.
The space L2 (Ω, R), respectively L2 (Ω, C), is a Hilbert space equipped with
the scalar product (5.13).
For the solution of PDEs so called Sobolev spaces turn out to be useful.
For p ∈ [1, ∞) and Ω bounded we define
W m,p (Ω, R) = cl·W m,p (C ∞ (Ω, R)),
where

uW m,p = ( ∂xj upLp )1/p
|j|≤m

and
W0m,p (Ω, R) = cl·W m,p (C0∞ (Ω, R)),
for general Ω. Since the sum in the definition of  · W m,p is finite there

are various equivalent norms such as uW m,p = |j|≤m ∂xj uLp . By con-
struction these spaces are Banach spaces, too. The spaces H m (Ω, R) =
W m,2 (Ω, R) and H0m (Ω, R) = W0m,2 (Ω, R) are Hilbert spaces equipped with
the scalar product
 & '

u, vH m = ∂xj u, ∂xj v L2 .
|j|≤m

By Sobolev’s embedding theorem [Alt16, Satz 8.8.], Sobolev spaces can

be embedded continuously into classical function spaces. We have
W m,p (Ω, R) → C n,α (Ω, R) if m − d/p > n + α,
i.e., there exists a C > 0, such that for all u ∈ W m,p
uC n,α ≤ CuW m,p
and in the equivalence class of u ∈ W m,p there is a representative u ∈
C n,α (Ω, R). For the proof of special cases see Lemma 5.1.27 and Lemma
5.2.3.
A different characterization of these spaces is (e.g., [Alt16, §1.25])
W m,p (Ω, R) = {u : Ω → R : ∂xα u ∈ Lp for |α| = 0, . . . , m, uW m,p < ∞},
where ∂xα u denotes the αth weak derivative of u. For Ω ⊂ Rd the function
∂xα u ∈ Lp (Ω, R) is called αth weak derivative of u ∈ Lp (Ω, R) if for all
φ ∈ C0∞ (Ω, R) we have
 
(∂xα u(x))φ(x) dx = (−1)|α| u(x)(∂xα φ(x)) dx.
Ω Ω
5.2. Basic function spaces and Fourier series 155

We define L∞ (Ω, R) as the space of all measurable functions u : Ω → R

for which
sup |f (x)| < ∞ for a null set N .
x∈Ω\N

This space is equipped with the norm

uL∞ = inf sup |f (x)| .
N is a null set x∈Ω\N

We introduce

m
uW m,∞ = ∂xα uL∞
|α|=0

and the space W m,∞ (Ω, R) as the space of all functions u : Ω → R for which
the weak derivatives ∂xα u exist for |α| = 0, . . . , m and for which uW m,∞ <
∞.
Remark 5.2.1. The concept of weak derivatives can be generalized to the
concept of distributional derivatives [RR04, Chapter 5]. A priori, the sets
C ∞ and C0∞ are just vector spaces. There is no norm for which these spaces
are complete. However, using inductive limits of semi-norms, C0∞ (Ω, R) can
be made to be a complete metric space D(Ω), called space of test functions,
where convergence un → u in D(Ω) means: a) There exists a compact
K ⊂ Ω such that supp(un ), supp(u) ⊂ K, b) limn→∞ ∂xα un (x) = ∂xα u(x)
uniformly in K for all α ∈ Nd . However, this convergence is not induced by
a norm.
The elements of the dual space of D(Ω, R) = C0∞ (Ω, R) are called dis-
tributions, i.e., a distribution T is a continuous linear map from D into the
real or complex numbers. This means that un → u in D implies T un → T u,
which is equivalent to the formulation that for all open bounded sets D there
is a constant C and a number m ∈ N such that
(5.14) |T (φ)| ≤ CφCbm for all φ ∈ C0∞ (D, R).

For a continuous function u ∈ Cb0 (Rd , R), or for u in one of the above other
spaces 
Tu (φ) = u(x)φ(x) dx
Rd
defines the so called associated distribution, which is then called regular.
For the distribution associated to ∂xα u we find

T∂xα u (φ) = (∂xα u(x))φ(x) dx
Rd

=(−1)|α| u(x)(∂xα φ(x)) dx = (−1)|α| Tu (∂xα φ).
Rd
156 5. PDEs on an interval

This property is taken to define the αth derivative of an arbitrary distribution

by
(∂xα T )(φ) = (−1)|α| T (∂xα φ).

The distributional derivative of a function u is not necessarily again a func-

tion, as the next example shows. If (∂xα Tu ) can be represented by a function
g, i.e., ∂xα Tu = Tg , then g is the αth weak derivative of u.

Example 5.2.2. For u(x) = |x| we show that u ∈ H 1 ((−1, 1)) but u ∈
H 2 ((−1, 1)) by computing the weak derivatives ∂x u, ∂x2 u. For φ ∈ C0∞ ((−1, 1))
0 1 0 1
we have Tu (∂x φ) = −1 −x∂x φ dx + 0 x∂x φ dx = −1 φ dx + 0 −φ dx =
−Tg (φ) with

−1, x < 0,
g(x) = ∂x u(x) =
1, x > 0.
0 1
Similarly, Tu (∂x2 φ) = −1 ∂x φ(x) dx+ 0 −∂x φ(x) dx = 2φ(0) = 2δ0 (φ) where
δ0 is called the Dirac δ distribution. Thus, ∂x2 u ∈ L2 as there is no function
1
g such that −1 g(x)φ(x) dx = φ(0).

In order to solve nonlinear PDEs two additional properties have to be

satisfied by the function spaces in use. First, the values of the functions on
the boundary have to be well defined. For X = L2 (Ω) a function u ∈ L2 (Ω)
is only unique up to a null set in Ω ⊂ Rd . Since a smooth boundary ∂Ω
is a null set, boundary conditions in L2 are not well defined. Secondly, in
the nonlinearity we have products of functions of X, i.e., with u ∈ X, also
u2 should be in X, i.e., X should be an algebra. For u ∈ L2 in general
we do not have u2 ∈ L2 . However, for m sufficiently large (depending on p
and the space dimension d), H m or more generally W m,p are algebras, and
point-wise values are defined, or at least the boundary conditions can be
fulfilled in a generalized sense.
We close this subsection with the proof of a very simple version of
Sobolev’s embedding theorem and an example indicating which of the func-
tion spaces are suitable for solving PDEs and which are not.

Lemma 5.2.3. Let −∞ < a < b < ∞. Then H 1 ((a, b)) ⊂ C 0,1/2 ((a, b))
and
 
1
(5.15) uL∞ ≤ 2uL2
2
uL2 + ∂x uL2 ,
b−a
√
(5.16) |u(x) − u(y)| ≤ x − y∂x u2L2 .
5.2. Basic function spaces and Fourier series 157

Proof. Since C 1 ((a, b)) is dense in H 1 ((a, b)) w.r.t. the  · H 1 -norm, it is
sufficient to prove (5.15) and (5.16) for u ∈ C 1 ((a, b)). We have
 x  
2 d s−a 2
u (x) = u (s) ds
ds x − a
a x  x
1 2 s−a
= u (s) ds + 2u(s)∂x u(s) ds
a x−a a x−a
1
≤ u2L2 + 2uL2 ∂x uL2
x−a
and similarly
 b  
d s−x 2 1
2
u (x) = u (s) ds ≤ u2L2 + 2uL2 ∂x uL2 .
x ds b − x b − x
Hence, u2 (x) ≤ min{ x−a
1 1
, b−x }u2L2 + 2uL2 ∂x uL2 . For the second esti-
mate we use the Cauchy-Schwarz inequality, namely
 y   y
  

|u(x) − u(y)| =  ∂s u(s) ds ≤ 1|∂s u(s)| ds ≤ |x − y|∂x uL2 . 
x x

We already stated that C k -spaces are in general not optimal concerning

the regularity of solutions of PDEs. More life is given to this statement by
the following example [Sal08, Example 8.2]. This gives a motivation for
the use of Sobolev spaces, in particular for the use of H m -spaces for which
Hilbert space methods are available.
Example 5.2.4. For 0 < α < 2π let Ωα := {(r, φ) : 0 < r < 1, −α/2 <
θ < α/2} be the two-dimensional sector with opening angle α. Consider the
Dirichlet boundary value problem
(5.17) −Δu = 0 in Ωα , u|∂Ω = gα (r, φ) on ∂Ωα ,
with gα (r, φ) = cos(πφ/α) for r = 1, g(r, φ) = 0 else, where (r, φ) are polar
coordinates. Identifying R2 with C we find that f (z) = z π/α is holomorphic
in Ωα , and thus
u(r, φ) = Re(f (z)) = rπ/α cos(πφ/α)
is harmonic in Ωα and satisfies the boundary conditions. Thus, it is the
unique solution of (5.17).
Clearly, u ∈ C ∞ (Ω), and we now consider the regularity of u up to the
boundary and compare it with Sobolev regularity. Let α = π, otherwise
u(x, y) = x. We find
1 π 2 2(π/α−1)
|∇u|2 = (∂r u)2 + (∂ φ u) 2
= r ,
r2 α2
 2 1
and this is in C 1 (Ω) only for α ≤ π. But Ωα |∇u|2 dx = πα 0 r2π/α−1 dr =
π/2 independent of α, and thus u ∈ H 1 (Ω) for all α. Next, |∂x2i u| ∼
158 5. PDEs on an interval

rπ/α−2 for r → 0, thus, u ∈ C 2 (Ω) for α ≤ π/2, and u ∈ H 2 (Ω) for α ≤ π,

i.e., if the sector is convex. By setting g(r, φ) = rj g(φ) with j ≥ 2 and
v = u− g we find g ∈ C ∞ (Ωα ) and −Δv = −Δu + Δ g = Δ g =: f in
Ωα , and v|∂Ωα = 0. For the last system the Lax-Milgram theorem [Eva98,
§6.2.1] guarantees u ∈ H 1 for general spatial domains and regularity theory
[Eva98, §6.3] guarantees u ∈ H 2 for convex spatial domains.

0.5

0.5
0
-0.5
-0.5 0 0.5 -1
1

Figure 5.3. z → (Rez)3/4 solves the boundary value problem (5.17) if

α = 4π/3. The derivative is unbounded at the origin.

5.2.2. Fourier series. PDEs with periodic boundary conditions for the
spatial coordinates can be transferred to ODEs in R∞ with the help of
Fourier series. For notational simplicity we restrict ourselves first to the
torus Td = Rd /(2πZ)d . Let Cper ∞ be the space of functions u : Td → Rd , with
∞
a C periodic extension u  : Rd → Rd satisfying
u (x1 + 2π, x2 , . . . , xd ) = u
(x1 , x2 , . . . , xd ) = u (x1 , x2 + 2π, . . . , xd )
(x1 , x2 , . . . , xd + 2π).
= ... = u
We define
m ∞
Hper = clos·H m (Td ) (Cper ).
The question is if and in what sense a function can be represented by
its Fourier series, or equivalently, in which norms Fourier series converge.
In L2 we have a simple answer which follows from the general theory of
orthonormal systems, and which for convenience we summarize here.
Definition 5.2.5. Let H be a Hilbert space with scalar product
·, · : H ×
H → C. A (finite or infinite) system (φj ) in H, j = 1, . . . , N or j ∈ N, is
called orthogonal system if
φi , φj  = 0 for i = j. It is called orthonormal
if additionally
φj , φj  = 1. It is called a complete orthonormal system
(complete ONS) or Hilbert basis if
u, φj  = 0 for all j implies u = 0 for
u ∈ H.

In Hilbert spaces H the following holds.

5.2. Basic function spaces and Fourier series 159

Lemma 5.2.6. a) If un → u and vn → v in H, then

un , vn  →
u, v.
b) (Pythargoras) For φ1 , . . . , φn ∈ H with
φi , φj  = 0 for i = j we have
φ1 + . . . + φn 2 = φ1 2 + . . . + φn 2 .
c)
∞ If (φj )j∈N is an orthonormal system ∞ in H, and uj a sequence in C, then
u φ
j=1 j j converges if and only if |u
j=1 j | j j∈N ∈  .
2 converges, i.e., if (u ) 2

 If (φj )j∈N 2is an orthonormal

d) (Bessel’s inequality) sequence in H, then for
all u ∈ H we have ∞ j=1 |
u, φj  | ≤ u 2.

Proof. a) By Cauchy-Schwarz we have

|
un , vn  −
u, v | = |
un , vn − v +
un − u, v |
≤ un |vn − v + un − uv,
where supn∈N un  < ∞ for convergent series (un )n∈N .
b) Direct calculation for the finite n sums. 2 n
c) For all m ≤ n we have  j=m uj φj  = j=m |uj | by b). Thus,
2
∞ ∞
j=1 |uj | is a Cauchy se-
2
j=1 uj φj is a Cauchy sequence if and only if
quence.
d) For N ∈ N we have
( )
N 
N N
0≤ u−
u, φj  φj , u −
u, φj  φj = u2 − |
u, φj  |2 ,
j=1 j=1 j=1
N
and hence j=1 |
u, φj  |2 ≤ u2 , which implies convergence of the series
and Bessel’s inequality. 
Lemma 5.2.7. The following statements are equivalent:
(i) (φj )j∈N is a complete ONS. 
(ii) For all u ∈ H we have u = ∞ j=1
u, φj  φj .
(iii) For all u, v ∈ H we have Parseval’s identity
∞

(5.18)
u, v =
u, φj 
v, φj  .
j=1

(iv) For all u ∈ H we have Bessel’s equality

∞
(5.19) u =
2
|
u, φj  |2 .
j=1

Proof. (i)⇒(ii). For u ∈ H we have convergence of ∞ j=1 |
u, φj  | by d)
2
∞
and convergence of j=1
u, φj  φj to some v ∈ H by Lemma 5.2.6 c). By
Lemma 5.2.6 a) we have
∞


u − v, φj  =
u, φj  −
u, φn 
φn , φj  =
u, φj  −
u, φj  = 0,
n=1
160 5. PDEs on an interval

and since (φj ) is complete this implies v = u. (ii)⇒(iii) again follows from
Lemma 5.2.6 a), and (iv) follows from (iii) with v = u. Finally, (iv)⇒(i)
since
u, φj  = 0 for all j and (iv) imply u = 0, hence u = 0. 
Due to the equivalence of (5.18) and (5.19), often both are called Par-
seval’s identity. Clearly, Lemma 5.2.6 and Lemma 5.2.7 also holds if se-
quences (φj )j∈N are replaced by sequences (φj )j∈Z , (φj )j∈Nd , (φj )j∈Zd , with
the respective replacements in the sums. The most important example are
classical Fourier series.

Theorem 5.2.8. a) The functions φk = 1 ik·x

2π e with k ∈ Zd are a complete
ONS in L2 (Td ) w.r.t. the inner product

1

u, vL2 = u(x)v(x) dx.
(2π)d Td

For u ∈ L2 (Td ) we have L2 -convergence of the Fourier series, i.e., for

 
1
SN (x) = k e
u ik·x
, with k =
e
u ik·x
, uL2 = u(x)e−ik·x dx
(2π)d Td
|k|≤N

we haveu − SN L2 → 0 as N → ∞. This convergence is abbreviated as

u(x) = k∈Zd uk eik·x .
b) For all φ ∈ TN = span{eik·x : |k| ≤ N } we have u − SN L2 ≤
u − φL2 , i.e., SN is the best approximation of u in TN in the quadratic
mean.
c) We have Parseval’s identity

 1
|
uk |2 = u2L2 .
(2π)d
k∈Zd

d) There exists a C > 0 such that if u ∈ C m is 2π-periodic in each

uk | ≤ C|k|−m .
direction, then |

1, if k = m,
Proof. a) By direct calculation we find
φk , φm  = i.e.,
0, else,
the (φk ) are an ONS. The completeness of this ONS can be shown with the
Weierstraß approximation theorem, see [Alt16, Satz 7.10].
b) follows since SN is the orthogonal projection of u on TN .
c) Parseval’s identity can be computed directly for finite sums. Going
to the limits shows the assertion.
5.2. Basic function spaces and Fourier series 161

d) Through integration by parts we find


1
k =
e , uL2 =
u ik·x
u(x)e−ik·x dx
(2π)d Td

1
= (−1)n (∂ m u(x))e−ik·x dx.
(2π)d k m Td x


The map u → ( u)k∈Zd will be abbreviated with F . By c), F is an
isometric isomorphism from L2 to 2 . Its inverse ( uk )k∈Z → u is denoted
by F −1 . By d) the smoothness of u is related to the decay of its Fourier
coefficients.

Formally we have ∂x u(x) = uk eikx , or equivalently F (∂x u) =
k∈Z ik
uk )k∈Z . It follows, that F is in fact an isomorphism between the Sobolev
(ik
m and the spaces of sequences 
spaces Hper 2,m which have been introduced in
−1
§5.1. Moreover, F maps 1,m to Cb . m

Lemma 5.2.9. Let m ∈ N0 . a) There exists a C > 0, such that for all
 ∈ 1,m
u
uCbm ≤ C
u1,m .

 ∈ 2,m
b) There exist C1 , C2 > 0, such that for all u

C1 
u2,m ≤ uH m ≤ C2 
u2,m .

c) There exist C1 , C2 > 0, such that for all u ∈ Hper

m

C1 uH m ≤ 
u2,m ≤ C2 uH m .

Proof. For notational simplicity we consider d = 1 and u with u 0 = 0.

Moreover, we first consider F and F −1 on the dense subspaces Cper∞ respec-

tively the space of finite sequences. The results then follow by continuous
extension, see the subsequent Lemma 5.2.10.
a) We have
 
   
 j 
uCbm ≤ C sup sup ∂x k eikx  ≤ C sup sup
u |k|j |
uk | |eikx |
x∈R 0≤j≤m  k∈Z
 x∈R 0≤j≤m k∈Z
≤ C
u1,m .

Continuity in respect to the differentiability of u follows from the uniform

and absolute convergence of the series.
162 5. PDEs on an interval

b) and c) The second estimate in b) and the first estimate in c) follow

from
 2
m m    
 j 
u2H m = ∂xj u2L2 = ∂x k eikx  dx
u
T d  
j=0 j=0 k∈Z
 2
m    m 
 j ikx 
=  uk (ik) e  dx ≤ 2π |
uk |2 |k|2j ≤ C
u22,m .
T d  
j=0 k∈Z j=0 k∈Z

The first estimate in b) and the second estimate in c) follow from

 2
   

u2,m =
2
|
uk | |k| =
2 2m  e −ikx
u(x)dx  |k|2m
 
k∈Z k∈Z Td
 2

by parts   
=  e −ikx
∂x u(x)dx |k|2(m−1) = . . .

k∈Z Td
 2
  
=  e ∂x u(x)dx
−ikx m

k∈Z Td

Parseval 1
= F (∂xm u)22,0 = ∂ m u2 2 ≤ u2H m .
2π x L


Lemma 5.2.10. Let X be a metric space, A ⊂ X a dense set, and Y a

complete metric space. Then every uniformly continuous function f : A → Y
possesses a unique uniformly continuous extension f : X → Y .

Proof. The condition that the extension must be continuous leads to the
only possible extension of f , namely
f(x) = lim f (x ).
x ∈A,x →x

It remains to prove the existence of this limit, i.e., to prove that f is well
defined. In order to do so, let (xn )n∈N be a sequence with xn ∈ A and
limn→∞ xn = x. Hence, (xn )n∈N is a Cauchy sequence in X and from the
uniform continuity it follows that the image sequence (f (xn ))n∈N is a Cauchy
sequence in Y . Since Y is complete, we have the existence of
y = lim f (xn )
n→∞

in Y . Obviously the limit y is independent of the chosen sequence. It is an

easy exercise to prove the uniform continuity of f. 
We give a number of remarks and further results about Fourier series
which will be useful later.
5.2. Basic function spaces and Fourier series 163

Remark 5.2.11. (Hausdorff-Young) ßThe discrete Fourier transform F

is continuous from Lp to q with 1/p+1/q = 1 for p ≤ 2. The discrete inverse
Fourier transform F −1 is continuous from q to Lp with 1/p + 1/q = 1 for
q ≤ 2, but not for q > 2, cf. [Duo01, Corollary 1.20]. This can be shown
with the so called Riesz-Thorin interpolation between the inequalities from
Theorem 5.2.8 c) and Lemma 5.2.9 a).

Remark 5.2.12. Lemma 5.2.9 suggests to define non-integer Sobolev spaces

by Fourier series, i.e., for θ ∈ R let
θ
Hper = F −1 2,θ with uH θ = 
u2,θ .
We will come back to this definition in §6.2.1.

Remark 5.2.13. (Real Fourier series) Besides the complex Fourier ex-
pansion also real Fourier polynomials and series of the form
∞

a0
u(x) = + [ak cos(k · x) + bk sin(k · x)],
2 d
k∈N

with ak , bk ∈ R, are in use, where


2
ak = u(x) cos(k · x) dx, k ≥ 0,
(2π)d Td

2
bk = u(x) sin(k · x) dx, k ≥ 1.
(2π)d Td
k and (ak , bk ) are
The relations between u
1 1 1
0 = a0 ,
u k = (ak − ibk ),
u −k = (ak + ibk ),
u
2 2 2
k + u
ak = u −k , uk − u
bk = ( −k )i, k ∈ N.
For u(x) ∈ R we have u −k . In this book we prefer the concise complex
k = u
notation.

Remark 5.2.14. (General periodic boxes) Let L1 , . . . , Ld > 0, Ω =

(0, L1 ) × · · · × (0, Ld ). As in Theorem 5.2.8 we may expand u ∈ L2 (Ω) as
∞  
2πk1 2πkd
u(x) = k eiωk ·x with ωk =
u ,..., ,
d
L1 Ld
k∈Z

where 
1
k =
u u(x)e−iωk ·x dx = (F u)k .
L1 L2 · · · Ld Ω
Again, F is an isomorphism between Hper
m (Ω) and 
2,m .
164 5. PDEs on an interval

Remark 5.2.15. Let Ω ⊂ Rd be a bounded domain and let (ϕj )j∈N be

a complete orthonormal system in L2 (Ω), i.e., every u ∈ L2 (Ω)possesses
a unique representation as convergent series in L2 (Ω), i.e., u = j∈N cj ϕj
with cj ∈ C. Then F , defined by (F u)j = cj , is an isomorphism between
L2 (Ω) and 2 . However, in general F is not an isomorphism between H m (Ω)
and 2,m . The set {sin nx : n ∈ N} is a basis of L2 ((0, π)), but not a basis
of H 1 ((0, π)). Since H 1 ⊂ Cb0 in H 1 , only functions u with u(0) = u(π) = 0
can be approximated. In L2 the two points x = 0, π are a null set.
Remark 5.2.16. Point-wise convergence of Fourier series is a rather deli-
cate issue. For instance, the Fourier series of u ∈ L1 may diverge almost ev-
erywhere [Kol27], while for u ∈ L2 we have convergence almost everywhere
[Car66]. No necessary and sufficient conditions are known for the point-wise
convergence of the Fourier series of a function u. However, there are various
sufficient conditions, for instance if u is piecewise C 1 , then Sn (x) → u(x) at
points of continuity. More generally,
1
Sn (x) → (u(x+) + u(x−)),
2
where u(x+) and u(x−) denote the right and the left limit of u in x. Thus,
Sn (x0 ) converges to the mean of u at jump points x0 . This convergence
comes with notable oscillations (≈ 19%) to the left and right of x0 , which is
known as Gibbs phenomenon, see Figure 5.4.

Figure 5.4. Fourier expansion

 of u(x)= − 1, x ∈ (−π, 0), u(x)=1, x ∈
(0, π) yields u(x) = π4 ∞ n=1 2n−1 sin(2n − 1)x. The figure shows the
1

partial sums s1 , s5 , s13 and thus illustrates the Gibbs phenomenon.

Fourier series allow giving simple proofs of classical inequalities when

the functions involved are spatially periodic.

Lemma 5.2.17. (Poincaré’s inequality) For u ∈ Hper (T , R) with

1 d

Td u dx = 0 we have
 
(5.20) |u|2 dx ≤ |∇u|2 dx.
Td Td
5.2. Basic function spaces and Fourier series 165

Proof. Parseval’s identity gives

   
|u|2 dx = 2π |
uk |2 ≤ 2π |k| |
2
uk | =
2
|∇u|2 dx. 
Td Td
k∈Zd \{0} k∈Zd \{0}

Next we have the following version of Sobolev’s embedding theorem in

d space dimensions.

Lemma 5.2.18. For m − d/2 > n there exists a C > 0, such that
uC n (Td ) ≤ CuH m (Td ) .

Proof. The assertion follows from Lemma 5.2.9 and Lemma 5.1.27. 
Analytic properties of the solution operator of a linear evolution equation
can be established with the help of Fourier series.

Example 5.2.19. We consider the solution operator T (t) defined via the
solution u(x, t) = T (t)u0 )(x) of the linear heat equation ∂t u = ∂x2 u, with x ∈
[0, π], under Dirichlet boundary condition u(0, t) = u(π, t) = 0 to the initial
value u(x, 0) = u0 (x). In order to prove that (T (t))t≥0 is a C0 -semigroup in
L2 ((0, π)) and in H m ((0, π))∩H01 ((0, π)) for every m ∈ N we make an odd 2π-
periodic extension of the functions with u(0, t) = u(π, t) = 0. The semigroup
in the space of 2π-periodic functions is denoted again by T (t). We proved
in §5.1.2 that T(t) = F T (t)F −1 defined by (T(t) u(0))k∈Z = (e−k t u
2
k (0))k∈Z
(0) ∈ 2,m we have
is continuous in 2,m , i.e., for every u

T(t) (0)2,m → 0
u(0) − u for t→0.
Due to the isomorphism property of F between Hper
m and 
2,m , cf. Lemma
5.2.9 b) and c), it follows that
T (t)u(0) − u(0)Hper
m ≤ C1 F −1 T (t)u(0) − u
(0)2,m

= C1 T (t)
u(0) − u(0)2,m → 0 for t → 0 ,
m . The restriction of x to [0, π] gives
i.e., T (t) is a C0 -semigroup in Hper
the result. Moreover, from Example 5.1.21 it is known that for r ≥ 0 the
semigroup T(t) can be estimated by

T(t)  max(1, t−r/2 )

u(0)2,m+r ≤ C u(0)2,m .
Using again that F is an isomorphism between Hper
m and 
2,m shows

T (t)u0 Hper (t)

m+r ≤C1 T  max(1, t−r/2 )
u(0)2,m+r ≤ C1 C u(0)2,m
≤C max(1, t−r/2 )u0 Hper
m ,

 2.
with C = C1 CC
166 5. PDEs on an interval

Example 5.2.20. We consider ∂t2 u = ∂x2 u with x ∈ (0, 2π) and periodic
boundary conditions. We rewrite this equation as first order system for
z = (∂t u, ∂x u) and obtain
 
d 0 ∂x
z(x, t) = Az(x, t), with A = ,
dt ∂x 0
or in Fourier space
 
d z (k, t), = 0 ik
z(k, t) = A with A .
dt ik 0
The general solution is given by

z(x, t) = c1 (k)eik(t+x) z1 + c2 (k)eik(−t+x) z2 =: etA z(·, 0)(x),
k∈Z
where    
c1 (k) 1 1 1
= z(k, 0).
c2 (k) 2 1 −1
Hence, we have a uniformly bounded C0 -semigroup for z in, e.g., H m × H m ,
which however is not smoothing.

5.2.3. Some nonlinear PDE examples. We start this section with a

version of the local existence and uniqueness theorem, Theorem 5.1.23, in
physical space. In order to use the results from §5.1 we refrain from greatest
θ in accordance with
generality and restrict ourselves to Sobolev spaces Hper
Lemma 5.2.9. We consider
d
(5.21) u = Λu + N (u), u|t=0 = u0 ,
dt
with Λ generating a C0 -semigroup which satisfies
(5.22) etΛ uHper
θ ≤ Cθ−r eβt t−α uHper
r

for θ ≥ r with constants Cθ−r , β, and α ∈ [0, 1). Moreover, let N be a

locally Lipschitz-continuous map from H θ into H r .
Similar to Definition 5.1.19 we define
Definition 5.2.21. a) A function u ∈ C([0, T0 ], Hper
θ ) which satisfies
 t
tΛ
(5.23) u(t) = e u0 + e(t−τ )Λ N (u(τ )) dτ
0
is called a mild solution of (5.21).
θ ), with Λu ∈ C([0, T ], H θ ), is called a
b) A function u ∈ C 1 ([0, T0 ], Hper 0 per
θ for each t ∈ [0, T ].
strong solution of (5.21), if (5.21) holds in Hper 0

Theorem 5.2.22. For all C1 > 0 there exists a T0 > 0 such the following
holds. For u0 ∈ H θ with u0 H θ ≤ C1 there exists a unique solution u ∈
C([0, T0 ], H θ ) of (5.21) with u|t=0 = u0 .
5.3. The Chafee-Infante problem 167

Proof. This a direct consequence of Theorem 5.1.23, Remark 5.1.24, and

Lemma 5.2.9. 
There are other straightforward generalizations from the finite- to the
infinite-dimensional situation.
Theorem 5.2.23. Consider (5.21), where (5.22) is satisfied for a β < 0.
Then the fixed point u∗ = 0 is asymptotically stable.

Proof. The proof goes line for line as the proof of Theorem 2.3.4 a). 
In order to check the assumptions for system (5.21) for a concrete non-
linear PDE we have to handle products of functions in physical space. The
θ -spaces are closed under multiplication if θ is sufficiently big.
Hper
Lemma 5.2.24. For all θ > d/2 there exists a C > 0, such that for all
u, v ∈ Hper
θ we have

uvH θ ≤ CuH θ vH θ .

Proof. This follows from Lemma 5.1.28 by using the isomorphism F :

θ → .
Hper 
2,θ
We come back to the PDEs introduced in Example 5.1.29. We have
already proved the local existence and uniqueness of solutions of the Fourier
transformed versions. The isomorphism property between 2,θ and Hper θ

gives the following result.

Theorem 5.2.25. For the KPP equation, the Allen-Cahn equation, the
NLS equation, the Burgers equation, and the GL equation with 2π-periodic
boundary conditions we have the local existence and uniqueness of solutions
θ if θ > 1/2, i.e., for all C > 0 there exists a T > 0 such the follow-
in Hper 0
ing holds. For u0 ∈ Hper with u0 H θ ≤ C there exists a unique solution
θ

u ∈ C([0, T0 ], Hper
θ ) with u|
t=0 = u0 , respectively, A ∈ C([0, T0 ], Hper ) with
θ

A|T =0 = A0 .

The θ in the last theorem can be made smaller by using the smoothing
properties of the semigroup, cf. §6.2.1. Moreover, the smoothing estimate
(5.11) can be used to show that solutions to the KPP equation, the Allen-
Cahn equation, and the GL equation become arbitrary smooth and even
analytic for t > 0. This is done for instance in §5.3.3 or §6.2.2.

5.3. The Chafee-Infante problem

After having discussed the local existence and uniqueness theory of PDEs on
an interval we now consider the qualitative behavior of solutions in a specific
example, namely the Chafee-Infante problem [CI75]. The presentation is
based on [Hen81, §5.3].
168 5. PDEs on an interval

The Chafee-Infante problem is to find the attractor of a semi-linear par-

abolic PDE, the Allen-Cahn equation,
(5.24) ∂t u = ∂x2 u + αu − u3 ,
with α ∈ R, u = u(x, t) ∈ R, t ≥ 0, and x ∈ (0, π), under Dirichlet boundary
conditions u(0, t) = u(π, t) = 0. Our goal is to characterize the attractor
of this system for different values of α. This PDE can be interpreted as an
infinite-dimensional gradient system. Similar to finite-dimensional gradient
systems, see §2.4.5, this fact restricts the elements of the attractor in the
following to fixed points and heteroclinic connections.

5.3.1. Local and global existence of solutions. As phase space we use

H01 = H01 (0, π). Solving the Allen-Cahn equation with 2π-periodic bound-
ary conditions and restricting to the invariant subspace of odd functions is
the same as solving the Allen-Cahn equation with Dirichlet boundary con-
ditions. Hence, Theorem 5.2.25 applies and we have the local existence and
uniqueness of solutions in H θ if θ > 1/2.
Theorem 5.3.1. For all C > 0 there exists a T0 > 0 such that for all
u0 ∈ H01 with u0 H 1 ≤ C there exists a unique solution u ∈ C([0, T0 ], H01 )
of the Allen-Cahn equation (5.24) with u|t=0 = u0 .

To prove the global existence of solutions it is sufficient to bound the

H 1 -norm. We prove more, namely the existence of an absorbing set for
(5.24).
Theorem 5.3.2. (Global existence and existence of an absorbing
set) For all α ∈ R there exists a R > 0 such that for all C1 ≥ 0 we have
a T > 0 such that the followings holds. If u0 ∈ H01 satisfies u0 H 1 ≤ C1 ,
then the associated solution satisfies u(t) ∈ B = {u ∈ H01 : uH 1 ≤ R} for
all t ≥ T .

Proof. Again the solutions are extended to odd 2π-spatially periodic solu-
tions. Then we have
  2π
d 2π 2
(5.25) u dx = 2 −(∂x u)2 + αu2 − u4 dx
dt 0 0
  2π
d 2π
(5.26) (∂x u)2 dx = 2 −(∂x2 u)2 + α(∂x u)2 − 3u2 (∂x u)2 dx.
dt 0 0
If α < 0 all terms on the right-hand side are negative and we have
lim sup u(t)H 1 = 0.
t→∞
In order to obtain estimates which are also good for small α ≥ 0 we split the
parameter regime α ≥ 0 in two parts. First let α ∈ [0, 1/2]. Adding (5.25)
5.3. The Chafee-Infante problem 169

and (5.26) yields


d 2π 2
u + (∂x u)2 dx
dt 0
 2π
=2 −(∂x2 u)2 + (α − 1)(∂x u)2 − 3u2 (∂x u)2 + αu2 − u4 dx
0
 2π
≤2 −(∂x u)2 /2 + αu2 − u4 dx
0
 2π
≤2 −(∂x u)2 /2 − u2 /2 + 1/8 + α2 /2 dx
0
 2π
≤− u2 + (∂x u)2 dx + π.
0
This immediately shows that
lim sup u(t)2H 1 ≤ π.
t→∞

For α > 1/2 we consider


d 2π
2αu2 + (∂x u)2 dx
dt 0
 2π
=2 −(∂x2 u)2 − α(∂x u)2 − 3u2 (∂x u)2 + 2α2 u2 − 2αu4 dx
0
 2π
≤2 −α(∂x u)2 + 2α2 u2 − 2αu4 dx
0
 2π
≤2 −α(∂x u)2 − 2α2 u2 + 2α3 dx.
0
 2π
Hence, for E = 0 2αu2 + (∂x u)2 dx we have E ≤ −2αE + 8πα3 and thus
8πα3
E(t) ≤ e−2αt E(0) + (1 − e−αt ).
2α

Hence, lim supt→∞ E(t) ≤ 4πα2 . Since uH 1 ≤ ( 2αu2 + (∂x u)2 dx)1/2 for
α > 1/2, we are done. 

5.3.2. Existence of the attractor. The existence proof of attractors in

finite dimensions uses the argument that a bounded sequence contains a
convergent subsequence. In infinite dimensions this is in general no longer
true. Hence, this compactness argument has to be recovered by using the
smoothing properties of the solution operator St with St u0 = u(·, t).
Lemma 5.3.3. For t > 0 fixed the solution operator St maps bounded balls
of H01 into bounded balls of H 2 ∩ H01 .
170 5. PDEs on an interval

Proof. We consider the variation of constant formula and estimate

 t
u(t)H 2 ≤T (t)u0 H 2 + T (t − τ )u3 (τ )H 2 dτ
0
 t
−1/2
(5.27) ≤Ct u0 H 1 + C(t − τ )−1/2 dτ sup u(τ )3H 1 < ∞,
0 τ ∈[0,t]

where T (t) is the semigroup from Example 5.2.19. 

Since H2 ∩ H01
is compactly embedded in H01
we have compactness of
the operator St in H01 . Theorem 5.3.2 thus shows that the Chafee-Infante
problem (5.24) defines a dissipative dynamical system such that Theorem
2.4.4 applies.
Theorem 5.3.4. For the Chafee-Infante problem (5.24) there exists a non-
empty, compact, time-invariant set A = ω(B) ⊂ H01 , the global attractor,
for which
dist(u(t, B), A) = sup inf u(t, b) − aH 1 → 0 for t → ∞,
b∈B a∈A
where B is the absorbing set from Theorem 5.3.2.

Proof. For convenience we repeat the main steps from the proof of Theorem
2.4.4. The attractor is defined by
*
A= At
t≥0

with At = closH 1 (St (B)). Since B is positively invariant, the family (At )t≥0 ,
satisfies At1 ⊂ At2 for t1 > t2 . Hence, A ⊂ A0 is bounded. Since St is a
compact operator for t > 0, the set At is compact for t > 0. Since (At )t≥0
is a decreasing family of compact non-empty sets, the attractor A = ∩t≥0 At
is non-empty and compact.
We skip the proof of the time invariance and restrict ourselves to the
attractivity which is proved by contradiction. We assume that B is not
attracted by A. Then there exists a δ > 0, sequences tn → ∞ and un ∈ B,
such that dist(Stn (un ), A) > δ > 0 for all n ∈ N. For a small t > 0 the
sequence Stn −t (un ), (n ∈ N) is bounded. Since St is a compact operator
there exists a subsequence such that vj = Stnj (unj ) converges towards a w
for j → ∞. Therefore, w ∈ A which contradicts the above assumption that
the sequence is bounded away from A. 

5.3.3. The choice of regularity does not matter (much). As already

said, major differences between finite-dimensional systems and infinite-di-
mensional systems are due to the fact that in finite dimensions all norms
are equivalent, whereas in infinite dimensions there are infinitely many
non-equivalent norms and so infinitely many possible non-equivalent phase
5.3. The Chafee-Infante problem 171

spaces. More or less all definitions in the theory of dynamical systems, such
as continuity of solutions w.r.t. time, stability of solutions, etc. depend on
the chosen norm. Therefore, we expect that the choice of a suitable phase
space in infinitely many dimensions in general plays a crucial role. It is the
purpose of this section to explain that for systems with smoothing proper-
ties this often is not the case. If there is a global bound in one H θ -space,
then it does not matter which H θ -space is chosen as long as these spaces
are connected with a smoothing estimate.
The estimate (5.27) can be generalized to
 t
u(t)H θ+1 ≤T (t − τ )u(τ )H θ+1 + T (t − s)u3 (s)H θ+1 ds
τ
 t
≤C(t − τ )−1/2 u(τ )H θ + C(t − s)−1/2 ds sup u(s)3H θ
τ s∈[τ,t]
−1/2
≤C(t − τ ) Cθ (τ ) + 2C(t − τ ) 1/2 3
Cθ (τ ) ,
where Cθ (τ ) = sups∈[τ,∞] u(s)H θ . Hence

(5.28) Cθ+1 (t) ≤ C(t − τ )−1/2 Cθ (τ ) + 2C(t − τ )1/2 Cθ (τ )3 .

From Theorem 5.3.2 we know that C1 (0) < ∞ and that lim supτ →∞ C1 (τ ) ≤
R. In H 2 we have the local existence and uniqueness of solutions, i.e., for
u0 ∈ H 2 with u0 H 2 ≤ C  there is a T2 > 0 and a C 2 < ∞ such that the
2 . Moreover, we
solutions exist for all t ∈ [0, T2 ] and supt∈[0,T2 ] u(t)H 2 ≤ C
get
−1/2 1/2
C2 (T2 + τ ) ≤ CT2 C1 (τ ) + 2CT2 C1 (τ )3 < ∞.
Combining the last two estimates shows that
2 , C2 (T2 )) < ∞.
sup u(t)H 2 ≤ max(C
t∈[0,∞)

Moreover, choosing t − τ = 1 in (5.28) yields

lim sup u(t)H 2 ≤ CR + 2CR3 .
t→∞

Hence, we have established an absorbing set in H 2 . With the same procedure

we show the existence of an absorbing set in H 3 , etc.. As a consequence the
attractor exists in each H θ and the choice of phase space is not crucial with
this respect.

5.3.4. Characterization and bifurcation of the attractor. Here we

give a characterization of the attractor for different values of α. In a first
step we write (5.24) as a gradient system, cf. §2.4.5. We have
∂t u = ∂x2 u + αu − u3 = −β∂u V (u)
172 5. PDEs on an interval

with potential
 π
1 α 1
V (u) = (∂x u(x))2 − u(x)2 + u(x)4 dx
0 2 2 4
and β a linear map defined below. In order to justify this formula, first
recall that for a function V : Rd → R we have
V (u + εv) = V (u) + εaT v + O(ε2 ) = V (u) + ε
a, v + O(ε2 ) for all v ∈ Rd ,
where
u, v = uT v is the scalar product between the vectors u and v, i.e.,
the derivative is defined as an element of the dual space of Rd . However, it
can be identified with Rd through the map
β : Lin(Rd , R) → Rd ,
a, · → a.
For a map V : X → R where the function space X is equipped with the
scalar product  π

u, v = u(x)v(x) dx
0
we define the map
β : Lin(X, R) → X,
a, · → a.
This is well defined since in Hilbert spaces the dual space Lin(X, R) can be
identified with X by the Riesz representation theorem [Alt16, Satz 4.1].
Using the boundary conditions and integration by parts we find
V (u + εv) − V (u)
lim
ε π
ε→0
1 α 1
= lim ε−1 ( (∂x (u + εv))2 − (u + εv)2 + (u + εv)4 )
ε→0 0 2 2 4
1 α 1
− ( (∂x u)2 − u2 + u4 ) dx
 2 2 4
π
=− (∂x2 u + αu − u3 )v dx
0
and so by comparison
β∂u V (u) = −(∂x2 u + αu − u3 ).
Therefore, (5.24) is a gradient system in H01 , and hence the function t →
d
V (u(t)) decreases along solutions u = u(t), i.e., dt V (u(t)) ≤ 0, where equal-
ity only holds in fixed points. Consequently, no non-trivial periodic solution
can occur. Moreover, V is bounded from below, since
 π  π 2
1 α 1 α πα2
V (u) = (∂x u(x)) − u(x) + u(x) dx ≥ −
2 2 4
dx = − .
0 2 2 4 0 4 4
Similar to the finite-dimensional situation, cf. Theorem 2.4.15, the attractor
consists of the fixed points and their unstable manifolds, cf. [Rob01, The-
orem 10.13]. In case that only finitely many fixed points exist, the attractor
5.3. The Chafee-Infante problem 173

consists of these fixed points and their heteroclinic connections. This can be
seen directly. In a gradient system every solution must end in a fixed point.
Solutions in the attractor must also start in one of the finitely many fixed
points. This follows from the fact that backwards in time the system in the
attractor is a gradient system, too. The potential is given by −V and it is
bounded on the attractor.
We compute the fixed points, or stationary solutions, of the PDE, which
satisfy
∂x2 u + αu − u3 = 0.
Due to the boundary conditions u(0) = u(π) = 0 in the (u, u )-plane we
have to find solutions which start from the v = u -axis, end on this axis,
and need for this part of the orbit the ’time’ x = π. For all α > 0 the phase
portrait looks qualitatively the same. The periodic orbits around the origin
have a periodicity which is i) minimal at the origin, namely the periodicity
√
of the linearization, 2π/ α, ii) infinity at the heteroclinic orbits, and iii)
which increases strictly monotonic with the distance from the origin. Thus,
non-trivial equilibria of (5.24) can only exist for α > 1 since for α ≤ 1 the
solutions are too slow to make half of the periodic orbit in a time π.
Using i)-iii) the complete bifurcation picture can be established in a
rigorous way. The number of solutions with u(0) = u(π) = 0 changes for
√
mπ/ α = π with m ∈ N, an integer multiple of half the minimal period.
As a consequence, for α ∈ (−∞, 1] we have one equilibrium, the origin; for
α ∈ (1, 4] we have 3 equilibria, the origin, and two equilibria called u±1 ;
for α ∈ (4, 9] we have 5 equilibria, . . . ; and for α ∈ (m2 , (m + 1)2 ] we have
2m + 1 equilibria, the origin, u±1 , . . ., and u±m .
Hence, for fixed α there are only finitely many fixed points which are
elements of the attractor. In order to understand the dynamics in the at-
tractor, i.e., to find the heteroclinic connections between the fixed points,
we analyze the linearization at the fixed point u ≡ 0, i.e.,
∂t u = ∂x2 u + αu,

with u(0, t) = u(π, t) = 0, or equivalently, with u(x, t) = n (t) sin(nx),
n∈N u
d
n = (α − n2 )
u un .
dt
Therefore, the linear operator
Λ· = ∂x2 · +α·
with Dirichlet boundary conditions has eigenvectors u(x) = sin mx with
associated eigenvalues λ = α − m2 for m ∈ N. Equivalently the infinite-
dimensional diagonal matrix
 nm )n,m∈N = ((α − m2 )δnm )n,m∈N
(Λ
174 5. PDEs on an interval

has the eigenvectors em defined by (em )n = δmn with associated eigenvalues

λ = α − m2 .
Im u

-15 -8 -3 Re
1 4 9 α

Figure 5.5. Left: the spectrum of the operator Λ· = ∂x2 ·+α· under
Dirichlet boundary conditions for α = 1. Right: the bifurcation
diagram. At the parameter values α = n2 unstable equilibria bifur-
cate via a pitchfork bifurcation from the trivial branch u(α) ≡ 0.

Hence, for α ∈ (−∞, 1) the origin is asymptotically stable. For α > 1 the
origin is unstable, with a one-dimensional unstable manifold for α ∈ (1, 4],
with a two-dimensional unstable manifold for α ∈ (4, 9], and with an m-
dimensional unstable manifold for α ∈ (m2 , (m + 1)2 ].
For α ∈ (1, 4) the one-dimensional unstable manifold of the origin ends
in the stable equilibria u±1 . For α ∈ (4, 9) the equilibria u±1 , u±2 lie on
the two-dimensional unstable manifold of the origin. Since u±2 bifurcates
from the unstable origin, these fixed points are also unstable and their one-
dimensional unstable manifold ends in u±1 .
The reasons are as follows. Since for fixed α the fixed points uj are
isolated and since the linearization only has real eigenvalues due to the gra-
dient structure, no eigenvalue of the linearization around the equilibria uj
crosses the imaginary axis after the bifurcation when α is increased. There-
fore, the dimension of the unstable manifold of uj is the same as at their
bifurcation point from the trivial branch. The fixed points u±1 bifurcating
at α = 1 are always stable. The fixed points u±2 bifurcating at α = 4 have
a one-dimensional unstable manifold which ends in the fixed points u±1 .
The fixed points u±3 bifurcating at α = 9 have a two-dimensional unsta-
ble manifold and so heteroclinic connections to the fixed points to u±1 and
u±2 exist. Figure 5.6 sketches the dynamics in the attractor of (5.24). For
α ∈ (n2 , (n + 1)2 ) we have an attractor of dimension n consisting of finitely
many fixed points and heteroclinic orbits between these fixed points, in par-
ticular, it contains the n-dimensional unstable manifold of the origin. A local
bifurcation analysis via center manifold reduction can be found in §13.2.1.
Further Reading. Our point of view of PDEs over bounded sets as count-
ably many ODEs is similar to [Hal88, Rob01, KP13], while [Paz83] gives
5.3. The Chafee-Infante problem 175

Figure 5.6. The finite-dimensional attractors consisting of fixed

points and connecting orbits for α < 1, 1 < α < 4, 4 < α < 9, and
9 < α < 16. Only selected heteroclinic orbits have been plotted.
The attractor always contains the n-dimensional unstable manifold
of the origin.

an excellent and concise account of the general semigroup approach, see

also [RR04, Chapter 11]. Fourier series can be found in most textbooks
on analysis and functional analysis, and in many books on PDE; we recom-
mend [Olv14, Chapter 3] for an introduction with a PDE point of view,
and [Duo01] for a concise but comprehensive treatment. Classical books
on linear functional analysis, covering much wider ground than what is used
here and in the following chapters are [Yos71, RS75a, Kat95]; our fa-
vorites are [Rud73, Wer00, Alt16]. A concise introduction to nonlinear
functional analysis is [AA11]. Our presentation of function spaces follows
[Alt16, Wlo87], but the same material can be found in many textbooks, for
instance, from a PDE perspective, in [Str92, RR04, Eva98, Sal08]. Com-
prehensive treatments of distributions and Sobolev spaces, including various
versions of Poincaré inequalities and Sobolev imbeding theorems, are given
in [Hör83, Maz11], and [Tay96, Chapters 3,4,13]. See also [Geo15] for
a concise introduction aimed at graduate students, and Section 7.3 of this
book for the case of unbounded domains.

Exercises
5.1. Prove that the space c00 = {u : Z → R : un = 0 for finitely many n} equipped
with the 1 -norm is not complete.
d2
5.2. Consider un = −ωn2 un , n ∈ N, with un (t) ∈ R, and ωn ∈ R. Write the
dt2
equation as first order system and find some phase space where the infinitely many
first order ODEs define a C0 -semigroup. Under which additional assumptions on
the ωn is the semigroup uniformly continuous, differentiable, or analytic?

5.3. Work out the details for Example 5.1.15.

176 5. PDEs on an interval

5.4. Let (H,

·, ·) be a Hilbert space and (en ) an orthonormal basis of H. Let
λn > 0, λn ≤ λn+1 and limn→∞ λn = ∞. Finally, let
 2
V := {u ∈ H : λn
u, en  < ∞}.
n∈N

Show that a(u, v) = n∈N λn
u, en 
v, en  defines a scalar product in V such that
V is a Hilbert space, and V ⊂ H compact.
5.5. Let (H,
·, ·) be a Hilbert space with an orthonormal basis (en )n∈N . Let
λn > 0, λn ≤ λn+1 and limn→∞ λn = ∞. For t ≥ 0 let T (t) : H → H be defined
by T (t)u = n∈N e−λn t
u, en  en . Show that (T (t)) is a C0 -semigroup, and that

limt0 1t (T (t)u − u) =: Au exists iff n∈N λ2n
u, en 2 < ∞.
 t
Finally, for v = n∈N λn
u, en  en show that T (t)u − u = 0 T (s)v ds.
5.6. Prove local existence and uniqueness of solutions for ∂t un = −n4 un + u3n , with
n ∈ Z, in spaces p,θ . Do we have global existence and uniqueness of solutions?
5.7. Consider the discrete NLS equation
i∂t un = ε(un+1 − 2un + un−1 ) + |un |2 un ,
with un (t) ∈ C and n ∈ Z for 0 ≤ ε ≤ 1.
a) Show the local existence and uniqueness of solutions in 2 . Note that the linear
part is not diagonal, but bounded. Show that the system conserves the 2 -norm.
Conclude the global existence and uniqueness of solutions in 2 from this fact.
b) For ε = 0 find non-trivial solutions U0 of the form u0 (t) = r0 eiωt with r0 ∈ R
and un = 0 for all other n ∈ Z. Use the implicit function theorem for instance in
∞ to prove that for ε > 0 there are solutions Uε of the form un (t) = rn eiωt nearby
U0 . What additional information is gained if ∞,θ instead of ∞ is used?
5.8. Show that any f ∈ C 1 (Rd , Rd ) is locally Lipschitz continuous.
5.9. a) Let I ⊂ R be a compact interval and α ∈ (0, 1). Show that C 0,α (I) is a
Banach space.
b) Let I = [0, 1] and 0 < β < α < 1. Show that C 0,α (I) ⊂ C 0,β (I) as a proper
subset.
5.10. Let Ω = (0, 1). Find sequences (un ), un : Ω → R such that:
(a) (un ) bounded in H 1 (Ω), but (un ) does not converge in L2 (Ω).
(b) un → 0 in L2 (Ω) but un (x) → 0 for all x ∈ Ω.
(c) un L2 = 1 for all n ∈ N and un → 0 weakly in L2 (Ω).
Remark. Since H 1 (Ω) ⊂ L2 (Ω) compact, in (a) we always have unk → v in L2 (Ω)
for a subsequence unk .
R
(0). Show that u(x) = ln | ln |x|| ∈ H 1 (Ω) \ L∞ (Ω).
2
5.11. Let Ω = B1/2

5.12. Let d ∈ N and Ω = B1 (0) in Rd . For which α do we have (a) |x|α ∈ H 1 (Ω);
(b) (sin |x|)α ∈ H 1 (Ω); (c) (ln |x|)α ∈ H 1 (Ω) ?
 1/2
5.13. For H = L2 (0, 1) define F : H → R by F (u) = 0 u(x) dx. Do we have
F ∈ H  ? If so, find a representation
·, vL2 of F with v ∈ H.
5.3. The Chafee-Infante problem 177

5.14. For the following PDEs with x ∈ (0, 2π) and periodic boundary conditions
investigate whether the solution operator defines a C0 -semigroup in L2per ((0, 2π), R)
with smoothing properties
a) ∂t u = ∂x4 u, b) ∂t u = −∂x4 u, c) ∂t u = ∂x3 u.
5.15. Consider the complex GL equation
∂t u = (1 + iα)∂x2 u + Ru − (1 + iβ) |u|2 u
with 2π-periodic boundary conditions, u(x, t) ∈ C, and α, β ∈ R. In case |β| < 1/3
prove the global existence of solutions in H 1 for all R ∈ R.
5.16. Consider ∂t u = ∂x2 u + u3 for t ≥ 0, x ∈ (0, π) and u(x, t) ∈ R with boundary
condition u(0, t) = u(π, t) = 0 and initial condition u(x, 0) = φ(x), cf. [Hen81,
Page 49]. Prove that there are solutions which converge
π in finite time towards ∞.
Hint: Derive a differential inequality for s(t) = 0 sin(x)u(x, t)dx. With Hölder’s
π
inequality we obtain s(t) ≤ 22/3 ( 0 sin(x)u3 (x, t)dx)1/3 and so dtd
s ≥ −s + s3 /4.
5.17. Write ∂t u = −∂x4 u + sin(u), with u(x, t) ∈ R, x ∈ R, and t ≥ 0, subject to
periodic boundary conditions u(x, t) = u(x + 2π, t) as a gradient system.
5.18. Consider the Cahn-Hilliard equation ∂t u = ∂x2 (−γ∂x2 u−u+u3 ), with u(x, t) ∈
R, γ > 0, and 2π-periodic boundary conditions.
d
 2π
a) Prove that dt C = 0, where C(t) = 0 u(x, t) dx.
 2π 1 2 
b) Let F (u) = 0 4 (u − 1)2 + γ2 |∂x u|2 dx and show that dt d
F = − |∇w|2 dx
with w = u3 − u − γ∂x2 u.
c) Find the possible ω-limit sets.
 Chapter 6

The Navier-Stokes
equations

6.1. Introduction
In this chapter we give an introduction to the Euler and Navier-Stokes equa-
tions, which over unbounded domains will also play a role in subsequent
chapters. The global existence and uniqueness of solutions of the three-
dimensional (3D) Navier-Stokes equations is one of the seven so called ’one
million dollar’ or millennium problems in mathematics presented by the Clay
Mathematics Institute in the year 2000. There are a number of reasons for
this choice. On the one hand, the solution of this problem would allow us to
understand and simulate the motion of fluids more rigorously. On the other
hand, the 3D Navier-Stokes equations are interesting PDEs which resisted
so far all attempts to prove the global existence and uniqueness of solutions.
Their history goes back a long way. The equations describing the motion
of non-viscous fluids are called Euler equations and have been derived by
Leonhard Euler (1707–1783). The Navier-Stokes equations generalize the
Euler equations and include the case of viscous fluids. They have been
derived independently by a number of people, including Claude-Louis Navier
(1785–1839), George Stokes (1819–1903), Simeon Poisson (1781–1840) and
Jean Claude Saint-Venant (1797–1886).
First we recall the derivation of the Navier-Stokes equations, following
[Fow97, §6]. Then we focus on the analysis of the Navier-Stokes equa-
tions in Ω = Td = Rd /(2πZ)d , i.e., Ω = [0, 2π)d with periodic boundary
conditions. After the presentation of some local existence and uniqueness
results we formulate the global existence question. The local existence and

179
180 6. The Navier-Stokes equations

uniqueness of solutions in some phase space X is obtained by a fixed point

argument in C([0, T0 ], X) for a T0 > 0, using the variation of constant for-
mula. A good choice of the phase space X turns out to be essential in this
construction. The background of the millennium problem is the fact that in
infinite-dimensional spaces there are infinitely many non-equivalent norms.
For the 3D Navier-Stokes equations so far in no phase space simultaneously
the global existence of solutions and their uniqueness can be shown.

6.1.1. Derivation of the Navier-Stokes equations. The Navier-Stokes

equations describe the velocity and the pressure field of an incompressible
fluid. By Newton’s law (force=mass×acceleration) the N molecules of the
fluid satisfy the system of ODEs

mẍj = Fj (x1 , . . . , xN )

for j = 1, . . . , N . The motion of the fluid is completely determined by the

evolution of this system. However, the system is pretty useless due to the
very large number N . Therefore, the fluid is modeled as a continuum. In
doing so we have to guarantee that no molecules are lost, i.e., that mass is
conserved.
The velocity field of the continuum at a position x ∈ Rd at a time t
is denoted by u(x, t) ∈ Rd for d = 2, 3. With ρ = ρ(x, t) ∈ R we denote
similarly the density of the fluid. The Navier-Stokes equations consist of two
equations, a scalar one for the conservation of mass and a second equation
with d components for the conservation of the momentum. In general, by the
internal friction of the fluid heat will be produced which leads to a coupling
of the Navier-Stokes equations with a heat equation. However, here we will
neglect this aspect.
Conservation of mass. We consider a fixed test volume V with surface
S. The total mass in V can only change by the flow through the boundary
S, i.e.,
    
d
d
ρ dV = ∂t ρ dV = − ρu · n dS = − ρuj nj dS
dt V V S S j=1
  
d
=− div(ρu) dV = − ∂xj (ρuj ) dV,
V V j=1

where we used the Gauss integral theorem and where n(x) = (n1 , . . . , nd )(x)
is the outer unit normal in the point x at the boundary S. Since this relation
holds for all test volumes V the integrands must be equal, i.e.,

(6.1) ∂t ρ + div(ρu) = 0.
6.1. Introduction 181

Conservation of momentum. Similarly, the momentum of a test volume

V can only change by the flow through the boundary and by forces, for
instance friction forces, on the surface of the test volume. For these forces
f we assume the existence of a matrix σ = (σij )i,j=1,...,d , the so called stress
tensor, which relates the direction of the outer normal n with the direction
and magnitude of the resulting force f , i.e.,

fi = σij nj .
j=1,...,d

For examples see below. Hence, we obtain for the change of the momentum
   d  d
d
ρui dV = − (ρui )uj nj dS + σij nj dS.
dt V S S
j=1 j=1

Application of the Gauss integral theorem and the above arguments yield
∂t (ρui ) + div(ρui u) = div(σi· ).
Using conservation of mass gives

∂t (ρui ) = ρ∂t ui + ui ∂t ρ = ρ∂t ui − ui ∂xj (ρuj ),
j=1,...,d

thus, in vector notation,

(6.2) ρ[∂t ui + (u · ∇)ui ] = ∇ · σi·
or, in coordinates,
⎡ ⎤

d 
d
ρ ⎣∂t ui + (uj · ∂xj )ui ⎦ = ∂xj σij .
j=1 j=1

Constitutive laws. In order to obtain a closed set of equations from (6.1)

und (6.2) we need to know how the stress tensor σ depends on the velocity
u and the density ρ. Such a relation σ = σ(u, ρ) is called a constitutive
law and depends on the fluid under consideration, i.e., the function differs
strongly for instance between water and honey. It is possible that σ(t) is
not only a function of (u, ρ)(t), but depends on the whole history of (u, ρ),
cf. [Ren00].
i) For a non-viscous fluid, i.e., for a fluid without internal friction, the
surface forces f are only due to pressure. Hence, the forces f are parallel to
the outer normals n and their magnitude does not depend on the direction,
i.e.,

d 
d
(6.3) fi = σij nj = − pδij nj = −pni ,
j=1 j=1
182 6. The Navier-Stokes equations

which yields the constitutive law

σij = −pδij ,
where δij = 1 for i = j and δij = 0 else. Since

d 
d
∂xj σij = − ∂xj (pδij ) = −∂xi p
j=1 j=1

we obtain for non-viscous fluids

1
(6.4) ∂t u + (u · ∇)u = − ∇p.
ρ
The system consisting of (6.1) and (6.4) is still not closed. We have to relate
p with the velocity u and density ρ. A typical choice would be p = cργ with
some constants c > 0 and γ ≥ 1. In case of incompressible fluids, i.e.,
ρ = const., we obtain Euler’s equations
∂t u + (u · ∇)u = − ρ 1 ∇p,
(6.5)
∇ · u = 0.
ii) For a viscous fluid the constitutive law is given by
σij = −pδij + τij ,
where τij will model the internal friction. In order to find a model for τij as
simple as possible we consider a stationary constant density shear flow with
velocity field (u1 (x2 ), 0), see Figure 6.1. The friction forces act on the top
surface and bottom surface, and they are proportional to the difference of
these velocities. Hence, for an infinitesimal small test volume we find that
the surface force f is proportional to the so called strain ∂x2 u1 . Since the
friction forces are perpendicular to the top surface we find
τ12 = μ∂x2 u1 ,
where μ > 0 is called the dynamic viscosity. Due to the isotropy of the fluid
σ must be symmetric. This finally leads to

d
τij = 2με̇ij + λ ε̇kk δij
k=1
where
1
ε̇ij = (∂xj ui + ∂xi uj ).
2
The constants μ for the shear and λ for the compressions are called the
Lamé-constants.
Remark 6.1.1. There is no reason why in our modeling σ should only
depend linearly on the first derivatives of u. An answer to the millennium
problem can lead to some corrections at this point of the modeling.
6.1. Introduction 183

V
x2

x1
Figure 6.1. Shear flow: the internal friction is proportional to ∂x2 u1 .

The Navier-Stokes equations. For air, compressibility is an important

issue. However, for fluids such as water, ρ can be considered to be a constant,
i.e., ∂t ρ = 0. As in (6.5) the conservation of mass then simplifies to


d
∂xj uj = ∇ · u = divu = 0.
j=1
d d
From k=1 ε̇kk = k=1 ∂xk uk = 0 we obtain τij = 2με̇ij . Moreover, we find


d 
d 
d 
d 
d
∂xj τij = ∂xj (∂xj ui + ∂xi uj ) = ∂x2j ui + ∂xi ( ∂xj uj ) = ∂x2j ui ,
j=1 j=1 j=1 j=1 j=1

such that
1
∂t u + (u · ∇)u = − ∇p + νΔu,
ρ
where ν = μ/ρ is called the kinematic viscosity. The Navier-Stokes equations
are then given by
1
∂t u + (u · ∇)u = − ∇p + νΔu,
(6.6) ρ
∇ · u = 0.

In order to handle these equations as a dynamical system boundary

conditions have to be added. At rigid boundaries, a viscous fluid satisfies u =
0. At free surfaces, boundary conditions involve for instance the prescription
of stresses. In this section we consider the Navier-Stokes equations with
periodic boundary conditions. These have no physical meaning, but allow
us to focus on the equations itself.
In order to eliminate the physical units from the Navier-Stokes equations,
let U be a typical velocity and l be a typical length of the flow. We set

u = U u∗ , x = lx∗ , p = ρU 2 p∗ , t = lt∗ /U,

184 6. The Navier-Stokes equations

and obtain after dropping the ∗ the Navier-Stokes equations in dimensionless

form
1
∂t u + (u · ∇)u = −∇p + Δu, ∇ · u = 0,
R
where R = U l/ν is called the Reynolds number. The larger R, the more
complex the flow. Until further notice we assume w.l.o.g. for our purposes
that R = 1.

6.1.2. The vorticity, and some explicit solutions. The vorticity ω of

the 3D velocity field u is defined by
⎛ ⎞
∂x2 u3 − ∂x3 u2
ω = ∇ × u = ⎝∂x3 u1 − ∂x1 u3 ⎠ ,
∂x1 u2 − ∂x2 u1
while for 2D flows the vorticity is the scalar ω = ∂x1 u2 − ∂x2 u1 . Applying
the curl-operator ∇× to the Navier-Stokes equations gives in R3 that
(6.7) ∂t ω = νΔω − (u · ∇)ω + (ω · ∇)u.
Specializing (6.7) to two-dimensional flows shows ω ⊥ ∇u, and so in R2 we
have
(6.8) ∂t ω = νΔω − (u · ∇)ω.
The pressure gradient has vanished from (6.7) and (6.8). On the other hand,
the velocity u still appears and has to be reconstructed from ω by solving
the PDEs
ω =∇×u and ∇ · u = 0.
If Ω = Rd the solution is given by the Biot-Savart law [FLS64, II-14-10]
from magnetostatics, cf. Exercise 6.3.
There are major differences between the 2D case and the 3D case. Beside
the diffusion and transport of vorticity which appear in 2D and 3D, in 3D
there is also the production term (ω·∇)u for vorticity. Hence, it is likely that
the differences in the global existence and uniqueness question in 2D and 3D
are not only an artificial functional analytic problem. In fact, experiments
and simulations, cf. the discussion in [GW06a], show that in 3D smaller
and smaller vortices are created, whereas in 2D the smaller vortices vanish
and are eaten up by the larger ones.
In Rd or Td we have that ω = 0 is a solution of the vorticity equation.
The fact that vorticity is preserved by the motion of the fluid together with
the incompressibility of the fluid allows us to construct a number of non-
trivial solutions for the Navier-Stokes equations. The existence of a potential
Φ : Rd → R with u = ∇Φ follows from ω = ∇ × u = 0. The potential Φ
satisfies ΔΦ = 0 due to the incompressibility of the fluid. Such flows are
6.1. Introduction 185

called potential flows. However, in general the boundaries of the physical

domain will create vorticity.
Remark 6.1.2. The conditions
(6.9) ∇ × u = 0, and ∇·u=0
can be interpreted in two space dimensions, i.e. u = (u1 , u2 ), as the Cauchy-
Riemann differential equations of a complex valued function z → w(z) de-
fined through w(x1 + ix2 ) = u1 (x1 , x2 ) − iu2 (x1 , x2 ). From complex analysis
it is well known that the complex differentiability of w, together with (6.9),
implies the analyticity of w, respectively u.

Here are a number of examples for potential flows. Further interesting

and more complicated exact solutions of the Navier-Stokes equations can for
instance be found in [MB02].
Example 6.1.3. We already encountered the explicit constant shear flow
solution (2D)
   
u1 (x2 ) cx2
u(x, t) = = , p(x, t) = p0 ,
0 0
with p0 some constant. This is an example of a parallel or laminar flow
with non-vanishing vorticity. Further examples are Couette and Poiseuille
flow. The latter describes the flow in an infinitely long pipe Ω = R × Σ,
where Σ ⊂ Rd is a bounded cross-section with rigid boundary conditions
u|∂Σ⊂Rd = 0. As an explicit example we again consider the 2D case and set
Σ = (−1, 1). Then
 
u1 (x2 )
u(x, t) = , with u1 (x2 ) = c(x22 − 1), c ∈ R,
0
is an exact solution, see Figure 6.2a), with the pressure given by p(x, t) =
p0 − Rc x1 .

An important property of all parallel flows is that the nonlinear term

(u · ∇)u of the Navier-Stokes equations drops out. As a consequence, in
Rd parallel flows can always be superimposed. This also works in general
domains if the boundary conditions permit it.
Example 6.1.4. In 2D so called irrotational strain flows are given by
 
−x1 γ2
u(x, t) = γ , p(x, t) = p0 − (x21 + x22 ),
x2 2
see Figure 6.2b), while vortices are given by
 
−x2 ω2
u(x, t) = ω0 , p = p0 + 0 (x21 + x22 ),
x1 2
186 6. The Navier-Stokes equations

see Figure 6.2c). A 3D flow generalizing the 2D strain flow is the irrotational
stationary jet
⎛ ⎞
−γ1 x1
1
u(x, t) = ⎝ −γ2 x2 ⎠ , p(x, t) = p0 − (γ12 x21 + γ22 x22 + (γ1 + γ2 )2 x23 ).
2
(γ1 + γ2 )x3

a) b) c)
x 2 ,u 2 x 2 ,u 2 x 2 ,u 2

x 1 ,u 1 x 1 ,u 1
x 1 ,u 1

Figure 6.2. Three planar exact solutions: Poiseuille flow, strain flow,
a vortex.

To get a feeling for the behavior of general solutions of the Navier-Stokes

equation we recommend to do some numerical experiments. For this we refer
to [Uec09] and the matlab scripts provided and explained therein, including
some brief discussion of turbulence.

6.2. The equations on a torus

6.2.1. Local existence and uniqueness. We remind the reader of basic
problems with existence and uniqueness of solutions of ODEs and PDEs. In
Example 2.2.4 we saw that the scalar equation u̇ = |u| with the initial
condition u(0) = 0 has infinitely many different solutions. Two examples
are u(t) = 0 and u(t) = t2 /4. In Example 2.2.5 we saw that the unique
solution u(t) = tan(t) of u̇ = (1 + u2 ) with the initial condition u(0) = 0
becomes unbounded for t = π/2.
We study the local existence and uniqueness of solutions of the Navier-
Stokes equations
(6.10) ∂t u + (u · ∇)u = −∇p + Δu, ∇ · u = 0.
We follow the formulation of the millennium problem [Fef06] and consider
the Navier-Stokes equations in Td , i.e., in [0, 2π)d with periodic boundary
conditions. The phase space is chosen in such a way that its elements sat-
isfy the boundary conditions. The Navier-Stokes equations have the special
difficulty that the second equation, ∇ · u = 0, is without a time derivative,
6.2. The equations on a torus 187

and that the variable p occurs without time derivative at all. This problem
is solved by prescribing the equation ∇ · u = 0 as additional condition in the
definition of the phase space. The term −∇p in the first equation will be
interpreted as projection P onto the divergence free vector fields such that
the Navier-Stokes equations can be written as
∂t u = P Δu − P (u · ∇)u.
For periodic boundary conditions we will have P Δ = ΔP such that we
finally have to consider ∂t u = Δu − P (u · ∇)u in the space of divergence free
vector fields {u : u = P u}.
The Navier-Stokes equations are semi-linear parabolic differential equa-
tions such that for the construction of local solutions in time we use again
the variation of constant formula
 t
(6.11) u(t) = e u(0) −
tΔ
e(t−s)Δ (P (u · ∇)u)(s) ds
0

and the scheme introduced in §5.1.4. The semigroup etΔ generated by the
linear part is smoothing, i.e., u0 ∈ L2 implies that tm/2 ∂xm u(t) is bounded
in L2 (Td ) for every t > 0 and m ∈ N. Semi-linear means here that the
nonlinearity only contains terms with less derivatives than in the linear part,
i.e., for the Navier-Stokes equations first derivatives in (u · ∇)u compared
with the second order derivatives in Δu. More precisely, we will prove that
etΔ maps H m into H m+1 with a singularity t−1/2 and that P (u · ∇)u is a
bilinear map from H m+1 × H m+1 → H m . Then all assumptions following
(5.21) are satisfied and the local existence and uniqueness Theorem 5.2.22
will apply. Hence, the major step is to give a precise definition of P and to
investigate its analytic properties.
The projection on the divergence free vector fields. We define the
projection P via the solution v = P f of the system of PDEs
(6.12) v + ∇p = f, ∇ · v = 0,

where f : Rd → Rd should be thought as a placeholder for the nonlinear

terms (u · ∇)u. For notational simplicity we carry out the analysis only in
case x ∈ R2 with 2π-periodic boundary conditions. In order to solve (6.12)
we make an expansion in Fourier series
  
vj (x) = vj,k eik·x , p(x) = pk eik·x , fj (x) = fj,k eik·x ,
k∈Z2 k∈Z2 k∈Z2

with vj,k , fj,k , pk ∈ C for j = 1, 2. Plugging this into (6.12) yields

(6.13) v1,k + ik1 pk = f1,k , v2,k + ik2 pk = f2,k , ik1 v1,k + ik2 v2,k = 0.
188 6. The Navier-Stokes equations

In case |k| = 0 we find the solution

⎛ ⎞ ⎛ ⎞⎛ ⎞
v1,k k22 −k1 k2 −ik1 f1,k
⎝ v2,k ⎠ = 1 ⎝ −k1 k2 k12 −ik2 ⎠ ⎝ f2,k ⎠ .
k12 + k22
pk −ik1 −ik2 1 0
For the subspace k1 = k2 = 0 there are two possibilities.
Case i). We prescribe the periodicity of the pressure p. Then we have
that vj,0 = fj,0 in (6.13) and that p0 is arbitrary which is no problem since
only ∇p occurs in the Navier-Stokes equations. This choice can lead to a
non-vanishing mean flow.

Case ii). We require that the mean flows Ω vj (x1 , x2 ) dx vanish for
j = 1, 2, i.e., v1,0 = v2,0 = 0. In order to do so we consider a pressure

2
p(x, t) = αj xj + p(x, t),
j=1

where p(x, t) is 2π-periodic w.r.t. the xj . Then ∂xj p(x, t) = αj + ∂xj p(x, t)
and so vj,0 +αj = fj,0 . Thus, to a fj,0 we always find an αj such that vj,0 = 0.
Example 6.2.1. To illustrate the difference between Case i) and Case ii),
we consider the vector field
 
−2(x1 −π)2 −2(x2 −π)2 2 + tanh(x2 − π)
f (x1 , x2 ) = e .
0
In Figure 6.3 we show the different effects of choosing i) or ii).

Figure 6.3. Illustration of the difference between cases i) and ii), con-
cerning the boundary conditions for the pressure in the definition of the
projection u = P f for f from Remark 6.2.1, via a plot of the pressure
function. In the left panel we require a periodic pressure, giving a mean
flow in P f , while in the right panel we require zero mean flow, giving a
linear growth of the pressure p.
6.2. The equations on a torus 189

Choosing between i) and ii) is a question of modeling. i) has the disad-

vantage that if a constant mean force is added to the Navier-Stokes equations
then this choice leads to unbounded growth of (laminar) mean flows. There-
fore, in the following we opt for ii), i.e., vj,0 = 0 for j = 1, 2 and define the
projection P as direct sum of the projections Pk , i.e., vk = (Pf)k = Pk fk ,
where
    
v1,k 1 k22 −k1 k2 f1,k
(6.14) v0 = 0 and = 2
v2,k k + k2
1 2
−k1 k2 k12 f2,k

for k = 0. In physical space we define P by P = F −1 P F .

Lemma 6.2.2. The projection P is a bounded linear map in 1,m and in
2,m , i.e., for all m ∈ R there exists a C > 0 such that
Pf1,m ≤ Cf1,m and Pf2,m ≤ Cf2,m .
m , i.e., for all
Hence, the projection P is also a bounded linear map in Hper
m ∈ R there exists a C > 0 such that
P f H m ≤ Cf H m .

Proof. We find
Pf1,m = (Pk fk )k∈Z2 1,m ≤ sup Pk  (fk )k∈Z2 1,m ≤ Cf1,m .
k∈Z2
The proof for 2,m works exactly the same. Using that Fourier transform is
an isomorphism between 2,m and Hperm , cf. Lemma 5.2.9, yields

P f H m ≤ C1 P f2,m ≤ C1 Cf2,m ≤ C1 CC2 f H m . 

Remark 6.2.3. a) Lemma 6.2.2 is valid in arbitrary dimensions d ≥ 2.
b) In Fourier space we have vk ∈ Vk := {vk ∈ Cd : k · vk = 0}, due to
∇ · v = 0. The pressure gradient ∇p defines in each Cd a vector i
pk k, with
pk ∈ C, which is orthogonal to Vk . This property can be generalized to
general domains Ω ⊂ Rd , cf. §6.3.

The phase space and the fixed point argument. In the following we
will solve the Fourier transformed Navier-Stokes equations in two classes of
phase spaces, namely
div = {
1,m  = Pu
u ∈ (1,m )d : u }, div = {  = P u
u ∈ (2,m )d : u
2,m },
and the Navier-Stokes equations in physical space in the class of phase spaces
div,m
Hper = {u ∈ (Hper
m d
) : u = P u}.
In case of periodic boundary conditions we have P Δ = ΔP . Hence, for u
with u = P u the solution operator of the linearized Navier-Stokes equations
is given coordinate-wise by the solution operator etΔ of the linear diffusion
190 6. The Navier-Stokes equations

equation ∂t u = Δu. In one space dimension this operator has been discussed
a number of times, cf. Example 5.1.21 and Example 5.2.19. The statements
made above about this operator transfer line to line from R1 to Rd .

Theorem 6.2.4. The solution operator (e−|k| t )k∈Zd of the linearized Navier-
2

Stokes equations ∂t u k = −|k|2 u k in Fourier space defines a C0 -semigroup

in 1,m and 2,m for all m ∈ R. The associated solution operator etΔ =
div div

F −1 (e−|k| t )k∈Zd F of the linearized Navier-Stokes equations in physical space

2

div,m
defines a C0 -semigroup in Hper . Moreover, for all m ∈ R and r ≥ 0 there
exists a C > 0, such that
(e−|k| t )k∈Zd u ≤C max(1, t−r/2 )
2
div udiv ,
1,m+r 1,m

−|k|2 t −r/2
(e div
)k∈Zd u ≤C max(1, t udiv ,
)
2,m+r 2,m
−r/2
e uH div,m+r ≤C max(1, t
tΔ
)uH div,m .
per per

Since we already analyzed P in §6.2.1 it remains to bound the nonlinear

div,m m−1 if m is
term (u·∇)u). It is easy to see that it is smooth from Hper to Hper
sufficiently large. Hence, with the previous estimates the local existence and
div,m
uniqueness of solutions easily follows in every Hper if m is sufficiently large.
However, since we are also interested in the global existence and uniqueness
of solutions, and since it is more easy to obtain a priori estimates on the
div,m
solutions in Hper -spaces for small m, we would like to have m as small as
possible. Hence, we spend a little bit more time at this point to optimize
the estimates.
We use the variation of constant formula (6.11) to prove local existence
and uniqueness of solutions. The key ingredients are the smoothing proper-
ties of the semigroup etΔ summarized in Theorem 6.2.4 and the Lipschitz-
continuity of the nonlinear terms P (u · ∇)u). Since a singularity t−1+δ , with
δ > 0, is integrable, our approach also works if the nonlinearity P ((u · ∇)u)
div,m div,m−2+δ
is Lipschitz-continuous from Hper into Hper . By using the incom-
d
pressibility j=1 ∂xj uj = 0 the l component of (u · ∇)u can be written
th

as

d 
d 
d 
d
uj ∂xj ul = uj ∂xj ul + ul ∂xj uj = ∂xj (uj ul ),
j=1 j=1 j=1 j=1

or equivalently in vector notation as

(6.15) (u · ∇)u = ∇ · (uuT ).
Using this representation we have to establish
∇ · uv T H m−2+δ ≤ CuH m vH m ,
6.2. The equations on a torus 191

or equivalently
(6.16) uv T H m−1+δ ≤ CuH m vH m
for a δ > 0 and all m > m∗ for a m∗ ∈ R. To find the minimal m∗ we start
with the following lemma.
 ∈ R, r ≥ 0, and δ > 0 there exists a C > 0 such
Lemma 6.2.5. For all m
that
u ∗ v2,m ≤ C (
 u2,m+r


v  + 
u 
v 2,m+r

).
2, d
2 −r+δ 2, d
2 −r+δ

Proof. We define (k) = (1 + |k|)m ,  

1 (k) = (1 + |k|)m+r , and 2 (k) =
−r
(1 + |k|) , for which we have the inequality
 
(k) ≤(1+|k|)m ≤ C((1+|k − l|)m+r (1+|l|)−r +(1+|l|)m+r

(1+|k − l|)−r )
1 (k − )
≤C( 2 () + 2 (k − )
1 ()).
Using this and Lemma 5.1.26 yields

u ∗ v2,m = u ∗ v)2 = 2 (
( u1 ) ∗ (
v 2 ) 2 + 2 (
u2 ) ∗ (
v 1 ) 2
≤2
u1 2 
v 2 1 + 2u2 1 
v 1 2
 
=2C  u2,m+r


v 1,−r + 2 u1,−r  v 2,m+r

 
 
≤2C C u2,m+r 
v  d +  u d v 2,m+r
 2, 2 −r+δ 2, 2 −r+δ 

for a δ > 0 according to Sobolev’s embedding, cf. Lemma 5.1.27. 

For the validity of (6.16) we have to choose r = 1 − δ for a δ > 0 such
that m∗ = d/2−1 and use that Fourier transform is an isomorphism between
m .
2,m and Hper
Remark 6.2.6. (The critical Sobolev number m∗ for 1,m ) For all m ≥
0 there exists a constant C such that (1+|k +l|)m ≤ C((1+|k|)m +(1+|l|)m )
for all k and l. Thus, we have  u ∗ v1,m ≤ C
u1,m 
v 1,m which yields
∗
m ≥ m = 0.

Therefore, we obtain the following local existence and uniqueness result.

0 ∈ div ∗
Theorem 6.2.7. a) Let u 2,m for m > m = d/2 − 1. Then there exists
u0 div ) > 0 and unique mild solution u
a T0 = T0 (  ∈ C([0, T0 ], div
2,m ) of the
2,m
Fourier transformed Navier-Stokes equations (6.11) with u |t=0 = u 0 . The
same is true if div is replaced by  div and m ≥ 0.
2,m 1,m

b) Let u0 ∈ Hper div,m

for m > m∗ = d/2 − 1. Then there exists a T0 =
div,m
T0 (u0 H div,m ) > 0 and unique mild solution u ∈ C([0, T0 ], Hper ) of the
per
Navier-Stokes equations (6.11) with u|t=0 = u0 .
192 6. The Navier-Stokes equations

6.2.2. Analyticity of solutions. As pointed out in §5.3.3 the solutions of

semi-linear equations where the semigroup is smoothing are infinitely often
differentiable for every t > 0. The scheme which has been explained in §5.3.3
can also be applied to the Navier-Stokes equations if m > m∗ . However, the
step size for increasing the differentiability has to be decreased from 1 to
δ/2. Actually, the solution becomes analytic in a strip along the real axis in
the complex plane if m > m∗ . In order to prove this we define
Definition 6.2.8. For β ≥ 0 let

ω1,β = {
u : Zd → C : 
uω1,β < ∞} and 
uω1,β = |
uk |eβ|k| .
k∈Zd

 ∈ ω1,β then u = F −1 u
Lemma 6.2.9. If u  is analytic in a strip
Sβ = {z = (z1 , . . . , zd ) ∈ Cd : max |Imzj | < β}
j=1,...,d

in Cd and supz∈Sβ |u(z)| ≤ 

uω1,β .

Proof. The estimate

 
(6.17) sup |u(z)| ≤ uk | sup |eikz |) ≤
(| uk |eβ|k| ) ≤ 
(| uω1,β < ∞
z∈Sβ z∈Sβ
k∈Zd k∈Zd

shows that the function u(z) = k∈Zd u k eikz is the uniform limit of the an-

alytic functions z → sN (z) = |k|≤N u k eikz in their domain of convergence
Sβ and so the function u is analytic in Sβ . 
For every t > 0 and β ≥ 0 the linear Navier-Stokes semigroup T(t)
defined by (T(t)u)k = e−|k| t u
2
k maps 1 into ω1,β due to
 
T(t) |e−|k| t u
k eβ|k| | ≤ sup |e−|k| t eβ|k| |
2 2
uω1,β ≤ |
uk | ≤ C(β, t)
u1 ,
k∈Zd k∈Zd k∈Zd

where supk∈Zd |e −|k|2 t

eβ|k| | ≤ C(β, t) < ∞, since e−|k| t eβ|k| → 0 for |k| →
2

∞ if t > 0. The constant C = C(β, t) satisfies C(β, t) → ∞ for t →

0 with a non-integrable singularity which makes this estimate useless for
nonlinear problems.
√ However, this singularity can be avoided if we choose
β proportional to t.
Lemma 6.2.10. There exists a constant C < ∞ such that the semigroup
T(t) defined by (T(t)
u)k = e−|k| t u
2
k satisfies for all t ≥ 0 that
T(t)
uω √ ≤ C
u1 .
1, t

Proof. The assertion follows since

 √ 
|e−|k| t e t|k| | ≤ |e−|s| e|s| | ≤ C < ∞
2 2

k∈Zd s∈Rd
can be bounded independently of t. 
6.2. The equations on a torus 193

This estimate can be used to prove the analyticity of the solutions of

Navier-Stokes equations w.r.t. x in S√t ⊂ C for t > 0. According to Lemma
(t) ∈ ω1,√t for t > 0.
6.2.9 this assertion follows if u

Lemma 6.2.11. For every M > 0 there exists a T0 > 0 such that the map
 t
−k2 t
e−k (t−τ ) P(k)ik(
2
F (
u)(k, t) = e (k, 0) +
u u∗u
T )(k, τ ) dτ
0
is a contraction in
u : Z × [0, T0 ] → C : 
X = { u−u
lin X ≤ M },
lin (k, t) = e−k t u
2
uX = supt∈[0,T0 ] 
where  u(t)ω √ and where u (k, 0).
1, t

u ∗ vω √ ≤ 
Proof. It is easy to see, cf. Exercise 6.4, that  uω √ 
v  ω √
1, t 1, t 1, t
which implies
(6.18) u ∗ vX ≤ uX vX .
√ √ √
Using this estimate and t − s ≥ t − s shows that
  t √
e−k (t−τ ) P(k)ik(
2
F (
u) − ulin X ≤ sup | T )(k, τ ) dτ |e t|k|
u∗u
t∈[0,T0 ] k∈Z 0
 t √ √ √
|e−k P(k)ik( t|k| − τ |k| τ |k|
2 (t−τ )
≤ sup u∗u
T )(k, τ )e e e | dτ
t∈[0,T0 ] k∈Z 0
 t √ √
|e−k P(k)ike t−τ |k| τ |k|
2 (t−τ )
≤ sup u∗u
( T )(k, τ )e | dτ
t∈[0,T0 ] 0 k∈Z
 t √  √
sup |e−k t−τ |k| τ |k|
2 (t−τ )
≤C sup ike | dτ sup |(
u∗u
T )(k, τ )e |
t∈[0,T0 ] 0 k∈Z s∈[0,T0 ] k∈Z
1/2
≤CT0 
u2X < ∞.
Thus, we proved that F maps the space X into itself if T0 > 0 is sufficiently
small. The proof of the contraction property works the same way. 
Corollary 6.2.12. For all C > 0 there exists a T0 > 0 such that the solu-
tions u of the Navier-Stokes equations are analytic w.r.t. x in S√t ⊂ C for
all t ∈ [0, T0 ] if 
u(0)1 ≤ C.
Remark 6.2.13. It is easy to see that Theorem 5.1.23 can be generalized to
ω1,β -spaces and the Fourier transformed Navier-Stokes equations such that
there is local existence and uniqueness in ω1,β , too.
Remark 6.2.14. In general it cannot be expected that in nonlinear prob-
lems the strip of analyticity is arbitrarily wide. An explicit, but typi-
cal, example is the z → tanh(z) equilibrium of the Allen-Cahn equation
194 6. The Navier-Stokes equations

∂t u = ∂x2 u + u − u3 in the subsequent §7.2. The function tanh has singulari-

ties in the complex plane, due to tanh(iy) = tan(y), for z = iπ/2 + ikπ with
k ∈ Z.

6.2.3. Global existence in 2D. In order to prove the global existence and
uniqueness of solutions of the Navier-Stokes equations in a phase space X
we need a local existence and uniqueness result in X and a priori bounds for
the solutions in X. Then as explained already a number of times the local
existence and uniqueness result can be applied again and again to construct
a solution for all t ≥ 0.
Bounds for the L2 -norm. The L2 -norm of the solutions u of the Navier-
Stokes equations can be bound in every space dimension. By using integra-

tion by parts and the incompressibility dj=1 ∂j uj = 0 we obtain
  d   d
1 d
uj uj dx = uj ∂t uj dx
2 dt Td Td
j=1 j=1
 
d 
d 
d
= uj (∂x ∂x uj − ∂j p − u ∂x uj ) dx
Td j=1 =1 =1

= − (∂x uj )(∂x uj ) dx.
Td
Poincaré’s inequality, cf. Lemma 5.2.17, implies
Lemma 6.2.15. For all d ≥ 2 we have
1 d
u2L2 ≤ −u2L2 , and so u(t)L2 ≤ e−t u(0)L2 .
2 dt
Bounds for the H 1 -norm. In R2 with periodic boundary conditions also
the H 1 -norm
can be bound. By using integration by parts and the incom-
2
pressibility j=1 ∂j uj = 0, in R2 we find after some explicit calculation, cf.
Exercise 6.5, that
  2 2  2
(6.19) (∂xj u )∂xj (um ∂xm u ) = 0.
T2 j=1 =1 m=1

Therefore, we find as above

  2 2   2 
2 
2
1 d
(∂xj u )(∂xj u ) dx = − (∂xj ∂xk u )(∂xj ∂xk u ) dx.
2 dt T2 T2
j=1 =1 j=1 k=1 =1
Again with Poincaré’s inequality, cf. Lemma 5.2.17, we find
  2 2   2 
2
1 d
(∂xj u )(∂xj u ) dx = − (∂xj u )(∂xj u ) dx.
2 dt T2 T2
j=1 =1 j=1 =1

Combining this estimate with the L2 -estimate yields

6.2. The equations on a torus 195

Lemma 6.2.16. For d = 2 (and only for d = 2 and periodic boundary

conditions) we have
1 d
u(t)2H 1 ≤ −u(t)2H 1 and so u(t)H 1 ≤ e−t u(0)H 1 .
2 dt
Combining this a priori estimate with the previous local existence and
uniqueness result in R2 yields
div,1
Theorem 6.2.17. Let u0 ∈ Hper (T2 ). Then there exists a unique mild
div,1 2
solution u ∈ C([0, ∞), Hper (T )) of the Navier-Stokes equations (6.11) with
u|t=0 = u0 . Moreover, for t > 0 the solution is an analytic function.

The same is true for every Hper div,m

(T2 ) with m > m∗ = 0. For m ∈
(m∗ , 1) the result follows from local existence and uniqueness of solutions in
div,m
such Hper (R2 ) and due to u(t)H div,m (T2 ) ≤ u(t)H div,1 (T2 ) for such m.
per per
For m > 1 we have
(6.20) uv T H m ≤ CuH m vH m
div,m
and the local existence and uniqueness in Hper (T2 ). Arguing as in §5.3.3
yields an a priori-bound for u(t)H div,m (T2 ) in terms of u(t − T0 )H div,1 (T2 )
per per
for all t ≥ T0 . Since u(t − T0 )H div,1 (T2 ) is globally bounded and decays to
per
div,m
zero, the local existence and uniqueness result in Hper (T2 ) can be applied
again and again to construct a solution for all t ≥ 0. Therefore, we have the
following theorem about the asymptotic stability of the zero solution.
Theorem 6.2.18. Let u0 ∈ Hper div,m
(T2 ) with m > m∗ = 0. Then there
div,m 2
exists a unique mild solution u ∈ C([0, ∞), Hper (T )) of the Navier-Stokes
equations (6.11) with u|t=0 = u0 . Moreover, for t > 0 the solution is an
analytic function and satisfies
lim u(t)H div,m (T2 ) = 0
t→∞ per

with some exponential rate.

6.2.4. The millennium problem. Figure 6.4 displays the Sobolev num-
bers for the local existence and uniqueness and the a priori estimates.

d=2 d=3
0 1 0 1/2
Figure 6.4. A priori estimates in L for d = 3 and H 1 for d = 2. Local
2

existence and uniqueness holds in H m for m > d/2 − 1. For d = 2

there is no gap between the exponents m for which we have a priori
estimates and the ones for which we have local existence and uniqueness
of solutions. For d = 3 a gap remains, and global existence of smooth
solutions cannot be concluded.
196 6. The Navier-Stokes equations

In R3 we have no a priori estimate for the spaces where we have local

existence and uniqueness. Hence, global existence of smooth solutions is
unclear. This question is exactly the content of the millennium problem
formulated in [Fef06]:

Millennium problem of the Clay-foundation. Prove (or disprove) the

global existence and uniqueness of solutions of the Navier-Stokes equations
in three space dimensions. For instance, show T0 = ∞ in Theorem 6.2.7,
i.e., close the gap between the a priori estimates and the local existence- and
uniqueness theorem, respectively.

Partial results are already known, from which we list only the two abso-
lute basic ones.
• Jean Leray [Ler34] proved the global existence of so called weak so-
lutions, cf. §7.4.2, of the Navier-Stokes equations. These solutions
are very rough and they are not unique.
• For small initial conditions due to the linear stability of the ori-
gin one easily obtains an a priori estimate and thus it follows
limt→∞ u(t)H div,m (T3 ) = 0 with some exponential rate for every
per
m > m∗ = 1/2.

6.2.5. Some qualitative theory. The 2D Navier-Stokes equations with-

out forcing are a bit boring since u(t)H 1 → 0 for t → ∞, cf. Lemma 6.2.16.
Thus, we show the existence of a global attractor for the two-dimensional
Navier-Stokes equations with forcing, i.e.,
(6.21) ∂t u = Δu − ∇p − (u · ∇)u + f, ∇ · u = 0,

with 2π-periodic boundary conditions and external (time-independent) force

f ∈ L2 with f0 = 0. As above we find
1 d 1 1
u2H 1 ≤ −u2H 1 + uL2 f L2 ≤ − u2H 1 + f 2L2
2 dt 2 2
using uL2 f L2 ≤ 12 u2L2 + 12 f 2L2 , and so

u(t)H 1 ≤ e−t u0 H 1 + f L2 (1 − e−t ).

Hence, the set
B = {u ∈ H 1 : uH 1 ≤ 2f H 1 }
is absorbing, i.e., attracts balls of finite size in finite times. Moreover, it is
(positively) invariant under the flow of (6.21), i.e., u0 ∈ B implies u(t, u0 ) ∈
B for all t ≥ 0.
6.3. Other boundary conditions and more general domains 197

Since the embedding H 1 → L2 is compact (cf. Theorem 5.1.33), as in

Theorem 5.3.4 we thus obtain the existence of the global attractor (in L2 ),
given by
*
A= At , with At = St (B).
t≥0

Theorem 6.2.19. The 2D Navier-Stokes equations (6.21) have a non-

empty, compact, time-invariant set A ⊂ L2 , the global attractor, with
distL2 (St (B), A) = sup inf a − bL2 → 0 as t → ∞.
b∈St (B) a∈A

Remark 6.2.20. With a little more work it can be shown that A is a Hper 1

attractor for (6.21), i.e., A is compact in H 1 and attracts in H 1 , i.e.,

distH 1 (St (B), A) = sup inf a − bH 1 → 0 as t → ∞.
b∈St (B) a∈A
2 .
To show this, use a priori estimates to obtain an absorbing set in Hper

6.3. Other boundary conditions and more general domains

In this section we consider the Navier-Stokes equations with Dirichlet bound-
ary conditions u|∂Ω = 0 in an open domain Ω ⊂ Rd with smooth boundary.
In order to prove the local existence and uniqueness of solutions we general-
ize our previous approach and recall the basics of analytic semigroup theory,
cf. [Hen81]. We recall that the resolvent set of a (bounded or unbounded)
linear operator A : D(A) ⊂ X → X is defined as the set of all λ ∈ C for
which (λ − A) has a bounded inverse (λ − A)−1 : X → X, the resolvent.
Definition 6.3.1. A closed and densely defined operator A in a Banach
space X is called sectorial if there exists an a ∈ R, a φ ∈ (0, π2 ) and an
M ≥ 1, such that the sector
Sa,φ = { λ : φ ≤ | arg(λ − a)| ≤ π, λ = a }
is a part of the resolvent set of A, and such that for all λ ∈ Sa,φ we have the
estimate
M
(λ − A)−1 X→X ≤ .
|λ − a|

The negative of a sectorial operator generates an analytic semigroup.

Definition 6.3.2. A C0 -semigroup T (t) of bounded linear operators is called
analytic if t → T (t)u is analytic for 0 < t < ∞ and all u ∈ X.

The following theorem gives an explicit construction of the semigroup

generated by the negative of a sectorial operator.
198 6. The Navier-Stokes equations

Theorem 6.3.3. Let A be a sectorial operator. Then −A generates an

analytic semigroup with the representation

−tA 1
e := (λ + A)−1 eλt dλ,
2πi Γ
where Γ is a curve in the resolvent set ρ(−A) with arg λ → ±θ for |λ| → ∞
and a θ ∈ ( π2 , π). The semigroup can be extended analytically into the sector
{t = 0 : | arg t| ≤ ε} for a ε > 0. If Re λ > a for λ in the spectrum σ(A)
then
C
e−tA  ≤ Ce−at und Ae−tA  ≤ e−at
t
for t > 0 and a constant C. Moreover,
d −tA
e = −Ae−tA .
dt
Remark 6.3.4. a) For t > 0 the integral

1
(λ + A)−1 eλt dλ
2πi Γ
is well defined since Re λ → −∞ for |λ| → ∞ and (λ + A)−1 X→X ≤ |λ−a| M
.
Since the resolvent is holomorphic on the resolvent set, due to Cauchy’s
theorem of complex analysis the value of the integral is independent of the
special choice of Γ.
b) The estimate Ae−tA X→X ≤ Ct e−at implies that e−tA maps the space
X into the domain of definition D(A). Since An e−tA = (Ae−A n )n we even
t

have
  
n −tA −A nt n C −a t n C
A e X→X ≤ Ae X→X ≤ e n ≤ n e−at ,
t t
such that e−tA : X → D(An ) for t > 0, and the semigroup is smoothing.
Remark 6.3.5. a) The concept of sectorial operators is very robust under
perturbations. Let A be a sectorial operator with A(λ−A)−1  ≤ C for all λ
in a chosen sector. Moreover, let B be a linear operator with D(B) ⊃ D(A)
satisfying
Bx ≤ εAx + Kx
with ε, K some constants. If εC < 1, then also A + B is a sectorial operator.
For a self-adjoint A it is sufficient that ε < 1. Hence, it is sufficient to check
the assumptions for the principal part of a given operator. Such estimates
can often be found in the existing literature.
b) The most essential remark is that, due to Parseval’s identity, in a
Hilbert space every self-adjoint operator which is bounded from below is a
sectorial operator.
6.3. Other boundary conditions and more general domains 199

In order to apply the previous ideas to the Navier-Stokes equations we

first have to get rid of the pressure term and of the equation ∇ · u = 0.
 R ) with ∇ · u = 0 and u · n|∂Ω = 0. Then for
Therefore, let u ∈ C 1 (Ω, d

φ ∈ C (Ω, R) we have Ω u · ∇φ dx = 0. On the other hand a vector field

1

u which is orthogonal to {∇φ : φ ∈ C 1 (Ω, R)} satisfies ∇ · u = 0 and

u · n|∂Ω = 0.
We define L2p to be the closure of {∇φ : φ ∈ C 1 (Ω, R)} and L2div to be
the closure of {u ∈ C 1 (Ω, Rd ) : ∇ · u = 0, u · n|∂Ω = 0}. Then L2p and L2div
are orthogonal subspaces of L2 with L2 = L2P ⊕ L2div . We introduce P to be
the orthogonal projection on the subspace L2div .
As before we write the Navier Stokes equations as ∂t u = −Au + N (u)
with Au = −P Δ under Dirichlet boundary conditions and N (u) = −P ((u ·
∇)u). It is easy to see that A is a self adjoint and positive definite operator
which immediately implies that A is a sectorial operator, too. Hence, −A is
the generator of an analytic semigroup, cf. Remark 6.3.5 b).
Again u is called mild solution if u satisfies the variation of constant
formula
 t
u(t) = e−tA u0 + e−(t−τ )A N (u(τ )) dτ.
0
In order to control the nonlinear terms we need so called X α -spaces. We
consider a sectorial operator A with Re σ(A) > δ > 0. For a given sectorial
operator A this can always be achieved by considering A  = A + βI for a
suitable chosen β > 0.

Definition 6.3.6. For α > 0 define


−α 1
A =− λ−α (λI − A)−1 dλ,
2πi Γ
where Γ is a curve asymptotically coming from e−iθ ∞ and asymptotically
going to eiθ ∞ with π2 − δ < θ < π running between the origin and σ(A).
The branch of the function λ → λ−α is chosen in such a way that the slit in
the complex plane where λ → λ−α is not analytic coincides with the negative
real axis, see [Gam01] for an introduction to complex analysis.

Since (λI − A)−1  ≤ |λ|

C
we have the convergence of the integral for
α > 0. There exists a C ≥ 0, such that A−α  ≤ C for 0 < α ≤ 1.
Moreover, we have A−α A−β = A−(α+β) if α, β ∈ (0, 1].
Since Re σ(A) > δ > 0 we have the injectivity of A−n . Since A−n =
A−n+α A−α for n > α we have the injectivity of A−α . Therefore, A−α :
−α
X → R(A ) is bijective and we have a (non bounded) inverse. Other
representation formulas for Aα can be found for instance in [Hen81, §1.4]
200 6. The Navier-Stokes equations

Definition 6.3.7. We set Aα = (A−α )−1 for α > 0. The domain of defi-
nition is given by D(Aα ) = R(A−α ). We introduce X α = D(Aα ) equipped
with the norm
uX α = Aα uX .

For α ∈ (0, 1) we have D(A) ⊂ D(Aα ) since R(A−1 ) ⊂ R(A−α ), and so

D(Aα ) is dense in X. Due to the construction of the operators we have that
e−tA uX α ≤ Mα t−α e−δt uX .
In order to proceed with the local existence uniqueness of solutions as above
we need the Lipschitz-continuity of the nonlinear terms N (u) from X α to
X. The following lemma reduces this proof to the proof of the Lipschitz-
continuity from W k,q to Lp , cf. [Hen81, Theorem 1.6.1].
Lemma 6.3.8. Let Ω ⊂ Rd be an open set with smooth boundary, let 1 ≤
p < ∞, and let A be a sectorial operator in X = Lp (Ω) with D(A) = X 1 ⊂
W m,p (Ω) for a m ≥ 1. Then
X α ⊂ W k,q or X α ⊂ C ν (Ω)
for α ∈ [0, 1] if k − d/q < mα − d/p, q ≥ p or 0 ≤ ν < mα − d/p.

In order to prove that D(A) = X 1 ⊂ W 2,2 , for f ∈ L2 one has to find

solutions u ∈ W 2,2 of the elliptic problem
−Δu + ∇p = f, ∇·u=0
in Ω with Dirichlet boundary conditions for u. The existence of such so-
lutions is implied by elliptic regularity theory [ADN59, ADN64]. Since
D(A) ⊂ W 2,2 it follows for d = 3 by Lemma 6.3.8 that for α ∈ (1/2, 1)
that X α ⊂ W 1,q provided 1/q > (5 − 4α)/6 and that for α ∈ (3/4, 1) that
X α ⊂ L∞ . Therefore
N (u)X = N (u)L2 ≤ uL∞ ∇uL2 ≤ Cu2X α
for α ∈ (3/4, 1). For d = 2 we find α ∈ (1/2, 1). Hence, we have the
Lipschitz-continuity of the polynomial N (u) from X α to X for d = 3 if
α ∈ (3/4, 1) and for d = 2 if α ∈ (1/2, 1).
Theorem 6.3.9. Let α ∈ (3/4, 1) if d = 3 or α ∈ (1/2, 1) if d = 2. For
u0 ∈ X α there exists a T0 > 0 such that the Navier-Stokes equations possess
a unique mild solution u ∈ C([0, T0 ], X α ) with u|t=0 = u0 .
1,div
In the proof of the global existence of solutions u ∈ C([0, T0 ], Hper ) of
the Navier-Stokes equations in d = 2 space dimensions in §6.2.3 we used
(6.19) which only holds in R2 and periodic boundary conditions. In this
section we prove the global existence without using (6.19). The method gives
6.3. Other boundary conditions and more general domains 201

weaker estimates but it is more general and periodic boundary conditions

are not needed; nevertheless we keep them for simplicity.
With the Gagliardo-Nirenberg inequality Lp -norms can be estimated by
Lq -Norms and gradients, for p > q. Such inequalities are called interpolation
estimates since a norm in the middle (Lp ) is interpolated with the help of
a weaker norm (Lq ) and a stronger norm (gradient). Here we give a simple
version. See, e.g., [Hen81, Page 37] for a general version, and Exercise 6.6
for the proof of an even weaker version.
Lemma 6.3.10. (A simple Gagliardo-Nirenberg inequality) For d =
1, 2, 3, 4 we have
1−d/4 d/4
uL4 (Td ) ≤ CuL2 (Td ) uH 1 (Td ) .

With the help of Lemma 6.3.10 we can proceed as follows. The L2 -

estimate in domains with general Lipschitz-continuous boundary and Dirich-
let boundary conditions works exactly the same as before. For the H 1 -norm
we estimate again as before
 
1d
(6.22) |∇u| dx ≤ −
2
|Δu|2 dx + g(u)
2 dt Td Td
 
 
with g(u) =  (Δu) · ((u · ∇)u) dx . For periodic boundary conditions and

Td
d = 2 we have g(u) = 0. For general boundary conditions and/or d = 3 the
best estimate known is
 1/2
g(u) ≤ ΔuL2 (u · ∇)uL2 ≤ ΔuL2 |∇u| |u| dx
2 2
Td
≤ ΔuL2 uL4 ∇uL4
where we used the Cauchy-Schwarz inequality. The Gagliardo-Nirenberg es-
timate and the Poincaré-inequality uH m ≤ C |α|=m ∂xα uL2 , see §5.2.2,
give
 
1−d/4 d/4 1−d/4 d/4
g(u) ≤ΔuL2 CuL2 uH 1 C∇uL2 ∇uH 1
1+d/4 1−d/4
≤CΔuL2 uL2 ∇uL2 .
In order to balance the factor Δu2L2 on the right-hand side against −Δu2L2
in (6.22) we use Young’s inequality
ap bq
(6.23) ab ≤ + for a, b ≥ 0 and p, q > 1 with 1/p + 1/q = 1.
p q
Case d = 2: We choose ε = (4/3)3/4 and obtain
C 1/2 3/2
g(u) ≤ uL2 ∇uL2 εΔuL2
ε
202 6. The Navier-Stokes equations

 4
1 C 1/2 3ε4/3
≤ uL2 ∇uL2 + Δu2L2
4 ε 4
≤ Δu2L2 + Cu2L2 ∇u4L2
using (6.23) with q = 4/3 and p = 4. Hence, we obtain
d
∇u2L2 ≤ a(t)∇u2L2 , where a(t) = 2Cu2L2 ∇u2L2 .
dt
From
1 d
u2L2 = −∇u2L2 ,
2 dt
cf. §6.2.3, follows u(t)L2 = e−t u0 L2 , and then
 t
∇u(τ )2L2 dτ = u0 2L2 − u(t)2L2 .
0
Note, that this estimate does not imply a uniform bound for h(t) = ∇u(t)L2 ,
but its square-integrability. As a consequence,
 t  t
0≤ a(t) ≤ C sup u(τ )L22
∇u2L2 dτ < ∞
0 τ ∈(0,t) 0

uniformly for all t ≥ 0 and so

t
∇u(t)2L2 ≤ e 0 a(τ ) dτ
∇u0 2L2 ≤ M ∇u0 2L2
for a M ≥ 0 independent of t ≥ 0. Therefore, the local solution can be
extended to a global solution, i.e., we obtain the global existence in 2D.
Case d = 3: Similar to the 2D case we use the Gagliardo-Nirenberg estimate
and Young’s inequality with q = 8/7 and p = 8 and obtain
 
 
g(u) =  (Δu) · ((u · ∇)u) dx ≤ ΔuL2 uL4 ∇uL4

Td

7/4 1/4 1/4 8
≤CΔuL2 ∇uL2 uL2 ≤ C ∇uL2 uL2 + Δu2L2 .
If we proceed as above it follows
d
∇u2L2 ≤ a(t)∇u2L2 , where a(t) = 2Cu2L2 ∇u6L2 .
dt
The equation for the dissipation only guarantees that the function h(t) =
∇u(t)L2 is in L2 ((0, T0 )), but not in L6 ((0, T0 )), and therefore we cannot
proceed as above. For a further discussion we refer to [Wie99, Con01].
Remark 6.3.11. Since so far it cannot be proved that for d = 3 unique
global solutions exist, it appears to be nonsense to discuss their long time
dynamics. Nevertheless, if one assumes the existence of global strong so-
lutions, then the concept of attractor is again quite useful. In particular,
there are a number of estimates for the dimension of attractors (which is
finite) for the Navier-Stokes equations in two and three space dimensions,
6.3. Other boundary conditions and more general domains 203

cf. [Tem97]. However, a finite-dimensional attractor by no means implies

that the dynamics is “simple”. In fact there is a lot of theory on turbulent
flows and also on so called fully developed turbulence which is mainly based
on methods from statistical physics, cf. [FRMT01].

Further Reading. Classical books about local existence and uniqueness

of the solutions of the Navier-Stokes equations are [vW85, Tem01]. More
background on the derivation and applications of the Navier-Stokes equa-
tions and related equations can be found in [Fow97]. Concise treatments
of these equations as a dynamical system, embedded in the general exis-
tence theory of semilinear parabolic equations can be found for instance in
[Hen81, DG95, Tem97, Rob01]; the latter three are also recommended
for the so called Galerkin method as an alternative to semigroup methods for
proving local existence in the Navier-Stokes equations and general parabolic
systems.
An excellent textbook, going way beyond the brief summary given here,
is [MB02]. An essential reference for steady problems in exterior domains,
including the necessary function spaces and inequalities, is [Gal11]. The
dynamics and stability of vortices is treated in [MB02, GW05, GW06b].
See [WW15, Chapter 1] for a very accessible account on metastable states
and the finite dimensionality of the global attractor for the 2D Navier-Stokes
equations. An enlightening essay about the Navier-Stokes Millenium prob-
lem is [Tao09], emphazising the scale invariance, see Exercise 6.7. Finally,
[Lem16] gives an impressive overview about the state of the art of the math-
ematical analysis of the Navier-Stokes equations, and an excellent review of
the Millennium problems, with focus on the Navier-Stokes equations.

Exercises
6.1. (a) Show that u : R2 → R2 with ∇ × u = 0 and ∇ · u = 0 is equivalent to the
Cauchy-Riemann differential equations for w(z) = u1 (x, y) − iu2 (x, y), z = x + iy.
(b) Sketch the flow belonging to w(z) = z 2 and calculate the associated pressure.
3 
6.2. (d’Alembert Paradox) Let a > 0, U ∈ R3 , and φ(x) = 2x a
3 + 1
U, x
for x ∈ Ω = R3 \ Ba (0). Sketch u = ∇φ, show that div u = 0, and calculate the
drag f = − Ba (0) φn dS.

6.3. For given ω = ω(x), x ∈ T2 , find in Fourier space an explicit solution u of

∇×u=ω and ∇ · u = 0.

6.4. Show that 

u ∗ vω √ ≤ 
uω √ 
v ω √ .
1, t 1, t 1, t

6.5. Show the enstrophy identity (6.19).

204 6. The Navier-Stokes equations

6.6. Prove the following weak form of the Gagliardo-Nirenberg inequality for (2π)d -
periodic functions. For d = 1, 2, 3 and δ > 0 we have
1−d/4−δ d/4+δ
uL4 (Td ) ≤ CuL2 (Td ) uH 1 (Td ) .
Hint: Use uLp ≤ C uLq for 1/p+1/q = 1 and q ∈ [1, 2] and the Hölder inequality
   
|ak bk ck | ≤ ( |ak |p1 )1/p1 ( |bk |p2 )1/p2 ( |ck |p3 )1/p3
k∈Z k∈Z k∈Z k∈Z

with 1/p1 + 1/p2 + 1/p3 = 1.

6.7. Let (u, p) : Rd × [0, T ) → Rd × R be a solution of the d-dimensional Navier-
Stokes equations over Rd . For l > 0 set ul (x, t) = lα1 u(lβ x, lγ t) and pl (x, t) =
lα2 p(lβ x, lγ t). Find α1,2 , β, γ, and δ such that (ul , pl ) is again a solution of the
Navier-Stokes equation. Discuss how the energy and energy dissipation behave
under this scaling.
 Chapter 7

Some dissipative PDE

models

This is the first chapter of Part III of this book. Here and in the remainder
of this book we consider PDEs on unbounded domains. In order to avoid
dealing with far away boundaries, whose influence on the solutions in the
interior of the domain is small at least for a long time, we idealize the
large domain to an unbounded domain. For instance, instead of x from
some large interval (−L, L) we consider x ∈ R. From a didactic point
of view the consideration of unbounded domains has certain advantages.
Since we do not have to deal with boundary conditions which are often a
source of functional analytic difficulties, this idealization allows to explain
genuine PDE phenomena such as transport, diffusion and dispersion. Hence,
it allows us to keep the functional analytic tools at a minimum. Unbounded
domains are easy in this respect.
On the other hand, compared to PDEs over bounded domains there are
new fundamental and challenging open questions, mainly due to the fact that
PDEs on unbounded domains define dynamical systems with uncountably
many modes (degrees of freedom). In contrast to the situation of countably
many modes considered in Chapters 5 and 6, a separation of the uncountably
many modes into single modes is a highly singular action from a functional
analytic point of view, and therefore in general of little use. The recovery
of compactness by smoothing properties is no longer true, and therefore
finite-dimensional attractors in general cannot be expected.
To illustrate our point of view, the following example shows that also
for PDEs defined on a very large domain in space the interpretation as
countably many ODEs is no longer a big help.

205
206 7. Some dissipative PDE models

Example. Consider the linear wave equation ∂t2 u = ∂x2 u for t ∈ R, x ∈

(−L, L), L > 0 very large, u(x, t) ∈ R, with Dirichlet boundary conditions
u(−L, t) = u(L, t) = 0. We consider two special classes of solutions, first
the oscillations of the eigenfunctions,
u(x, t) = sin(nπt/(2L)) sin(nπ(x − L)/(2L)), n ∈ N,
and secondly the traveling wave solutions
u(x, t) = f (x − t) + g(x + t)
with f and g arbitrary smooth functions with compact support
supp(f ) = {x : f (x) = 0} ⊂ [−1, 1] and supp(g) ⊂ [−1, 1].
As long as |t| < L − 1 this is a solution of the PDE, i.e., for a very large
time interval traveling wave solutions play a role. An expansion of these
solutions in eigenfunctions is of no use.
In this Chapter 7 we start with some scalar model problems. These are
the Kolmogorov-Petrovsky-Piskounov (KPP) or Fisher equation in §7.1, the
Allen-Cahn equation in §7.2, and the Burgers equation in §7.4. Moreover,
there is the method oriented §7.3 about Fourier transform. We keep the
exposition rather brief and aim for a basic understanding of the various
models. We are interested in the local existence and uniqueness of solutions
and in the existence and stability of special solutions, which are important
for the underlying physical processes which are described by the models.
To construct these special solutions we often use the ODE methods from
Part I. In order to make this part more self-contained we recall a number of
definitions and constructions which are only small adaptions of respective
concepts from Part I. In Chapter 8 we consider with the NLS, KdV, and the
GL equation the three canonical modulation equations whose dynamics we
will recover in more complicated PDEs in Part IV of this book. Part III of
this book is closed with Chapter 9 about reaction-diffusion systems.
For each of the equations considered in this Part there already exists
much literature on various levels, see the “further reading” at the end of
each chapter.

7.1. The KPP equation

The Kolmogorov-Petrovsky-Piskounov (KPP) equation [KPP37] or Fisher
equation [Fis37]
(7.1) ∂t u = ∂x2 u + u − u2 ,
with t ≥ 0, x ∈ R, and u = u(x, t) ∈ R, occurs as a model for various sys-
tems in nature, for instance for chemical reactions or population dynamics.
The equation consists of two parts, namely the diffusion term ∂x2 u and the
7.1. The KPP equation 207

nonlinear reaction term u−u2 . Therefore, it brings together PDE with ODE
dynamics.
Inserting u(x, t) = v(t) into (7.1) gives the one-dimensional ODE

(7.2) v̇ = v − v 2 .

The 1D phase portrait shows that the fixed point v = 0 is unstable and
that the fixed point v = 1 is asymptotically stable. The term +v in the
KPP equation describes exponential growth for small v and the term −v 2
represents saturation. For instance, a population of animals or a chemical
reaction initially increases with some exponential rate until the growth is
saturated by the available food or the missing reactant. If v(0) > 0, then
limt→∞ v(t) = 1.
Before we combine the ODE dynamics coming from the reaction term
u − u2 with the dynamics coming from the diffusion term ∂x2 u we discuss the
modeling and the properties of linear diffusion in the next two subsections.

7.1.1. The modeling of diffusion. Diffusion occurs in various situations.

We explain three such situations, namely Brownian motion, a discrete ran-
dom walk, and Fourier’s law.
Brownian motion and diffusion. The term Brownian motion is named
after Robert Brown who in 1827 described the irregular motion of pollen
particles suspended in water. In [Ein05] Einstein studied Brownian motion
the following way. Consider a long, thin tube filled with clear water, into
which we inject at time t = 0 a unit amount of ink, at the location x = 0.
Let u(x, t) denote the density of ink at position x ∈ R and time t ≥ 0.
Suppose that the probability that an ink particle moves from x to x + y in a
time τ is translational invariant, i.e., does not depend on x. This probability
is denoted by ρ(y, τ ). Then

u(x, t + τ ) = u(x − y, t)ρ(y, τ ) dy
R  
1 2
= u(x, t) − (∂x u(x, t))y + (∂x u(x, t))y + . . . ρ(y, τ ) dy.
2
R 2

Now ρ(−y, τ ) = ρ(y, τ ) by symmetry such that R yρ(y, τ ) dy = 0. Next
assume that the variance is linear in τ , i.e.,

(7.3) y 2 ρ(y, τ ) dy = 2Dτ
R

for a D > 0. Then

1
(u(x, t + τ ) − u(x, t)) = D∂x2 u(x, t) + h.o.t.
τ
208 7. Some dissipative PDE models

Under the assumption that all higher moments of ρ decay faster than τ for
τ → 0, in the limit τ → 0 we obtain the linear diffusion equation
(7.4) ∂t u = D∂x2 u.
Einstein derived the relation 2D = RT /(NA ν) where R is the gas constant,
T the temperature, NA the Avogadro number, and ν a friction coefficient.
A discrete random walk. We consider a two-dimensional rectangular
lattice, comprising the sites {(mδx, nδt) : m = 0, ±1, ±2, . . . ; n = 0, 1, 2, ...}.
A particle starting in x = 0 at a time t = 0 decides at each time nδt to
move an amount δx to the left or to move an amount δx to the right, both
possibilities with probability 1/2. Denote by p(m, n) the probability that
the particle is at the position mδx at the time nδt. Then p(0, 0) = 1 and
p(m, 0) = 0 for m = 0. Also, p(m, n + 1) = 12 (p(m − 1, n) + p(m + 1, n)),
hence
1
p(m, n + 1) − p(m, n) = (p(m − 1, n) − 2p(m, n) + p(m + 1, n)).
2
2
Now assume that (δx) /(δt) = 2D which corresponds to (7.3) above. Then
1 D
(p(m, n + 1) − p(m, n)) = (p(m − 1, n) − 2p(m, n) + p(m + 1, n)),
δt (δx)2
and sending δ → 0 again yields the linear diffusion equation (7.4).
Fourier’s law. Let u : R3 → R be the temperature inside a material body
and let V ⊂ R3 be a test volume with surface S. Then
  
d
u dV = − j · n dS = − div j dV ,
dt
V S V

where j = j(x, t) ∈ R3 is the heat flow. Since this is true for all test volumes
V we find
∂t u + div j = 0 .
It is reasonable to assume that the heat flow from warm to cold is propor-
tional to the negative temperature gradient, i.e., j = −D∇u. This so called
Fourier’s law again yields the heat equation
∂t u = D div ∇u = DΔu.

7.1.2. Diffusion on the real line. Throughout this subsection we con-

sider the linear diffusion equation with diffusion coefficient D = 1, i.e.,
(7.5) ∂t u = ∂x2 u,
which always can be achieved by a rescaling of time or space. We already
observed the dissipative character of this equation in previous sections where
we found the solutions u(x, t) = e−k t sin(kx) which decay to zero for |k| > 0
2

with some exponential rate.

7.1. The KPP equation 209

For t > 0 the general formula for the solutions of the linear diffusion
equation (7.5) is given by
 ∞
1 (x−y)2
(7.6) u(x, t) = √ e− 4t u(y, 0) dy.
4πt −∞
The derivation of this formula is given subsequently, but also by a different
method in §7.3. The existence of this integral is guaranteed for t > 0 if for
instance supy∈R |u(y, 0)| < ∞.
From (7.6) we immediately obtain the estimate
 ∞
1
sup |u(x, t)| ≤ √ |u(x, 0)| dx,
x∈R 4πt −∞
i.e., solutions to spatially localized initial conditions decay uniformly towards
zero with a rate t−1/2 . Since mass is conserved, i.e.,
 ∞  ∞
u(x, t) dx = u(x, 0) dx
−∞ −∞
for all t ≥ 0, this is how diffusion is expected to work. The conservation of
mass follows for instance with the use of the solution formula from
 ∞  ∞  ∞ 
1 − (x−y)2
u(x, t) dx = √ e 4t dx u(y, 0) dy
−∞ −∞ −∞ 4πt
 ∞
= 1 · u(y, 0) dy.
−∞
The decay happens in a universal manner. The initial condition
u(y, 0) = δ0 (y),
with δ0 the ”δ-distribution in x = 0”, cf. Example 5.2.2, leads to the self-
similar solution
1 − x2
(7.7) u(x, t) = √ e 4t .
4πt
The solution only exists for t > 0. This is a general rule. For an arbitrary
initial condition the diffusion equation cannot be solved backwards in time.
Since this solution is the starting point of the construction of the general
 solution. In order to derive
solution formula it is also called fundamental
(7.7) we make the ansatz u(x, t) = v √xt and find v = − 12 ξv which is
ξ2

solved by v = ce− 4 with a constant c ∈ R. Since with u(x, t) = v √xt

also ∂x u = √1t v √xt is a solution of (7.5) we find (7.7).
In lowest order self-similar behavior is also observed for general spatially
localized initial conditions, namely,
 
A∗ x
u(x, t) = √ v √ + O(1/t) with v(ξ) = e−ξ /4
2
(7.8)
t t
210 7. Some dissipative PDE models

Figure 7.1. Self-similar decay of the special solution u(x, t) =

x2
√ 1 e− 4t for t ∈ {1, 2, 5, 10}.
4πt

and with a constant A∗ ∈ R depending on the initial conditions. This is

explained in detail in §14.
Diffusion is smoothing. We obtain the estimate
 1  ∞ (x − y) (x−y)2 
 
sup |∂x u(x, t)| = sup  √ − e− 4t u(y, 0) dy 
x∈R x∈R 4πt −∞ 2t
 1  ∞ √  
 s 
− √ e−s u x − 2 ts, 0 ds
2
= sup  √
x∈R π −∞ t
 ∞
1 2 C
se−s ds sup |u(x, 0)| ≤ √ sup |u(x, 0)|,
2
≤√ √
t π 0 x∈R t x∈R
with a constant C independent
√ of t and of u(·, 0), where
√ we made the trans-
formation s = (x − y)/(2 t), i.e., ds = − dy/(2 t). This can easily be
generalized.
Theorem 7.1.1. Let u = u(·, t) be a solution of the linear diffusion equa-
tion. Then for all n ∈ N there exists a C > 0 such that for all t > 0
∂xn u(·, t)C 0 ≤ Ct−n/2 u(·, 0)C 0 .
b b

Finally for every t0 > 0 and x0 ∈ R the function (x, t) → u(x, t) can be
expanded in a convergent power series around (x0 , t0 ), i.e., u is an analytic
function and can be extended into the complex plane. See §6.2.2.

In order to handle the linear diffusion equation ∂t u = ∂x2 u with dynam-

ical systems methods we have to choose a suitable phase space. We already
know that in infinite dimensions the choice of phase space and associated
0
norm is fundamental. A first choice is X = Cb,unif (R, R), the space of uni-
formly bounded and uniformly continuous functions u : R → R, equipped
with the norm
u(t)C 0 = sup |u(x, t)| .
b,unif
x∈R
7.1. The KPP equation 211

Lemma 7.1.2. The curve t → u(t, u0 ) is continuous in X if u0 ∈ X.

Proof. Since u(t + s, u0 ) = u(t, u(s, u0 )) it is sufficient to prove the continu-

ity of the orbit t → u(t) for t  0 in the space X. With H(z) = √14π e−z /4
2

we estimate
u(t, u0 ) − u0 C 0 = sup |u(x, t) − u(x, 0)|
b
x∈R
 ∞ 1   
 x−y 
= sup  √ H √ (u(y, 0) − u(x, 0)) dy 
x∈R −∞ t t
 ∞ √  
 
= sup  H(z) u(x − tz, 0) − u(x, 0) dz 
x∈R −∞
 
≤ sup | . . . | dz + sup | . . . | dz = s1 + s2 .
x∈R |z|≤R x∈R |z|≥R

For a given ε > 0 we have to find a t0 > 0 such that for all t ∈ (0, t0 ) we
have s1 + s2 < ε. We can estimate

s2 ≤ 2 H(z) dz sup |u(x, 0)| < ε/2
|z|≥R x∈R

by choosing an R > 0 sufficiently large due to the definition of H. Next, we

estimate
 ∞
√
s1 ≤ H(z) dz sup |u(x + tz, 0) − u(x, 0)|
−∞ x∈R,|z|≤R
√
= sup |u(x + tz, 0) − u(x, 0)|.
x∈R,|z|≤R

Since x → u(x, 0) is uniformly continuous for all ε > 0 there exists a δ > 0
such that for all y ∈ R with |y| < δ we have |u(x + y, 0) − u(x, 0)| < ε/2.
Choosing t0 > 0 so small that t0 R < δ we are done. 
The deeper reason why X = Cb0 would not be a good choice is explained
in the following remark, but for a slightly simpler PDE.

Remark 7.1.3. If we consider the translation semigroup T (t) : u(·) →

u(· + t) which is the solution operator of the transport equation ∂t u = ∂x u
in the space X = Cb0 (R, R), we have that T (t) is not a C0 -semigroup, cf.
Definition 5.1.9, since for u(x) = sin(x2 ) which is an element of Cb0 , but
0
not of Cb,unif we always have u(·) − u(· + t)C 0 = 2, if t > 0. The same
b
is true for the linear diffusion equation. The deeper reason for the non-
continuity of both semigroups in Cb0 is the fact that the domain of definition
Cb1 , respectively Cb2 , is not dense in Cb0 , cf. the theorem of Hille-Yosida, cf.
[Paz83, Section 1.3, Theorem 3.1]. In order to have a C0 -semigroup we
have to restrict to X = Cb,unif0 (R, R), which excludes the counter-example
212 7. Some dissipative PDE models

u(x) = sin(x2 ), where the faster and faster oscillations for |x| → ∞ destroy
the uniform continuity w.r.t. x and t.
Remark 7.1.4. With (x, t) → u(x, t) a solution of (7.5) also (x, t) → u(x +
y, t) is a solution of (7.5), i.e., every solution shifted by y is a solution,
too. More abstractly, the solution operator and the translation operator
commute. With t → u(x, t) a solution, obviously every derivative t →
∂xn u(x, t) and every integral is a solution, too.

x2
Figure 7.2. The solutions u(x, t) = √ 1 e− 4t , ∂x u and its first integral
4πt
for t = 1.

Moreover, every linear combination and every convergent series of solu-

tions, or every convergent integral over a set of solutions are again solutions.
These properties are used in the following paragraph.
Our starting point to derive the solution formula (7.6) is the explicit
solution (7.7). With this solution also
 ∞
1 (x−y)2 c(y) − (x−y)2
u(x, t) = √ e− 4t and u(x, t) = √ e 4t dy
t −∞ t
are solutions of (7.5). From the limit t → 0 and Lemma 7.1.2 we find that
c(y) = u(y,0)
√
4π
. The solution formula can be interpreted as linear combination
of fundamental solutions, i.e., of diffusion processes starting in every point
x ∈ R with a δ-distribution as initial condition.

7.1.3. The reaction-diffusion equation. The KPP equation is obtained

by adding the diffusion term to the ODE (7.2). Thus, the KPP equation
describes for instance a chemical reaction or the evolution of a disease in
a large, here in an infinitely extended, domain, where the concentration
u = u(x, t) spreads into space by diffusion.
In order to handle the KPP equation as an abstract dynamical system
we choose the same phase space as for the linear diffusion equation, namely
0
X = Cb,unif (R, R). Solutions u = u(t) ∈ X satisfy the KPP equation only
in a weak sense.
7.1. The KPP equation 213

Definition 7.1.5. Solutions of (7.5) are called mild solutions of (7.5) if

they satisfy the variation of constant formula
 t
(7.9) u(t) = T (t)u0 + T (t − τ )N (u)(τ ) dτ,
0

where T (t) is the solution operator of the linear diffusion equation, i.e.,
 ∞
1 (x−y)2
(T (t)u0 )(x) = √ e− 4t u0 (y)dy,
4πt −∞
and where N (u)(τ ) represents the reaction term u(x, τ ) − u(x, τ )2 .

Lemma 7.1.6. If u ∈ C([0, T0 ], Cb,unif

2 ) ∩C 1 ([0, T0 ], Cb,unif
0 ) solves the varia-
tion of constant formula (7.9) then u is also a classical solution of the KPP
equation, i.e., u = u(x, t) can be inserted into the KPP equation.

Proof. See Exercise 7.5. 

There is local existence and uniqueness of mild solutions for the KPP
equation in X.
Theorem 7.1.7. For all C0 > 0 there exists a T0 > 0 such that for u0 ∈ X
with u0 X ≤ C0 there exists a unique solution u ∈ C([0, T0 ], X) of the KPP
0
equation (7.5) with u|t=0 = u0 and X = Cb,unif (R, R).

Proof. The proof of the local existence and uniqueness of solutions is based
on the contraction mapping principle and on the variation of constant for-
mula. Fix C1 > 0. For sufficiently small T0 > 0 the right-hand side of the
variation of constant formula (7.9) defines a contraction F in the complete
metric space
M = C 0 ([0, T0 ], {u(t) ∈ X : u(t) − T (t)u0 X ≤ C1 })
with the metric d(u, v) = u − vM induced by the norm
uM = sup u(t)X .
t∈[0,T0 ]

We have C2 := supu∈M uM ≤ C0 + C1 . In order to prove the contraction

mapping property we need three estimates. For t ≥ 0, u0 ∈ X and u, v ∈ M
we have
(E1)T (t)u0 X ≤ u0 X ,
(E2)N (u)M ≤ C2 + C22 ,
(E3)N (u) − N (v)M ≤ (1 + 2C2 )u − vM .
214 7. Some dissipative PDE models

Estimate (E1) follows from

  ∞ 
 1 (x−y)2 
T (t)u0 X 
= sup  √ e − 4t
u0 (y)dy 
x∈R 4πt −∞
  ∞ 
 1 (x−y)2 

≤ sup  √ e − 4t
dy  · sup |u0 (y)|
x∈R 4πt −∞ y∈R
= sup |u0 (y)| = u0 X .
y∈R

Estimate (E2) follows from

 
N (u)M = sup sup u(x, t) − u(x, t)2 
t∈[0,T0 ] x∈R

≤ sup sup |u(x, t)| + sup sup |u(x, t)|2 ≤ C2 + C22 .

t∈[0,T0 ] x∈R t∈[0,T0 ] x∈R

Estimate (E3) follows from

 
N (u) − N (v)M = sup sup u(x, t) − u(x, t)2 − v(x, t) + v(x, t)2 
t∈[0,T0 ] x∈R
≤ sup sup (|u(x, t) − v(x, t)| + |u(x, t) + v(x, t)| |u(x, t) − v(x, t)|)
t∈[0,T0 ] x∈R
≤ u − vM + (uM + vM ) u − vM ≤ (1 + 2C2 ) u − vM .

i) F maps M into M , since

 t
F (u) − (T (t)u0 )t≥0 M = sup  T (t − τ )N (u)(τ ) dτ X
t∈[0,T0 ] 0
 t
≤ sup T (t − τ )N (u)(τ )X dτ
t∈[0,T0 ] 0
 t
≤ sup N (u)(τ )X dτ
t∈[0,T0 ] 0
 t  
≤ sup dτ N (u)M ≤ C2 + C22 T0 ≤ C1
t∈[0,T0 ] 0

for T0 > 0 sufficiently small.

ii) F is a contraction in M , since
 t 
 
F (u) − F (v)M = sup   T (t − τ )(N (u) − N (v))(τ ) dτ 

t∈[0,T0 ] 0 X
 t
≤ sup T (t − τ )(N (u) − N (v))(τ )X dτ
t∈[0,T0 ] 0
 t
≤ sup (N (u) − N (v))(τ )X dτ
t∈[0,T0 ] 0
7.1. The KPP equation 215

 t
≤ sup dτ (N (u) − N (v))M
t∈[0,T0 ] 0
1
≤ T0 (1 + 2C2 )u − vM ≤ u − vM
2
for T0 (1 + 2C2 ) ≤ 1/2. Thus, for T0 > 0 sufficiently small there is a fixed
point u = F (u), which is a mild solution of the KPP equation. 
Remark 7.1.8. The KPP equation is a semi-linear parabolic equation
[Hen81], i.e., the semigroup T (t) generated by the operator Λ = ∂x2 is
smoothing and the nonlinearity N contains only derivatives of strictly lower
order than in Λ. Like for the diffusion equation the solutions of the KPP
equation or more general of a semi-linear parabolic equation on the real line
are analytic for all t > 0. The proof is similar to the one given in §6.2.2.
7.1.4. The maximum principle. Second order scalar parabolic and ellip-
tic PDEs have a special property which helps a lot in their analysis, namely
the maximum principle.
Theorem 7.1.9. Let u1 , u2 be bounded solutions of the KPP equation (7.1)
in X for t ∈ [t0 , t1 ]. From u1 (x, t0 ) ≤ u2 (x, t0 ) for all x ∈ R it follows that
u1 (x, t) ≤ u2 (x, t) for all x ∈ R and all t ∈ [t0 , t1 ].
Idea of the proof. We show a slightly modified statement, namely that
u1 (x, t0 ) < u2 (x, t0 ) for all x ∈ R implies u1 (x, t) < u2 (x, t) for all x ∈ R
and all t ∈ [t0 , t1 ].
If this is wrong, then there exists a t∗ ∈ (t0 , t1 ) and a x0 ∈ R such that
u1 (x0 , t∗ ) = u2 (x0 , t∗ ). Since the solutions change smoothly in x and t in
this point we necessarily have ∂x u1 (x0 , t∗ ) = ∂x u2 (x0 , t∗ ) and generically
∂x2 u1 (x0 , t∗ ) < ∂x2 u2 (x0 , t∗ ). Then it follows
∂t (u2 (x0 , t∗ )−u1 (x0 , t∗ )) = ∂x2 (u2 (x0 , t∗ ) − u1 (x0 , t∗ ))
+ (u2 (x0 , t∗ )−u1 (x0 , t∗ ))−(u22 (x0 , t∗ )−u21 (x0 , t∗ ))
= ∂x2 (u2 (x0 , t∗ ) − u1 (x0 , t∗ )).
In the generic case ∂x2 (u2 (x0 , t∗ ) − u1 (x0 , t∗ )) > 0, we have again u2 (x0 , t) >
u1 (x0 , t) for t ∈ [t0 , t1 ], i.e., u1 can never be larger than u2 . For a complete
proof see for instance [RR04, Theorem 4.26]. 

Figure 7.3. The maximum principle: the difference of the solutions

for t = t∗ . For t > t∗ we have u2 − u1 > 0 again.

A direct consequence is the following lemma.

216 7. Some dissipative PDE models

Lemma 7.1.10. Solutions that start at t0 with values in [0, 1] will stay with
their values in [0, 1] for t ≥ t0 , i.e., 0 ≤ u(x, t0 ) ≤ 1 for all x ∈ R implies
0 ≤ u(x, t) ≤ 1 for all x ∈ R and all t ≥ t0 .

Proof. Since u ≡ 0 and u ≡ 1 are solutions of the KPP equation, the

assertion follows from the maximum principle. 
Hence, it makes sense to talk about concentrations. A second conse-
quence of the maximum principle is the global existence and uniqueness of
mild solutions of the KPP equation. The estimate from Lemma 7.1.10 is
called a priori estimate. A priori estimates and local existence and unique-
ness of solutions gives global existence and uniqueness of solutions. See e.g.
§2.2, §5.3.1, or §6.2.3.
Theorem 7.1.11. For u0 ∈ X with 0 ≤ u0 (x) ≤ 1 there exists a unique
solution u ∈ C([0, ∞), X) of the KPP equation with u|t=0 = u0 and X =
0
Cb,unif (R, R).

Proof. For all initial conditions with values in [0, 1] the local existence and
uniqueness theorem 7.1.7 guarantees the existence for all t ∈ [0, T0 ] with for
instance T0 = 1/6. By the last lemma we have 0 ≤ u(x, T0 ) ≤ 1. Then
we start the KPP equation again, but now with initial condition u(T0 ).
The local existence and uniqueness theorem guarantees for the same reason
solutions for all t ∈ [T0 , 2T0 ]. Repeating the argument shows the assertion.

Theorem 7.1.12. For u0 ∈ X with 0 ≤ u0 (x) ≤ 1 there exists a unique
solution u ∈ C([0, ∞), X) of the KPP equation with u|t=0 = u0 and X =
0
Cb,unif (R, R).

Proof. For all initial conditions with values in [0, 1] the local existence and
uniqueness theorem 7.1.7 guarantees the existence for all t ∈ [0, T0 ] with for
instance T0 = 1/10. By the last lemma we have 0 ≤ u(x, T0 ) ≤ 1. Then
we start the KPP equation again, but now with initial condition u(T0 ).
The local existence and uniqueness theorem guarantees for the same reason
solutions for all t ∈ [T0 , 2T0 ]. Repeating the argument shows the assertion.


7.1.5. Stability, instability, and fronts. We are interested in the qual-

itative behavior of solutions of the KPP equation. Due to the chemical or
biological motivation of the KPP equation we restrict to solutions u with
0 ≤ u(x, t) ≤ 1. As usual we start with the fixed points, i.e., with the
stationary solutions of the KPP equation. They satisfy
0 = ∂x2 u + u − u2 .
7.1. The KPP equation 217

By having a look at the phase portrait in Figure 7.4 we see that the only
equilibria having only values in [0, 1] are the fixed points u = 0 and u = 1.

u’
0

−1

−1 0 1 2
u

Figure 7.4. The phase portrait for the stationary solutions of the
KPP equation.

Next we are interested in the stability of these stationary solutions. We

have exactly the same definition as for ODEs, cf. Definition 2.3.3, except
that the ODE phase space Rd has to be replaced by the PDE phase space
X.

Definition 7.1.13. A fixed point u∗ is called stable for the KPP equation
in X if for any ε > 0 there is a δ > 0 such that u0 − u∗ X < δ implies
u(t, u0 ) − u∗ X < ε for all t ≥ 0. Otherwise, it is called unstable. A stable
fixed point is called asymptotically stable in X if additional limt→∞ u(t, u0 ) =
u∗ holds.

For ODEs, in most cases the eigenvalues of the linearized system deter-
mine stability or instability, cf. Theorem 2.3.4. For u∗ = 0 we find the linear
operator (∂x2 + 1)· which has the eigenfunctions eikx for k ∈ R and the spec-
trum (−∞, 1]. For u∗ = 1 we find the linear operator (∂x2 − 1)· which has the
eigenfunctions eikx for k ∈ R and the spectrum (−∞, −1]. Hence, we expect
that u∗ = 0 is unstable, whereas u∗ = 1 is expected to be asymptotically
stable. However, instead of using these spectra, here we use the maximum
principle to study stability and instability.

Theorem 7.1.14. The fixed point u∗ = 0 is unstable in X.

Proof. Let u0 = δ/2 for 0 < δ  1. Since u0 is constant in space, no

diffusion occurs and the dynamics is determined by the ODE dt d
v = v − v2.
Therefore, we have limt→∞ u(t, u0 ) = 1 and hence u(t, u0 ) leaves every ε-
neighborhood of u∗ = 0 independent of how small δ > 0 was at the begin-
ning. 

Theorem 7.1.15. The fixed point u∗ = 1 is asymptotically stable in X.

218 7. Some dissipative PDE models

Proof. Choose δ = min{ε, 1/10} and let u0 ∈ X with u0 − u∗ X < δ. Let
v− = 1 − δ and v + = 1 + δ such that v− ≤ u0 (x) ≤ v+ for all x ∈ R. By the
maximum principle we have for the associated solutions
v− (t) ≤ u(x, t) ≤ v+ (t)
for all x ∈ R and t ≥ 0. Since limt→∞ v− (t) = limt→∞ v+ (t) = 1 we can
conclude
lim u(t, u0 ) − u∗ X = 0,
t→∞
i.e., the asymptotic stability of u∗ . 

Figure 7.5. By the maximum principle the solution u is confined

by the spatial constants.

If a chemical reaction (which is described by the KPP equation) is started

locally in space, i.e., if we have a small positive perturbation with compact
support of the unstable equilibrium u∗ = 0, then we expect two reaction
fronts to be created, one moving to the left and one moving to the right. In
the following we are interested in the velocities of these reaction fronts.
Thus, we consider solutions of permanent form, i.e. u(x, t) = v(x − ct) =
v(ξ), with velocity c, satisfying limξ→−∞ v(ξ) = 1 and limξ→∞ v(ξ) = 0.
Inserting this into the KPP equation gives the ODE
−c∂ξ v = ∂ξ2 v + v − v 2
which we write as first order system ( = ∂ξ )
v =w,
w = − cw − v + v 2 .
We find that the qualitative form of the phase portraits in Figure 7.6 changes
when c is increased.
For all values of c = 0 we find a heteroclinic connection between the fixed
points (v, w) = (0, 0) and (v, w) = (1, 0). Since the heteroclinic connections
are in the intersection of the one-dimensional unstable manifold of (1, 0) and
the two-dimensional stable manifold of (0, 0) this connection is very robust
under perturbations. However, for |c| < 2 these heteroclinic connections
spiral into (v, w) = (0, 0) which can be seen by considering the linearization
(7.10) v = w, w = −cw − v.
7.1. The KPP equation 219

2
w 1

w
0

−1
−2

−1 0 1 −0.5 0 0.5 1 1.5

v v

1
1
0.5
v

0.5

v
0
0 5 10 0
x−t 0 5 10 15
x−2t

Figure 7.6. The phase portraits for the front solutions for the KPP
equation for c = 1 (left) and c = 2 (right), with the associated
fronts.

We find the eigenvalues

√
c2 − 4 −c ±
λ1,2 = .
2
For c ∈ (0, 2) we have complex conjugate eigenvalues with Re λ1,2 < 0
and Im λ1,2 = 0, and for c ≥ 2 we have real negative eigenvalues. Thus,
monotonic fronts which only take values in [0, 1] can only exist for c ≥ 2.
The goal of Exercise 7.6 is an analytical existence proof of these fronts.
Remark 7.1.16. There is a very general result that heteroclinic connections
for the ODE which is satisfied by the spatially homogeneous solutions imply
the existence of fronts with large c in the PDE, cf. [AJMR95]. In detail,
for the KPP equation the front satisfies
∂ξ2 u + c∂ξ u + u − u2 = 0.
We rescale time ξ = cζ, such that ∂ξ = ε∂ζ where ε = 1/c is small. Then
(7.11) ε2 ∂ζ2 v + ∂ζ v + v − v 2 = 0
which, for ε → 0, converges towards the ODE of the pure reaction, i.e.,
(7.12) ∂ζ v + v − v 2 = 0.
The theory of singularly perturbed ODEs [Fen79] guarantees that the hete-
roclinic connection between the fixed points of (7.12) persists for small ε > 0
in (7.11).
220 7. Some dissipative PDE models

In order to control the velocity of the reaction fronts created by small

positive spatially localized initial conditions we use the maximum principle.
To an initial condition u0 with compact support and values in [0, 1) we find
an exact front u− = u− (x) moving to the left and an exact front u+ = u+ (x)
moving to the right with minimal velocity c = 2 satisfying
u− (x) ≥ u0 (x) ≤ u+ (x)
for all x ∈ R. From the maximum principle we obtain
u− (x + 2t) ≥ u(x, t) ≤ u+ (x − 2t).
Therefore, the reaction fronts cannot move faster than the exact fronts with
the minimal velocities c = ±2. In [Bra83] it has been shown that the upper
bound is a good approximation and that e.g. the reaction front moving to
the right moves like
xf (t) = 2t − c1 ln t + c2 (t),
with limt→∞ c2 (t) = c∗2 < ∞, and c1 , c∗2 ∈ R some constants. See Figure 7.7.

Figure 7.7. Solutions to localized initial conditions are confined

by two fronts with minimal velocity.

In the following we restrict to front solutions with velocities |c| ≥ 2 which

allows us to argue with the maximum principle.
Theorem 7.1.17. The front solutions are unstable in X.

Proof. Let uf be a front solution with velocity c, and u0 (x) = min(uf (x), δ/2).
Then uf − u0 X < δ with u(t, uf )[x] = uf (x − ct), but limt→∞ u(t, u0 ) = 1
due to the maximum principle, since δ/2 ≤ u0 (x) ≤ 1, and hence
lim u(t, uf ) − u(t, u0 )X = 1 > ε
t→∞
independent of how small δ > 0 is chosen. See Figure 7.8. 

uf

δ/2
0

0
Figure 7.8. Instability of front solutions w.r.t. perturbations in Cb,unif .
7.1. The KPP equation 221

Remark 7.1.18. On the other hand, by the maximum principle the front
solutions are stable w.r.t. smaller sets of perturbations. If the small pertur-
bation v of the front solution uf is contained between two translates of the
same front then by the maximum principle it will stay there for all times.
In detail, if
uf (x − x− ) ≤ uf (x) + v(x, 0) ≤ uf (x + x+ )
for all x ∈ R, then
uf (x − ct − x− ) ≤ uf (x − ct) + v(x, t) ≤ uf (x − ct + x+ )
for all t ≥ 0 and all x ∈ R. See Figure 7.9.

Figure 7.9. Stability of the fronts w.r.t. smaller sets of perturba-

tions. The perturbations are confined by parallel fronts with the
same velocity.

This leads to the following generalization of the definition of stability.

Definition 7.1.19. A fixed point u∗ is called (X1 , X2 )-stable, if for all ε > 0
there exists a δ > 0 such that u0 − u∗ X1 < δ implies u(t, u0 ) − u∗ X2 < ε
for all t ≥ 0.

Definition 7.1.19 makes sense due to the fact that in infinite dimensions
there are infinitely many non-equivalent norms. For instance, the fronts are
2
0
(X1 , Cb,unif ) stable, where uX1 = supx∈R |u(x)ex |, see, e.g., [BK00, §9.3].
Note that e−x decays much faster than the difference of two fronts which
2

converge towards each other with some exponential rate.

Example 7.1.20. Consider
∂t u = ∂x2 u + c∂x u + u.
0
Obviously u = 0 is unstable in Cb,unif . However, in a weighted space stability
can be obtained. The transform u(x, t) = v(x, t)eβx yields
∂t v =∂x2 v + 2β∂x v + β 2 v + c∂x v + cβv + v
=∂x2 v + (2β + c)∂x v + (β 2 + cβ + 1)v .
Hence, v = 0 may expected to be stable if β 2 + cβ + 1 < 0 which  can be sat-
isfied for |c| > 2 by choosing β ∈ (β− , β+ ) with β± = −c/2 ± c2 /4 − 1. In-
deed, setting γ1 = 2β+c, γ2 = β 2 +cβ+1 and v(x, t) = w(x+γ1 t, t) we obtain
2
∂t w(x, t) = ∂x2 w(x, t) + γ2 w(x, t) and from this w(x, t) = eγ2 t et∂x w0 (x) → 0
222 7. Some dissipative PDE models

0
in Cb,unif as t → ∞. In summary, we have asymptotic (X, X) stability of
u = 0, where uX = supx |u(x)e−βx |, i.e., u0 X ≤ δ ⇒ u(t)X → 0 as
t → ∞. Note that this does not mean that u stays bounded. What happens
is that for c > 2 and hence β < 0 the (growing) mass is transported to
−∞ where it vanishes in the X-norm. This is sometimes called convective
stability.
Remark 7.1.21. The idea of weighted variables from Example 7.1.20 can
be used to prove the nonlinear stability of the front solutions uf for |c| > 2.
In a co-moving frame ξ = x − ct the deviation v = u − uf from the front
satisfies
∂t v(ξ, t) = ∂ξ2 v(ξ, t) + c∂ξ v(ξ, t) + (1 − 2uf (ξ))v(ξ, t) − v(ξ, t)2 .
The weighted variable w(ξ, t) = v(ξ, t)e−βξ satisfies
∂t w(ξ, t) = Lw w(ξ, t) − eβξ w(ξ, t)2 ,
with
Lw w(ξ, t) = ∂ξ2 w(ξ, t) + (2β + c)∂ξ w(ξ, t) + (β 2 + cβ + 1 − 2uf (ξ))w(ξ, t).
Since uf ≥ 0 we can find a β < 0 such that w = 0 is linearly stable, e.g., in
H 1 . However, the nonlinearity is not well defined in H 1 due to the factor
eβξ . Therefore, the equation for v and w have to be combined, namely
∂t v(ξ, t) =Lv v(ξ, t) + (2 − 2uf (ξ))eβξ w(ξ, t) − v(ξ, t)2 ,
∂t w(ξ, t) =Lw w(ξ, t) − v(ξ, t)w(ξ, t),
with
Lv v(ξ, t) = ∂ξ2 v(ξ, t) + c∂ξ v(ξ, t) − v(ξ, t).
Since the semigroups generated by Lv and Lw decay with some exponential
rate and since (2−2uf (ξ))eβξ is bounded, the asymptotic stability of v=w=0
follows, i.e., uf is asymptotically stable w.r.t. perturbations v which are
small w.r.t. the norm vH 1 + veβx H 1 . This idea goes back to [Sat76].

For an extensive discussion of traveling wave solutions in parabolic sys-

tems see [VVV94]. An overview about tools for investigating the stability
of pulses and fronts can be found in [San02].

7.2. The Allen-Cahn equation

Closely related to the KPP equation is the Allen-Cahn equation
(7.13) ∂t u = ∂x2 u + u − u3 ,
with t ≥ 0, x ∈ R, and u = u(x, t) ∈ R, a prototype of a reaction-diffusion
equation with bistable behavior, cf. also §5.3. The equation again consists
7.2. The Allen-Cahn equation 223

of two parts, namely the diffusion term ∂x2 u and the nonlinear reaction term
u − u3 . Inserting u(x, t) = v(t) into (7.13) gives the one-dimensional ODE
(7.14) v̇ = v − v 3 .
The phase portrait shows that the fixed point v = 0 is unstable and that the
two other fixed points v = ±1 are asymptotically stable. The last two fixed
points can be interpreted as two stable phases. Hence, in the Allen-Cahn
equation both phases have the tendency to grow. At their interfaces both
phases try to move the interfaces towards the other phase. The Allen-Cahn
equation therefore describes interface motion between two stable phases with
applications in material science, image processing, biology and geology.
The local existence and uniqueness of solutions follows exactly as for
the KPP equation. As a consequence of the dynamics of (7.14) and the
maximum principle we have the following global existence und uniqueness
result for mild solutions.
Theorem 7.2.1. For u0 ∈ Cb,unif 0 there exists a unique mild solution u ∈
C([0, ∞), Cb,unif ) of the Allen-Cahn equation with u|t=0 = u0 .
0

Time-independent solutions u(x, t) = v(x) satisfy 0 = ∂x2 v + v − v 3 .

In the (v, v )-phase plane, we find a family of periodic solutions around
v = (0, 0) and two heteroclinic connections, cf. Exercise 2.10 and §5.3.4.
We do not further discuss the periodic solutions since they turn out to be
unstable as stationary solutions of the Allen-Cahn equation (7.13). The
heteroclinic connection between (−1, 0) and (1, 0) is called kink and the
connection between (1, 0) and (−1, 0) is called anti-kink. In contrast to
the stable/unstable steady state connection the stable/stable steady state
connection exists exactly for one velocity, here c = 0. The reason for this
is that the velocity is needed as a parameter to get an intersection of the
one-dimensional unstable manifold of (−1, 0) and the one-dimensional stable
manifold of (1, 0).
The stability of these kink solutions w.r.t. small spatially localized pertur-
bations can again be seen with the help of the maximum principle. If the
small perturbation v of the kink solution ukink is contained between two
translates of the kink then by the maximum principle it will stay there for
all times. In detail, if
ukink (x − x− ) ≤ ukink (x) + v(x, 0) ≤ ukink (x + x+ )
for all x ∈ R, then
ukink (x − x− ) ≤ ukink (x) + v(x, t) ≤ ukink (x + x+ )
for all t ≥ 0 and all x ∈ R. See Figure 7.10.
224 7. Some dissipative PDE models

-1

Figure 7.10. Stability of the kink w.r.t. to small spatially localized

perturbations. The perturbations are confined by parallel kinks.

Further remarks. We expect that due to asymptotic stability of the

states for x → ±∞ the kink solutions are stable w.r.t. to a much larger class
of perturbations. In the following we state a number of results concerning
the long time behavior of solutions, but refer to the literature for proofs and
more details, in particular to [ER99].
Linearizing the Allen-Cahn equation around the kink-solution ukink gives
∂t v = ∂x2 v + v − 3u2kink v = Lv + Bv
where Lv = ∂x2 v−2v and Bv = −3(u2kink −1)v. The operator B is a relatively
compact perturbation of the operator L since it contains less derivates than
the elliptic operator −L and since it decays to 0 with some exponential
rate for |x| → ∞. As a consequence, the essential spectra in the sense
of [Hen81, Chapter 5] of L + B and L coincide, i.e., they only differ by
discrete eigenvalues, cf. [Hen81, Chapter 5, Theorem A.1]. We find σ(L) =
{−k 2 − 2 : k ∈ R} for the spectrum of L in various spaces, such as Cb,unif
0 or
2
L .
For the computation of the discrete spectrum we remark that L + B is
self-adjoint in L2 , i.e., ((L + B)u, v)L2 = (u, (L + B)v)L2 . Hence, we only
have real eigenvalues. There is one eigenvalue zero corresponding to the
derivative ∂x ukink of the kink-solution. Since this eigenfunction possesses
no zeroes and since the eigenvalue problem comes from the linearization
of a scalar translational invariant reaction-diffusion problem it must be the
eigenfunction to the largest eigenvalue. Hence, all other eigenvalues must be
strictly negative. The nonlinear stability of the kink-solution follows with
the help of the center manifold theorem [Hen81, Ch. 6]. Since the operator
0
L + B is sectorial both in Cb,unif or L2 the position of the spectrum gives
the growth rates of the associated semigroups, cf. [Hen81, Theorem 1.3.4].
Thus, we have
Theorem 7.2.2. Let X = Cb,unif 0 or L2 . The family of translates of the
kink solutions is asymptotically stable. In detail, there are δ0 > 0, β > 0
and C > 0 such that the following holds. For u0 ∈ X with u0 −ukink X ≤ δ0
7.3. Intermezzo: Fourier transform 225

there is a γ ∈ R such that

u(t, u0 ) − ukink (· + γ)X ≤ Ce−βt .

Solutions of the Allen-Cahn equation with more than one interface are
the next non-trivial set of solutions which can be understood completely.
These solutions consist of almost flat regions, where they are close to the
exponentially stable states ±1. In between there are the interface regions,
where the solutions connect the almost flat regions.
We define classes of functions with m interfaces
Im = {u ∈ Cb0 is an interface function with m zeros, lim u(x) = 1 or − 1}.
x→±∞

Then from the analysis in [CP89, CP90] it is known that the interfaces
annihilate. Moreover, it is known that this happens in a very robust way,
i.e., small perturbations do not change the dynamics. With the help of the
maximum principle the following global result can be obtained
lim u(·, t) ∈ I0 , if u(·, 0) ∈ I2m , lim u(·, t) ∈ I1 , if u(·, 0) ∈ I2m+1 ,
t→∞ t→∞
with I0 = {−1, 1}. The time for the annihilation of the interfaces goes
exponentially (O(eCL )) with the distance L of the interfaces. It has been
shown that by separating the interfaces in the right way, i.e., making the
distances bigger and bigger for x → ±∞ it is possible to have an annihilation
of interfaces for all times, i.e., infinitely many annihilations.
Comparing to the Allen-Cahn equation on a bounded domain where it
has a finite dimensional attractor, cf. §5.3, we see that large time asymptotics
over unbounded domains can be fundamentally different from the bounded
domain case.

Figure 7.11. Annihilation of two interfaces to a solution with 3 interfaces

7.3. Intermezzo: Fourier transform

The analysis of the KPP and AC equations and of many more equations
of similar form, heavily relies on the explicit solution formula for the linear
diffusion equation. In general, explicit solution formulas of linear partial
differential equations are not available. A powerful tool to analyze linear
226 7. Some dissipative PDE models

partial differential equations with constant coefficients on the real line is

Fourier transform. We first define it for functions in L1 (Rd ). We shall
find that Fourier transform extends to an unitary operator in L2 (Rd ) and
finally to tempered distributions. We use Fourier transform to character-
ize the Sobolev spaces H s (Rd ) and introduce some further function spaces
needed in the sequel. In applications we will mostly focus on the case of one
unbounded direction, but the generalization to Rd will be straightforward.
Fourier transform is treated in most books on analysis, and in most intro-
ductory books on PDEs. There are also a number of books specialized on
Fourier transform, see for instance [Duo01]. We recall the main concepts
and adapt them to our purposes.

7.3.1. Fourier transform in L1 . Before the definition of Fourier trans-

form we recall without proof some important theorems from integration
theory and basic facts about Lp -spaces.
Definition 7.3.1. For 1 ≤ p < ∞ let Lp (Rd , Y ) be the space of measurable
functions with values in a Banach space Y such that the Lp -norm which
defined by
 1/p
p
(7.15) f L =
p f (x)Y dx
Rd

is finite.

If f1 , f2 ∈ Lp and f1 = f2 almost everywhere (a.e.), i.e., except on a set

of measure zero, then f1 − f2 p = 0. We start with two convergence results
in case Y = R, namely the lemma of Beppo-Levi [Alt16, Folgerung 1.12 c)]
and Lebesgue’s dominated convergence theorem [Alt16, Satz 1.21].
Lemma 7.3.2. Assume that fn ∈ L (R , R) for n ∈ N with 0 ≤1 fnd % f for
1 d

n → ∞ a.e., and that lim supn→∞ fn (x) dx<∞. Then f ∈ L (R , R) and

fn − f L1 → 0 for n → ∞.
Theorem 7.3.3. Let fn , f : Rd → R be measurable functions and let g ∈
L1 (Rd , R). If |fn | ≤ g for all n ∈ N a.e. and if fn → f a.e. for n → ∞,
then f ∈ L1 (Rd , R) and fn − f L1 → 0 for n → ∞.

We further need Fubini’s theorem about iterated integrals [Alt16, Satz

4.10], and the fundamental lemma of variations [Alt16, Lemma 2.21].
Theorem 7.3.4. Let Y be a Banach space and 1 ≤ p < ∞. For f ∈
Lp (Rd1 × Rd2 , Y ) we have the existence of

F (x1 ) = f (x1 , x2 ) dx2 ,
R d2
7.3. Intermezzo: Fourier transform 227

for almost all x1 ∈ Rd1 , and F ∈ Lp (Rd1 , Y ). Moreover, by symmetry

  
f (x1 , x2 ) d(x1 , x2 ) = f (x1 , x2 ) dx2 dx1
Rd1 ×Rd2 R d1 R d2
 
= f (x1 , x2 ) dx1 dx2 .
R d2 R d1

Lemma 7.3.5. Let Ω ⊂ Rd be an open set and Y a Banach space. For

g ∈ L1 (Ω, Y ) we have the equivalence of (i) Ω ζ(x)g(x)dx = 0 for all ζ ∈
C0∞ (Ω) and (ii) g = 0 a.e. in Ω.

For the rest of the section we restrict to Y = R. Allthough f ∈ Lp (Rd )

in general is not defined point-wise, it is clear that to have the existence of
the integral in (7.15), in some sense |f (x)| should decay to zero sufficiently
fast as |x| → ∞. This can be made precise as follows.
Lemma 7.3.6. If f ∈ L1 (R), then limn→∞ f (x + n) = 0 for almost all
x ∈ R.

Proof. Since for all x ∈ R there exists a z ∈ Z with

 x + z ∈ [0, 1] we may
assume x ∈ [0, 1]. Let fk (x) = f (x + k), gn (x) = nk=1 |fk (x)|. Then
 1  1n n  1

gn (x) dx = |fk (x)| dx = |f (x + k)| dx
0 0 k=1 k=1 0
 n 
= |f (x)| dx ≤ |f (x)| dx
1 R
1
is bounded by a constant M < ∞. Thus, gn ≤ gn+1 and 0 gn (x) dx ≤ M ,
i.e., (gn )n∈N is monotonically increasing with a uniformly bounded L1 -norm.
By the Beppo Levi Lemma 7.3.2 there exists a g ∈ L([0, 1]) with gn → g
a.e., and hence |fk | → 0 a.e.. 
Fourier transform can be interpreted as an expansion into the eigen-
functions eik·x , k ∈ Rd , of spatially homogeneous linear partial differential
operators on Rd . We start with a definition for L1 functions which is later
on extended to L2 and tempered distributions.
 :
Definition 7.3.7. (Fourier transform) For u ∈ L1 (Rd ) we define u
Rd → C by

1
(7.16) (k) =
u e−ik·x u(x) dx,
(2π)d Rd
 = F u the Fourier transform of u.
and call u

 is bounded with the estimate

Clearly, u

1
(7.17) sup |
u(k)| ≤ |u(x)| dx.
k∈R (2π)d Rd
228 7. Some dissipative PDE models

The proof of the continuity of u  is the goal of Exercise 7.11. Similarly, it

is easy to prove that decay of u corresponds to smoothness of u . In the
following α = (α1 , . . . , αd ) ∈ Nd0 is some multi-index, |α| = α1 + . . . + αd ,
and
xα = xα1 1 · · · xαd d and ∂xα = ∂xα11 · · · ∂xα11 .

Theorem 7.3.8. (Smoothness and decay) a) If u ∈ Cbm fulfills ∂xα u ∈ L1

for |α| ≤ m then F (∂xα u)(k) = (ik)α u
(k) for |α| ≤ m. Moreover, there exists
a constant C such that for all k ∈ Rd :
u(k)| ≤ C(1 + |k|)−m .
|
b) If vα (x) = xα u(x) satisfies vα ∈ L1 for |α| ≤ m, then u
 ∈ Cbm and
F (vα ) = (−i∂k ) u
α .
Proof. Using integration by parts the estimate in a) is straightforward. 
Using the estimate (7.17) immediately yields the so called roll off formula
which subsequently will turn out to be very useful.

Lemma
 7.3.9. For u, v ∈ L 
1 we have u v, u
v ∈ L 1 and (x)v(x) dx =
Rd u
R d u(x)
v (x) dx.
We will also use the following formulas which again follow directly from
the definition of Fourier transform.
Lemma 7.3.10. a) Let u ∈ L1 . For a ∈ Rd and ua (x) = u(x − a) we have
a (k) = e−ik·a u
u (k).
(rk).
b) For r > 0 and v(x) = u(x/r) we have v(k) = rd u
Above we noted that Fourier transform can be interpreted as an expan-
sion into the eigenfunctions eik·x , k ∈ Rd , of spatially homogeneous linear
partial differential operators on Rd . The coefficients of this expansion define
the Fourier transform. Hence, we expect that the reconstruction formula for
u from u is as follows.
Theorem 7.3.11. (Inverse Fourier transform) Assume u ∈ L1 (Rd ) and
 ∈ L1 (Rd ). Then
u 
u(x) = (k) dk.
eik·x u
Rd
 → u is called inverse Fourier transform and is denoted by u =
The map u
F −1 u
.
The proof is given below. Note that u ∈ L1 does not imply u  ∈ L1 . The
−1
inversion formula means that F and F map the subspace L ∩F (L1 ) ⊂ Cb0
1

bijectively into itself. However, since a characterization of F (L1 ) ⊂ Cb0 is a

non-trivial task, this is not very satisfactory.
7.3. Intermezzo: Fourier transform 229

7.3.2. Fourier transform in S, L2 , and general Lp -spaces. The goal is

to find spaces where the analytic properties of Fourier transform are better
than in L1 . Our starting point are the relations
(k) and
F (∂xα u)(k) = (ik)α u F −1 ((i∂k )α u
)(x) = xα u(x).
We look for a subspace of L1 which is mapped into itself by F , differentiation,
and multiplication with polynomials. Such a subspace is the Schwartz space
of quickly decaying functions.
Definition 7.3.12. The Schwartz space S = S(Rd ) is defined by
S = {u ∈ C ∞ (Rd ) : xα ∂xβ u(x) is bounded for any α, β ∈ Nd0 }.
Theorem 7.3.13. For u ∈ S we have u
 ∈ S and vice versa.

Proof. Let u ∈ S, and let α, β ∈ Nd0 be some multi-indices. Clearly vβ (x) =

xβ u(x) satisfies vβ ∈ S, hence u  ∈ C |β| (Rd ) and ∂kβ u
 = (−1)|β| F (vβ ) by
Theorem 7.3.8. Since also ∂xα (vβ ) ∈ S ⊂ L1 we have F (∂xα (vβ )) ∈ Cb0 , i.e.,
k → k α ∂kβ u
(k) is in Cb0 . 
Theorem 7.3.14. F is a bijection on S, i.e., F −1 F |S = F F −1 |S = I.

Proof. It will turn out in §7.3.4 that the Fourier transform of the
δ-distribution is the constant function with value 1. With this knowledge
the formal argument is as follows.
   
−1 1 −ik·y
(F F u)(x) = e ik·x
e u(y) dy dk
Rd (2π)d Rd
   
1 ik·(x−y)
= e dk u(y) dy
Rd (2π)d Rd

= δ(x − y)u(y) dy = u(x).
Rd
In order to make this formal argument rigorous we use an approximation
of the δ-distribution and the roll-off formula. Hence, we set v(k) = e−|k| /2
2

and let vr (k) = v(k/r) for r > 0. For r → ∞ and fixed k ∈ Rd we have
vr (k) → 1 monotonically from below.
For u ∈ S we set ux (k) = u(k − x) and find
  
ik·x
e u (k)vr (k) dk = u/
−x (k)vr (k) dk = u−x (k)
vr (k) dk
Rd Rd Rd
 
= u−x (k)rd v(rk) dk = u−x (m/r)v (m) dm
Rd Rd

= (2π)−d/2 u−x (y/r)v(y) dy
Rd
230 7. Some dissipative PDE models

by Lemma 7.3.10 and since F (e−|x| /2 )(k) = (2π)−d/2 e−|k| /2 , cf. Exercise
2 2

7.12. Now
 
ik·x
e u (k)vr (k) dk → (k) dk as r → ∞
eik·x u
Rd Rd
by the Lebesgue dominated convergence theorem 7.3.3 since the integrand
is dominated by |u| ∈ L1 . Similarly,
 
−d/2 −d/2
(2π) u−x (y/r)v(y) dy → (2π) u(x) v(y) dy = u(x)
Rd Rd

as → ∞ since the integrand is dominated by u∞ v ∈ L1 .

 r ik·x Hence
Rd e (k) dk = u(x).
u 
The inversion formula is true in much larger sets. It is first extended
from S to L1 .

Proof of Theorem 7.3.11. Let u, u (x) = Rd eik·x u
 ∈ L1 and set u (k) dk.
By Theorem 7.3.8 we have u  ∈ Cb and we want to show that u
0  = u. By
Theorem 7.3.14, for any φ ∈ S there exists a v ∈ S with v = φ. Then, by
Lemma 7.3.9 and again by Theorem 7.3.14 we find
 
u(y)
v (y) dy = (y)v(y) dy
u
Rd Rd
   
= (y)
u e v(k) dk dy =
ik·y
v(k)
u(k) dk.
Rd Rd Rd
Hence  
(u − u
)
v dx = (u − u
)φ dx = 0,
Rd Rd
 = F −1 u
and since φ was arbitrary we find u  = u by Lemma 7.3.5. 
The following crucial result allows us to bound the norm of the Fourier
transform and of its inverse simultaneously. This will allow us to go from
physical space to Fourier space and back, and to subsequently extend F to
L2 (Rd ).
Theorem 7.3.15. (Parseval’s identity for Schwartz functions) For
u, v ∈ S we have

u, vL2 = (2π)d

u, vL2 ,

where
u, vL2 = Rd u(x)v(x) dx, and in particular
(7.18) uL2 = (2π)−d/2 uL2 .


Proof. Let u, v ∈ S and set g = v which implies (2π)d g = v. Then, by

Lemma 7.3.9,
 

u, vL2 = u(x)v(x) dx = (2π)d u(x)
g (x) dx
Rd Rd
7.3. Intermezzo: Fourier transform 231


=(2π)d (x)g(x) dx = (2π)d

u u, vL2 .
Rd

Since S is dense in
L2 we find that F can be extended to an isomorphism
in L which is isometric up to a factor (2π)d .
2

Theorem 7.3.16. There exists a unique linear bounded invertible map F :

L2 → L2 which equals F on S. Its inverse F −1 equals F −1 on S.
Remark 7.3.17. There exist u ∈ L2 (R) which are not in L1 (R), for instance
u(x) = (1+|x|)−1 . For these u the Fourier transform F u cannot be computed
by the formula (7.16). However, with
 n
1
n (k) =
u e−ik·x u(x) dx
(2π)d −n
we always have 
u−u n L2 → 0 again by the Lebesgue dominated conver-
gence theorem 7.3.3.

Similar to Remark 5.2.11 from  uL∞ ≤ uL1 and  uL2 =

(2π) −d/2 uL2 , by the so called Riesz-Thorin interpolation F can be es-
timated on Lp (Rd ) for 1 ≤ p ≤ 2, cf. [Duo01, Corollary 1.20].
Theorem 7.3.18. (Hausdorff-Young) For 1 ≤ p ≤ 2 there exists a Cp >

 ∈ Lp (Rd ), 1/p + 1/p = 1, and
0 such that for f ∈ Lp (Rd ) we have u
uLp ≤ Cp uLp .


7.3.3. Convolution. The convolution of two functions is formally defined

by 
(u ∗ v)(x) = u(x − y)v(y) dy.
Rd
It occurs as the Fourier transform of the product of two functions, in detail

−d
F (uv)(k) = (2π) e−ik·x u(x)v(x) dx
R
d
 
−d −ik·x
= (2π) u(x)e v()e dl dx
il·x
Rd Rd
 
1
= v() d
u(x)e−i(k−)·x dx d
(2π)
R R
d d

= u(k − ) v () dl = (u ∗ v)(k)

Rd
by Fubini’s theorem 7.3.4 if for instance u, v ∈ S. The identity can be
extended to u, v ∈ L1 by the dominated convergence theorem, similar to the
proof of Theorem 7.3.11. For later purposes we also need the formula
(7.19) F −1 (
uv) = (2π)−d u ∗ v
232 7. Some dissipative PDE models

which can be obtained in the same way. We have the following estimate
Lemma 7.3.19. (Young’s inequality for convolutions) For 1 ≤ p ≤ ∞
and f ∈ Lp (Rd ), g ∈ L1 (Rd ) we have f ∗g ∈ Lp and f ∗gLp ≤ f Lp gL1 .

Proof. For p = ∞ we have Rd |f (x − y)g(y)| dy ≤ f L∞ gL1 < ∞ a.e..
Thus let 1 ≤ p < ∞. The map (x, y) → f (x − y)g(y) is measurable as a
product of measurable functions. By Hölder’s inequality we obtain
 

|f (x − y)g(y)| dy = |f (x − y)||g(y)|1/p|g(y)|1/p dy
Rd Rd
 1/p  1/p
≤ |f (x − y)| |g(y)| dy
p
|g(y)| dy
Rd Rd

for 1/p + 1/p = 1, which implies

  p  
p/p
|f (x − y)g(y)| dy dx ≤ g1 |f (x − y)|p dx|g(y)| dy
Rd Rd Rd Rd
1+p/p
= gL1 f pLp < ∞.
Taking the pth root and using 1 + p/p = p yields the result. 
Example 7.3.20. For the linear diffusion equation ∂t u = ∂x2 u we once more
derive the solution formula (7.6), i.e.,
 ∞
1 (x−y)2
u(x, t) = √ e− 4t u(y, 0) dy,
4πt −∞
now using the Fourier transform. Inserting the approach
 ∞
u(x, t) = (k, t)eikx dk,
u
−∞

into the linear diffusion equation ∂t u = ∂x2 u gives

 ∞
(k, t) + k 2 u
(∂t u (k, t))eikx dk = 0,
−∞

and so for each k ∈ R the ODE ∂t u  = −k 2 u

, which is solved by u
(k, t) =

G(k, t) 
u(k, 0) with G(k, t) = e −k 2 t . Using (7.19) we then have
u(x, t) =(F −1 u  t)
)(x, t) = (F −1 G(·, u(·, 0))(x, t)
 ∞
1
= G(x − y, t)u(y, 0) dy.
2π −∞
 is given by
The inverse Fourier transform of G
π − x2
G(x, t) = e 4t ,
t
7.3. Intermezzo: Fourier transform 233

see Exercise 7.12, which yields the assertion. We have the following com-
mutative diagram
G(t)
u0 −→ u(·, t)
−1
↓F ↑F

G(t)
0
u −→ u
(·, t)

7.3.4. Fourier transform of tempered distributions. So far we de-

fined Fourier transform for u ∈ Lp with p ∈ [1, 2]. These spaces are still
too small for many applications since one wants to use Fourier transform
0
for instance also in spaces as Cb,unif . Since any bounded function defines
a (regular) distribution, the basic idea is to define the Fourier transform
of distributions by rolling off the transform to the test function. In detail,
recall that a classical function u can be interpreted as a regular distribution
Tu by 
Tu (φ) = u(x)φ(x) dx for all φ ∈ D,
R
cf. Remark 5.2.1. We defined derivatives ∂ j T for T ∈ D = L(D, K) via
∂ j T (φ) = (−1)j T (∂xj φ),
which corresponds to the formula of integration by parts. Similarly, for
T ∈ D we might want to define F T by
 for all φ ∈ D.
(F T )(φ) = T (φ)
However, for φ ∈ D the problem is that φ does not have compact support and
hence is not a test function, since first the Paley-Wiener theorem, cf. Lemma
6.2.9, implies that a compact support of φ implies the analyticity of φ, and
secondly that the identity theorem for complex power series implies that an
analytic function with compact support must be identical zero. Therefore,
we shall restrict the definition of Fourier transform to the subspace S ⊂ D .
The Schwartz space S can be equipped with some topology, namely it can
be made a Fréchet-space with the semi-norms pβ,m , where
pβ,m (φ) := sup |∂xβ φ(x)(1 + |x|)m | < ∞.
x∈Rd
An equivalent characterization is
φk → φ in S ⇔ xα ∂xβ φk (x) → xα ∂xβ φ(x) uniformly on Rd for any α, β ∈ Rd .
For details see [Wlo87]. The space of tempered distributions S is defined as
the dual space of S consisting of all continuous linear functionals T : S → C.
Definition 7.3.21. T ∈ S is called a tempered distribution if there exists
m ∈ N, β ∈ Nd , and a C > 0 such that
|T (φ)| ≤ Cpβ,m (φ) for all φ ∈ S.
234 7. Some dissipative PDE models

Since D ⊂ S we have S ⊂ D as proper subsets.

Example 7.3.22. A continuous function u which satisfies supx∈R |u(x)| =
Cu < ∞ can be interpreted as tempered distribution Tu since
 
|Tu (φ)| ≤ |u(x)φ(x)| dx ≤ Cu |φ(x)| dx
R R

≤ Cu (1 + x2 )−1 dx p0,2 (φ) ≤ Cp0,2 (φ)
R

for all φ ∈ S. More general, Tu : S → C with φ → Rd u(x)φ(x) dx defines
a tempered  distribution if for instance for u ∈ L1loc there exists an N ∈ N
−N
such that R |u(x)|(1 + x) dx < ∞ or if u ∈ Lp (Rd ) for some p ≥ 1. On
the other hand, u(x) = ea|x| with a > 0 gives a Tu ∈ D \ S .

Since F maps S to S we may now define

Definition 7.3.23. The Fourier transform F is defined for tempered dis-
tributions T via
(F T )(φ) = T (F φ) for all φ ∈ S.

This extended Fourier transform F : S → S is the adjoint of the clas-

sical Fourier transform F : S → S. Again F : S → S is an isomorphism.
Example 7.3.24. For u(x) = eik·x we have F Tu = δk , the δ-distribution in

 dξ = φ(k).
k ∈ R, since (F Tu )(φ) = eik·ξ φ(ξ)
Remark 7.3.25. Associated
 to the δ-distribution sometimes the δ-function
formally satisfying R δ0 (x)φ(x) dx = φ(0) is introduced. To show that δ0 is
not a regular distribution assume that δ0 = Tu for some u ∈ Lploc . If φ ∈ D
then also ψ = x2 φ is a test function and we have

0 = ψ(0) = δ0 ψ = u(x)x2 φ(x)2 dx
R

for all φ, hence = 0 a.e., hence u = 0 a.e., which contradicts δ0 = 0.

u(x)x2

Nevertheless the notation R δ0 (x)φ(x) dx = φ(0) for δ0 φ is often used.
The motivation is that δ0 can be approximated by regular distributions. For
instance the fundamental solution
1
e−|x| /4t
2
G(x, t) = 1/2
(4πt)
of the linear heat equation ∂t u = ∂x2 u approximates the δ0 -distribution in
the sense of distributions for t → 0, i.e.,

G(x, t)φ(x) dx → φ(0)
R
7.3. Intermezzo: Fourier transform 235

for all φ ∈ D and t → 0. To show this let ε > 0 and φ ∈ D. Then there
exists a δ > 0 with |φ(x) − φ(0)| < ε/2 for |x| < δ, and thus
   
   
 G(x, t)φ(x) dx − φ(0) =  G(x, t)(φ(x) − φ(0)) dx
   
R
R 
≤ . . . dx + . . . dx < ε/2 + ε/2
|x|<δ |x|>δ

for t > 0 sufficiently small, where we used in the first integral that
R G(x, t) dx = 1.

7.3.5. Sobolev spaces. For the handling of nonlinear PDEs, Lp -spaces are
of no use since they are not closed under multiplication. In §5.2.1 we defined
the Sobolev spaces H m (Ω) over some bounded domain Ω, and in §5.2.2 we
showed that for suitable boundary conditions there is an easy alternative
characterization using Fourier series, see in particular Lemma 5.2.9. Here
we use Fourier transform to study the Sobolev spaces over Rd . Again we
usually do not distinguish real-valued from complex-valued or vector-valued
functions.
Definition 7.3.26. For m ∈ N define the Sobolev space
W m,p (Rd ) := {u ∈ Lp (Rd ) : ∂xα u ∈ Lp (Rd ) for |α| ≤ m},
where ∂xα with multi-index α denotes the distributional derivative of order
α. It is equipped with the norm
⎛ ⎞1/p

uW m,p = ⎝ ∂xα upLp ⎠ .
|α|≤m

Of particular importance are

 the Hilbert spaces H m (Rd ) = W m,2 (Rd ) with
scalar product
u, vH m = |α|≤m
∂xα u, ∂xα vL2 .

In Theorem 7.3.8 we have seen that smoothness of u is related to decay

 and vice versa. As in Remark 5.2.12, this can be used to generalize
of u
Sobolev spaces H m to non-integer m.
Definition 7.3.27. Let ρ(x) = (1 + |x|2 )1/2 . For m ≥ 0 define the weighted
L2 -space
L2m := {u ∈ L2 : uL2m := uρm L2 < ∞}
and the Sobolev space H m := {u ∈ L2 : u
 ∈ L2m } with norm uH m = 
uL2m .

For m ∈ N0 this definition coincides with the previous one. Next, sim-
ilar to Lemmas 5.2.3 and 5.2.24 we obtain the following embedding and
multiplication results.
236 7. Some dissipative PDE models

Lemma 7.3.28. We have H 1 (R) ⊂ C 0,1/2 (R) and

(7.20) u2∞ ≤ 2uL2 ∂x uL2 ,


(7.21) |u(x) − u(y)| ≤ |x − y|∂x uL2 .

Proof. Since C 1 (R)∩H 1 (R) is dense in H 1 (R) it is sufficient to prove (7.20)

and (7.21) for u ∈ C 1 (R) ∩ H 1 (R). We have
 x  x
d 2
u2 (x) = u (s) ds = 2 u(s)∂s u(s) ds ≤ 2uL2 ∂x uL2 .
−∞ ds −∞

Using the Cauchy-Schwarz inequality yields the second estimate

 y   y
  

|u(x) − u(y)| =  ∂s u(s) ds ≤ 1|∂s u(s)| ds ≤ |x − y|∂x uL2 ,
x x

which immediately implies the embedding H 1 (R) ⊂ C 0,1/2 (R). 

Lemma 7.3.29. For u, v ∈ H m (Rd ), m > d/2, we have u, v ∈ L∞ and

uvH m ≤ CuH m vH m .

Proof. For ρ(k) = (1 + |k|2 )1/2 we have ρ−m ∈ L2 if m > d/2, and thus

(7.22) uL1 = Cρ−m ρm u

 L1 ≤ Cρ−m L2 
uL2m ≤ C
uL2m ,

which implies

u∞ ≤ C
uL1 ≤ C
uL2m ≤ CuH m .

Next ρm (k) ≤ C(ρm (k − ) + ρm ()) for a constant C, and thus by Lemma

7.3.19 we obtain

uvH m ≤C
u ∗ vL2m ≤ C(ρm u
L2 
v L1 + 
uL1 ρm vL2 )
uL2m 
≤2C v L2m ≤ 2CuH m vH m ,

where we used again (7.22). 

For later purposes we also need weighted Sobolev spaces.

Definition 7.3.30. For r, m ≥ 0 and ρ(x) = (1 + x2 )1/2 let

r
Hm := {u ∈ L2 : uρm H r < ∞}.

The fact that decay in Fourier space corresponds to smoothness in x-

space can now be quantified as follows.
r and H m .
Lemma 7.3.31. Fourier transform is an isomorphism between Hm r
7.4. The Burgers equation 237

7.4. The Burgers equation

The (inviscid) Burgers equation
1
(7.23) ∂t u = − ∂x (u2 ),
2
with t ≥ 0, x ∈ R, and u(x, t) ∈ R, appears as a model for the velocity field
of (an inviscid) compressible one-dimensional fluid or gas. Traffic problems
are very often modeled as a one-dimensional gas with the cars taking the
role of the gas particles. Therefore, the Burgers equation also occurs as a
model for traffic flows.
Example 7.4.1. Consider traffic flow with the following modeling assump-
tions: a1) There is only one lane. a2) The local (car)-density
number of cars at position x and time t
ρ(x, t) =
unit length
makes sense. In particular, we consider a long stretch of the road. a3) The
local speed of traffic v(x, t) is a function of ρ(x, t) only. (This is typically
revised if viscous or other effects are to be taken into account.)
Let x1 < x2 be arbitrary. With the flux function f (ρ) = ρv(ρ), where v
is the speed, we obtain the integral form of conservation of mass (cars)
 x2  x2
∂t ρ(x, t) dx = ∂t ρ(x, t) dx = f (ρ(x1 , t)) − f (ρ(x2 , t))
x1 x1
 x2
=− ∂x f (ρ(x, t)) dx.
x1
Since x1 and x2 are arbitrary we obtain
(7.24) ∂t ρ + ∂x (f (ρ)) = 0.
To close the equation we need a constitutive law v(ρ) for the dependence
of v on ρ. The following assumptions seem reasonable: There is a maximal
density ρmax with v(ρmax ) = 0 (traffic jam), and there is a maximal speed at
low density v(0) = vmax . W.l.o.g. we assume that ρmax = 1 and vmax = 1,
and set v(ρ) = 1 − ρ. Thus, f (ρ) = ρ(1 − ρ).

The inviscid Burgers equation (7.23) is a one-dimensional scalar conser-

vation law, which in general form reads
(7.25) ∂t u + ∂x (f (u)) = 0,
where f : R → R is called the flux function. This is an important class of
PDEs, e.g. [LeV92, Ser99, Ser00], [Deb05, §3-§7], or [Eva98, §3.4]. We
do not consider this class in this book, except for the brief section 7.4.2,
which however only serves to explain the nonlinear transport phenomena
associated with (7.25).
238 7. Some dissipative PDE models

If viscous effects are taken into account, one obtains the viscous Burgers
equation
1
(7.26) ∂t u = ν∂x2 u − ∂x (u2 ),
2
where ν > 0 is the viscosity parameter. This equation also occurs as a
modulation equation describing the behavior of, e.g., the free surface in
viscous flows [Uec07], or the evolution of wave numbers in stable periodic
pattern [DSSS09]. On the real line the parameter ν > 0 can be scaled to
1, hence we consider ν = 1 in the following.
Remark 7.4.2. a) Like for the KPP equation the maximum principle, cf.
Theorem 7.1.9, holds for the viscous Burgers equation, too. With the nota-
tion from page 215 we have
0 1
∂t u2 (x0 , t1 ) − u1 (x0 , t1 )
0 1
= ∂x2 u2 (x0 , t1 ) − u1 (x0 , t1 )
0 1
− ∂x u2 (x0 , t1 ) − u1 (x0 , t1 ) (u2 (x0 , t1 ) + u1 (x0 , t1 ))
0 1
− u2 (x0 , t1 ) − u1 (x0 , t1 ) ∂x (u2 (x0 , t1 ) + u1 (x0 , t1 ))
0 1
= ∂x2 u2 (x0 , t1 ) − u1 (x0 , t1 ) ,
and so the same argument applies.
b) The Burgers equation possesses the so called boost invariance u → u + c
and x → x − ct, i.e., if we add a constant to a solution and go into a co-
moving frame, then we again obtain a solution. Therefore, for a number
of
 purposes it is sufficient to restrict to, e.g., solutions with zero mean, i.e.
R u(x, t) dx = 0.

7.4.1. A local existence and uniqueness result. In order to handle

the viscous Burgers equation (7.26) or simply the Burgers equation as a
dynamical system we again choose the phase space
0
X = Cb,unif (R, R).
We have the local existence and uniqueness of mild solutions in X, and by
the maximum principle also the global existence and uniqueness of solutions
in X of Burgers equation.
Definition 7.4.3. u ∈ C([0, T0 ], Cb,unif
0 ) is called mild solution with initial
condition u0 of the Burgers equation (7.26) if u satisfies

1 T  
u(t) = T (t)u0 + T (t − s)∂x −(u(s)2 ds
2 0
where T (t) is the solution operator of the linear diffusion equation which has
been introduced in Definition 7.1.5.
7.4. The Burgers equation 239

Theorem 7.4.4. For all C0 > 0 there exists a T0 > 0 such that the following
holds. For u0 ∈ X with u0 X ≤ C0 . Then there exists a unique solution
u ∈ C([0, T0 ], X) of the Burgers equation (7.26) with initial datum u0 .

Proof. The proof of Theorem 7.1.7 has to be modified as follows. As before

we set
M = C 0 ([0, T0 ], {u(t) ∈ X : u(t) − T (t)u0 X ≤ C1 }),
but the variation of constant formula is then given by
 t
(7.27) u(t) = T (t)u0 + T (t − τ )∂x N (u)(τ ) dτ
0
where N (u) represents the term −u2 (x, t)/2. Additionally to T (t)∂x =
∂x T (t), see Remark 7.4.5, we need the estimates
T (t)u0 X ≤ u0 X ,
T (t)∂x u0 X ≤C(1 + t−1/2 )u0 X ,
1
N (u)M ≤ u2M ≤ C22 ,
2
N (u) − N (v)M ≤u + vM u − vM /2 ≤ C2 u − vM ,
which follow almost line for line as for the KPP equation. The second
estimate is not sharp. The one on the right-hand side is kept to have the
estimates which occur in the general case, cf. (5.7) and Remark 7.4.6.
With these estimates we obtain that the map F defined by the right-
hand side of (7.27) maps M into M , since
 t
F (u) − (T (t)u0 )t≥0 M = sup  T (t − τ )∂x N (u)(τ ) dτ X
t∈[0,T0 ] 0
 t
≤ sup T (t − τ )∂x N (u)(τ )X dτ
t∈[0,T0 ] 0
 t 
≤ sup 1 + (t − τ )−1/2 N (u)(τ )X dτ
t∈[0,T0 ] 0
 t 
≤ sup 1 + (t − τ )−1/2 dτ N (u)M
t∈[0,T0 ] 0

1/2
≤C22 T0 + 2T0 /2 ≤ C1
for T0 > 0 sufficiently small. The map F is a contraction in M , since
 t
F (u) − F (v)M = sup  T (t − τ )(N (u) − N (v))(τ ) dτ X
t∈[0,T0 ] 0
 t
≤ sup T (t − τ )(N (u) − N (v))(τ )X dτ
t∈[0,T0 ] 0
240 7. Some dissipative PDE models

 t 
≤ sup 1 + (t − τ )−1/2 (N (u) − N (v))(τ )X dτ
t∈[0,T0 ] 0
 t 
≤ sup 1 + (t − τ )−1/2 dτ (N (u) − N (v))M
t∈[0,T0 ] 0
 1
1/2
≤ T0 + 2T0 C2 u − vM ≤ u − vM
2
for T0 > 0 sufficiently small. Hence, there is a unique fixed point u of the
map F which is a mild solution of the Burgers equation. 

Remark 7.4.5. The operator T (t)∂x is applied to functions in Cb,unif 0 al-

1
though it is a priori only defined for functions in Cb,unif . However, there
is a continuous extension of this operator to a bounded linear operator on
0
Cb,unif 1
. The reason for this is as follows. On Cb,unif we have T (t)∂x = ∂x T (t)
such that on the dense subspace (Cb,unif ,  C 0
1 ) we have
b,unif

T (t)∂x uC 0 = ∂x T (t)uC 0 ≤ Ct−1/2 uC 0 .

b,unif b,unif b,unif

Therefore, due to Lemma 5.2.10 this bounded linear operator for t > 0 can
be extended to a bounded linear operator from the dense subspace to the
full space by
T (t)∂x u = lim ∂x T (t)un
n→∞

for u ∈ Cb,unif
0 , and un ∈ Cb,unif
1 with un → u in Cb,unif
0 for n → ∞.

Remark 7.4.6. Like the KPP equation, the Burgers equation is a semi-
linear parabolic equation [Hen81], i.e., the semigroup T (t) generated by
the operator −A = ∂x2 is smoothing and the nonlinearity N contains only
derivatives of lower order than A. As already explained in §5.2.3 the scheme
and the estimates of the last proof work for equations of the form

∂t u = ∂x2 u + f (u, ∂x u),

with f a smooth function. It is also the same scheme and the same estimates
as for the Navier-Stokes equations in §6.2.1. It does not work for

∂t u = ∂x2 u + f (u, ∂x u, ∂x2 u)

since then (1 + (t − τ )−1/2 ) in the proof of Theorem 7.4.4 has to be replaced

by (1 + (t − τ )−1 ) which is not integrable from 0 to t. As pointed out in
§5.3.3, the solutions of semi-linear parabolic equations are infinitely often
differentiable for every t > 0, and as pointed out in §6.2.2, they even become
analytic in a strip along the real axis in the complex plane for every t > 0.
7.4. The Burgers equation 241

7.4.2. Characteristics. In case of small diffusion we may expect that so-

lutions of the viscous Burgers equation behave similar to the solutions of the
inviscid Burgers equation. Thus, to get a first understanding of the dynam-
ical behavior of (7.26) we first consider the inviscid case (7.23) which can be
viewed as a nonlinear transport equation where the speed at (x, t) is given by
u(x, t). Therefore, we look for curves (t, x(t, x0 )) in R2 along which u is con-
stant. Differentiating u(x(t, x0 ), t) = u(x0 , 0) w.r.t. t gives (∂x u)ẋ+∂t u = 0.
Comparison with (7.23) shows that ẋ(t, x0 ) = u(x(t, x0 ), t) = u(x0 , 0). The
solution curves x(t, x0 ) = x0 + u(0, x0 )t of this ODE are called the charac-
teristics of (7.23). Along the characteristics the values of u are constant.
Obviously, even for smooth initial data x → u0 (x), characteristics can inter-
sect.
Example 7.4.7. Consider
⎧
⎨ 1, for x ≤ 0,
u0 (x) = 1 − x, for x ∈ [0, 1],
⎩
0, for x ≥ 1.
The point (x, t) = (1, 1) is the intersection point of the characteristics start-
ing in [0, 1] × {0}, see Figure 7.12.

Figure 7.12. Characteristics intersect and create a shock.

Once more, Example 7.4.7 shows that properties of dynamical systems

in infinite dimensions strongly depend on the phase space. The solutions
0
stay bounded in Cb,unif 1
, but explode in Cb,unif . Such solutions describe the
creation of a shock.
After the characteristics have intersected a new understanding of what
is meant by a solution is needed. We briefly review the concept of weak
solutions, see, e.g., [Eva98, §7] or [RR04, §3.4]. If the Burgers equation is
tested with a φ ∈ C0∞ (R × R+ ) we obtain via integration by parts
   
1
∂t u(x, t) + ∂x (u(x, t)2 ) φ(x, t) dx dt
R+ R 2
   
1
=− u(x, t)∂t φ(x, t) + u(x, t)2 ∂x φ(x, t) dx dt = 0.
R+ R 2
Hence, all derivatives fall on the smooth so called test function φ.
242 7. Some dissipative PDE models

Definition 7.4.8. u is called weak solution of the Burgers equation if

   
1 2
u(x, t)∂t φ(x, t) + u(x, t) ∂x φ(x, t) dx dt = 0
R+ R 2
for all φ ∈ C0∞ (R × R+ ).
Example 7.4.9. The shock

1, for x ≥ t/2,
u(x, t) =
0, for x < t/2,
is a weak solution of the inviscid Burgers equation (7.23) since
 ∞    ∞ ∞
1 2 1
u∂t φ + u ∂x φ dx dt = ∂t φ + ∂x φ dx dt
0 R 2 0 t/2 2
 ∞  2x  ∞
1
= ∂t φ(x, t) dt dx − φ(t/2, t) dt = 0
0 0 2 0
for all φ ∈ C0∞ (R × R+ ).

As the following example together with Exercise 7.4.9 shows, weak so-
lutions are in general not unique.

1, for x ≥ 0,
Example 7.4.10. Consider u0 (x) = For this initial con-
0, for x ≤ 0.
dition the characteristics leave a complete region in the (x, t)-plane empty.
See Figure 7.13. There are many possibilities to fill this region. A physically
realistic solution will be a so called rarefaction wave
⎧
⎨ 1, for x ≥ t,
u(x, t) = x/t, for x ∈ [0, t],
⎩
0 for x ≤ 0.
We remark that u(x, t) = limε→0 uε (x, t), where uε solves ∂t uε = ε∂x2 uε −
uε ∂x uε . Thus, u is called viscosity solution.

Figure 7.13. The characteristics do not enter parts of the (x, t)-
plane and create a rarefaction wave.
7.4. The Burgers equation 243

7.4.3. Special solutions and Cole-Hopf transformation. We are in-

terested in special solutions such as stationary or traveling wave solutions
of the viscous Burgers equation
∂t u = ν∂x2 u − u∂x u.
Obviously every constant in space is a stationary solution. In order to find
the traveling waves we make the ansatz u(x, t) = v(x−ct) = v(ξ) and obtain
−cv = νv − vv .
Integration yields
cv + νv − v 2 /2 + d = 0
with a constant d. This one-dimensional ODE has two fixed points v− and
v+ satisfying cv − v 2 /2 + d = 0. To given v− for ξ → −∞ and v+ for ξ → ∞
the system of linear equations
2 2
cv+ + d = v+ /2 and cv− + d = v− /2
always has unique solutions c and d if v− = v+ . However, looking at the
phase portrait of the one-dimensional ODE shows a heteroclinic connection
vh,ν between the fixed points with limξ→±∞ vh (ξ) = v± only if (v+ −v− )/c <
0. This heteroclinic connection gives a monotonic front u(x, t) = vh (x−ct) in
the Burgers equation, which is a diffusively smoothed shock. These solutions
describe for instance a front of gas with constant high density moving into
a domain with constant low density. In case (v+ − v− )/c > 0 there exist no
shocks, and the diffusively smoothed rarefaction wave from Example 7.4.10
occurs. In case v− = 1 and v+ = 0 we find d = 0 and c = 1/2. Since
νv = −v/2 + v 2 /2 we have vh,ν (ξ) = vh,1 (ξ/ν) such that for the vanishing
viscosity limit

0, for x ≥ 0,
lim vh,ν (ξ) =
ν→0 1, for x < 0,
i.e., the weak shock from Example 7.4.9 occurs as a vanishing viscosity limit.
With the help of the maximum principle a number of stability results
0
can be established. Recall that we have chosen X = Cb,unif (R, R) as phase
space.
Theorem 7.4.11. The fixed points u∗ ∈ X defined by u(x, t) = u∗ ∈ R are
stable, but not asymptotically stable.

Proof. For given ε > 0 choose δ = ε > 0. Then by the maximum principle
from u0 − u∗ X < δ, i.e., u∗ − δ < u0 (x) < u∗ + δ for all x ∈ R, we have
that u∗ − δ < u(x, t) < u∗ + δ for all x ∈ R and t > 0, i.e., u(t, u0 ) − u∗ X <
ε which shows the stability. Since there are infinitely many fixed points
{u∗ + δ : |δ| < δ0 } in every δ0 -neighborhood of u∗ in X, the fixed point u∗
cannot be asymptotically stable. 
244 7. Some dissipative PDE models

Remark 7.4.12. Like for the KPP equation again the maximum principle
can be used to prove the stability of the traveling waves under small localized
perturbations by confining the perturbation between two translates of the
0
traveling wave. The instability in Cb,unif comes from the fact that there are
nearby traveling waves with a different velocity.

Figure 7.14. Shocks are robust against small spatially localized per-
turbations. For the shock from Example 7.4.9 these perturbations are
transported by the characteristics into the shock with relative velocities
±1/2, where they finally vanish. This idea can also be used for analytic
stability proofs of viscous shocks, cf. [Kap94b].

With the help of the Cole-Hopf transformation the nonlinear viscous

Burgers equation (7.26) is transformed into the linear heat equation ∂t ψ =
∂x2 ψ. The idea behind this remarkable transformation is as follows. Substi-
tuting u = ∂x φ into (7.26) yields the so called integrated Burgers equation

1
(7.28) ∂t φ = ν∂x2 φ − (∂x φ)2 .
2
Next we let φ = g(ψ) with g to be chosen such that the quadratic term in
(7.28) vanishes. We obtain
1
g (ψ)∂t ψ = νg (ψ)(∂x ψ)2 + νg (ψ)∂x2 ψ − (g (ψ)∂x ψ)2
2
which we rewrite as
1
g (ψ)(∂t ψ − ν∂x2 ψ) = (νg (ψ) − g (ψ)2 )(∂x ψ)2 .
2
The condition νg (ψ) − 12 g (ψ)2 = 0 leads to g(ψ) = −2ν log ψ. Thus,
  x 
1 ∂x ψ(x, t)
ψ(x, t) = exp − u(y, t) dy and u(x, t) = −2ν ,
2ν a ψ(x, t)
where a ∈ R is arbitrary. Moreover, we have by construction ∂t ψ = ν∂x2 ψ.
Denoting the Cole-Hopf transformation by T we thus have the commutative
7.4. The Burgers equation 245

diagram
u0 nonlin. evolution in the Burgers eq. u(·, t)
−−−−−−−−−−−−−−−−−−−−−−−−−−−→ −1
↓T ↑T
ψ0 linear evolution in the heat equation ψ(·, t)
−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Example 7.4.13. We consider the explicit solution
 x
√ 1
e−y /4 dy,
2
(7.29) ψ(x, t) = 1 + zErf(x/ νt) where Erf(x) = √
4π −∞
which describes the diffusive mixing of the asymptotic states 1 for x < 0
and 1 + z for x ≥ 0. Here Erf is related to the more standard definition of
the so called error function
 x
2 1
e−t dt via Erf(x) = (1 + erf(x/2)).
2
erf(x) = √
π 0 2
We obtain
√
e−x /4νt
2
∂x ψ νz
(7.30) u(x, t) = −2ν = −√ √ .
ψ πt 1 + zerf(x/ νt)
This is illustrated in Figure 7.15 with the small diffusion coefficient ν = 0.01.
Initially the wave steepening by nonlinear transport dominates. However, for
t > 0 the solution is always smooth, and as t → ∞ the solution decays to zero
in L∞ with a rate O(t−1/2 ). Figure 7.16 shows the so called renormalized
asymptotic profiles
√
√ √ νz e−x /4ν
2

(7.31) tu(x t, t) = − √ √ =: fz∗ (x)

π 1 + zerf(x/ ν)
which are independent of t.

50

40
50
1
1 40 30

30
20
0
20 t
0
10
10
-3 0
-3 0 0 3 6 x
3 9

Figure 7.15. Diffusive mixing (7.29) in the linear heat equation

(left) and associated
∞ exact solution (7.30) of Burgers equation
(right), M = −∞ u(y, t) dy = 1, ν = 0.01. Note that for u we
start plotting at t = 1.
246 7. Some dissipative PDE models

1
√
ν=0.01 tu
ν=0.2
ν=1
0.5

0 √
−5 0 5 tx

Figure 7.16. Asymptotic Burgers profiles (7.31), normalized to

M = 1, i.e., z = e−1/2ν − 1.

More generally, solutions of the linear heat equation ∂t ψ = ∂x2 ψ to initial

conditions
lim ψ(x, 0) = 1 and lim ψ(x, 0) = 1 + z < 1
x→−∞ x→∞

satisfy under some additional assumptions, see §14, that

√ √
lim ψ(x t, t) = 1 + zerf(x/ ν)
t→∞
√
with a rate O(1/ t). This shows that solutions u to spatially localized
initial conditions in the Burgers equation satisfy
√ √
lim tu(x t, t) = fz∗ (x)
t→∞
√
with a rate O(1/ t). Therefore, the renormalized solutions converge to a
non-Gaussian limit, see again Figure 7.16. This will be reconsidered in more
detail in §14.

Further Reading. A standard reference for the maximum principle is

[PW84]. The long time dynamics in the KPP equation is dominated by
front solutions, and the KPP attractor is fully characterized in [HN01]. An
extensive review about front propagation, containing a wealth of references,
is [vS03], and an encyclopaedic book about traveling waves for parabolic
problems is [VVV94].
For more results on Fourier transform we again refer to [Duo01] and
[RS75b]. See also the “further reading” of Chapter 5.
The method of characteristics can be extended to general scalar first
order PDEs of the form
F (x1 , . . . , xd , u, ∂x1 u, . . . , ∂xd u) = 0,
with F : R2d+1 → R a smooth function. It can be shown that the solution
of this PDE is equivalent to the solution of an ODE in R2d+1 . See [CH89,
§2]. The shock speed in conservation laws is determined by the the Rankine-
Hugoniot conditions; see Exercises 7.22 and 7.23 for the scalar case, and,
e.g., [Whi99, LeV92, Ser00] for the case of systems.
7.4. The Burgers equation 247

Exercises
7.1. Derive the two-dimensional diffusion equation ∂t u = Δu, u = u(x, y, t), from
a two-dimensional random walk.
7.2. Compute u = u(z, t) with z = xr + ixi using the solution formula for linear
diffusion. Show the convergence of the integral for every t > 0 and z ∈ C.
7.3. Prove that the solutions u(t, u0 ) of the linear diffusion equation satisfy the
fundamental property of a dynamical system u(t + s, u0 ) = u(t, u(s, u0 )).
7.4. Use the maximum principle to give an alternative proof of (E1) on page 213.
7.5. Prove that if u ∈ C([0, T0 ], Cb,unif
2
) ∩ C 1 ([0, T0 ], Cb,unif
0
) solves the variation of
constant formula, then u is also a classical solution of the KPP equation.
7.6. Let R = {(v, w) : 0 < v < 1, −kv < w < 0}. Show that for suitable k the
vector field f = (w, −cw − v + v 2 ) for (7.10) points inwards on ∂R. Use this and
the Poincaré-Bendixon theorem to prove that for c = 2 there exists a monotonic
front connecting u = 1 and u = 0 for the KPP equation.
7.7. a) Discuss for which α, β there exist similarity solutions u(x, t) = tα v(x/tβ ) of
∂t u = ∂x2 u + f (x), x ∈ R, where (i) f (x) = 0, (ii) f (x) = 1, (iii) f (x) = x.
b) Find α, β such that ∂t u + up ∂x u + μ∂x3 u = 0 is invariant under u → au, x → aα x,
t → aβ t.
7.8. a) Solve explicitly ∂t u = ∂x2 u with IC u0 (x) = x2 .

b) Use a) to calculate explicitly s2 e−s ds.
2

c) Compare the solution in a) with the general solution formula.

Hints: For a) make an ansatz u(x, t) = a(t)x2 + b(t)x + c(t).
7.9. For d = 1, 2, 3 and δ > 0 prove the Gagliardo-Nirenberg inequality uL4 (Rd ) ≤
1−d/4−δ d/4+δ
CuL2 (Rd ) uH 1 (Rd ) . Hint: cf. Exercise 6.6

7.10. a) For u(x) = e−a|x| with a > 0 calculate u

(k).
(k) = π1 sinkak for k = 0 and u
b) Let u(x) = χ[−a,a] with a > 0. Show that u (0) = a
π.

 ∈ Cb0 .
7.11. Show that u ∈ L1 implies u
7.12. a) For u(x) = e−x √1 e−k /2
2 2
/2
(k) =
show u 2π
by solving the Fourier trans-
(k) = e−k t , t > 0.
2
formed ODE which is satisfied by u. Calculate u(x) for u
7.13. Consider the linear heat equation with initial conditions

a, for x < 0,
u0 (x) =
b, for x ≥ 0.
√ x
Prove that u(x, t) = a+(b−a)erf(x/ t), where erf(x) = √14π −∞ e−y /4 dy.
2

r
7.14. Prove that Fourier transform is an isomorphism between Hm and Hrm .
7.15. Let g(x) = e−x
2
/2
. Find g ∗ g.
7.16. For f ∈S, prove the Heisenberg uncertainty principle xf L2 kfL2 ≥
2 f L2 . Hint. Use Parseval, the relation between differentiation and Fourier trans-
1 2

form, and the Cauchy-Schwarz inequality.

248 7. Some dissipative PDE models

7.17. The scaling of a distribution is, as usual, defined by applying the scaling to
the test function. I.e., for k = 0, T ∈ D and φ ∈ D we set Sk1 T (φ) = T (φ(k·)),
Sk2 T (φ) = k1 T (φ(·/k)). Show that Sk1 δ = Sk2 δ.
D
7.18. Prove that (a) If un ∈ C 0 (R) and un → u uniformly in x, then un → u for
n → ∞.
D
(b) If un ∈ L2 (R) and un − uL2 → 0 , then un → u for n → ∞.
D D
(c) sin nx → 0 and n2 sin nx → 0 for n → ∞.
2 dn −x2
7.19. The Hermite functions are ψn (x) = (−1)n ex /2
e , n ∈ N0 .
dxn
i) Sketch ψ0 , ψ1 and ψ2 .
ii) Show the recursions ψn (x) = xψn (x) − ψn+1 (x) and F(ψn ) (k) = kψn (k) −
iψn+1 (k).
iii) Use i) to inductively show ψn = λn ψn with λn = (−i)n .
Remark. iii) means that all the ψn√ are eigenfunctions of F. They form a orthogonal
system in L2 (R),
ψn , ψm  = 2n n! πδnm , and the normalized hn = ψn /ψn  form
a complete orthonormal system in L2 (R), which is quite useful in various fields,
e.g., quantum mechanics.
7.20. Consider the scalar first order PDE
a(x, t, u)∂t u + b(x, t, u)∂x u = c(x, t, u).
Make the ansatz v(s) = u(x(s), t(s)) and find a system for (x(s), t(s), v(s)).
7.21. Show that u(x, t) = u0 (x0 ) = u0 (x − tu(x, t)) is an implicit representation
of the solutions of ∂t u + u∂x u = 0. Show by differentiation of this representation
w.r.t. x that for t = t∗ = −1/(inf x0 ∈R u0 (x0 )) a shock occurs, i.e., that u = u(x, t)
is no longer differentiable w.r.t. x.
7.22. Given ul > ur and u0 (x) = ul for x ≤ 0
and u(x) = ur for x > 0 calculate the
ul , for x ≤ αt,
shock speed α of the weak solution u(x, t) = of the inviscid
ur , for x > αt
Burgers equation.
7.23. Consider for a general scalar conservation law ∂t u = −∂x (f (u)) the weak
solution from Exercise 7.22. Derive the so called Rankine-Hugoniot jump condition
α = [f ]/[u] := (f (ur ) − f (ul ))/(ur − ul ) for the velocity of the shock.
7.24. Discuss the phase portrait of −cv  = v  − vv  in the (v, v  )-plane. For v− = 1
and v+ = 0 show that c = 1/2 and that v(ξ) = (1 + eξ )−1 is the explicit solution.
7.25. Consider the scalar conservation

law ∂t u + u3 ∂x u = 0, x ∈ R, t > 0. Find
1, 0 < x < 1,
the solution for u(x, 0) = g(x) =
0, else.

the viscous Burgers equation ∂t u = ν∂x u − 2 ∂x (u ) with u(x, 0) =
2 1 2
7.26. Solve
−1, for x < 0,
H(x) = Give a sketch of the solution.
1, for x ≥ 0.
 Chapter 8

Three canonical
modulation equations

In this chapter we consider another three scalar equations on the real line,
namely the Nonlinear Schrödinger (NLS) equation, the Korteveg-de Vries
(KdV) equation, and the Ginzburg-Landau (GL) equation. These equations
are the three most important so called modulation equations, sometimes
also called amplitude or envelope equations. By multiple scaling analysis
the NLS equation can be derived in order to describe the evolution of an
envelope of a spatially and temporarily oscillating wave packet. This will
be discussed in Chapter 11. Similarly, the KdV equation can be derived
for the description of long waves in dispersive media, see Chapter 12, and
the GL equation can be derived to describe the evolution of the envelope of
spatio-temporal pattern in dissipative systems, see Chapters 10.
Besides the basic theory and the occurring phenomena for each equation,
we introduce some general concepts, which are useful for the analysis of these
equations but also in other circumstances. These are for instance the method
of stationary phase in §8.1.4, the use of uniformly local Sobolev spaces in
§8.3.1, and a concept of attractors on unbounded domains in §8.3.4.
As already said, each of the equations considered in this Chapter plays a
very important role in the description of various phenomena in more compli-
cated systems, and thus this chapter is partly a preparation for the remain-
der of this book. On the other hand, as for the equations from Chapter 7,
for each of these equations there exists many papers, and often whole books
entirely devoted to the particular equation. Again we keep the exposition
rather brief, and give a list of further reading at the end of the chapter.

249
250 8. Three canonical modulation equations

8.1. The NLS equation

The Nonlinear Schödinger (NLS) equation
(8.1) ∂t u = iν1 ∂x2 u + iν2 |u|2 u,
with t ∈ R, x ∈ R, u = u(x, t) ∈ C, and coefficients ν1 , ν2 ∈ R, can be
derived by multiple scaling analysis in order to describe the evolution of the
envelope of a spatially and temporarily oscillating wave packet, as will be
discussed in Chapter 11. By rescaling u → β1 u, x → β2 x, and t → β3 t with
βj ∈ R for j ∈ {1, 2, 3} it can be transformed into
(8.2) ∂t u = −i∂x2 u + αi|u|2 u, α = ±1.
See Exercise 8.1. Two cases remain, namely α = −1 and α = 1. The case
α = −1 is called focusing and the case α = +1 is called defocusing. In
particular, the focusing NLS equation is widely used in nonlinear optics to
describe the evolution and interaction of optical pulses. The NLS equation
also plays a role in some theories about the occurrence of so called freak or
rogue waves, cf. [Osb10]. The variant
(8.3) ∂t u = −i∂x2 u + iV u + αi|u|2 u,
with V = V (x) some potential, is called the Gross-Pitaevsky equation and
can be derived for the description of Bose-Einstein condensates, cf. [Pel11].
In summary, the NLS equation is a widely used model in nonlinear physics.
The evolution of the NLS equation is not smoothing in the sense of the
KPP or the Burgers equation. Hence, the phase spaces are not connected
by smoothing. We can have global existence or stability in one space, but
explosion or instability in another space.

8.1.1. Nonlinear oscillations and pulse solutions. The NLS equation

consists of two parts, namely the dispersion part −i∂x2 u and the nonlinear
oscillation part αiu|u|2 .
Making the ansatz u(x, t) = v(t) = r(t)eiφ(t) for x-independent solutions
we find
∂t r = 0 and ∂t φ = αr2 ,
with the solutions r(t) = r(0) and φ(t) = φ(0) + αr2 (0)t, i.e., oscillations
where the frequency increases with r. Next, we search for solutions of the
form u(x, t) = v(t)eikx = r(t)eiφ(t) eikx . We find
∂t r = 0 and ∂t φ = k 2 + αr2 ,
with the solutions r(t) = r(0) and φ(t) = φ(0) + ω(k, r)t, where ω(k, r) =
(k 2 + αr2 ). Thus, we have solutions which are periodic in time and space,
namely
2 2
u(x, t) = u(x, t; k, r, φ0 ) = re(i(kx+φ0 +(k +αr )t)) .
8.1. The NLS equation 251

In the defocusing case, α = 1, where always ω(k, r) = k 2 + αr2 > 0, all

periodic waves travel left, whereas in the focusing case, α = −1, where
ω(k, r) = k 2 + αr2 can have either sign, the periodic waves can travel left
or right. For illustration, we sketch some of these nonlinear oscillations in
Figure 8.1. Interestingly, in the focussing case they are all unstable, which is
known as Benjamin–Feir instability. See Exercise 8.23 where we consider the
related stability question for spatially periodic solutions in the GL equation.

a) α = 1, k = 1 b) α = −1, k = 1 c) α = −1, k = 2

1 2 1 2 1 2

0 0 0
1 1 1
-1 -1 -1 t

-4 -2 0 -4 -2 0 0 -4 -2 0 0
0 2 4 2 4 2 4 x

Figure 8.1. Nonlinear oscillations (real part) for the NLS equation,
with r = 1. a) shows the defocusing case, where all waves travel
left, while b), c) show the focusing case, where waves can travel
left or right.

However, from the point of applications of the NLS equation, in particu-

lar due to its derivation for the description of modulations of electromagnetic
waves, pulse solutions are more interesting. In order to find them, we make
the ansatz
u(x, t) = B(x − ct)ei(qx−ωt+φ0 ) ,
with B(ξ) ∈ R, which yields
−iωB − cB = −iB + 2qB + iq 2 B + αiB 3 .
Separating real and imaginary part gives the relations c = −2q and
(8.4) 0 = B − (ω + q 2 )B − αB 3 .
This ODE can be discussed using the methods from §2.3.3, see in particular
Remark 2.3.21. One finds that for ω + q 2 > 0 and α = −1 there exist
solutions of (8.4) which are homoclinic to the origin. Figure 8.2 shows the
phase portrait. In fact, there exist explicit formulas for homoclinic solutions
of (8.4). Before computing these in Exercise 8.4, it is useful to have a look
at the symmetries of the NLS equation in Exercise 8.3.

8.1.2. Dispersion. The linear Schrödinger equation

(8.5) ∂t u = −i∂x2 u
is a prototype example of a linear dispersive equation. Dispersion means that
the phase velocity cp (k) = ω(k)/k of harmonic waves ei(kx−ω(k)t) depends
252 8. Three canonical modulation equations

Figure 8.2. Phase portrait for (8.4), ω + q 2 = 1 and α = −1.

in a non-trivial way on k. This has a number of consequences which we

explain in the following. For (8.5) we have solutions u(x, t) = ei(kx−ωt)
where ω = −k 2 . Hence, cp (k) = k and the amplitude of the harmonic waves
is conserved. This behavior is in contrast to the linear diffusion equation
∂t u = ∂x2 u where we have u(x, t) = eikx+λt , with λ = −k 2 , and thus all
spatially harmonic waves are damped with some exponential rate, except
for k = 0.
Like for the diffusion equation there exists an explicit solution formula
which can be derived by Fourier transform and which at first looks very
similar to the one of the diffusion equation.
Lemma 8.1.1. The initial value problem for the linear Schrödinger equation
∂t u = −i∂x2 u, u|t=0 = u0 , is solved by
 ∞
1 i(x−y)2
(8.6) u(x, t) = √ e− 4t u0 (y) dy.
−4πit −∞
The integral exists for u0 ∈ L1 (R) and extends to an isometry in L2 (R), i.e.,
u(·, t)L2 = u0 L2 .

Proof. Proceeding
 ik2 t ikxexactly as in Example 7.3.20 it remains to calculate
G(x, t) = R e e dk. We have

2 ix
∂x G(x, t) = ikeik t eikx dk = − G(x, t).
R 2t
This differential equation is solved by G(x, t) = Ce−ix /4t . The constant C
2

can be computed through

 
2 1 2 π iπ/4 π
C = G(0, t) = eik t dk = eiy dy = e = .
R t R t −it
 2
The integral R eiy dy exists as improper integral due to faster and faster
2
oscillations of eiy for |y| → ∞. The isometry follows immediately from
u(·, t)L2 = (2π)1/2 
u(·, t)L2 = (2π)1/2 
u(·, 0)L2 = u0 L2 . 
8.1. The NLS equation 253

Like for diffusion, from (8.6) we immediately obtain the estimate

 ∞
C
(8.7) sup |u(x, t)| ≤ √ |u(x, 0)| dx,
x∈R t −∞
i.e., solutions to spatially localized initial conditions decay uniformly towards
zero with a rate t−1/2 . But there are major differences: by diffusion energy
is lost, i.e., dtd
|u(x, t)|2 dx ≤ 0, while dispersion conserves energy, i.e.,

d
dt |u(x, t)|2 dx = 0, but spreads it all over the real line.
Dispersion smoothes solutions locally in space. We can compute
 ∞
1 −2i(x − y) − i(x−y)2
∂x u(x, t) = √ e 4t u(y, 0) dy < ∞
4πit −∞ 4t
∞
for all x ∈ R and all t > 0 if −∞ |y||u(y, 0)| dy < ∞, i.e., ∂x u(x, t) can be
computed point-wise, although the initial condition may only be continuous.
∞
More generally, if the moment −∞ |y|n |u(y, 0)| dy is finite, then ∂xn u(x, t) is
finite for all x ∈ R and all t > 0 .
The reason for this behavior is as follows: Consider initial conditions
which are slow modulations in space of an underlying carrier wave, i.e.,
initial conditions of the form
u(x, 0) = u0 (εx)eikx
where 0 < ε  1 is a small parameter and u0 a smooth spatially localized
function. The right-hand side of the Schrödinger equation applied to this
initial condition yields
 
i∂x2 (u0 (εx)eikx ) = ieikx −k 2 u0 (εx) + ε2iku 0 (εx) + ε2 u 0 (εx) .
This motivates us to make the ansatz
 
u(x, t) = B ε(x − cg t), ε2 t eik(x−cp t) ,
with constants cp and cg , and a function B = B(ξ, τ ) with ξ = ε(x − cg t)
and τ = ε2 t. Inserting this ansatz and computing the coefficients in front of
ε0 , ε1 , and ε2 shows that
−ikcp = −ik 2 , −cg ∂ξ B = −2k∂ξ B, and ∂τ B = i∂ξ2 B.
The constant cp = ω(k)/k = k with frequency ω = k 2 is called phase velocity.
The constant cg = ω (k) = 2k is called the group velocity. This calculation
shows that wave packets with carrier wave eikx with k large are transported
with large group velocity cg = 2k towards infinity. If the initial condition is
spatially localized, then nothing can come from infinity. Therefore, at fixed
x ∈ R only wave packets with low derivatives remain, and hence a local
smoothing occurs.
On the other hand this also has the consequence that if the initial con-
dition is not spatially localized, then packets with high derivatives can come
254 8. Three canonical modulation equations

from infinity. In fact, there is no local existence and uniqueness of solutions

0
in Cb,unif (R, R).

Remark 8.1.2. For the transport equation and the linear wave equation we
obtain ω(k) = ck with a constant c ∈ R and hence the group velocity ω (k) =
c is bounded and the transport equation and the linear wave equation can
0
be solved in Cb,unif (R, R).

It turns out that X = L2 (R, C) is a good choice for the handling of the
linear Schrödinger equation as a dynamical system.

Lemma 8.1.3. The curve of solutions t → U (t, u0 ) with U (t, u0 )(x) =

u(x, t) and u(x, 0) = u0 is continuous in X if u0 ∈ X. Moreover, U (t +
s, u0 ) = U (t, U (s, u0 )).

Proof. First of all we have the semigroup property

U (t + s, u0 ) =F −1 (k → eik
2 (t+s)
0 (k))
u
=F −1 (k → e ik2 t
(e ik2 s
0 (k))) = U (t, U (s, u0 ))
u

such that is sufficient to prove continuity in t = 0. For all ε > 0 we have to

find a t0 > 0 such that for all t ∈ (0, t0 ) we have
 
U (t, u0 ) − u0 L2 =
2
|u(x, t) − u(x, 0)| dx = 2π |
2
u(k, t) − u (k, 0)|2 dk
R R
  
ik2 t
=2π |(e − 1) 2
u(k, 0)| dk = 2π . . . dk + 2π . . . dk < ε2 ,
R |k|<L |k|≥L

where L is chosen so large that

 
2π . . . dk ≤ 4π |
u(k, 0)|2 dk < ε2 /2.
|k|≥L |k|≥L

For this L we find for the first integral

 
ik2 t
2π . . . dk ≤ 2π sup |e − 1| 2
(k, 0)|2 dk
u
|k|<L |k|<L R

2
≤ 2π|eiL t − 1|2 |
u(k, 0)|2 dk < ε2 /2
R

if t0 > 0 is sufficiently small. Therefore, we are done. 

The choice X = L2 (R, C)
is only a good choice for the handling of the
linear Schrödinger equation. For the NLS equation Sobolev spaces H m (R, C)
with m > 1/2 are more adequate.
8.1. The NLS equation 255

8.1.3. Local existence and uniqueness, and the Hamiltonian. In the

following we choose X = H m (R, C) with m > 1/2 as phase space for the
NLS equation.
Theorem 8.1.4. Let u0 ∈ X. Then there exist a T0 > 0 and a unique local
solution u ∈ C([0, T0 ], X) of the NLS equation with u|t=0 = u0 .
Proof. According to Exercise 8.6 the operator −i∂x2 generates a C0 -semi-
group e−it∂x : X → X. Moreover, by Lemma 7.3.29 the nonlinearity u →
2

iα|u|2 u is locally Lipschitz-continuous in X. The local existence thus follows

by the variation of constant formula and the contraction mapping theorem,
cf. the proof of Theorem 7.1.7. 
Remark 8.1.5. Like for the wave equation or the transport equation the
solution also exists backwards in time, i.e., u ∈ C([−T0 , T0 ], X)).
This local solution can be continued as long as the H m -norm of the
solution stays bounded. First we show that the L2 -norm is conserved along
solutions. We have
 
d d
uL2 =
2
uu dx = (∂t u)u + u(∂t u) dx
dt dt R R

 
= 2 Re u −i(∂x2 u − α|u|2 u) dx = 0.
R
The estimate for the H 1 -norm is related to the Hamiltonian structure of the
NLS equation. It turns out that the NLS equation, like the KdV equation in
the next section, is a completely integrable Hamiltonian system which can
be solved explicitly. See [DJ89, §6] for an overview. Here, we only show
that the NLS equation is a Hamiltonian system. With

1 1
(8.8) H(u) = |∂x u(x)|2 + α|u(x)|4 dx
R 2 4
we find
∂u H[v] = lim ε−1 (H(u + εv) − H(u))
ε→0

−1
= lim ε |∂x (u + εv)|2 /2 + α|u + εv|4 /4 − |∂x u|2 /2 − |u|4 /4 dx
ε→0 R

= Re (−∂x2 u + αu|u|2 )v dx
R
which maps v ∈ X = L2 (R, C) linearly into R, i.e., ∂u H is a linear map from
X to R and hence an element of the dual space. According to the Riesz
representation theorem [Alt16, Satz 4.1], in Hilbert spaces the dual space
can be identified with X by defining a map (the canonical isomorphism)
  
(8.9) β : Lin(X, R) → X, v →
u, vL2 = Re u(x)v(x) dx → u
R
256 8. Three canonical modulation equations

and therefore β∂u H = −∂x2 u + αu|u|2 . We finally have

∂t u = −i∂x2 u + αiu|u|2 = iβ∂u H(u) = Jβ∂u H(u)
where the operator Ju = iu is skew symmetric in X since
 

Ju, vL2 = Re iu(x)v(x) dx = − Re u(x)iv(x) dx = −
u, JvL2 .
R R

The Hamiltonian is well defined on H 1 since H ≤ ∂x u2L2 +u2C 0 u2L2 .

In the defocusing case the Hamiltonian is positive definite, i.e., H(u) > 0 for
u = 0, whereas in the focusing case it is indefinite. Hence, in the defocusing
case we get the H 1 -estimate for free since then
u(t)2H 1 ≤ H(t) + u(t)2L2 = H(0) + u0 2L2
for all t ∈ R and thus the solution exists globally with a uniform bound in
H 1 (R), and we already proved half of the following theorem.
Theorem 8.1.6. For u0 ∈ H 1 (R) the local solution of the NLS equation
exists globally in time and stays uniformly bounded in H 1 (R).

Proof. It remains to consider the focusing case. We have u(t)2L2 =

u(0)2L2 and H(u(t)) = H(u(0)) and thus
 
1 1
(8.10) |∂x u| dx = H(u(0)) +
2
|u|4 dx.
2 R 4 R
To estimate the last term we use the Gagliardo-Nirenberg estimate, cf.
Lemma 6.3.10,
  (2(q+1)−d(q−1))/4  d(q−1)/4
(8.11) |u| q+1
dx ≤ |u| dx
2
|∇u| dx
2
,
R R R
which for q = 3 and d = 1 yields
  3/2  1/2
(8.12) |u| dx ≤
4
|u| dx
2
|∂x u| dx
2
.
R R R

R |∂x u|
1 2 dx
Therefore, with φ(u) = 2 we have
3/2

0 ≤ φ(u(t)) ≤ H(u(0)) + u(0)L2 φ(u(t))
since the L2 -norm and the Hamiltonian are conserved. This inequality im-
mediately implies a uniform bound for φ(u(t)) and so also the H 1 -norm
stays bounded. 
Remark 8.1.7. Note that (8.11) depends on d. In fact, for x ∈ Rd with
d ≥ 2 the solutions of ∂t u = −iΔu − i|u|2 u can blow up in finite time T , i.e.,
u(t)H 1 → ∞ as t → T , cf. [SS99b, Fib15].
8.1. The NLS equation 257

As already said, we come back to the NLS equation in Part IV, while
here we close with some bounds for oscillatory integrals which are used in
the derivation of dispersive estimates.

8.1.4. The method of stationary phase. We explain a method which

gives more insight into dispersion and which especially allows to compute
decay rates like (8.7) from the Fourier representation u  t)
(k, t) = G(k, u(k, 0)
of the solution u also in situations where in physical space no explicit rep-
resentation formula for G such as (8.6) is known. Our simple approach is
based on the subsequent Lemma of van der Corput, cf. [Ste93].
Lemma 8.1.8. (Lemma of van der Corput) Let φ : R → R in C  with
|φ() | ≥ 1 for all θ ∈ [a, b]. For  = 1 assume the monotonicity of φ . Then
there exists a C > 0 such that
 b 
 
 itφ(θ)dθ  −1/
 e  ≤ C t .
a

Proof. For  = 1 we obtain

 b  b  
itφ(θ) 1 d itφ(θ)
I(t) = e dθ =
(e ) dθ
a a itφ (θ) dθ
  
1 1 b d 1
= e |a −
itφ(θ) b
eitφ(θ) dθ.
itφ (θ) it a dθ φ (θ)
This can be estimated by
 b 
1 1 d 1 
|tI(t)| ≤ + +  
|φ (b)| |φ (a)|  dθ φ (θ)  dθ
a
1 1 1
≤ + + ≤4
|φ (b)| |φ (a)| |φ (b)−1 − φ (a)−1 |
where we used the monotonicity of φ .
For all other  we use induction. Assume that |φ(+1) (θ)| ≥ 1 for all
θ ∈ [a, b]. Then there exists at most one point θ0 ∈ [a, b] with φ() (θ0 ) = 0
and we have |φ() (θ)| ≥ δ for |θ−θ0 | ≥ δ. If this point does not exist we apply
the Lemma after a possible rescaling for . Thus, we assume the existence
of such a θ0 and write I(t) = I1 (t) + I2 (t) where I1 (t) is the integral over
(a, θ0 − δ) ∪ (θ0 + δ, b) and I2 (t) over (θ0 − δ, θ0 + δ). We find
|I1 (t)| ≤ 2C (δt)−1/
since for θ ∈ (a, θ0 − δ) ∪ (θ0 + δ, b) we have |φ() (θ)| ≥ δ such that the
induction for  can be applied. Moreover, we have
 θ0 +δ
|I2 (t)| ≤ | eitφ(θ) dθ| ≤ 2δ.
θ0 −δ
258 8. Three canonical modulation equations

Choose δ = t−1/(+1) . Since (δt)1/ = (t−1/(+1) t)1/ = t−1/(+1) we have

I(t) ≤ C+1 t−1/(+1), with C+1 = 2(1 + C ). 

Remark 8.1.9. Since no compactness argument has been used in the proof
of Lemma 8.1.8 values a = −∞ and b = ∞ are allowed if the integrals exist.

We use Lemma 8.1.8 and Remark 8.1.9 to confirm the dispersive estimate
(8.7).

Example 8.1.10. The linear Schrödinger

∞ equation ∂t u = i∂x2 u with t, x ∈ R,
u(x, t) ∈ C is solved by u(x, t) = −∞ G(x − y, t)u(y, 0) dy with
 ∞  ∞
−ik2 t ikx
G(x, t) = e e dk = eitφ(k) dk
−∞ −∞

such that φ(k) = −k 2 + kx/t. The integral is estimated for every fixed
ξ = x/t. Since φ (k) = −2 we thus have supξ∈R |G(ξt, t)| ≤ Ct−1/2 . From
supξ∈R |G(ξt, t)| = supx∈R |G(x, t)| it follows

u(·, t)L∞ ≤ CG(·, t)L∞ u(·, 0)L1 ≤ Ct−1/2 u(·, 0)L1

due to Lemma 7.3.19.

Example 8.1.11. The Airy equation ∂t u = ∂x3 u with t, x, u(x, t) ∈ R is

∞
solved by u(x, t) = −∞ G(x − y, t)u(y, 0)dy with
 ∞  ∞
e−ik t eikx dk =
3
G(x, t) = eitφ(k) dk
−∞ −∞

such that φ(k) = −k 3 + kx/t. Since φ (k) = −6 we have like in Example

8.1.10 that G(·, t)L∞ ≤ Ct−1/3 and therefore

u(·, t)L∞ ≤ CG(·, t)L∞ u(·, 0)L1 ≤ Ct−1/3 u(·, 0)L1 .

The handling of oscillatory integrals by the stationary phase method is

a well developed theory, cf. [Ste93]. Even more complicated dispersive esti-
mates can be obtained with this approach, cf. [LP09]. Dispersive estimates
can be transferred into so called Strichartz estimates. In case of the linear
Schrödinger semigroup, for 2 ≤ q ≤ ∞, 2q + dp = d2 , and 1 ≤ p < ∞, we have
for instance
eitΔ/2 u0 Lq ((0,t),Lp ) ≤ Cu0 L2 .
Such estimates allow to prove local and global existence in lower regularity
spaces, cf. [Tao06].
8.2. The KdV equation 259

8.2. The KdV equation

The Korteweg-deVries (KdV) equation [KdV95]
(8.13) ∂t u = −∂x3 u + 6u∂x u,
with t ∈ R, x ∈ R, and u = u(x, t) ∈ R, can be derived as a modula-
tion equation from various models for the description of long unidirectional
waves, such as long wave length surface water waves, see Chapter 12. It
consists of two parts, namely the dispersive term −∂x3 u and the nonlinear
transport term 6u∂x u. By rescaling u, x, and t every other value for the
coefficients can be obtained. The choice in (8.13) is the one which is most
often used in the literature and is motivated by its derivation from the water
wave problem. Our presentation follows in big parts the textbook [DJ89].
The linearized KdV equation or Airy equation is given by
∂t u = −∂x3 u.
It possesses solutions u(x, t) = eikx+λt with λ = ik 3 , i.e., the amplitude of
harmonic waves is preserved. Moreover, the total energy is conserved by the
Airy equation since
  
d
u (x, t) dx = −2 u(x, t)∂x u(x, t) dx =
2 3
∂x (∂x u(x, t))2 dx = 0.
dt R R R
The Airy equation shows dispersion, i.e., energy is spread over the real line.
The estimate 
−1/3
sup |u(x, t)| ≤ Ct |u(y, 0)| dy,
x∈R R
can be established with the stationary phase method, cf. Example 8.1.11.
The decay rate t−1/3 is a consequence of λ (0) = 0, but λ (0) = 0. Alterna-
tively, it can be obtained by an explicit solution formula which can be build
with the help of self-similar solutions of the form u(x, t) = t−1/3 u(xt−1/3 ) as
fundamental solutions, cf. [Rau91].
The nonlinear transport term
∂t u = 6u∂x u
already appeared in the Burgers equation. We found that shocks may be
created in finite time. Like in the Burgers equation, the linear semigroup,
here generated by ∂x3 , inhibits the creation of shocks in the full KdV equation.
The local existence and uniqueness theory, cf. Theorem 8.2.1, for the
solutions of the KdV equation is more advanced due to the fact that the
KdV equation is a quasilinear hyperbolic equation for which the previous
approach with the variation of constant formula does not work. The solu-
tions are constructed with the help of the iteration scheme
(8.14) ∂t un+1 = −∂x3 un+1 + 6un ∂x un+1 .
260 8. Three canonical modulation equations

For the existence of solutions un+1 of the linear equation (8.14) for a given
function un , see, e.g., [Paz83, Chapter 5]. Next the convergence of the
sequence (vn )n∈N , with vn = un+1 − un , towards zero and the boundedness
of the sequence (un )n∈N is shown. Then we have, cf. [Kat81, Paz83,
KPV91],
Theorem 8.2.1. Let u0 ∈ X = H s (R, C) with s > 3/2. There exist a
T0 > 0 and a unique local solution u ∈ C([0, T0 ], X) of the KdV equation
(8.13) with u|t=0 = u0 .

With a more clever approach solutions can be constructed in Sobolev

spaces of smaller s, cf. [Tao06, §4] and [LP09, Ch. 7], which is a prerequisite
for the study of statistical aspects of the behavior of ensembles of solutions
of these equations with the help of invariant measures.

8.2.1. The solitary wave. The starting point of the KdV history is the
observation of a solitary wave in 1834 by the engineer Scott Russell in a
canal between Edinburgh and Glasgow. The solitary wave was created by a
stopping boat and was of 40 cm height and 10 m length. It traveled with a
speed of 16 km/h for many kilometers through the canal without changing
significantly its shape. Russell was the first to realize the importance of this
phenomenon. Motivated by his observation, in the following years he made
a number of experiments. After the publication of his results [Rus44], a
vigorous scientific debate followed on whether or not such a wave of perma-
nent form could exist. Airy [Air45] argued that even if dissipation, i.e., the
loss of energy, is neglected, dispersion, i.e., the spreading of energy, which is
concentrated in the middle of the solitary wave, in the linear problem will
destroy the solitary wave. It was finally accepted that such waves exist when
Boussinesq [Bou77] and Rayleigh [Ray76] found approximations to such a
wave by deriving the stationary KdV equation (8.15) with some perturba-
tion analysis which takes into account the nonlinear nature of the problem.
In 1895 Korteweg and de Vries [KdV95] derived the time-dependent KdV
equation. A rigorous proof that the full water wave problem possesses such
solitary waves remained open for another few decades and was given by
Friedrichs and Hyers [FH54] in 1954.
What already Boussinesq [Bou77] observed is that in the KdV equation
there is a balance between linear dispersion and nonlinear transport. This
balance creates a wave of permanent form, i.e., u(x, t) = v(x − ct) = v(ξ).
Inserting this ansatz into the KdV equation yields
−cv = −v + 6vv .
Integration w.r.t. ξ yields
−cv + v − 3v 2 = D
8.2. The KdV equation 261

with a constant D ∈ R. We are interested in localized waves and therefore

we have lim|ξ|→∞ v(ξ) = 0 and lim|ξ|→∞ v (ξ) = 0 which together yield
√
D = 0. By rescaling ξ = cζ and v = cv we obtain the first order system
(8.15) v = w and w = v + 3
v2,
which can be discussed in phase plane. In the phase portrait, Figure 8.3, we
find a homoclinic orbit vhom = vhom (ζ). Undoing the scalings yields in the
original variables a family of solitary waves
√ 
(8.16) u(x, t) = cvhom c(x − ct) ,
i.e., the higher the solitary wave, the smaller the width and the faster the
solitary wave. For an explicit formula see Exercise 8.12.

0.5

−0.5
−0.6 −0.4 −0.2 0 0.2

Figure 8.3. The phase portrait for the stationary KdV equation
with D = 0, and the solitary wave.

8.2.2. The KdV equation as a completely integrable system. The

KdV equation is a completely integrable Hamiltonian system which can
be solved explicitly. After showing that the KdV equation can be written
in Hamiltonian form we prove that the KdV equation possesses infinitely
many independent conservation laws, which is a necessary prerequisite for
complete integrability.
With 
1
H(u(t)) = (∂x u(x, t))2 + u(x, t)3 dx
R 2
we find
∂u H[v] = lim ε−1 (H(u + εv) − H(u))
ε→0

−1 1
= lim ε (∂x (u(x, t) + εv(x, t)))2 + (u(x, t) + εv(x, t))3
ε→0 R 2
1
− (∂x u(x, t))2 − u(x, t)3 dx
2
262 8. Three canonical modulation equations


= (−∂x2 u(x, t) + 3u(x, t)2 )v(x, t) dx,
R

which maps v ∈ X = L2 (R, C) linearly into R. Hence, ∂u H is a linear map

from X to R, i.e., an element of the dual space which again can be identified
with −∂x2 u + 3u2 by the canonical isomorphism β from (8.9) and therefore
β∂u H = −∂x2 u + 3u2 . We finally have
∂t u = −∂x3 u + 3∂x (u2 ) = ∂x β∂u H(u) = Jβ∂u H(u),
where the operator (Ju) = ∂x u is skew symmetric in X since
 

Ju, vL2 = (∂x u(x))v(x) dx = − u(x)∂x v(x) dx = −
u, JvL2 .
R R

If the KdV equation is a completely integrable Hamiltonian system, then

there must be infinitely many independent conserved quantities for the KdV
equation. An equation
∂t T (x, t) + ∂x X(x, t) = 0
is called a conservation law. We have
  
∂t T dx = ∂t T dx = − ∂x X dx = −X|∞
−∞ = 0,
R R R

and hence R T dx is a conserved quantity, i.e., it does not change in time.
For the KdV equation we find
∂t u + ∂x (∂x2 u − 3u2 ) = 0

and therefore R u(x, t) dx is independent of time. Next, we find that
 
3 1 2
∂t u + (∂x u)
2
 
9 4 1 2 2
+ ∂x − u + 3u ∂x u − 6u(∂x u) + (∂x u)(∂x u) − (∂x u) = 0,
2 2 2 3
2 2

and hence T3 = u3 + 12 (∂x u)2 yields the conserved quantity R T3 (x, t) dx
which is the Hamiltonian which we already know to be conserved. It turns
out that there are infinitely many conserved quantities:
T4 =5u4 + 10u(∂x u)2 + (∂x2 u)2 ,
T5 =21u5 + 105u2 (∂x u)2 + 21u(∂x2 u)2 + (∂x3 u)2 ,
..
.

Remark 8.2.2. For the linearized system we have that Tn+2 = (∂xn u)2 since
 
∂t (∂xn u)2 + ∂x ∂xn+2 u∂xn u − (∂xn+1 u)2 /2 .
8.2. The KdV equation 263

3
In Fourier space we find the solutions u (k, t) = eik t u
(k, 0). Thus, by intro-
ducing polar coordinates u (k, t) = r(k, t)e iφ(k,t) we find that r(k, t) = r(k, 0)
and φ(k, t) = φ(k, 0) + k 3 t mod 2π or, equivalently,
∂t r(k, t) = 0 and ∂t φ(k, t) = k 3 = const.
Thus, for the linearized system we found uncountably many action variables
r(k, ·) and uncountably many angle variables φ(k, ·), and consequently tori
of any dimension.
In the following we describe how a transformation discovered by Miura
[Miu68] and then generalized in [GGKM74] leads very easily to the con-
clusion that there are infinitely many conserved quantities for the KdV equa-
tion. The basic idea is that, given a transformation which maps solutions
of one equation to solutions of a second one, the existence of simple con-
served quantities for the first equation leads, via the transformation, to more
complicated conserved quantities for the second equation.
To a given u = u(x, t) we implicitly define w(x, t) via the formula
(8.17) u(x, t) = w(x, t) + ε∂x w(x, t) + ε2 (w(x, t))2 .
If w is smooth enough and ε is small, then we can invert this relation recur-
sively to obtain w in terms of u via the formula
(8.18) w =u − ε∂x u − ε2 (u2 + ∂x2 u) + ε3 (∂x3 u + 4u∂x2 u)
 
+ ε4 2u3 + 5(∂x u)2 + 6u∂x2 u + ∂x4 u + O(ε5 ).
Now we compute
∂t u + ∂x3 u − 6u∂x u =(∂t w − 6w∂x w − 6ε2 w2 ∂x w + ∂x3 w)
(8.19) + 2ε2 w(∂t w − 6w∂x w − 6ε2 w2 ∂x w + ∂x3 w)
+ ε∂x (∂t w − 6w∂x w − 6ε2 w2 ∂x w + ∂x3 w).
From this we immediately see that, if w satisfies the modified KdV equation
(8.20) ∂t w − 6w∂x w + 6ε2 w2 ∂x w − ∂x3 w = 0,
then u, defined by (8.17), satisfies the KdV equation. However, one also sees
immediately that the integral of w is a conserved quantity of (8.20) for all
values of ε, i.e., if we define Iε (t) = R w(x, t) dx, then Iε is a constant for
all values of ε. (We assume here that w is defined on the real line, and that
w and its derivatives go to zero as |x| tends to infinity. Similar results hold
for x running over a finite interval with periodic boundary conditions.) But
this immediately implies that, if we use (8.18) to expand Iε in powers of ε,
then the coefficients in this expansion must also be constant in time. Since
these coefficients will be expressed as integrals of u and its derivatives they
will give us (infinitely many) conserved quantities for the KdV equation.
For the first few of these we find:
264 8. Three canonical modulation equations


• K0 = R u(x, t) dx. The conservation of this quantity follows im-
mediately from the form of the KdV equation.

• K1 = R ∂x u(x, t) dx = 0, if we assume that u and its derivatives
tend to zero as |x| tend to infinity. Thus, we gain no new informa-
tion from this quantity and in fact, all the integrals coming from
the odd powers of ε turn out to be “trivial”. Therefore, we ignore
them and focus just on the even powers of ε.
 
• K2 = R (u2 − ∂x2 u) dx = R u2 dx. That this is a conserved quan-
tity is again easy to see directly from the KdV equation, just by
multiplying the equation by u and integrating w.r.t. x.

• K4 = R ∂x (2u∂x − ∂x (∂x2 u − u2 )) + 2u(∂x2 u − u2 ) + (∂x u)2 dx =
(3u2 + (∂x u)2 ) dx.
In [ZF71] it has been proved that the KdV equation is a completely inte-
grable Hamiltonian system. In particular, there exists a canonical transfor-
mation such that, w.r.t. the new coordinates, the Hamiltonian is a function
only of the action variables and hence the action variables remain constant
in time.

8.2.3. Solitons. In the experiments made by Russell from 1834 to 1844

he was also interested in the interaction of solitary waves. He made some
sketches indicating that solitary waves after some nonlinear interaction seem
to have gone through each other as if there had been no interaction, i.e., the
interaction of two solitary waves leads to two solitary waves with the same
shapes and velocities as before the interaction. By some first computer ex-
periments in the 1950s and 1960s this behavior was observed for a number
of systems. The first one was the Fermi-Pasta-Ulam system [FPU55], a
nonlinear lattice differential equation which in contrast to the observed be-
havior was created to understand why heat conductivity in solids is finite.
Zabusky and Kruskal [ZK65] observed that the KdV equation shows this
behavior, too. Although the solitary waves Aβj interact nonlinearly, after
the interaction they seem to have gone through each other without changing
their shape, i.e., asymptotically

N
 
AN −Sol. (x, t) ∼ Aβj x − βj2 t ± γj + δj ,
j=1

for t → ±∞, with β1 > . . . > βN and γj , δj ∈ R some phase-shifts. It is the

purpose of this section to explain where this behavior comes from.
The Lax pair formulation. The key observation for the understanding
of all the remarkable properties of the KdV equation is the formulation of
Lax [Lax68] which allows to transfer the KdV equation into a system of
8.2. The KdV equation 265

explicitly solvable equations. We write the KdV equation as an abstract

evolutionary system
∂t u = N (u),
and interpret this equation as an evolution equation for operators acting in
some Hilbert space. Both sides are interpreted as multiplication operators,
namely

2
2
L (R) → L2 (R), L (R) → L2 (R),
∂t u : and N (u) :
ψ → (∂t u)ψ, ψ → (N (u))ψ.
In a first step we complicate the situation heavily and assume that the
evolutionary system can be written as
∂t L(t) = [M, L](t) := M (t)L(t) − L(t)M (t)
with self-adjoint L(t) = L∗ (t) and skew-adjoint M (t) = −M ∗ (t).
Example 8.2.3. Consider L(t)ψ(x) = −∂x2 ψ(x)+u(x, t)ψ(x). Then for each
fixed t we have ∂t L(t)ψ(x) = ∂t u(·, t)ψ(x). The operator L(t) is self-adjoint
in L2 (R) since

(L(t)ψ, φ)L2 = (−∂x2 ψ(x) + u(x, t)ψ(x))φ(x) dx
R
= ∂x ψ(x)∂x φ(x) + u(x, t)ψ(x)φ(x) dx = (ψ, L(t)φ)L2 ,
R
i.e., t is considered as fixed parameter in this calculation.

Hence, with L(t) from the example the left-hand side of the KdV equa-
tion and the left-hand side of the Lax pair formulation coincide. Before
we explain how to find operators M (t) such that also the right-hand sides
coincide we explain first consequences of a Lax pair representation.
Lemma 8.2.4. If the self-adjoint operator L=L(t) satisfies ∂t L(t)=[M, L](t),
with M (t) = −M ∗ (t) a skew-adjoint operator, then the eigenvalues of L(t)
are independent of t.

Proof. In the following ψ(t) denotes a L2 -normalized eigenfunctions of L(t)

associated to the eigenvalues λ(t). Differentiating the eigenvalue problem
L(t)ψ(t) = λ(t)ψ(t)
w.r.t. t, where λ(t) ∈ R, yields
∂t L(t)ψ(t) + L(t)∂t ψ(t) = ∂t λ(t)ψ(t) + λ(t)∂t ψ(t).
Using the right-hand side of the Lax pair formulation gives
∂t λ(t)ψ(t) =(M (t)L(t) − L(t)M (t))ψ(t) + L(t)∂t ψ(t) − λ(t)∂t ψ(t)
(8.21) =M (t)λ(t)ψ(t) − L(t)M (t)ψ(t) + L(t)∂t ψ(t) − λ(t)∂t ψ(t)
266 8. Three canonical modulation equations

=(L(t) − λ(t))(∂t ψ(t) − M (t)ψ(t)).

Using the fact that L is self-adjoint and the normalization
ψ(t), ψ(t)L2 = 1
we find
∂t λ(t) =
ψ(t), (L(t) − λ(t))(∂t ψ(t) − M (t)ψ(t))L2
=
(L(t) − λ(t))ψ(t), ∂t ψ(t) − M (t)ψ(t)L2
=
0, ∂t ψ(t) − M (t)ψ(t)L2 = 0.

Lemma 8.2.5. If the self-adjoint operator L=L(t) satisfies ∂t L(t)=[M, L](t),
with M (t) = −M ∗ (t) a skew adjoint operator, and if the eigenvalues of L(t)
are semi-simple, then there exists a real-valued function c = c(t) such that
(8.22) ∂t ψ(t) − M (t)ψ(t) = c(t)ψ(t).

Proof. From (8.21) it follows that (L(t) − λ(t))(∂t ψ(t) − M (t)ψ(t)) = 0. If

the eigenvalues of L(t) are semi-simple, then ∂t ψ(t) − M (t)ψ(t) must be a
multiple of ψ(t). Hence, there exists a c = c(t) such that (8.22) holds. 
The eigenvalues of L(t) from Example 8.2.3 are semi-simple. Since
M (t)L(t)−L(t)M (t)=(M (t)−c(t))L(t)−L(t)(M (t)−c(t))
we can always redefine M (t) such that ψ(t) satisfies ∂t ψ(t) = M (t)ψ(t).
We summarize what we found so far. If the KdV equation can be written
as a Lax pair formulation then the eigenvalues of L(t) are conserved quan-
tities for the KdV equation although L(t) contains the solution u explicitly.
Moreover, the associated eigenfunctions ψ(t) satisfy ∂t ψ(t) = M (t)ψ(t). It
will turn out subsequently that although M (t) also contains u explicitly this
equation can be solved without the knowledge of u. Finally, we will be able
to reconstruct u(·, t) out of the λs and the eigenfunctions ψ(t).
The KdV hierarchy. In order to find the Lax pair formulation for the
KdV equation we have to find operators L(t) and M (t) with
∂t L(t)ψ = (∂t u)ψ and (M (t)L(t) − L(t)M (t))ψ = (−∂x3 u + 6u∂x u)ψ.
Especially, all derivatives ∂x of M (t) and L(t) which fall on ψ have to cancel
such that the commutator of M (t) and L(t) is a pure multiplication operator.
Example 8.2.6. We consider the symmetric operator
L(t)· = −∂x2 · +u(x, t)·
and the skew symmetric operator M = ∂x . We find
(M (t)L(t)−L(t)M (t))ψ(x) =∂x (−∂x2 + u(x, t))ψ(x)+(∂x2 − u(x, t))∂x ψ(x)
=(∂x u(x, t))ψ(x).
8.2. The KdV equation 267

Hence, the PDE

∂t u = ∂x u,
which possesses the solution u(x, t) = u(x + t, 0) can be solved alternatively
with a Lax pair formulation. The eigenfunctions ψ = ψ(x, t) of
−∂x2 ψ(x, t) + u(x, t)ψ(x, t) = −∂x2 ψ(x, t) + u(x + t, 0)ψ(x, t) = λ(t)ψ(x, t)
evolve according to the above theory as a solution of
∂t ψ(x, t) = ∂x ψ(x, t).
which is solved by ψ(x, t) = ψ(x + t, 0). This result is consistent with
the above theory. The eigenvalues of the Schrödinger operator L(t) are
independent of t since the potential u is only translated in time. Hence, the
eigenfunctions are also translated in time. For the reconstruction of u from
the λs and the ψs see §8.2.3.

It turns out that for the operator L(t) from Example 8.2.6 an operator
M (t) can be found such that L and M give a Lax pair formulation of the
KdV equation. In order to derive the KdV equation we make the ansatz
M (t)· = −α∂x3 · +U ∂x · +∂x (U ·) + A
with α ∈ R, U = U (x, t) and A = A(x, t). We find
(M (t)L(t) − L(t)M (t))
=(−α∂x3 + ∂x3 U + ∂x2 A + 2(∂x u)U ) − (3α∂x2 u + 4∂x2 U + 2∂x A)∂x
− (−3α∂x u + 4∂x U )∂x2 .
In order to have a multiplication operator the coefficients in front of ∂x and
∂x2 in the second and third line have to vanish identically. From the third
line we obtain U = 3αu/4 and then from the second line that A = A(t) is
independent of x, such that finally
α
M (t)L(t) − L(t)M (t) = (−∂x3 u(x, t) + 6u(x, t)∂x u(x, t)).
4
Choosing α = 4 gives the right-hand side of our KdV equation.

Summary. If u = u(x, t) solves the KdV equation (8.13), then the eigen-
values λ = λ(t) of the eigenvalue problem
−∂x2 ψ(x, t) + u(x, t)ψ(x, t) = λ(t)ψ(x, t)
are independent of time and the associated eigenfunctions evolve according
to
∂t ψ(x, t) = −4∂x3 ψ(x, t)+3u(x, t)∂x ψ(x, t)+3∂x (u(x, t)ψ(x, t))+A(t)ψ(x, t).
268 8. Three canonical modulation equations

For the same L(t), beside the two M (t)s from above there are infinitely
many skew symmetric operators M (t) which lead to a multiplication oper-
ator for M (t)L(t) − L(t)M (t) giving a “hierarchy of KdV equations”. The
next one would be
∂t u = −∂x5 u + 10u∂x3 u + 20∂x u∂x2 u − 30u2 ∂x u.
The Hamiltonian of the nth equation in KdV hierarchy is given by the nth
conserved quantity of the KdV equation obtained via the Miura transfor-
mation, cf. [New85].
The evolutionary problem ∂t ψ(t) = M (t)ψ(t). As the next step on
our way to solve the KdV equation with the Lax pair formulation we have
to solve the evolutionary problem ∂t ψ(t) = M (t)ψ(t) for the eigenfunctions
ψ(t). It is given by
∂t ψ(x, t) = − 4∂x3 ψ(x, t)+3u(x, t)∂x ψ(x, t)+3∂x (u(x, t)ψ(x, t))+A(t)ψ(x, t)
(8.23) =2(u(x, t)+2λ)∂x ψ(x, t) − ∂x u(x, t)ψ(x, t)+A(t)ψ(x, t),
where we used the eigenvalue problem to simplify the system slightly. At
first sight it seems that we need to know the solution u = u(x, t) of the KdV
equation to solve this problem which would make the approach pretty use-
less. However, it turns out that ∂t ψ(t) = M (t)ψ(t) can be solved explicitly
without the knowledge of u. In order to explain why this is the case we
have to look more closely at the eigenfunctions ψ(t) of the self-adjoint linear
operator L(t) defined by
L(t)ψ(x, t) = −∂x2 ψ(x, t) + u(x, t)ψ(x, t),
where u = u(x, t) is a solution of the KdV equation. There are finitely many,
say N , discrete eigenvalues, 0 > λ1 > λ2 > . . . > λN and a continuum [0, ∞)
of spectral values which are denoted by λ(k) = k 2 for k ∈ R. Corresponding
to the eigenvalues there are eigenfunctions ψk ∈ L2 for k = 1, . . . , N and
ψ(k) ∈ L∞ \ L2 for k ∈ R.
Example 8.2.7. We compute the eigenvalues and eigenfunctions of Lψ =
ψ − u(x)ψ in case u(x) = −U0 δ(x), with U0 a positive constant and δ
the Dirac delta-distribution, cf. Example 5.2.2. Integration of −∂x2 ψ −
U0 δ(x)ψ = λψ gives
 
−∂x ψ|x=− − U0 δ(x)ψ(x) dx = λψ(x) dx,
− −

which, since ψ ∈ C 0 , for ρ → 0 yields the jump condition

−[∂x ψ] − U0 ψ(0) = 0, [u] := lim u(ε) − lim u(−ε).
ε→0 ε→0
8.2. The KdV equation 269

For x = 0 we have to solve −∂x2 ψ = λψ. For λ > 0, i.e., λ = k 2 we have

ψ ∼ e±ikx . We make the ansatz

−ikx
 e + b(k)eikx for x > 0,
ψ(x, k) =
a(k)e−ikx for x < 0.
Continuity in x = 0 leads to 1 + b = a and the jump condition to
−ik + bik − (−ika) = −U0 (1 + b) = U0 a(k)
which finally leads to
U0
b(k) = − .
U0 + 2ik
For λ < 0, i.e., λ = −κ2n we have ψ ∼ e±κn x . Therefore, we make the ansatz

αn e−κn x for x > 0,

ψn (x) =
βn e+κn x for x < 0.
Continuity yields αn = βn . Normalisation gives
 0  ∞
2 2κn x
αn e dx + αn2 e−2κn x dx = 1,
−∞ 0

and hence αn = κ2n . The jump condition then gives

[ψn ] = − κn αn − κn αn = − U0 αn ,
and therefore κ1 = 12 U0 , i.e., we have only negative eigenvalues.

Since the eigenfunctions satisfy a second order scalar ODE they are
completely determined
 by their asymptotic behavior for x → ±∞. For ψk
normalized by |ψk (x)|2 dx = 1 we have

−κ x
e k x → ∞,
ψk ∼ ck
eκk x x → −∞,
where κ2k = −λk and κk > 0. For ψ(k) we have

−ikx
e + b(k)eikx x → ∞,
ψ(k) ∼ −ikx
a(k)e x → −∞.
The question is if it possible to derive some ODEs for the so called scattering
data (λk (t), ck (t)) for k = 1, . . . , N and (λ(k, t), a(k, t), b(k, t)) for k ∈ R
from the PDEs ∂t ψ(t) = M (t)ψ(t) for the eigenfunctions ψ(t). It turns out
that this is possible, and that these ODEs can be solved explicitly without
knowing the evolution of the KdV solution u = u(x, t).
i) We start with the discrete spectrum. Multiplying (8.23) by ψk (x, t)
gives that A(t)ψk (x, t)2 equals
1
∂t (ψk (x, t)2 ) + ∂x (u(x, t)ψk (x, t)2 − 2(∂x ψk (x, t))2 − 4λk ψk (x, t)2 ),
2
270 8. Three canonical modulation equations


by straightforward calculation.
 Performing the integration R . . . dx and
2
using the normalization R ψk (x, t) dx = 1 yields
 
1 d
ψk (x, t) dx = 0 = A(t) ψk (x, t)2 dx = A(t),
2
2 dt R R
i.e., A(t) = 0, such that ψk (t) satisfies
(8.24) ∂t ψk (x, t) = 2(u(x, t) + 2λk )∂x ψk (x, t) − (∂x u(x, t))ψk (x, t).
As already said, ψk (x, t) is fully determined by its asymptotics ck (t)e−κk x
for x → ∞. Since u(x, t) → 0 for x → ∞ we find by multiplying (8.24) with
eκk x in the limit x → ∞ that
ċk (t) = −4λk κk ck (t) = 4κ3k ck (t).
This equation is independent of the evolution of u = u(x, t) and can be
solved explicitly, namely
3
ck (t) = e4κk t ck (0),
where ck (0) has to be computed from the KdV initial condition u = u(x, 0).
The asymptotics
ψk (x, t) ∼ ck (t)e−κk x = ck (0)e4κk t−κk x = ck (0)e−κk (x−4κk t)
3 2

for x → ∞ describes a translation of the eigenfunction ψk (x, t) with velocity

4κ2k .
ii) We come now to the continuous spectrum. Inserting the asymptotic
behavior
ψ(k; x, t) ∼ e−ikx + b(k, t)eikx
for x → ∞ into (8.23), and using again u(x, t) → 0 for x → ∞, gives for
the coefficient in front of e−ikx that 0 = 4(−ik)λ + A(t) = −4ik 3 + A(t) and
for the coefficient in front of eikx that ḃ(k, t) = 4ikλb(k, t) + A(t)b(k, t) =
(4ik 3 + A(t))b(k, t) such that finally
ḃ(k, t) = 8ik 3 b(k, t).
Again this equation can be solved explicitly, namely
3
b(k, t) = e8ik t b(k, 0),
where b(k, 0) has to be computed from the KdV initial condition u = u(x, 0).
With the same argument the evolution equation for a(k, t) can be derived.
Inserting the asymptotic behavior
ψ(k; x, t) ∼ a(k, t)e−ikx
for x → −∞ into (8.23), using u(x, t) → 0 for x → −∞, and A(t) = 4ik 3
gives for the coefficient in front of e−ikx that
ȧ(k, t) = −4ikλa(k, t) + A(t)a(k, t) = (−4ik 3 + A(t))a(k, t) = 0,
8.2. The KdV equation 271

such that a(k, t) = a(k, 0).

Remark 8.2.8. The scattering data variables can be interpreted as ac-
tion and angle variables of a completely integrable Hamiltonian system.
The eigenvalues λk and λ(k) are the conserved quantities, i.e., the action
variables, whereas the ck (t) and b(k, t) behave linearly, i.e., are the angle
variables.

The inverse scattering transform. Two steps remain. First we have

to compute the initial scattering data
(λk (0), ck (0), b(k, 0), a(k, 0))
from the initial condition u = u(x, 0) of the KdV equation. Secondly, we
have to reconstruct the solution u = u(x, t) of the KdV equation from the
scattering data
3 3
(λk (t), ck (t), b(k, t), a(k, t)) = (λk (0), e4κk t ck (0), e8ik t b(k, 0), a(k, 0)).
Both problems are well studied in quantum mechanics since the operator
L(t) is the Schrödinger operator with potential u = u(x, t) describing the
wave function of a non-relativistic electron. We recall the parts of scattering
and inverse scattering theory which are necessary for the construction of
u = u(x, t) from the data (λk , ck (t), λ(k), a(k, t), b(k, t)) by the Marchenko-
or Gelfand-Levitan equation [GL51, Mar55]. For each fixed time t the
solution u = u(x, t) is given by
d
u(x) = −2 K(x, x),
dx
with K(x, z) satisfying the linear integral equation
∞
K(x, z) + F (x + z) + K(x, y)F (y + z) dy = 0
x
where 

n
1
F (x) = c2j e−κj x + eikx b(k) dk.
2π
j=1 R
If we denote the calculation of the scattering data by T and the solution
of the Marchenko or Gelfand-Levitan equation, i.e., the inverse scattering
transform, by T −1 then we obtain the following commutative diagram.

u0 nonlinear
−−−−−−−−− evolution
−−−−−−−in−−KdV
−−→ u(·, t)
−1
↓T ↑T
scattering data scattering data
(λk (0), ck (0))k=1,...,N linear evolution of scattering data (λk (t), ck (t))k=1,...,N
(λ(k, 0), b(k, 0))k∈R −−−−−−−−−−−−−−−−−−−−−−−−−→ (λ(k, t), b(k, t))
k∈R
272 8. Three canonical modulation equations

That this procedure is really useful is explained in the next example where
we compute 2-soliton solutions of the KdV equation.
Example 8.2.9. We consider the situation of two negative eigenvalues λ1 =
−1 and λ2 = −4, such that κ1 = 1 and κ2 = 2. Next we choose the initial
scattering data
√ √
c1 (0) = 6, c2 (0) = 2 3, and b(k, 0) = 0
for all k ∈ R. Such a potential is called reflectionless. We find
√ √
c1 (t) = 6e4t , c2 = 2 3e32t , and b(k, t) = 0
for all k ∈ R and t ∈ R. The function F in the Marchenko or Gelfand-
Levitan equation is then given by
F (x) = 6e8t−x + 12e64t−2x .
In order to solve this equation we make the ansatz
K(x, z) = 1 (x)e−z + 2 (x)e−2z .
With Fj (x) = c2j e−κj x we obtain
0 =1 e−z + 2 e−2z + F1 (x)e−z + F2 (x)e−2z
∞
 
+ 1 (x)e−y + 2 (x)e−2y )(F1 (z)e−y + F2 (z)e−2y dy.
x

The coefficients in front of e−z and e−2z are then given by

1 + 6e8t−x + 31 e8t−2x + 22 e8t−3x =0,
2 + 12e64t−2x + 41 e64t−3x + 32 e64t−4x =0.
The solution of this linear system is
   
1 = 6 e72t−5x − e8t−x /D and 2 = −12 e64t−2x + e72t−4x /D,
where D = 1+3e8t−2x +3e64t−4x +e72t−6x . Finally, we obtain the remarkable
explicit solution
 
u(x, t) = − 2∂x 1 e−x + 2 e−2x
3 + 4 cosh(2x − 8t) + cosh(4x − 64t)
(8.25) = − 12 .
(3 cosh(x − 28t) + cosh(3x − 36t))2
This 2-soliton solution describes the reflectionless interaction of two solitons,
cf. Exercise 8.14 and Figure 8.4.

The previous computation can easily be generalized to N negative eigen-

values in case of a reflectionless potential, i.e., b(k, 0) = 0. The explicit
solution formula obtained in this way describes a so called N -soliton. For
8.2. The KdV equation 273

Figure 8.4. Density plot of the 2-soliton (8.25).

t → ±∞ the complicated solution separates in N individual solitary waves,

cf. Exercise 8.14.
Remark 8.2.10. The procedure can also be used to characterize the as-
ymptotic behavior for t → ∞ for arbitrary initial conditions u0 satisfying
R u0 (x)(1 + |x|) dx < ∞, cf. [EvH81]. Suppose that we have N negative
eigenvalues. Each eigenvalue corresponds to a soliton in the KdV equation.
Then for large t the solution u consists of N -solitons which are ordered
w.r.t. their height which as we have seen is proportional to the velocity of
the solitons. Behind the solitons we find a dispersive rest decaying with a
rate t−1/3 which is described by the coefficients b(k, t).

8.2.4. Stability of the solitary wave. We cannot expect naive stability

of a solitary wave uc since according to formula (8.16) for t = 0 there will
always be a nearby solitary wave uc∗ with different velocity c∗ which will
run away from uc for t → ∞. However, using the Hamiltonian structure
of the KdV equation so called orbital stability can be established [Ben72].
Orbital stability means that for all ε > 0 there exists a δ > 0 such that
(8.26) u0 (·) − uc (·, 0)H 1 ≤ δ ⇒ inf u(·, t) − uc (x + x0 , t)H 1 ≤ ε,
x0 ∈R

i.e., the solution u(·, t) stays close to a translate of uc for all times.
The argument is similar to the one in Theorem 4.1.6 where the stability
of a fixed point has been shown using the positive definiteness of the second
derivative of the Hamiltonian at the fixed point. See Figure 8.5.
In case of orbital stability the argument is a little bit more complicated.
First of all the solitary waves uc are only constrained minima, i.e., they
minimize H in the invariant subspaces of the functions with the same L2 -
norm. Moreover, in each of these subspaces there is a one-dimensional family
of minima due to the translation invariance of the problem. In the following
we sketch the idea to overcome these difficulties.
274 8. Three canonical modulation equations

o * +

Figure 8.5. Cartoon picture for stability. o is the minimum of a

conserved quantity M , hence perturbations ∗ and + can’t get away
from o.

We use Lagrange multipliers for the constraint and consider


1 c
M (u) = H(u) + cE(u) = (∂x u)2 + u3 + u2 dx
R 2 2
d
which is a constant of motion, i.e., dt M (u(t)) = 0. We find

1 c
M (uc + εv) = (uc + εv )2 + (uc + εv)3 + (uc + εv)2 dx
2 2
R
= M (uc ) + ε (−u c + 3u2c + cuc )v dx
R
 
2 1 2 2 c 2 3
(8.27) +ε v + 3uc v + v dx + ε v 3 dx
R 2 2 R
= M (uc ) + ∂u M (uc )[εv] + ∂u2 M (uc )[εv, εv] + ∂u3 M (uc )[εv, εv, εv].

Since uc satisfies the traveling wave equation u c − cuc − 3u2c = 0 for the KdV
equation we see that ∂u M (uc ) = 0 and that uc is a stationary point of M .
Next we restrict u0 to the energy surface of the traveling wave uc , i.e., we
restrict ourselves to the consideration of initial conditions u0 with E(u0 ) =
E(uc ) =: Ec . It can then be shown that uc is a minimum of M in this
surface, and moreover that there exist C1 , C2 > 0 such that
(8.28) 0 ≤ C1 inf u(· + x0 ) − uc (·)2H 1 ≤ M (u) − M (uc ) ≤ C2 u − uc 2H 1
x0 ∈R

for all sufficiently small u − uc H 1 , see [Ben72]. Next we have the positive
definiteness of the second variation

1 2 c
2
∂u M (uc )[v, v] = v + 3u c v 2 + v 2 dx
R 2 2
of M (uc ), cf.(8.27), in the subspace of constant L2 -norm. From (8.28) we
get orbital stability of uc in the energy surface E(u0 ) = Ec . This is called
in the following conditional orbital stability.
8.3. The GL equation 275

From E(uc ) = 8c3/2 /3 it follows that for |E(u0 ) − Ec | < δ there exists a
c∗ with |c − c∗ | ≤ Cδ and E(u0 ) = Ec∗ . Hence, uc − uc∗ H 1 < ε/2 if δ > 0
is sufficiently small. Using this and the conditional orbital stability of uc∗
yields
inf u(·, t)−uc (·+x0 , t)H 1
x0 ∈R
≤ inf u(·, t) − uc∗ (· + x0 , t)H 1 + inf uc∗ (·, t) − uc (· + x0 , t)H 1
x0 ∈R x0 ∈R
≤ε/2 + ε/2
for δ > 0 sufficiently small, i.e., the orbital stability of uc . The orbital
stability of N -solitons of the KdV equation has been established in [MS93].

8.3. The GL equation

The Ginzburg-Landau (GL) equation
(8.29) ∂t u = (1 + iα)∂x2 u + u − (1 + iβ)|u|2 u,
with α, β, x ∈ R, t ≥ 0, and u(x, t) ∈ C, appears as a universal modu-
lation equation describing bifurcating solutions in spatially extended pat-
tern forming systems close to the first instability. Typical examples are
reaction-diffusion systems such as the Brusselator, or hydrodynamical sta-
bility problems such as the Couette-Taylor problem or Bénard’s problem.
This connection and more examples are discussed in Part IV of this book,
where the mathematical justification of this so called GL approximation is
one of the major topics.
However, the name of this equation comes from a completely different
application. It has been derived in a multi-dimensional real version
(8.30) ∂t u = Δu + u − |u|2 u,
with x ∈ R3 , by Ginzburg and Landau as a model for the description of
superconductivity [GL50]. In this theory u is an order parameter which
describes in each point the fraction of the superconducting phase. The most
prominent solutions of (8.30) are the so called GL vortices [BBH94]. Due
to the applications which we have in mind we focus on the 1D GL equation
over the real line and completely ignore any connection to superconductivity.
The dynamics of (8.29) differs strongly depending on the coefficients α
and β. It spans from very simple dynamics to spatio-temporal chaos [TZ10],
but for all α and β the origin is unstable with an (uncountable) infinite-
dimensional unstable manifold. There are already a number of overview
articles, cf. [vS95, AK02], about the GL equation itself. We concentrate
on topics which are of interest for Part IV of this book. These are the local
and global existence of solutions in uniformly local Sobolev spaces which
will be introduced in a first step. Then we will discuss the existence of an
276 8. Three canonical modulation equations

attractor and the linear stability analysis of spatially periodic pattern. On

unbounded domains there is another source of loss of compactness, namely
the transport to infinity and so the analysis of PART II of this book about
the existence of attractors has to be modified in several directions.
As we have seen, Sobolev spaces on Rd and Fourier transform fit very
well together. However, for many systems over unbounded domains Sobolev
spaces are not rich enough to contain all interesting solutions since for in-
stance they do not contain spatially periodic functions or fronts, or, more
generally, solutions which do not decay to zero for |x| → ∞. Therefore, an
0
alternative choice could be Cb,unif and generalizations. On the one hand
0
Fourier transform is available in Cb,unif -spaces, but on the other hand its
analytic properties turn out to be rather complicated, see §5.2.1. Therefore,
we look for spaces which contain all these non-decaying functions and for
which additionally Fourier transform can be used to control the solution
operators of the linearized equations. The so called uniformly local Sobolev
θ turn out to be a good choice.
spaces Hul
Thus, before discussing special solutions of the GL equation, and local
(and global) existence of general solutions, here we start with some func-
tional analysis to set up the spaces where the interesting solutions of the
GL equation, such as periodic solutions and fronts, live in. Then we con-
sider spatially periodic solutions, but we postpone a more comprehensive
discussion of special solutions to Part IV, in particular §10.2.
Remark 8.3.1. a) If α = β = 0, then (8.29) is called real GL equation.
If α = 0 or β = 0, then (8.29) is called complex GL equation or just GL
equation. For the real GL equation we have the invariant subspace of real-
valued functions u : R → R, which then fulfill the Allen-Cahn equation
∂t u = ∂x2 u + u − u3 .
Thus, the dynamics which we found for the Allen-Cahn equation in §7.2 is
also present in the real GL equation.
b) Like the NLS equation the GL equation is S 1 -symmetric, i.e., if u :
R → C is a solution of (8.29), then so is v = eiφ u with arbitrary φ ∈ R.
c) The GL equation (8.29) can also be considered as a two component
real-valued system either for the real and imaginary parts of u, cf. Exercise
8.17, or for the polar coordinates, see §8.3.6. There is no maximum principle
for (8.29).

8.3.1. Uniformly local Sobolev spaces Hul θ . We choose a positive weight

 ρ : R → (0, ∞) which is continuous, bounded, and has a finite in-

function
tegral R ρ(x) dx. For later purposes we also impose ρ ∈ C 2 (R, R) and
assume |ρ (x)|, |ρ (x)| ≤ ρ(x) for all x. It turns out that the subsequent
8.3. The GL equation 277

θ is rather independent of
definition of the uniformly local Sobolev spaces Hul
the particular choice of ρ. Therefore, we may fix ρ once and for all to
ρ(x) = 1/ cosh(x) or ρ(x) = 2/(2 + x2 ),
which both satisfy the conditions on ρ. We let
L2
2 (R) = { u ∈ L2 (R) : u 2 < ∞ },
ul loc L ul

where 
u2L2 = sup ρ(y + x)u(x)2 dx.
ul y∈R R
The statement that different weight functions lead to the same uniform space
with equivalent norms is made rigorous with the following lemma.
Lemma 8.3.2. The norm  · Lul is equivalent to the norm  · ∗ defined by
 1/2
y+1/2
u∗ = sup |u(x)|2 dx .
y∈R y−1/2

Proof. Let α = minx∈[−1/2,1/2] ρ(x). Then

 1/2  1/2
y+1/2 y+1/2
α |u(x)| dx
2
≤ ρ(y + x)|u(x)| dx2
y−1/2 y−1/2
 1/2
≤ 2
ρ(y + x)|u(x)| dx .
R
Taking the supremum w.r.t. y shows αu∗ ≤ uLul . On the other hand
we have
⎛ ⎞1/2
 1/2   y+1/2
ρ(y + x)|u(x)|2 dx =⎝ ρ(y + x)|u(x)|2 dx⎠
R y∈Z y−1/2
⎛ ⎞1/2
  y+1/2
≤⎝ sup ρ(y + x) |u(x)|2 dx⎠
y∈Z x∈[−1/2,1/2] y−1/2

  1/2 ⎛ ⎞1/2
y+1/2 
≤ sup |u(x)|2 dx ⎝ sup ρ(y + x)⎠
y∈R y−1/2 y∈Z x∈[−1/2,1/2]
 1/2
y+1/2
≤C sup |u(x)| dx
2
y∈R y−1/2

since C = ( y∈Z supx∈[−1/2,1/2] ρ(y + x))1/2 < ∞ due to the assumptions on
ρ. This shows uLul ≤ αu∗ . 
278 8. Three canonical modulation equations

In the rest of this section we omit the domain R as all function spaces
considered here are defined over the real line. With the translation operator
Ty : L2
2 → L
ul
22 , (T u)(·) → u( · + y), our final space of uniformly local L2
ul y
functions is given as

L2ul = { u ∈ L2
2 : T u − u 2 → 0 as y → 0 }.
ul y L ul

This construction is very similar to the construction of the space of uniformly

continuous functions, cf. Exercise 8.20. For θ ∈ N we define the associated
Sobolev spaces H 3 θ θ
ul and Hul by requiring that the first m distributional
derivatives lie in the spaces as well. Some first properties of these spaces are
summarized in the following lemma, cf. [MS95, Lemma 3.1].

Lemma 8.3.3. a) L2ul is a closed subspace of L2

2 .
ul
3θ θ
b) The spaces H and H are dense in L . 2
ul ul ul
c) L2ul is a proper subset of L2
2 .
ul

Proof. a) To show that L2ul is a closed subspace let (un )n∈N be a Cauchy
sequence in L2 with limit u ∈ L2
ul
2 . Then
ul

Ty u−uL2 ≤ Ty u−Ty un L2 +Ty un −un L2 +un −uL2 = s1 +s2 +s3 .
ul ul ul ul

Since un approximates u there exists an n such that s1 = s3 < ε/3. For the
corresponding un ∈ L2ul we find a δ > 0 such that for all y ∈ (−δ, δ) we have
s2 < ε/3.
3
b) Let u ∈ H 1 . With part a) we conclude
ul

Ty u−u2L2 = sup{ ρ(x)(u(x+y+z)−u(x+z))2 dx : z ∈ R } ≤ C|y|u2H 1 .
ul
R ul

3 3
ul ⊂ Lul and hence Hul ⊂ Hul ⊂ Lul for θ ∈ N. To see that Hul
Thus, H 1 2 θ θ 2 1

is dense in L2ul choose any u ∈ L2ul . As y → Ty u ∈ L2ul is continuous the

h
Riemann integral vh = h1 0 Ty u dy exists and vh − uL2 → 0 for h → 0.
  ul
1 as v (x) = 1 h u(x + y) dy = 1 x+h u(z) dz. Iterating
Moreover, vh ∈ Hul h h 0 h x
θ is dense in L2 .
this process it follows that Hul ul
c) It suffices to consider the example u(x) = n for x ∈ [n, n + 1/n2 ],
n ∈ N, and u(x) = 0 else. This u lies in L2
2 \ L2 . Hence, L2 is a proper
ul ul ul
22
subspace of L . 
ul
Compactness. Over bounded domains, smoothness yields (pre)com-
pactness by Sobolev embeddings, see §5.3. This is not true over unbounded
domains.
8.3. The GL equation 279

Example 8.3.4. The embedding H 1 (R) → L2 (R) is not compact, since the
sequence (un )n∈N , with un (x) = u0 (x + n) and u0 (x) = e−x is bounded in
2

H 1 , but does not contain a convergent subsequence.

Instead, we have a form of localized compactness which plays a role in

§8.3.4 where we construct an attractor for the GL equation on the real line
based on Hulθ spaces, cf. [MS95, Theorem 3.2].

Theorem 8.3.5. Let θ > 1 and B ⊂ Hul

1 be any set which is bounded in
θ
Hul . Then B is pre-compact in

Hρ1 = {u : R → R : u, ∂x u ∈ L2loc , uHρ1 = uρH 1 < ∞}.

Proof. For every ε > 0 we have to show that B admits in Hρ1 a finite covering
by balls of radius less than ε. We decompose every u ∈ B into u = v + w
with v = uχβ and w = u(1 − χβ ), where the smooth cut-off function χβ
vanishes for |x| ≤ β and equals 1 for |x| ≥ β + 1. Then vHρ1 < ε for β
sufficiently large, as ρ decays for |x| → ∞ and u varies in a bounded set.
Moreover, w ∈ H θ ([−β − 1, β + 1]) which can be embedded compactly into
H 1 ([−β − 1, β + 1]). Since for functions with support [−β − 1, β + 1] the
norms in H θ and Hρθ are equivalent, there is a finite covering of this set by
balls BHρ1 (wi , ε) with i = 1, . . . , m < ∞. Thus, ∪i=1,...,m {BHρ1 (wi , 2ε)} is a
finite covering of B, since u − wi Hρ1 ≤ vHρ1 + w − wi Hρ1 . 
Multiplier theory in Hul θ -spaces. An important tool for studying trans-
θ is multiplier theory which uses
lational invariant operators on the spaces Hul
q
Fourier transform methods. An operator M : Hul → Hulθ is called multiplier

4 : R → C such that M u = F −1 (M
if there exists a function M 4F u), i.e., if the
associated operator in Fourier space is a multiplication operator. Taking the
Fourier transform of a function u ∈ L2l,u gives a tempered distribution. This
allows easily to define operators via their action in Fourier space. However,
it is rather complicated to estimate the norm of the operators in physical
space. In order to do so we proceed as follows. We will make use of the
fact that Fourier transform is an isomorphism between Hnm and Hm n , cf.

Definition 7.3.30 and Lemma 7.3.31. This is the case for almost all opera-
tors considered so far in this book. We have the following classical result of
multiplier theory.
4(k) ∈ C 0 (R, C).
Lemma 8.3.6. Let q, θ ≥ 0 and wθ−q (k) = (1 + k 2 )(θ−q)/2 M b
4F u) is well defined with the estimate
Then M : H q → H θ , u → F −1 (M
M uH θ ≤ C(q, θ)wθ−q C 0 (R,C) uH q ,
b

4.
where C(q, θ) does not depend on M
280 8. Three canonical modulation equations

Proof. The statement follows immediately from

M u θ ≤ CM4u 0 ≤ Cwθ−q  0 
uH 0 ≤ Cwθ−q 
H Hθ Cb (R,C) q Cb0 (R,C) uH q .

θ -spaces is as follows, cf. [Sch94a, Lemma 5],
The similar result in Hul
which also contains an extension of the result from Lul (R, C) to Lul (R, H),
with H some Hilbert space.
4(k) ∈ C 2 (R, C).
Lemma 8.3.7. Let q, θ ≥ 0 and wθ−q (k) = (1 + k 2 )(θ−q)/2 M b
q
Then Mul : Hul → Hul , u →
θ −1
 F (M4F u) is well defined with the estimate
Mul uH θ ≤ C(q, θ)wθ−q C 2 (R,C) uH q ,
ul b ul

4.
where C(q, θ) does not depend on M

Proof.
 We choose χ ∈ C0∞ such that its support is contained in [−1, 1] and
j∈Z χ(x + j) ≡ 1. The operator Mul then can be defined as follows. For
q
u ∈ Hul we set uj (x) = u(x − j)χ(x). Since uj ∈ H2q we find vj = M uj ∈ H2θ
according to
M uj H θ ≤ CM 4uj H 2 ≤ Cwθ−q C 2 
uj Hq2 ≤ uj H2q .
2 θ b
 θ . But,
Now let M u = j∈Z Tj vj . Clearly this sum does not converge in Hul
since Tj vj is concentrated around x = j and decays like 1/(1 + (x − j)2 ) it is
 θ with norm
easy to see that j∈Z Tj vj converges locally to a function in Hul
≤ C(q, θ)wθ−q C 2 (R,C) uH q . 
b ul

8.3.2. Local existence and uniqueness of solutions. We already con-

sidered the local existence and uniqueness of solutions of the GL equation in
§5.2.3 in case of periodic boundary conditions. In exactly the same way we
prove here the local existence and uniqueness of solutions in Sobolev spaces
H θ with θ > 1/2 on the real line.
Theorem 8.3.8. Let θ > 1/2 and u0 ∈ H θ . Then there exists a T0 =
T0 (u0 H θ ) > 0 and unique mild solution u ∈ C([0, T0 ], H θ ) of the GL
equation with u|t=0 = u0 .

Proof. The semigroup T (t) : H θ → H θ generated by (1 + iα)∂x2 is defined

by T (t) = F T(t)F −1 , where T(t)[k] = e−(1+iα)k t . We have the estimate
2

T (t)uH θ = T(t)
uL2 ≤ T(t)C 0 
uL2 ≤ 
uL2 = uH θ .
θ b θ θ

The nonlinear term N (u) = u − (1 + iβ)|u|2 u

is locally Lipschitz-continuous
from H θ to H θ if θ > 1/2. Thus, the right-hand side of the variation of
constant formula
 t
u(t) = T (t)u(0) + T (t − τ )N (u)(τ ) dτ
0
8.3. The GL equation 281

is a contraction in the space C([0, T0 ], H θ ) if T0 > 0 is sufficiently small. 

Remark 8.3.9. Following the proof of Lemma 6.2.5 we have the estimate
  L2 ≤ C(
u ∗ v ∗ w uL2 
v L1−r w
 L1−r
θ θ+2r

uL1−r 
+ v L2 w
 L1−r +
uL1−r 
v L1−r w
 L2 ).
θ+2r θ+r

uL1−r ≤ C
Since  uL2 for every δ > 0 and since Fourier transform is
1/2−r+δ
an isomorphism between L2s and H s we can choose θ = −1 − 2δ, r = 1/2 + δ
to establish the local Lipschitz-continuity of the nonlinear term N (u) =
u − (1 + iβ)|u|2 u from L2 into H −1−2δ for every δ > 0. Since the semigroup
T (t) is smoothing with a singularity t−1/2−δ from H −1−2δ to L2 the right-
hand side of the variation of constant formula is a contraction in the space
C([0, T0 ], L2 ), too, if T0 > 0 is sufficiently small. Hence, Theorem 8.3.8 is
already true for every θ ≥ 0.

However, H θ as a phase space does not contain spatially periodic func-

tions or fronts and is therefore too small for our purposes. Thus, we next
0
choose Cb,unif as phase space. In case of α = 0 the semigroup T (t) of the GL
equation coincides with the semigroup of the linear diffusion equation. Since
the nonlinear term N (u) = u−(1+iβ)|u|2 u is locally Lipschitz-continuous in
0
Cb,unif the right-hand side of the variation of constant formula is a contrac-
0
tion in the space C([0, T0 ], Cb,unif ) if T0 > 0 is sufficiently small. Therefore,
we have
Theorem 8.3.10. Let α = 0 and u0 ∈ Cb,unif 0 . Then there exists a T0 =
T0 (u0 C 0 ) > 0 and unique mild solution u ∈ C([0, T0 ], Cb,unif
0 ) of the
b,unif
GL equation with u|t=0 = u0 .

We refrain from generalizing the result to the case α = 0 since Cb,unif

0

is not suitable for establishing the existence of an absorbing set. Instead we

choose the uniformly local Sobolev spaces Hul θ . For the local existence and

uniqueness theory we need

Lemma 8.3.11. The space Hul θ is an algebra and can be continuously em-
0
bedded in Cb,unif if θ > 1/2.

Proof. We write u, v ∈ Hul θ as u(x) =
n∈Z un (x) where suppun (x) ∈
(n − 1,
 n + 2), and similarly for v. For the product we find u(x)v(x) =
l∈Z n∈Z ul (x)vn (x) where ul (x)vn (x) = 0 if |l − n| > 2 due to disjoint
supports. Hence
uvH θ [k,k+1] =(uk−1 + uk + uk+1 )(vk−1 + vk + vk+1 )H θ [k,k+1]
≤(uk−1 H θ [k,k+1] + uk H θ [k,k+1] + uk+1 H θ [k,k+1] )
× (vk−1 H θ [k,k+1] + vk H θ [k,k+1] + vk+1 H θ [k,k+1] )
282 8. Three canonical modulation equations

≤9uH θ vH θ
ul ul

for θ > 1/2 and as a consequence uvH θ ≤ 9uH θ vH θ . Similarly, the
ul ul ul
embedding follows. 
Then we have the following local existence and uniqueness result.

Theorem 8.3.12. Let θ > 1/2 and u0 ∈ Hul θ . Then there exists a T =
0
T0 (u0 H θ ) > 0 and unique mild solution u ∈ C([0, T0 ], Hul
θ ) of the GL
ul
equation with u|t=0 = u0 .

Proof. The nonlinear term N (u) = u − (1 + iβ)|u|2 u is locally Lipschitz-

θ . The linear operator (1 + iα)∂ 2 is the generator of a
continuous in Hul x
semigroup which is given in Fourier space by T(t)[k] = e−(1+iα)k t . In order
2

to get a bound for the semigroup T (t) = F T(t)F −1 in Hulθ we use Lemma

8.3.7. We have supk∈R |e−(1+iα)k t | ≤ 1 and for the second derivative

2

sup |(−4(1 + iα)2 k 2 t2 − 2(1 + iα)t)e−(1+iα)k t | = O(t)

2

k∈R

such that T(t)C 2 is bounded for finite t. Hence, the right-hand side of the
b
θ ) if
variation of constant formula is a contraction in the space C([0, T0 ], Hul
T0 > 0 is sufficiently small. Therefore, we are done. 
The local existence and uniqueness theorem can be improved from θ >
1/2 to θ ≥ 0. In order to do so we combine Remark 8.3.9 and Lemma 8.3.11.
Using the inequality of Remark 8.3.9 yields
 
 1 
  1 1


uvwH θ [k,k+1] = ( uk+j )( vk+j )( wk+j )

 j=−1 j=−1 j=−1  θ
H [k,k+1]

1 
1 
1
≤ uk+j1 vk+j2 wk+j3 H θ [k,k+1]
j1 =−1 j2 =−1 j3 =−1


1 
1 
1
≤ uk+j1 vk+j2 wk+j3 H θ
j1 =−1 j2 =−1 j3 =−1


1 
1 
1
≤C uk+j1 L2 vk+j2 L2 wk+j3 L2
j1 =−1 j2 =−1 j3 =−1
≤ 9CuL2 vL2 wL2 ,
ul ul ul

and so uvwH θ ≤ 9CuL2 vL2 wL2 if θ = −1 − 2δ for every δ > 0.

ul ul ul ul
As a consequence, it follows
8.3. The GL equation 283

Theorem 8.3.13. Let θ ≥ 0 and u0 ∈ Hul θ . Then there exists a T =

0
T0 (u0 H θ ) > 0 and unique mild solution u ∈ C([0, T0 ], Hul
θ ) of the GL
ul
equation with u|t=0 = u0 .

Remark 8.3.14. The smoothing properties of the semigroup allows us to

establish as before that the solutions are analytic for every t > 0. Again
the global existence of solutions or the existence of an absorbing set in
0
Cb,unif , respectively L2ul , implies the global existence of solutions or the
existence of an absorbing set in Cb,unifm θ , also for every
, respectively Hul
m ∈ N, respectively θ > 0.

8.3.3. Global existence for the GL equation. In order to establish the

existence of an attractor of a dynamical system one needs the existence of
an absorbing set in the chosen phase space X, cf. §2.4.2. If such a set exists
the solutions must exist globally in time, which follows by combining the
local existence and uniqueness of solutions with a priori estimates for the
norm u(t)X . According to Remark 8.3.14 it is sufficient to establish these
0
a priori estimates in Cb,unif , respectively L2ul .
0
The real case in Cb,unif . Let α = β = 0 and assume that u0 ∈ Cb,unif
0 .
We introduce polar coordinates u(x, t) = r(x, t)e iφ(x,t) and find

∂x r
(8.31) ∂t φ = ∂x2 φ + 2 + ∂x φ, ∂t r = ∂x2 r + r − r3 − r(∂x φ)2 .
r
Since supx∈R |u(x, t)| = supx∈R r(x, t) = R(t) it is sufficient to get a bound
for r. Since (∂x φ)2 ≥ 0 we obtain

∂t r ≤ ∂x2 r + r − r3 .

The maximum principle then implies ∂t R ≤ R − R3 such that

lim sup sup |u(x, t)| = lim sup R(t) ≤ 1.

t→∞ x∈R t→∞

Hence, we have the following global existence and uniqueness result

Theorem 8.3.15. Let α = β = 0 and u0 ∈ Cb,unif 0 . Then there exists a

unique mild solution u ∈ C([0, ∞), Cb,unif ) of the GL equation with u|t=0 =
0

u0 and lim supt→∞ u(·, t)C 0 ≤ 1.

b,unif

The real case in L2ul . Let α = β = 0 and assume that u0 ∈ L2ul . In order
to find the bound in L2ul we consider weighted energy estimates for Ey (t) =
 2 −1 and ρ (x) = ρ(x + y). We
R u(x, t)u(x, t)ρy (x) dx with ρ(x) = 2(2 + x ) y
284 8. Three canonical modulation equations

find through integration by parts and by using |ρ | ≤ ρ that

 
d
Ey (t) = (∂x u + u − |u| u)uρy dx +
2 2
(∂x2 u + u − |u|2 u)uρy dx
dt R R
 
=2 (−|∂x u| + |u| − |u| )ρy dx + 2 |u|2 ρ y dx
2 2 4

R  R

≤2 (2|u|2 − |u|4 )ρy dx ≤ 2 (4 − 2|u|2 )ρy dx

R √
R

≤8 ρ dx − 4Ey (t) = 16π 2 − 4Ey (t).

R
√
Thus, lim supt→∞ Ey (t) ≤ 4π 2, and since y ∈ R was arbitrary we finally
have
5
√
lim sup u(t)L2 = lim sup sup Ey (t) ≤ 4π 2.
1/2
ul
t→∞ t→∞ y∈R

The general case in L2ul . Now let α, β ∈ R be arbitrary. Again it is

sufficient to bound Ey (t), for which we find

d
Ey (t) = ((1 + iα)∂x2 u + u − (1 + iβ)|u|2 u)uρy dx
dt R

+ ((1 + iα)∂x2 u + u − (1 + iβ)|u|2 u)uρy dx
R
  √
= 2 (−|∂x u| + |u| − |u| )ρy dx + 2 |u|2 ρ y dx ≤ 16π 2 − 4Ey (t)
2 2 4
R R

as before since the imaginary parts cancel. Therefore, as before

5
√
lim sup u(t)L2 = lim sup sup Ey (t)1/2 ≤ 4π 2.
ul
t→∞ t→∞ y∈R

Using the variation of constant formula and the smoothing properties of the
semigroup we find that

lim sup u(t)H 1 = O(|β|)

ul
t→∞

for |β| → ∞. See Exercise 8.21.

Remark 8.3.16. First a priori estimates for the size of the solutions have
been established in [BCD+ 90] leading to the terminology of soft and hard
turbulence regimes in the (α, β)-plane. Later on the estimates have been
improved in [Mie98] where also estimates for higher space dimensions can
be found.
8.3. The GL equation 285

8.3.4. Attractors on unbounded domains. In §2.4 we introduced the

notion of an attractor and studied simple ODE cases. We then trans-
ferred this to the case of (dissipative) PDEs over spatially bounded do-
mains, see §5.3. A key property was compactness, which for ODEs with
finite-dimensional phase space simply follows from boundedness. For dissi-
pative PDEs over spatially bounded domains the compactness is obtained
from boundedness of orbits, smoothing properties of the semigroup, and
compactness of embeddings. This no longer works for PDEs over spatially
unbounded domains. In particular, for translational invariant problems we
cannot expect a compact embedding, cf. Example 8.3.4, and as a conse-
quence no compact attractor exists in any translational invariant norm.
Example 8.3.17. Consider the linear heat equation ∂t u = ∂x2 u in the space
0
Cb,unif (R, R). We are interested in the attractor A in the positively invariant
set B = {u ∈ Z : u∞ ≤ R} for an R > 0. The only solutions which stay in
B for all negative times are the constants, thus A = {u ≡ θ : θ ∈ [−R, R]} is
the only candidate for an attractor. Consider now the solution St (u0 ) with
the initial condition u0 (x) = (2/π) arctan(x). Using the explicit solution
 2
formula u(x, t) = R G(x − y, t)u0 (y)dy with G(x, t) = 2√1πt e(x−y) /(4t) we
find that for all t we have u(x, t) → ±1 for x → ±∞. Hence, we always
have distC 0 (St (u0 ), A) = 1, and attractivity of A cannot hold in the uniform
b
norm of Z. However, we have convergence to 0 on every finite interval I,
i.e., supx∈I |St (u0 ) − 0| → 0 for t → ∞.

Hence, we can only expect attractivity in a spatially localized norm. We

closely follow the setup introduced in [MS95] to show the existence of such
an attractor. We start with some abstract theory which will be applied to
the GL equation and later in Part IV to more complicated pattern forming
systems.
We are interested in attractors for semigroups (St )t≥0 on a Banach space
(Z,  · ). Our basic assumption is that St (u) ∈ Z depends continuously on
(t, u) ∈ [0, ∞) × Z. As before a subset B ⊂ Z is called positively invariant
for (St ) if St (B) ⊂ B for all t > 0, and it is called an absorbing set for (St ), if
it is bounded, positively invariant and every bounded set B ⊂ Z is absorbed
into B in finite time, i.e., there exists t > 0 such that St (B) ⊂ B.
Beside the ’uniform’ norm  ·  we work with a ’spatially localized’ norm
 · ρ such that uρ ≤ u. We denote by Zρ the set Z, but equipped with
the weaker topology induced by the norm  · ρ .
Definition 8.3.18. Let B ⊂ Z be positively invariant. A subset A ⊂ B is
called an (Z, Zρ )-attractor for St in B if the following conditions hold:
a) A is nonempty, closed, bounded in Z, and compact in Zρ .
b) A is invariant under St , i.e., St (A) = A for all t > 0.
286 8. Three canonical modulation equations

c) Every B ⊂ B which is bounded in Z is attracted to A in the distance

induced by the norm of Zρ , i.e.,
distZρ (St (B), A) := sup inf St (b) − aρ → 0 for t → ∞.
b∈B a∈A

If A is an (Z, Zρ )-attractor in B = Z, then it is called the (global) attractor

of St .

If Z = Zρ in the sense of topological spaces this definition coincides with

our previous Definition 2.4.3. It is standard to conclude the uniqueness of
a (Z, Zρ )-attractor, since any second candidate A2 would be attracted to A
as well as attract A towards itself. Using the invariance and closedness we
obtain
distZρ (A, A2 ) = distZρ (A2 , A) = 0,
which implies A = A2 . We assume that the ’uniform’ and the ’spatially
localized’ norm are connected as follows.
(A1) The translations Ty are continuous w.r.t. the norm  · ρ and
u = sup{Ty uρ : y ∈ R}.
For notational convenience we let Zu = (Z,  · ) and Zρ = (Z,  · ρ ) to
indicate that Z is equipped with different norms. However, we mostly omit
the subscript u to denote the topological space Z = Zu . The uniform space
Zu is the original Banach space, whereas Zρ is only a normed space and is
not necessarily complete. In fact, our interest lies precisely in those cases
where the translations Ty are not uniformly bounded in the ρ-norm, which
implies that Zρ is not complete.
We use the abbreviation BZ (r, u0 ) for the closed ball of radius r in Z
with center u0 , i.e., BZ (r, u0 ) = {u ∈ Z : u − u0  ≤ r}. Moreover we let
BZ (r) = BZ (r, 0). As a direct consequence of (A1) we obtain the following
result which will be needed in Theorem 8.3.22.
Lemma 8.3.19. If A ⊂ Z is contained in BZ (r) for some r > 0, then
ρ
A = closureZρ (A) ⊂ BZ (r).

Proof. Let (un )n∈N ∈ B be a sequence with limit u in Zρ . Since un  < r
and since Ty is continuous in Zρ we obtain for fixed y
Ty uρ ≤ Ty un ρ + Ty un − Ty uρ ≤ r + εn
with εn → 0 for n → ∞. Thus, u ≤ r which is the desired result. 
0
Example 8.3.20. Consider Z = Cb,unif and let Zρ be equipped with the
norm
uρ = sup |u(x)/(1 + x2 )|.
x∈R
8.3. The GL equation 287

Convergence in Zu is uniform convergence, whereas convergence in Zρ means

uniform convergence on each compact interval. Consider u : x → tanh(x)
and let A = {Ty u : y ∈ R}. Then A is closed in Zu , but its closure in Zρ is
ρ
A = A ∪ {v−1 , v1 }, where v±1 ≡ ±1.

These two norms allow us to define two different distances between sets,
distZρ and distZu .
Definition 8.3.21. We define
distZρ (b, A) = inf a∈A b − aρ ,
distZu (b, A) = inf a∈A b − a = inf a∈A supy∈R Ty b − Ty aρ .
For both distances we let dist(B, A) = supb∈B dist(b, A) for B ⊂ Z.

With these preparations we can show the existence of an

(Zu , Zρ )-attractor [MS95, Theorem 2.6].
Theorem 8.3.22. Let Zu , Ty , and Zρ be given as above such that (A1)
holds. Moreover, let St be a C0 -semigroup on Zu which is translational
invariant (Ty St = St Ty ) and has a nonempty, bounded, and positively in-
variant set B ⊂ Zu . Assume that the following additional assumptions hold:
(A2) (localized continuity) For each t ≥ 0 the evolution operator St is
continuous from Zρ into itself.
(A3) (compactness) For all subsets B ⊂ B there is a t0 > 0 such that
St0 (B) is pre-compact in Zρ .
Then there exists an unique (Zu , Zρ )-attractor A for St in B. If, addition-
ally, B is an absorbing set, then A is the global attractor. Moreover, A has
the following properties:
(i) A is translationally invariant, i.e., Ty A = A for all y ∈ R.
(ii) Every B ⊂ B which is bounded in Z is attracted to A w.r.t. the distance
distZρ , i.e.,
distZρ (St (B), A) := sup inf St (b) − aρ → 0 for t → ∞.
b∈B a∈A

Proof. W.l.o.g. we can assume that the set B is translational invariant,

i.e., Ty B = B for all y ∈ R. If not, take B1 = ∪y∈R Ty B instead of B. The
attracting set is defined by
* ρ
A= At with At = St (B) .
t≥0

As B is positively invariant, the family (At )t≥0 is a decreasing family, i.e.,

ρ
At1 ⊂ At2 for t1 > t2 . Hence, A ⊂ A0 and A0 = B is bounded by Lemma
8.3.19. Therefore, A is bounded in Zu . Moreover, from (A3) the set At0 , and
hence all At for t ≥ t0 are compact in Zρ . Thus, (At )t≥t0 forms a decreasing
288 8. Three canonical modulation equations

family of compact and nonempty sets in Zρ , and A = ∩t≥t0 At is nonempty

and compact in Zρ . As A is closed in Zρ it is also closed in Zu . This proves
part a) in the Definition 8.3.18 for a (Zu , Zρ )-attractor.
As Ty St (B) = St (Ty B) = St (B) we find by taking the closure in Zρ and
by using the boundedness of Ty the relation Ty At = At . This implies (i).
The more difficult part of the proof is to show that A is in fact an
attractor. It remains to show the time invariance and the attractivity, i.e.,
part b) and c) in Definition 8.3.18.
1) (Time invariance) Let v ∈ St (A), i.e., there is a u = limtn →∞ Stn (un )
in Zρ with un ∈ B such that v = St (u). Because of (A2) (continuity of St in
Zρ ), we have Stn (St (un )) = St (Stn (un )) → St (u) = v as n → ∞ in Zρ . As
St (un ) ∈ B we conclude v ∈ A and hence St (A) ⊂ A.
For the opposite direction (and the attractivity discussed below) the
compactness in Zρ plays a crucial role. Let v ∈ A, then there exist tn → ∞
and un ∈ B with tn < tn+1 and v = limtn →∞ Stn (un ) in Zρ . For any
t > 0 we wish to show v ∈ St (A). From (A3) the set {Stn −t (un ) : tn −
t ≥ t0 (B)} ⊂ St0 (B) (B) is pre-compact in Zρ . Therefore, for a subsequence
wi = Stni −t (uni ) → w in Zρ . Applying the continuous map St we find
v = St (w). As wj ∈ Atni −t for j ≥ i and all At are closed, w lies in all At
and hence in A. Thus, v = St (w) ∈ St (A) and A ⊂ St (A) is proved.
2) (Attractivity in Zρ ) We use the compactness to give a proof by con-
tradiction. Let B ⊂ B be arbitrary. Assume that B is not attracted to
A, then there exist C > 0, sequences tn → ∞ and un ∈ B such that
distZρ (Stn (un ), A) > C > 0 for all n ∈ N. Because of compactness there is
a subsequence such that vi = Stni (uni ) converges in Zρ to w. As w lies in
A this is a contradiction. Hence, we have shown distZρ (St (B), A) → 0 for
t → ∞, for all B ⊂ B. 

Remark 8.3.23. We remark that due to the translation invariance of A we

have convergence in every translated Zρ -norm, too, i.e.,

dist∗Zρ (St (B), A) = sup sup inf Ty St (b) − Ty aρ → 0 for t → ∞.
b∈B y∈R a∈A

We further remark that if Zρ1 and Zρ2 define equivalent norms in Zu via
(A1) then for the attractors constructed via Zρi we have A1 = A2 .

Example 8.3.24. We return to Example 8.3.17 and consider the linear

diffusion equation with Z, Zρ , and A = {u ≡ θ : θ ∈ [−R, R]}. Hence, obvi-
ously all assumptions of Theorem 8.3.22 are met and we have attractivity in
Zρ as well as in the intermediate distance dist∗Zρ . To substantiate the theory
0 ∈ B.
we consider solutions u(t) = St (u0 ) with arbitrary initial conditions u√
Using the explicit solution formula we easily find ∂x u(t)∞ ≤ CR/ t for
8.3. The GL equation 289

all u0 ∈ B. Choosing θ(y) = u(t, y), we obtain

1
Ty u(t) − θ(y)ρ ≤ sup 2
|u(t, x+y) − u(t, y)|
x∈R 1 + x
1 C √
≤ sup 2
√ |x|u0  ∞ ≤ C/ t.
x∈R 1 + x t
√
Thus, we conclude distZρ (St (B), A) = dist∗Zρ (St (B), A) ≤ CR/ t. Hence,
for this linear case we obtain an explicit decay rate towards the attractor.

Due to the smoothing and global existence results for the GL equation
(8.29) we then have
Theorem 8.3.25. For every θ ≥ 0 the GL equation (8.29) has a global
θ , H θ )-attractor A which satisfies attractivity in dist∗ , is translational
(Hul ρ G Hθ ρ
invariant, and invariant under the rotations Rφ : A → eiφ A.

8.3.5. GL rolls and their linearized stability. We now turn to spatially

periodic solutions of the GL equation (8.29), so called rolls, and their the
spectral stability. See also the exercises for further explicit solutions of the
GL equation. The rolls can easily be found by the ansatz
u(x, t) = aei(qx+ωt) , a ∈ C, q ∈ R,
0 1
which yields iωa = −(1 + iα)q 2 + 1 − (1 + iβ)|a|2 a. Separating real and
imaginary part gives the conditions
(8.32) |a|2 = 1 − q 2 and ω = −αq 2 − β|a|2 .
Thus, we have a family of stationary solutions parameterized by the phase
of a ∈ C and the wave number q ∈ R with q 2 < 1. We set a = reiθ with
r2 = 1 − q 2 and write uq,θ for these rolls.
The spectral stability of a fixed uq,θ is determined by the linearization
around uq,θ which reads
∂t v = (1 + iα)∂x2 v + v − (1 + iβ)(2|uq,θ |2 v + u2q,θ v).
Due to the occurrence of u2q,θ this linear system has x-dependent coeffi-
cients. However, due to the S 1 -symmetry of (8.29) the x-dependence can
be removed by using the ansatz
v = vei(qx+ωt) .
We obtain
(8.33) ∂t v = (1 + iα)(∂x2 v + 2iq∂x v) − (1 + iβ)r2 (v + v)
by using (8.32). The real and imaginary part of v = vr + ivi satisfy
          
vr 1 −α 2 vr −α −1 vr vr
(8.34) ∂t = ∂x + 2q − 2r 2
.
vi α 1 vi 1 −α vi βvr
290 8. Three canonical modulation equations

This system is solved by v(x, t) = eλt+ix w with w ∈ C2 , which yields the

algebraic system L(λ, q, )w = 0 with
 2 
− − 2qαi − 2r2 − λ α2 − 2qi
(8.35) L(λ, q, ) = .
−α2 + 2qi − 2r2 β −2 − 2qαi − λ
In order to find non-trivial solutions we need to solve det L(λ, q, ) = 0. For
fixed  we have two eigenvalues λ1,2 () which we order such that Reλ1 () >
Reλ2 ().
We always have λ1 (0) = 0. Eckhaus [Eck65] recognized that depending
on q an instability close to  = 0 can occur. As an example we start with
α = β = 0 where we have

λ1 () = −(r2 + 2 ) + r4 − 4q 2 2 .
To detect the instability close to  = 0 we expand λ1 () = 0 + λ 1 (0) +
1
2 λ1 (0) + O( ) and find
2 3

λ 1 (0) = 0 and λ 1 (0) = −2 + 4q 2 /(1 − q 2 ).

The condition for instability Reλ 1 (0) > 0 then yields
(8.36) q 2 > 1/3.
This instability is called sideband or Eckhaus instability [Eck65]. See Figure
8.6b).
For α, β = 0 the calculations become somewhat more involved. Exercise
8.22 yields λ 1 (0) = 2iq(β − α) and
2
(8.37) λ 1 (0) = − (1 + αβ − q 2 (3 + αβ + 2β 2 )).
1 − q2
Thus, we find Reλ1 () > 0 for small  if the generalized Eckhaus criterion
1 + αβ
(8.38) q2 >
3 + αβ + 2β 2
is satisfied. In particular, there are no stable rolls at all in the so called
Benjamin-Feir unstable regime, see Figure 8.6 c),
(8.39) αβ < −1.

8.3.6. Diffusive stability of Eckhaus-stable rolls. Even for spectrally

stable rolls we have spectrum up to the imaginary axis and so the principle of
linearized instability does not apply. However, it turns out that spectrally
stable equilibria are stable in the nonlinear system w.r.t. small spatially
localized perturbations. Here we give some heuristic arguments why this is
true for the Eckhaus-stable solution A = 1 for the real GL equation, i.e., in
case α = β = 0. Later on in §14 this will be explained in more detail.
8.3. The GL equation 291

(a) (b) (c)

0 Reλ1 3 Re
Reλ1
0 Reλ 2
Reλ2 2 Re
−1 Im
−1 1 Im

−2 0
−2
−1
−3 −3 −2

−2 −1 0 1 2
−2 −1 0 1 2 −2 −1 0 1 2

Figure 8.6. The curves of eigenvalues → λj ( ). (a) α √= β = 0 and

q = 0, stable case. (b) α = β = 0 and q = 0.7 > 1/ 3, Eckhaus
instability. (c) α = −2, β = 1 and q = 0, Benjamin-Feir instability.

We introduce polar coordinates u(x, t) = r(x, t) eiφ(x,t) and find r and φ

to satisfy
∂x r
(8.40) ∂t φ = ∂x2 φ − 2 ∂x φ,
 r 
(8.41) ∂t r = ∂x2 r + r 1−r2 −(∂x φ)2 .
The deviation (φ, s) from the stable equilibrium (1, 0) defined by r = 1 + s
satisfies
∂x s
(8.42) ∂t φ = ∂x2 φ − 2 ∂x φ,
1+s
(8.43) ∂t s = ∂x2 s − 2s − 3s2 − s3 − (∂x φ)2 − s(∂x φ)2 .
Therefore, on the linear level we have
∂t φ = ∂x2 φ and ∂t s = ∂x2 s − 2s,
i.e., φ behaves diffusively, while s is exponentially damped. For spatially
localized initial conditions we have that
φ∗
φ(x, t) = √ e−x /4t + O(1/t)
2

t
for a constant φ∗ ∈ R, such that φ ∼ t−1/2 , ∂t φ ∼ t−3/2 , and ∂x2 φ ∼ t−3/2
for t → ∞. It turns out that s is slaved by φ and that in lowest order s =
−(∂x φ)2 /2 ∼ t−2 for t → ∞. Therefore, (8.42) and (8.43) asymptotically
behave like
(8.44) ∂t φ = ∂x2 φ + O(t−7/2 ),
(8.45) O(t−1 ) = −2s − (∂x φ)2 + O(t−1 )
for t → ∞. Hence, all other terms vanish faster than those that we claimed
to describe the asymptotic behavior. This argument can be made rigorous
292 8. Three canonical modulation equations

and the asymptotics

φ∗
φ(x, t) = √ e−x /4t + O(1/t)
2
and s(x, t) = O(1/t2 )
t
can be established. This has been done first in [CEE92, BKL94]. This
kind of stability will be presented in detail in Chapter 14.

Further Reading. Besides its importance for applications, the NLS equa-
tion also plays a big role inside mathematics. Bourgain (in 1994) and Tao (in
2006) received Fields medals partly for work about the local and global exis-
tence and uniqueness of solutions in spaces of functions with low regularity,
see, e.g., [Bou99, Tao06]. There are numerous books and review papers
about the NLS equation, e.g. [AS81, DJ89, SS99b, BK00, Fib15]. A
major result is that the NLS equation is a completely integrable Hamil-
tonian system. There exist so called inverse scattering schemes, namely
the Ablowitz-Kaup-Newell-Segur (AKNS) scheme and the Zakharov-Shabat
(ZS) scheme which allow to solve the NLS equation explicitly, cf. [AS81,
DJ89]. The so called Birkhoff normal form for the NLS equation in case of
periodic boundary conditions is discussed in [GK14], which also contains a
modern account of the Hamiltonian formalism, of the Poisson bracket, and
of the Lax pair formalism for the NLS equation in terms of the ZS operator.
Besides the motivation by the water wave problem, cf. §8.2.1, the KdV
equation can be derived as an approximation to various other physical sys-
tems, see also [Cri95] and Chapter 12. The question occurs which phenom-
ena of the KdV equation are robust under perturbations. Viewing the KdV
equation as a completely integrable Hamiltonian system, this is analogous
to the questions studied by the Kolmogorov-Arnold-Moser (KAM) theory,
cf. §4.4, and has led to a development of KAM-like results for a number of
different PDEs like the KdV equation. If one considers the KdV equation
with periodic boundary conditions, temporally periodic or quasi-periodic
solutions will persist under small perturbations, cf. [KP03]. The situation
is more complicated and less well-understood for the equation on the whole
line. The local and global existence of low regularity solutions of the KdV
equation and related equations is a very active field of research. As already
said, see, e.g., [Tao06, §4], [LP09, Ch. 7], or [Koc15, Ch. 6] for modern
accounts.
The GL equation as a universal modulation equation occurs at the
end of the 1960s in a number of papers, cf. [NW69, Seg69, dES71],
cf. Chapter 10. There exist many further explicit solutions of (8.29), cf.
[AK02, Mie02]. A famous example are the so called Bekki-Nozaki holes,
see [BN85, vSH92]. There also exist stability results for such rather com-
plicated solutions of the GL equation, cf. [BNSZ14]. Definition 8.3.18 about
8.3. The GL equation 293

attractors on unbounded domains is a special case of that in [BV90, BV92].

The role of the GL equation in superconductivity, especially the occurrence
of vortices, is discussed in [BBH94, HT00, SS07, Ser15].
Exercises
8.1. Find the coefficients β1 , β2 , and β3 in terms of ν1 and ν2 for the transformations
which bring (8.1) into the form (8.2).
8.2. Let u = ur + iui and write the NLS equation as a system of two real equations
for (ur , ui ). Let u = reiφ and find the equations for (r, φ)(x, t).
8.3. Show that, additionally to the usual translation invariance, the NLS equation
is S 1 -, Galilei-, and scaling-invariant, i.e., if u = u(x, t) solves the NLS equation,
then, for all φ, c, η ∈ R,
v(x, t; φ) = u(x, t)eiφ , S 1 -invariance,
2
v(x, t; c) = u(x − ct, t)ei(c t−2cx)/4 , Galilei (or boost) -invariance,
v(x, t; η) = ηu(ηx, η 2 t), scaling-invariance,
solve the NLS equation, too.
8.4. Show that for arbitrary η, c, γ, x0 ∈ R the pulse solutions
√
u(t, x; η, c, γ, x0 ) = 2η sech(η(x − x0 − ct))ei((c −4η )t−2cx+γ)/4
2 2

are exact solutions of the focusing NLS equation.

1 − 4(1 + 2it) it
8.5. a) Show that u(x, t) = e and the Peregrine soliton uP (x, t) =
  1 + 2x2 + 4t2
4(1 + 2it)
1− eit are exact solutions of the focussing NLS.
1 + 4x2 + 4t2
b) Sketch the solutions (e.g., |u|) in the x–t plane. Does u have finite H? Does uP ?
8.6. Show that the curve of solutions t → u(t, u0 ) of the linear Schrödinger equation
is continuous in X = H m (R, C) with m > 1/2 if u0 ∈ X.
8.7. Prove by direct calculation the conservation of the Hamiltonian (8.8) for the
d
NLS equation, i.e., that dt H(u(t)) = 0.
8.8. Consider ∂t u = i∂x4 u, u = u(x, t) ∈ C. Show that:
4
a) This generates a C0 -semigroup u(t) = eit∂x u0 in L2 (R) with u(t)L2 = u0 L2 .
b) If u0 ∈ L1 (R), then u(t)L∞ ≤ Ct−1/4 u0 L1 .
8.9. Use Lemma 8.1.8 to establish a L1 -L∞ estimate for the solutions of the discrete
NLS equation i∂t un = −un+1 + 2un − un−1 with n ∈ Z.
a) Solve the discrete NLS equation by the ansatz un (t) = eikn eiω(k)t with k ∈ R.
b) Write down the solution in Fourier space as u  t)
(k, t) = G(k, u(k, 0).
c) Write down the solution in physical space as un (t) = m∈Z Gn−m (t)um (0),
 π itω(k) −ikn
1
with Gn (t) = 2π −π
e e dk, and estimate sup n∈Z |G n (t)| by separating the
interval [−π, π] in intervals where Lemma 8.1.8 can be applied.
d) Conclude that (un (t))n∈Z ∞ ≤ Ct−1/3 (un (0))n∈Z 1 .
8.10. Consider the traveling wave equation −cv + v  − 3v 2 = D for the KdV
equation. Depending on c, D, discuss the existence of periodic travelling waves,
and of solutions which have limξ→±∞ v(ξ) = D and v(0) < D.
294 8. Three canonical modulation equations


8.11. Prove directly that the energy R
u2 (x, t) dx is a conserved quantity for the
KdV equation.
√
8.12. Find α ∈ R such that u(x, t) = αc sech2 ( c(x − 4ct)) , with arbitrary c > 0
is an exact solution of KdV.
8.13. Let L(t)=L(t)T ∈Rd×d satisfy L(0) = U (t)T L(t)U (t) with U (t)U (t)T = I.
Show that L = L(t) satisfies
∂t L(t)=M (t)L(t)−L(t)M (t)
with
M (t)= − (∂t U (t))U (t)T =U (t)∂t U (t)T = − M (t)T ,
and that as a consequence the eigenvalues of L = L(t) are independent of t.
3 + 4 cosh(2ξ + 24t) + cosh(4ξ)
8.14. Consider the solution u(x, t) = −12 of the
(3 cosh(ξ − 12t) + cosh(3ξ + 12t))2
KdV equation, where ξ = x − 16t. Prove that  for t → ±∞  this solution
 separates
1 1
into two single waves, i.e., u(x, t) ∼ −8 sech 2ξ ∓ log 3 −2 sech η ± log 3
2 2
2 2
with η = x − 4t, where ∼ means asymptotically equal.

8.15. Show that xu + 3t2 u2 dx is conserved for the KdV equation.
8.16. Show that the Kadomtsev-Petviashvili (KP) equation ∂x (∂t u−6u∂x u+∂x3 u)+
3∂y2 u = 0 follows from the Lax pair L = −∂x2 + ∂y + u and M = −4∂x3 + 6u∂x +
x
3ux + 3 ∂y u dξ.
8.17. Let u = ur + iui and rewrite the GL equation (8.29) as a real system for
(ur , ui ).
8.18. Show that the GL equation ∂t u = ν2 ∂x2 u + ν0 u − ν3 |u|2 u, with νj ∈ C,
Reν0 , Reν2 , Reν3 > 0, and x ∈ R, can be rescaled to the standard form
∂t u = (1 + iα)∂x2 u + u − (1 + iβ)|u|2 u, α, β ∈ R.
8.19. a) Show local existence of solutions of the GL equation (8.29) for u0 ∈ H 1 (R).
b) Derive a priori estimates to show global existence. Do solutions stay bounded
in H 1 (R)?
8.20. Show that the space of uniformly continuous functions can equivalently be
0
characterized by Cb,unif (R, R) = { u ∈ Cb0 (R, R) : Ty u − uCb0 → 0 as y → 0 }.
8.21. Use the variation of constant formula and smoothing properties of the GL
semigroup to establish a bound lim supt→∞ u(t)Hul
1 = O(|β|) for |β| → ∞ from
 √
lim supt→∞ u(t)L2ul ≤ 4π 2 for the solutions of the GL equation.
8.22. Derive the expansion (8.37).
8.23. (Benjamin–Feir instability for the NLS equation) Discuss the sta-
bility of the solution a(t) = a0 e−ia0 t of the NLS i∂t u − ∂x2 u − |u|2 u = 0. For
2

the linearization around a, make an ansatz u(x, t) = a(t)(1 + b(x, t)) with b(x, t) =
b1 ei(ωt+kx) +b2 ei(ωt−kx) , and derive an algebraic system for the coefficients b1,2 ∈ C.
Show that this system possesses non-trivial solutions if ω 2 = 2k2 (a20 − k2 ), and
from this derive that a is unstable w.r.t. perturbations with wave numbers k with
k2 < a20 .
 Chapter 9

Reaction-Diffusion
systems

In §7.1 we considered with the KPP and the Allen-Cahn equation the sim-
plest examples of scalar nonlinear reaction-diffusion equations. In general,
chemical reactions involve a number of different species, and so we consider
in this section reaction-diffusion systems of the form
(9.1) ∂t u = DΔu + f (u),
with t ≥ 0, x ∈ Rn , u(x, t) ∈ RN , where f : RN → RN a smooth func-
tion, and D ∈ RN ×N is a positive (semi-)definite diffusion matrix. Of-
ten D is diagonal with strictly positive diagonal elements dj > 0, i.e.,
D = diag(d1 , . . . , dN ).
Famous examples of two-component reaction-diffusion systems, where
a, b, d, . . . denote parameters, are:
The Schnakenberg model [Sch79]
(9.2) ∂t u = Δu − u + u2 v, ∂t v = dΔv + b − u2 v.
The Brusselator [PL68]
(9.3) ∂t u = Δu + a − (b + 1)u + u2 v, ∂t v = dΔv + bu − u2 v.
The Lengyel-Epstein model [LE91]
4uv uv
(9.4) ∂t u = Δu + a − u − , ∂t v = dΔv + b(u − ).
1 + u2 1 + u2
The Gray-Scott model [GS83]
(9.5) ∂t u = Δu − uv 2 + f (1 − u), ∂t v = dΔv + uv 2 − (f + k)v.

295
296 9. Reaction-Diffusion systems

The Gierer-Meinhard model [GM72],

(9.6) ∂t u = Δu + a − u + u2 /v, ∂t v = dΔv + u2 − v.

The FitzHugh-Nagumo model [Fit69]

(9.7) ∂t u = Δu + u(1 − u)(u − a) − v, ∂t v = dΔv + bu − cv.

While models (9.2)-(9.6) essentially are all from chemistry, (9.7) models
the transmission of nerve impulses in some giant squid axon. Moreover,
the Kolmogorov models from ecology, i.e., the 2-species interactions from
Example 2.6.5, reappear as reaction-diffusion systems when combined with
diffusion of one or both of the species.
Typical questions about reaction-diffusion systems are again the local
and global existence of solutions in a suitable phase space, and the existence
and stability of special solutions such as pulses and fronts. Concerning
global existence, the crucial difference compared to scalar equations is that
the maximum principle is no longer valid for systems. However, it can often
be replaced by the method of invariant regions. Similarly, the existence of
special solutions becomes more complicated due to the fact that the traveling
wave system in general can no longer be written as a two-dimensional first
order system, such that phase plane methods are no longer available but
have to be replaced by more sophisticated methods, which often use so
called trapping regions.
In §9.1 we start with modeling aspects and the local existence and
uniqueness theory for reaction-diffusion systems, and discuss quadratic au-
tocatalysis as an example of a two-dimensional system of the form (9.1). It
turns out that when d1 = d2 such systems can be rescaled to the KPP equa-
tion from §7.1. We explain the concept of invariant regions which allows
to prove global existence and uniqueness of solutions for many reaction-
diffusion systems.
The FitzHugh-Nagumo system (9.7) and the Gray-Scott system (9.5)
are used as examples in §9.2 to discuss, on a heuristic level, some typical
phenomena of travelling pulses and pattern formation in reaction-diffusion
systems. However, our main objective is the Turing instability of spatially
homogeneous states, also called diffusion driven instability [Tur52, Mur89],
which we consider in §9.3. This is a famous explanation for morphogenesis,
which refers to pattern formation during growth in biological system. The
bifurcating pattern very often can be described by the GL approximation,
which will be studied in Part IV.
9.1. Modeling, and existence and uniqueness 297

9.1. Modeling, and existence and uniqueness

Assume that a chemical system consists of d chemical species α1 , . . . , αd
which react according to the law of mass action,
λ1 α1 + . . . + λd αd → μ1 α1 + . . . + μd αd .
The λi , μi are called stoichiometric coefficients. The reaction proceeds at
rate
ṙ = k[α1 ]λ1 · · · [αd ]λd
where ui = [αi ] is the dimensionless concentration of αi , k is a constant
which is often related to temperature T by Arrhenius law and which is
proportional to exp(−E/(RT )), where E is the so called activation energy
of the reaction, and R is the universal gas constant. Conservation of mass
yields
(9.8) u̇i = (μi − λi )ṙ.
An everyday example is 2H2 + O2 → 2H2 O. With α1 = H2 , α2 = O2 , α3 =
H2 O this yields λ = (2, 1, 0), μ = (0, 0, 2), and the reaction rate is propor-
tional to [H2 ]2 [O2 ]. Thus, the ODE system for u1 = [H2 ], u2 = [O2 ] and
u3 = [H2 O] reads
u̇1 = −2ku21 u2 , u̇2 = −ku21 u2 , u̇3 = 2ku21 u2 .
Clearly, the last equation decouples from the first two and is usually dropped,
since the product H2 O no longer takes part in the reaction, i.e., if u1 , u2
are known, then u̇3 = 2ku21 u2 can be solved by simple time integration.
Alternatively, mass conservation can be used.

9.1.1. Local existence and uniquenss. As explained above, in spatially

extended problems, i.e. x ∈ Rn , this yields systems of the form (9.1). Before
embarking on specific examples, we note the following local existence result,
where as phase space we always choose
0
X = [Cb,unif (R)]N , with norm u0 X = sup sup |uj (x)|.
j=1,...,N x∈R

Moreover, we restrict to diagonal diffusion matrices D = diag(d1 , . . . , dN ),

with di ≥ 0 the diffusion rate of the ith spezies, though the generalization
to general positive (semi-)definite D is straightforward.
Theorem 9.1.1. Let u0 ∈ X. Then there exists a t0 > 0 depending only on
u0 X such that (9.1) has a unique (mild) solution u ∈ C([0, t0 ], X).

Proof. The proof is very similar to the proof of Theorem 7.1.7. First we fix
a C1 > 0 and define the semigroup etDΔ by
2 2
etDΔ u = diag(etd1 ∂x u1 , . . . , etdN ∂x ud ),
298 9. Reaction-Diffusion systems

 − x−y
where etdj ∂x uj = √
2 1 2
e 4dj t uj (y) dy if dj > 0, cf. (7.6), and etdj ∂x uj =
4πdj t R
uj if dj = 0. Then etDΔ is a C0 -semigroup in X and etDΔ uX ≤ uX .
Next we use the variation of constant formula and show that
 t0
F (u) = etDΔ u0 + e(t−s)DΔ f (u(s)) ds
0
maps
M = C 0 ([0, t0 ], {u ∈ X : u − etDΔ u0 X ≤ C1 }),
equipped with the norm
uM = sup u(t)X ,
0≤t≤t0
into itself and is a contraction in M if t0 > 0 is sufficiently small. 
Remark 9.1.2. If dj > 0 for j = 1, . . . , N , then the solution u(·, t) becomes
analytic for t > 0 and hence a classical solution. If one or more of the dj
are zero, then in general the solutions only stay as smooth as the initial
conditions.

9.1.2. Quadratic autocatalysis. This section follows closely [BK00].

The reaction A+B → 2B, with reaction rate k[A][B], is called quadratic au-
tocatalysis, since the greater the concentration of the catalyst B, the faster
B is produced, i.e., B catalyses its own production, while it preys on A.
With a = [A] and b = [B] this yields
(9.9) ȧ = −kab, ḃ = kab,
which, augmented by diffusion, will be studied below. By mass conservation,
dt (a+b) = 0, hence a+b = a0 +b0 and eliminating a yields ḃ = kb(a0 +b0 −b).
d

Separation of variables yields

a0 (a0 + b0 ) b0 (a0 + b0 )ek(a0 +b0 )t
a(t) = , b(t) = ,
a0 + b0 ek(a0 +b0 )t a0 + b0 ek(a0 +b0 )t
such that a → 0 and b → a0 + b0 as t → ∞. For all b0 > 0, all of A is
converted into B as t → ∞. However, with diffusion the reduction to a scalar
equation only works if the diffusion constants of a and b are equal, and thus
the system form (9.9) must be kept. After suitable non-dimensionalization,
the one-dimensional reaction-diffusion system associated to (9.9) reads
(9.10) ∂t a = ∂x2 a − ab, ∂t b = d∂x2 b + ab,
where d = DB /DA is the ratio of the diffusion constants. With u = (a, b),
(9.10) is of the form (9.1) with D = diag(1, d) and f (u) = (−u1 u2 , u1 u2 ).
Local existence and uniqueness of solutions in X = Cb0 × Cb0 follows from
Theorem 9.1.1. To obtain global existence we need to bound u(t)X . Un-
fortunately, the maximum principle is no longer available, and thus it must
be substituted by different methods which we postpone to §9.1.3.
9.1. Modeling, and existence and uniqueness 299

We are interested in traveling waves for (9.10), where first we assume

that A and B diffuse with the same rate, such that d = 1. The ansatz
(a, b)(x, t) = (v1 (x − ct), v2 (x − ct)) for travelling waves yields
(9.11) ∂ξ2 v1 + c∂ξ v1 − v1 v2 = 0, ∂ξ2 v2 + c∂ξ v2 + v1 v2 = 0,
hence, with w = v1 + v2 , we find ∂ξ2 w + c∂ξ w = 0, which can be integrated
to w(ξ) = c0 + c1 e−cξ where c0 , c1 are constants of integration. Looking for
a front with w → 1 as ξ → ∞ and w bounded for ξ → −∞ we obtain c0 = 1
and c1 = 0, hence w ≡ 1, and (9.11) reduces to the single equation
∂ξ v2 + c∂ξ v2 + v2 (1 − v2 ) = 0
which is the traveling wave equation for the KPP equation and the analysis
of §7.1 applies. Thus, there are fronts for (9.11) connecting (v1 , v2 ) = (1, 0)
with (v1 , v2 ) = (0, 1) for all c ≥ 2.
If the diffusion constants of A and B are not equal, then the travelling
wave ansatz (a, b)(x, t) = (v1 (x − ct), v2 (x − ct)) yields
(9.12) ∂ξ2 v1 + c∂ξ v1 − v1 v2 = 0, d∂ξ2 v2 + c∂ξ v2 + v1 v2 = 0.
As before we assume that (a, b) → (1, 0) as ξ → ∞ (unreacted state far
ahead of the wave) and (a, b) → (0, 1) as ξ → −∞ (fully reacted state far
behind the wave). We may still add v1 and v2 and integrate once to obtain
∂ξ v1 + d∂ξ v2 + c(v1 + v2 ) = c,
but further integration and thus reduction to the KPP equation is no longer
possible. Instead, with v3 = ∂ξ v2 we obtain the first order system
(9.13) ∂ξ v1 = c(1 − v1 − v2 ) − dv3 , ∂ξ v2 = v3 , ∂ξ v3 = −(v1 v2 + cv3 )/d.
Thus, we are no longer able to use phase plane analysis to discuss the ex-
istence of a heteroclinic orbit for (9.13) connecting (1, 0, 0) with (0, 1, 0).
However, essentially the same method as for the KPP equation works. We
first consider the linearization at the fixed points and find that at (0, 1, 0)
(1)
we have a one-dimensional unstable manifold Mu , and at (1, 0, 0) a three-
(2) (1)
dimensional stable manifold Ms , such that we may expect that Mu and
(2) √
Ms intersect. Moreover, all eigenvalues at (1, 0, 0) are real for c ≥ 2 d
which hence yields the minimal front speed to obtain a monotonic front. Fi-
(1) (2)
nally, in order to show that Mu and Ms intersect we consider in Exercise
9.2 the wedge shaped trapping region
R = {(v1 , v2 , v3 ) : 0 ≤ v1 ≤ 1, 0 ≤ v2 ≤ 1, −cv2 /(2d) ≤ v3 ≤ 0},
and conclude on the existence of fronts for (9.10).
Thus, given initial conditions for (9.10) of the form a0 ≡ 1 (or some
constant) and b0 ≥ 0 some localized √ perturbation of 0 we expect a pair of
fronts with the minimal speed c = 2 d to emerge, one propagating to the
300 9. Reaction-Diffusion systems

left, one propagating to the right. Numerical simulations of (9.10) show that
this is indeed the case, see Figure 9.1. Note that the critical front speed only
depends on the diffusion coefficient of the autocatalyst B, which moreover
determines the steepness of the front. We remark that the shape of the front
vhet (ξ) can be obtained by considering asymptotic limits, in particular the
limit 0 < d  1, but for this and more analysis we refer to the literature,
e.g., [BK00].

Figure 9.1. Fronts (b-component) for (9.10) with d=1 (left) and
d=5 (right) for the initial condition (u0 , v0 )(x) = (1, 0.2 sech(x/2)).

9.1.3. Invariant regions. For reaction-diffusion systems the maximum

principle is no longer available. However, to show global existence in a sim-
ilar way as for the KPP equation it can often be substituted by the method
of contracting sets for the reaction term. This section follows [Smo94, §14],
which treats the problem on general (bounded) spatial domains Ω ⊂ Rm ,
but since we restrict to the case x ∈ R our presentation is somewhat shorter.
Definition 9.1.3. A closed bounded non-empty set B ⊂ RN is called con-
tracting for the vector field f if f (u) points inward on ∂B, i.e., if f (u)·ν(u) <
0 for u ∈ ∂B, where ν is the outer normal to B.
Contracting sets for the reaction f are often the first step to construct
invariant regions for the reaction-diffusion system (9.1).
Definition 9.1.4. A closed bounded non-empty set Σ ⊂ RN is called pos-
itively invariant region for (9.1), if u(x, 0) ∈ Σ for all x ∈ R implies
u(x, t) ∈ Σ for all t > 0 and x ∈ R.
The invariant regions we shall consider will be defined by functions gi ∈
C 2 (RN , R), namely
(9.14) i=1 {u ∈ R : gi (u) ≤ 0}.
Σ = ∩m N

Theorem 9.1.5. For u0 ∈ ∂Σ with gi (u0 ) = 0 for some i ∈ {1, . . . , N },

assume that the following holds.
9.1. Modeling, and existence and uniqueness 301

(1) ∂u gi (u0 ) is a left eigenvector of D, i.e., ∂u gi (u0 )D = μ∂u gi (u0 ) for

some μ ∈ R.
(2) If ∂u gi (u0 )D = μ∂u gi (u0 ) with μ = 0, then gi is quasi-convex at u0 ,
i.e., if ∂u gi (u0 ) · η = 0, then
∂u2 gi (v0 )η, η ≥ 0.
(3) ∂u gi (u0 ) · f (u0 ) < 0.
Then Σ defined by (9.14) is invariant for (9.1).

The idea of the proof is as for the maximum principle in §7.1.4. For a
d
solution u coming to the boundary for (x0 , t0 ) we have dt gi (u(x0 , t0 )) < 0
such that the solution goes back to the interior of Σ. 
Remark 9.1.6. Since D is diagonal, any left eigenvector ∂u g of D is also
a right eigenvector. If D = diag(d1 , . . . , dN ) with di = dj for i = j,
then the eigenvectors are obviously of the form e1 := (1, 0, . . . , 0), e2 :=
(0, 1, 0, . . . , 0), . . ., eN := (0, 0, . . . , 0, 1), which shows that each gi actually is
a function of only one variable. Essentially, this means that Σ is a rectangle
and that we consider the system in a “decoupled way”. If, e.g., di = dj for
exactly two i = j, then we have a two-dimensional eigenspace span{ei , ej }
and hence also more freedom for gi and gj .

From Remark 9.1.6 we have the following corollary.

Corollary 9.1.7. A rectangle
i=1 {ai ≤ ui ≤ bi }
Σ = ∩N
is invariant if f (u) points strictly inward on ∂Σ, i.e., if (3) in Theorem
9.1.5 holds.

Proof. For notational simplicity we write gi = g. We have, for u = u(x, t),

d
g(u) = ∂u g(u) · ∂t u = ∂u g(u) · (D∂x2 u + f (u)).
dt
By (1) we have, at u = u(x0 , t0 ), ∂u g(u)D = μ∂u g(u), hence
d
(9.15) g(u) = ∂u g(u) · (μ∂x2 u + f (u)).
dt
Thus, let μ > 0 and let h(x) = g(u(x, t0 )). Then h(x0 ) = h (x0 ) = 0 and
h (x0 ) ≤ 0 by assumption on x0 . Therefore, 0 ≥ h (x0 ) =
∂u2 g∂x u, ∂x u +
∂u g∂x2 u. By (2) we have
∂u2 g∂x u, ∂x u ≥ 0, thus ∂u g∂x2 u ≤ 0, hence again
dt g(u) ≤ ∂u g(u)f (u) < 0. 
d

Remark 9.1.8. Often, only rectangles Σ of the form (9.14) can be con-
structed such that f does not point outward on ∂Σ, i.e., f may vanish on
∂Σ or may be tangential to ∂Σ. Then, instead of (3) in Theorem 9.1.5 we
only have ∂u gi (u0 ) · f (u0 ) ≤ 0. However, under mild technical assumptions
302 9. Reaction-Diffusion systems

on f , which are fulfilled for smooth f , the assertion of Theorem 9.1.5 remains
true, i.e., Σ is positively invariant, see [Smo94, Thm 14.11].

9.2. Two classical examples

We discuss two specific examples of reaction-diffusion systems. The
FitzHugh-Nagumo (FHN) system models transmission of nerve pulses in
some giant squid axon. It is a simplification of the so called Hodgkin-
Huxley model for the transmission of a nerve impulse though the axon of a
giant squid’s neuron, see [BK00] for the modeling. Therefore, additional to
global existence, we shall in particular be interested in the existence of pulse
solutions. The Gray-Scott model in §9.2.2 is a famous model for cubic au-
tocatalysis which shows a multitude of different pattern. The details of the
analysis of both, the FHN pulses and the Gray-Scott pattern are rather dif-
ficult. Therefore, here we only explain the basic phenomena using heuristic
arguments and illustrations by numerical simulations.

9.2.1. The FitzHugh-Nagumo system. After suitable non-dimensiona-

lization, the FHN system reads
∂t u = ∂x2 u + f (u) − v,
(9.16)
∂t v = d∂x2 v + bu − cv,
where f (u) = u(1 − u)(u − a), with parameters a ∈ R, b, c > 0 and d ≥ 0,
where u = u(x, t) is the voltage in the axon, and v = v(x, t) is a so called
recovery variable. From the modeling point of view the case d = 0 where
only u diffuses makes sense. We also remark that here the modeling does
not give the constraints u, v ≥ 0.
The phase portrait for the ODE part
(9.17) u̇ = f (u) − v, v̇ = bu − cv,
of (9.16) is shown for some specific parameters in Figure 9.2. This is the
v bu−cv=0

a 1 u

f(u)−v=0

Figure 9.2. A contracting rectangle for (9.16)

9.2. Two classical examples 303

so called excitable case with exactly one stable fixed point. In the so called
oscillatory case the unique fixed point is unstable, while in the so called
bistable case there are three fixed points, two stable and one unstable (see
below).
In any case, there are large contracting sets, as illustrated by the rec-
tangle in Figure 9.2. Thus, using Theorem 9.1.1, and Corollary 9.1.7 we
conclude that for arbitrary (u0 , v0 ) ∈ X = [Cb,unif
0 (R)]2 we have a global
solution to (9.16).
From the modeling point of view, in suitable parameter regimes, (9.16)
should possess pulse solutions, i.e., solutions of the form
u, v)(x − ct)
(u, v)(x, t) = (
u, v)∗ as ξ → ±∞, where (
u, v)(ξ) → (
with ( u, v)∗ is a fixed point of (9.17).
These pulses are found in the regime b, c and d small. Then (9.17) is
often rescaled to the form
(9.18) εu̇ = f (u) − v, v̇ = u − γv,
where 0 < ε  1 is a small parameter. Since u̇ = 1ε (f (u) − v) in (9.18) is
large except if f (u) − v = 0, the u-dynamics is called fast in contrast to the
slow dynamics in v.
Excitable dynamics. If 0 < a < 1/2 and γ < 4/(a − 1)2 , then (u, v) =
(0, 0) is the only fixed point of (9.18). It is asymptotically stable, with
eigenvalues λ1 = −a/ε and λ2 = −1. However, as is evident from the phase
portrait in Figure 9.3, a finite perturbation with, e.g., u > 0 and v < 0 may
lead to a large excursion with u > 1 before the orbit returns to (0, 0) from
the left. This is called excitable behavior. In particular, this happens for
ε > 0 sufficiently small.
To understand this behavior we follow [Kee88, §12]. We first set ε = 0,
which formally yields
(9.19) v = f (u), v̇ = u − γv.
This is an implicit ODE for v. The algebraic equation v = f (u) has three
branches of solutions, namely
u = g− (v), u = g0 (v) and u = g+ (v),
see Figure 9.3. The lower/upper branches g−√(v) and g+ (v) exist only for
u < u∗− and u > u∗+ , where u∗± = 12 (a + 1 ± a2 − a + 1), and the middle
branch g0 (v) exists for u∗− ≤ u ≤ u∗+ . Equivalently, g− (v) and g+ (v) exist
for v > v− ∗ := f (u∗ ) and v < v ∗ := f (u∗ ), respectively. Thus, depending
− + +
on the branch, (9.19) has three different meanings, namely v̇ = g± (v) − v or
v̇ = g0 (v)−v. On the middle and upper branches, v increases monotonically,
while g− (v) − v changes sign at v = 0.
304 9. Reaction-Diffusion systems

0.2
f1=0
0.15 f =0
2
v bu−cv=0
g−(v) 0.1

0.05

g0(v) 0
a 1 u
f(u)−v=0 −0.05

−0.1
g+(v) 0 0.5 1

Figure 9.3. The mechanism of excitation for (9.16), explained

in the text, and the phase portrait for (9.18) for (a, γ, ε) =
(0.25, 1, 0.01)

On the other hand, in (9.18) we can make the change of variables

(U, V )(τ ) = (u, v)(ετ ) to obtain

(9.20) U̇ = f (U ) − V, V̇ = ε(U − V ).

Now setting ε = 0 the slow dynamics in V vanish, i.e., V = V0 some constant.

Then U̇ = f (U ) − V0 and given U0 = g0 (V0 ), U (τ ) converges to a stable root
of f (U ) − V0 , i.e.,

U (τ ) → g− (V0 ), if U0 < g− (V0 ),

(9.21)
U (τ ) → g+ (V0 ), if U0 > g+ (V0 ).

Given (u0 , v0 ) we can now (heuristically) piece together the solution of (9.18)
using the above two different scaling limits. Assume that u > 0 and v < v− ∗.

Then v first stays essentially constant, and u rapidly relaxes to u+ (v0 ),

according to (9.21). Then the slow dynamics in v kick in, i.e., v slowly
increases on the upper branch. This branch ceases to exist when v reaches
∗ , and the fast dynamics in u take over and rapidly take u to u (v ∗ ). After
v+ − +
this, the solution slowly relaxes along the lower branch to the equilibrium
(u, v) = (0, 0).
This was a formal discussion. Nevertheless, it agrees very well with the
phase portrait in Figure 9.4, and it can be made rigorous using so called
singular perturbation theory. Due to the rapid excitation followed by a slow
relaxation, this kind of dynamics is also called relaxation dynamics. If the
initial perturbation of (u, v) = (0, 0) is sufficiently small, e.g., u < a and
v = 0, then the solution directly relaxes back to (u, v) = (0, 0). This is
called sub-threshold dynamics.
9.2. Two classical examples 305

Before we discuss how this is related to the existence of traveling pulses

for the FHN PDE (9.16), we briefly explain two other regimes encountered
in the FHN ODE (9.18).

Remark 9.2.1. Bistable and oscillatory dynamics. If γ > 4/(a − 1)2

in (9.18), then the nullclines f (u) − v = 0 and u − γv intersect in three
points, namely5P = (0, 0), and Q = (u− , u− /γ) and R = (u+ , u+ /γ) with
2
(1+a)
u± = 1+a 2 ± 4 − (a + 1/γ). Now P and R are stable while Q is
unstable, see Figure 9.5. The two new fixed points are created in a saddle
(Q)-node (R) bifurcation at γ = 4/(a − 1)2 . The phase portrait shows that
for (u0 , v0 ) a sufficiently large perturbation of P the solution converges to R,
while for (u0 , v0 ) a sufficiently large perturbation of R the solution converges
to P .
If a < 0 and γ > 4/(a − 1)2 , then we again have the unique fixed point
(u, v) = (0, 0), but now it lies on the unstable middle branch. The above
(singular perturbation) analysis again applies and yields so called relaxation
oscillations.

0.2 0.3 u(t)

f1=0 f1=0
0.25 1
0.15 f =0 f =0
2 2
0.2
0.1
0.15 0.5

0.05 0.1

0.05 0
0
0
−0.05
−0.05 −0.5
−0.1 −0.1 0 1 2 3 4 5
0 0.5 1 −0.5 0 0.5 1

Figure 9.4. The bistable (a, γ, ε) = (0.25, 8, 0.01) and the oscilla-
tory (a, γ, ε) = (−0.25, 1, 0.01) case: phase portrait and relaxation
oscillations.

Travelling pulses for the FHN system. In the excitable regime we may
expect pulse solutions for the FHN system. The heuristic reason is as fol-
lows. Assume that we start with a (suitable, sufficiently large) perturbation
of (u, v) ≡ (0, 0), localized near some x0 . Then, locally, the ODE dynamics
wants to run the excitation. Meanwhile, diffusion pulls the solution beyond
the threshold for excitation also in a neighborhood of the primary perturba-
tion, and the process repeats. Thus, we may expect one or more (depending
on the details of the IC) excitation pulses to emerge, see Figure 9.5 for two
examples. Moreover, the pulses are asymptotically stable in the numerics,
with a large basin of attraction as we do not need any fine tuning of the
initial data.
306 9. Reaction-Diffusion systems

(a) (b) (c)

1
u
v

0.8

0.6
1200
1500
0.4
1
1000 0.5 800
1
0.2 0
0.5

0 500 400
0
-300
-200
-100
0
-300 -200 -100 0 -0.2 100
0 100 200 200 0
300 -300 -200 -100 0 100 200 300 300

Figure 9.5. Numerical simulation of (9.16) on a large domain with

periodic boundary conditions. (a) emergence of a pulse in the ex-
citable regime, IC (u0 , v0 )(x) = (sech(x/4), − sech(x/2) sin(x/2)),
parameters (a, b, c, d) = (1/4, 0.00075, 0.005, 0.1). (b) (u, v)|t=1775 .
(c) for, e.g., (u0 , v0 )(x) = (sech(x/10), − sech(x/5) sin(x/5)) we
obtain two counter-propagating pulses, which annihilate upon col-
lision. This behavior is typical but not the only possibility for
pulses in reaction-diffusion systems.

The analysis for the existence and in particular for the stability of the
FHN pulses is rather difficult, and we refer to the literature, see, e.g.,
[Jon84, Kue15].

9.2.2. The Gray Scott model. The Gray Scott model [GS83] is a famous
model for cubic autocatalysis of two species according to A + 2B → 3B, rate
k1 ab2 , where A is fed into the system at rate kf and B decays with rate k2 b,
and where the diffusion coefficients are DA and DB . After suitable non-
dimensionalization the system reads
∂t u = ∂x2 u − uv 2 + f (1 − u),
(9.22)
∂t v = d∂x2 v + uv 2 − (f + k)v,
with typically 0 < d < 1 in applications.
The nullclines of the ODE for (9.22) are given by
g1 = 0 : u = f /(f + v 2 ), g2 = 0 : u = (f + k)/v or v = 0.
We always have the stable fixed point P = (1, 0). For f > 4(f + k)2 there
occurs a saddle-node bifurcation with saddle Q = ((f + k)/v− , v− ) and
unstable node R = ((f + k)/v+ , v+ ) with

f f2
v± = ± − f.
2(f + k) 4(f + k)2
The radicand is positive for f > 4(f + k)2 , or equivalently k < −f +

f /4 =: fsn (f ). However, for k < kHopf (f ), R regains stability by a Hopf
9.3. The Turing instability 307

bifurcation such that R is surrounded by an unstable periodic orbit for

ksn (f ) > k > kHopf (f ). At (k, f ) = (1/16, 1/16) the two curves ksn (f ) and
kHopf (f ) meet. This is called a Bogdanov-Takens or co-dimension 2-point,
since in its neighborhood two parameters are needed to describe all possible
bifurcations. See Figure 9.6 for some ODE phase portraits.

v 0.5
0.5 f1=0
f1=0
f =0
f =0 0.4 2
0.4 2
f(u,v)=0
0.3 0.3
R
f<0 0.2 0.2

Q g>0 g(u,v)=0 0.1 0.1

g<0
0
p u 0 0.2 0.4 0.6 0.8 1 0
0 0.2 0.4 0.6 0.8 1

Figure 9.6. The schematic phase portrait for the Gray-Scott

Sytem (9.22) right after the saddle (Q)-node (R) bifurcation in
the first quadrant, and the phase portrait for (f, k) = (0.023, 0.05)
(ksn (f ) > k > kHopf (f )) and for (f, k) = (0.04, 0.05) (kHopf > k).

Thus, a (very rough) first characterization of the Gray-Scott system

(9.22) is: (a) excitable for k > kHopf (f ) and (b) bistable for k < kHopf (f ).
Thus, we again may expect pulses and fronts in cases (a) and (b), respec-
tively. Again, the actual proof of existence of these traveling waves is beyond
the scope of this book, not to mention their stability analysis. Therefore, we
content ourselves with a small sample of numerical simulations, see Figure
9.7. Some very interesting numerical simulations of the Gray-Scott system
leading to Sierpinsky-Gaskets of pulses can be found in [HO00], see also
[HO98].
Remark 9.2.2. a) From the phase portrait we see that the Gray Scott
model does not have contracting rectangles. Thus, we cannot obtain global
existence with the methods from §9.1.3. Nevertheless, global existence holds
and can be established using somewhat more involved (energy-type-) meth-
ods, see, e.g., [FMW92] for rather general results.
b) For specific parameter values in (9.22) some fronts and pulses can be
obtained explicitly. For this see [HPT00].

9.3. The Turing instability

The Turing instability [Tur52, Mur89] gives a mechanism for pattern for-
mation in a reaction-diffusion models. Our starting point is a system of
the form (9.1), where the reaction term f = fμ depends smoothly on a
parameter (vector) μ ∈ Rp , where w.l.o.g. we may restrict to p = 1. We
308 9. Reaction-Diffusion systems

7500
3000 1500
0.5 0.4
0.25 5000
2000 0.2 1000
0 0.5
0
0.25
2500
1000 500 0
-200
-100 -300
0 -200
0 -100
100 0 0
100
200 0 300 0 200

Figure 9.7. Numerical simulations of (9.22) on a large do-

main with periodic boundary conditions. (a) two counter-
propagating pulses that annihilate upon collision (d = 0.5, (f, k) =
(0.023, 0.05)). (b) Two fronts in the bistable regime (f, k) =
(0.04, 0.04). (c) self replicating pattern (SRP) (d = 0.5, (f, k) =
(0.038, 0.06)).

assume that at some fixed μ the system has a stable spatially homogeneous
equilibrium u∗ , i.e., u∗ is a stable fixed point of the ODE u̇ = f (u).
Intuitively one would expect that diffusion cannot destabilize a homoge-
nous state u∗ . However, Turing [Tur52] recognized that this intuition is
wrong if the diffusion matrix D is not a scalar multiple of the identity, and
it turns out that u0 may be unstable w.r.t. harmonic waves with some wave-
number kc = 0. An attempt for an intuitive explanation using fire as the
activator with fast diffusion and “sweating grasshoppers” as inhibitor with
slow diffusion is made in [Mur89].
The general situation is as follows. The linearization
∂t v = DΔv + fμ (u∗ )v
of (9.1) around u∗ has solutions of the form
v(x, t) = eikx+λ(k)t v(k) + c.c.,
where the eigenvalue λ(k) ∈ C and the vector v(k) ∈ CN are determined
from the N × N eigenvalue problem
(fμ (u0 ) − k 2 D)
v (k) = λ(k)
v(k),
i.e., there exist N curves k → λj (k, μ) with j = 1, . . . , N , which we order
such that Reλj (k, μ) ≥ Reλj+1 (k, μ).
The homogeneous state u∗ is unstable if there exists a k0 ∈ R with
Reλ1 (k0 ) > 0. The instability appears as some parameter μ of the system
is varied. For instance Reλ1 (k) < 0 for all k ∈ R if μ < μ0 , while at μ = μ0
we have Reλ1 (k0 ) = 0 for some k0 > 0 and Reλ1 (k) < 0 for k = ±k0 .
Finally, for μ > μ0 , Reλ1 (k) > 0 for all k in two unstable bands, i.e.,
±k ∈ (k− (ε), k+ (ε)), with ε2 = μ − μc .
9.3. The Turing instability 309

Thus, at criticality, i.e., at μ = μc , spatially periodic solutions may

bifurcate, which is called pattern formation. It is called a) (proper) Tur-
ing bifurcation if Im(λ1 (kc , μc )) = 0 and b) Turing-Hopf bifurcation if
Im(λ1 (kc , μc )) = 0. However, the instability may also appear at the wave
number k = 0, i.e., the homogeneous solutions u = u∗ itself may become
unstable as a fixed point of the reaction u̇ = fμ (u) as the parameter μ is
varied. It is called d) Hopf bifurcation if Im(λ1 (0, μc )) = 0.

a) Turing bifurcation b) Turing-Hopf bifurcation

Im λ 1
Im λ 1
k=k c
k=k c

Re λ 1 Re λ 1

c) stationary long wave bifurcation d) Hopf bifurcation

Im 1
Im λ 1

Re λ 1
Re λ1,2
Im λ 2
Figure 9.8. The four different instability scenarios

Example 9.3.1. We consider the so called Schnakenberg model [Sch79,

Mur89]
∂t u = ∂x2 u − u + u2 v,
(9.23)
∂t v = d∂x2 v + b − u2 v,
with parameter b > 0 and the ratio 0 < d = dV /dU of the diffusion constants.
Below we shall further assume that d > 1. To start the bifurcation analysis,
we first need to determine the linearized stability of the unique homogeneous
fixed point
(u∗ , v ∗ ) = (b, 1/b)
of (9.23). The new coordinates (u, v) = (u∗ , v ∗ ) + ( u, v) yield
   
u u
(9.24) ∂t = L(∂x ) u, v),
+ g(
v v
310 9. Reaction-Diffusion systems

with
   2  
u ∂x u+u  + b2 v 1
L(∂x ) = , u, v) = (2u0 u
g( v + u
 v0 + u
2
 v)
2
.
v d∂x2 v − 2
u − b2 v −1
Plugging the ansatz ( u, v)(x, t) = eikx+λ(k)t ϕ with ϕ = ϕ(k) ∈ C2 into the
u, v) = L(∂x )(
linearization ∂t ( u, v) yields the eigenvalue problem
 2 
−k + 1 b2
(9.25) ϕ = λϕ,
−2 −dk 2 − b2
with wave number k ∈ R as parameter. For notational convenience in the
following we drop the dependence on (b, d). The two curves of eigenvalues
λ1,2 (k) are determined by
(9.26) λ2 − λ[−(1 + d)k 2 + 1 − b2 ] + [dk 4 + (b2 − d)k 2 + b2 ] = 0.
The stability of the stationary point for the reaction system is determined
by λ1,2 (0). We obtain

λ1,2 (0) = (1 − b2 )/2 ± (1 − b2 )2 /4 − b2
such that (b, 1/b) is unstable for b ∈ (0, 1). We have Im(λ1,2 (0, b)) = 0 for
65 5 7
√ √
b∈ (3 − 5)/2, (3 + 5)/2 .

The homogeneous solution of (9.23) is asymptotically stable if Re(λ1,2 (k)) <

0 for all k ∈ R. Depending on the parameters, instability can set in in the
two ways a) and d) described in Figure 9.8. See Figure 9.9 for details.

1
4 Reλ1
Imλ1 (k)
Imλ1
0
2
Reλ
2

Imλ2 (k)-1 0

-2 Reλ1,2 (k) −2

-3 −4

-2 0 2 −2 −1 0 1 2

Figure 9.9. The curves of eigenvalues in the Hopf and in the Turing
case. Here (b, d) = (0.7, 1) and (b, d) = (3, 60), respectively.

The general two-dimensional situation. Even for the simple model

(9.23) the discussion of the dispersion relation (9.26) is rather unpleasant.
Thus, we derive necessary and sufficient conditions for the occurrence of a
Turing instability in a general two-component system
(9.27) ∂t u = ∂x2 u + f (u, v), ∂t v = d∂x2 v + g(u, v).
9.3. The Turing instability 311

Reλ1
0.2 Reλ2
0

-0.2

-0.4

-0.6

-0.8

-1
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 9.10. Transitional case: b = 1, d = 5.83.

We assume that the ODE

(9.28) u̇ = f (u, v), v̇ = g(u, v)
has some steady state (u∗ , v ∗ ). The eigenvalues λ of
 
∂(f, g) fu fv
A= = ,
∂(u, v) gu gv (u∗ ,v∗ )

satisfy
λ2 − (fu + gv )λ + (fu gv − fv gu ) = 0.
Hence, the steady state (u∗ , v ∗ ) is stable for the ODE (9.28) if
(9.29) tr A = fu + gv < 0 and det A = fu gv − fv gu > 0.
The eigenvalues λ of
 
∂x2 0
+A
0 d∂x2
satisfy
λ2 + λ(k 2 (1 + d) − (fu + gv )) + h(k 2 ) = 0,
with h(k 2 ) = dk 4 −(dfu +gv )k 2 +det A. Since already fu +gv < 0, a necessary
condition for Turing instability of the steady state (u∗ , v ∗ ) is h(k 2 ) < 0 for
some k, which requires dfu + gv > 0, hence, by (9.29) d = 1 and fu gv < 0.
Ultimately, we need hmin < 0, where, by calculus,
h(k 2 ) = hmin = det A − (dfu + gv )2 /(4d) at k 2 = km
2
= (dfu + gv )/(2d).
At the bifurcation point the condition hmin = 0 defines via |A| = (dfu +
gv )2 /4d a critical diffusion ratio as a root of d2c fu2 +2(2fv gu −fu gv )dc +gv2 = 0.
The critical wave number kc is given by

(9.30) kc2 = (dfu + gv )/(2d) = (fu gv − fv gu )/dc .
312 9. Reaction-Diffusion systems

Lemma 9.3.2. For a Turing-instability to occur in a system of the form

(9.27), the conditions
fu + gv < 0, fu gv − fv gu > 0, dfu + gv > 0,
(dfu + gv )2 − 4d(fu gv − fv gu ) > 0
are necessary and sufficient, with kc given by (9.30).
Definition 9.3.3. For a reaction-diffusion system ∂t u = D∂x2 u + f (u) with
a spatially homogeneous equilibrium u∗ which is asymptotically stable for
the ODE u̇ = f (u), the range of parameters for which there exists a k = 0
with Reλ(k) > 0 and Im(λ(k))=0 is called the Turing space for u∗ .

Lemma 9.3.2 gives criteria for the onset of a Turing instability in a

two-component reaction-diffusion system with diagonal diffusion. See Ex-
ample 9.3.4 for the Schnakenberg model, and the exercises for further exam-
ples and generalizations. For N ≥ 3-component reaction-diffusion systems,
characterizations of Turing-instabilities as in Lemma 9.3.2 become rather
complicated; see [HNM14] for some partial results. Moreover, in practice,
non-diagonal diffusion matrices occur, for which it is often more convenient
to find Turing instabilities numerically.
Example 9.3.4. For (9.23) we obtain fu = 1, fv = b2 , gu = −2, gv = −b2
and fu gv − fv gu = b2 . Thus Lemma 9.3.2 yields b > 1, b2 > 0, db > b3
and (db − b3 )2 > 4db4 . These are partly redundant, √and reduce to b > 1
0) and d > (3 + 8)b2 ≈ 5.83b2 . For
(for the stability of (b, 1/b) at k =5
√ 2 2
d = dc = (3 + 8)b we obtain kc = db c . For instance, for b = 1.5 we obtain
dc ≈ 13.12 and kc ≈ 0.63, cf. Figure 9.9.

For x ∈ Rn , n ≥ 2, the curves R  k → λj (k), j = 1, . . . , d, become

surfaces Rn  k → λj (k), which only depend on |k| due to the rotational
invariance of the Laplacian, e.g., for n = 2 we have a circle of critical wave-
vectors k ∈ R2 with |k| = kc .
Two-component systems with diagonal diffusion have played a major role
in pattern formation for reaction-diffusion systems in the last decades. The
examples at the beginning of Chapter 9 are very rich in “pattern formation”
in particular for x ∈ R2 , i.e., in suitable parameter regimes they show a
wealth of spatially non-homogeneous stationary solutions, some of which
show remarkable similarities to animal coat pattern (stripes on zebras, spots
on leopards, stripes and spots on fishes). The mathematical analysis of
these pattern starts with the investigation of the linear instabilities of the
homogeneous states as described above, and then proceeds with a weakly
nonlinear analysis via center manifold theory or the theory of modulation
equations as considered in Part IV.
9.3. The Turing instability 313

Further Reading. Besides [Mur89, Smo94, BK00], good introduc-

tory textbooks about reaction-diffusion systems, their mathematical analy-
sis and their various types of solutions are [Gri96, Nis02]. Recent text-
books about reaction-diffusion systems and background in applications are
[GR12, MK15]. The importance of reaction-diffusion systems in ecosys-
tem modelling is stressed in [Mer15], which also includes a very accessible
account on pattern formation and amplitude equations.
A relatively recent development are Turing(-like) bifurcations in infinite
time horizon distributed optimal control problems for parabolic PDEs of re-
action diffusion type. For these, the steady version of the so called canonical
system, derived via Pontryagin’s maximum principle, see [GCF+ 08], is an
elliptic system, which under certain condition has bifurcating spatially pe-
riodic solutions. In [BX08, BX10], this is called optimal diffusion induced
instability. In [Uec16, GU17], the optimality of these patterned canonical
steady states, and the associated optimal paths, have been studied numeri-
cally for some model problems. One result is that in certain systems a pat-
terned harvesting in an otherwise spatially uniform system gives a higher
yield; this may lead to many applications in the area of spatial resource
economics, in particular w.r.t. to natural reserves, and also in the field of
climate-economic models [BEX14].
A thorough treatment of (different versions of) the FitzHugh-Nagumo
system, and many other systems of excitable or oscillatory type with multiple
time scales is given in [Kue15], including many references. Much recent
work on the existence and dynamics for pulses in reaction-diffusion systems
focusses on singularly perturbed problems, and on heterogeneous media, see,
e.g., [NTYU07, DvHX16, DV15] and the references therein. Another
important class of reaction-diffusion systems arise in chemotaxis [Hor03,
Hor04] and are of the form ∂t u =div(D(u)∇u) + f (u). For such systems,
the diffusion coefficients depends on u which makes them quasilinear. The
local existence theory for such systems is more complicated than for (9.1),
requiring so called maximal regularity methods [Lun95, Ama95]. Blow-
up of solutions is more frequent than in the semilinear case, see for instance
[SGKM95]. See also [BBTW15, TW15] for coupling with fluid equations
and related modeling issues, and for results on equations of mixed parabolic-
elliptic type, that also frequently arise in applications.

Exercises
9.1. Show that any quadratic autocatalytic system (i.e., arbitrary k, DA , DB > 0)

∂t a = Da ∂x2 a − kab, ∂t b = Db ∂x2 b + kab

can be rescaled to the form ∂t u=∂x2 u−uv, ∂t v=d∂x2 v+uv.

314 9. Reaction-Diffusion systems

9.2. Show that the vector field f defined by (9.13) points inward on ∂R. Use this to
prove that there exists a heteroclinic orbit vhet (ξ) connecting (0, 1, 0) and (1, 0, 0).
9.3. Consider the reaction–diffusion–systems
∂t u = ∂x2 u + uM (u, v), ∂t v = d∂x2 v + vN (u, v),
associated to the Kolmogorov form of the equations for 2–species interaction in
mathematical ecology, see Example 2.6.5. For each type, (PP), (C) and (S), con-
struct (weakly) invariant regions, either abstractly, or for the concrete examples
given in Example 2.6.5.
9.4. Show that (9.18) is equivalent to (9.17) with ε = b, γ = c/ε and τ = εt.
9.5. Let the linearization of a system (9.27) around (u, v)∗ be given by
 
1 + ∂x2 4
wt = Aw with A = .
−1 −3 + d∂x2
Use Lemma 9.3.2 to find dc such that (u, v)∗ is Turing unstable for d>dc , and
calulate kc .
9.6. Sketch the Turing space for the unique fixed point of (simplified) Gierer–
Meinhard system
∂t u = ∂x2 u + a − u + u2 /v, ∂t v = d∂x2 v + u2 − v, a, d > 0.
9.7. In general, reaction diffusion systems may include so called cross-diffusion,
where the diffusion of one species directly influences some other species. For sim-
plicity we restrict to
   
u 1 0
(9.31) ∂t U = DΔU + F (U ), where U = , D= .
v d3 d4
Thus, if d3 > 0, then “cross–diffusion” in ∂t U = DΔU yields a decrease/increase
of v at the maxima/minima of u. Assume that (9.31) has a homogeneous steady
state U∗ . Show that U∗ is Turing unstable if
fu + gv < 0, fu gv − fv gu > 0, d4 fu − d3 fv + gv > 0,
(d4 fu − d3 fv + gv ) − 4d4 (fu gv − fv gu ) > 0,
2

d4 fu − d3 fv + gv
with critical wave number kc given by kc2 = .
2d4
 Chapter 10

Dynamics of pattern
and the GL equation

This is the first chapter of Part IV of this book. Part IV is about modula-
tion theory which is applied to physically realistic systems over unbounded
domains. Like for systems over bounded domains, which very often can be
reduced to finite-dimensional ODEs, also for problems over unbounded do-
mains often only a subset of the uncountable many degrees of freedom plays
a role. These still uncountable many degrees of freedom are described by
simpler PDEs, which in this book are called modulation equations. They
can be derived via perturbation theory. In particular, the Korteveg-de Vries
(KdV), the Nonlinear Schrödinger (NLS), and the Ginzburg-Landau (GL)
equations from Chapter 8, play an important role as modulation equations,
and one unifying theme of the remaining chapters of this book is what can
(and what can not) be deduced about the original full system from these
reduced PDEs.
Part IV consists of five chapters. The first three chapters are about the
derivation and justification of such approximations. Here, in Chapter 10 we
discuss the GL approximation, in Chapter 11 the NLS approximation, and
in Chapter 12 the KdV approximation. The last two chapters handle aspects
of the existence and stability theory of special solutions of nonlinear PDEs
on unbounded domains which are related to modulation theory. We start in
Chapter 13 with the construction of solutions via the spatial dynamics and
invariant manifold approach. The final Chapter 14 is about the stability of
solutions via the diffusive or dispersive stability approach.

315
316 10. Dynamics of pattern and the GL equation

The plan of this chapter is as follows. In §10.2 we introduce the GL ap-

proximation for the Swift-Hohenberg equation and justify it with a first ap-
proximation result. The universality of the GL approximation is explained
in §10.3. We develop an abstract approximation theory in §10.4, which
is subsequently applied to a number of examples, such as the Kuramoto-
Shivashinsky equation, or in §10.5 to reaction-diffusion systems, where we
also extend our theory to Hopf bifurcations at the wave number k = 0.
In §10.6 we consider convection problems and extend our theory to planar
systems and systems with a time-periodic forcing. In §10.7 we consider
with the Couette-Taylor problem another classical hydrodynamical stability
problem and extend our theory to Turing-Hopf bifurcations at wave numbers
k = 0. In §10.8 we discuss the attractivity of the set of GL described solu-
tions, and the comparison of the original pattern forming system attractor
and the associated GL attractor. We close this chapter with an outlook on
the Ginzburg-Landau approximation in stochastic PDEs, lattice differential
equations, and delay equations.

10.1. Introduction
In §9.3 we already discussed, on a linear level, Alan Turing’s remarkable
observation [Tur52] that diffusion can destabilize ODE fixed points in the
associated PDE, i.e., destabilize spatially homogeneous fixed points u∗ of
two-component reaction-diffusion systems of the form
∂t u = DΔu + f (u),

with u(x, t) ∈ R2 , x ∈ Rd , a reaction f : R2 → R2 , Δ = dj=1 ∂x2j , and
D ∈ R2×2 a diagonal matrix with entries d1 , d2 > 0. Until the work of Turing
it was believed that whatever the reaction is, diffusion will homogenize the
reactants in space. Turing found that this not true and that even simple
reaction-diffusion systems can lead to a spatial structure of the solutions
similar to the pattern on animals such as tigers, zebras, leopards and many
others.
As already explained in §9.3 the idea behind this observation is surpris-
ingly simple. The linearization
(10.1) ∂t v = DΔv + f (u∗ )v
around u∗ is solved in case x ∈ R by
(10.2) v(x, t) = eikx+λt v,
with k ∈ R, λ ∈ C, and v ∈ C2 , where λ and v are determined by the
eigenvalue problem
v + f (u∗ )
−|k|2 D v = λ
v.
10.1. Introduction 317

For fixed k ∈ R this is a two-dimensional eigenvalue problem, and thus we

obtain two curves of eigenvalues k → λ1,2 (k) and two curves of eigenvectors
k → v1,2 (k). If d1 = d2 , then obviously the maximum of the curves λ1,2
is always obtained at the wave number k = 0. However, if d1 = d2 , then
the maximum can be obtained at a non-zero wave number k = kc = 0. See
Figure 9.8 a) and b), and Lemma 9.3.2.
If a control parameter is varied in such systems, then an instability can
occur leading to linear growth of spatially periodic modes in some wave
number range. Using 0 < ε2  1 for the distance to the bifurcation point in
parameter space we may assume that this growth rate is of order O(ε2 ), and
generically, i.e., except for special cases where certain coefficients vanish, the
growth of these modes will be saturated by the nonlinear terms at an O(ε)
amplitude. Therefore, we have a so called weakly nonlinear regime.
The weakly nonlinear methods for the analysis of pattern forming sys-
tems differ fundamentally between finite and infinite domains. For finite
domains, bifurcating pattern near the bifurcation point can be analyzed via
the Lyapunov-Schmidt, cf. §3.1.3, or the center manifold reduction, cf. §3.2.
After the reduction they can be described by the solutions of some ODEs.
See also the subsequent §13.2. However, in many cases the wave length of
the pattern is much smaller than the physical domain, and the description
of the bifurcating pattern via Lyapunov-Schmidt reduction or center man-
ifold theory becomes increasingly useless for larger and larger domains, as
more and more modes become unstable. This can be seen as follows. If
we have a large bounded domain with suitable boundary conditions, e.g.,
x ∈ (0, lπ) with large l ∈ N and homogeneous Neumann boundary condi-
tions ∂n u = 0, then we have discrete eigenvalues Z/l  k → λj (k) with λj (k)
from Lemma 9.3.2. For large l the discrete eigenvalues fill up the continuous
curve R  k → λj (k), and Lemma 9.3.2 and formula (9.30) give good ap-
proximations for the occurence of the Turing instability and the critical wave
number kc . As a consequence the dimension of the center manifold and of
the reduced equations grow. Since in parallel the size of the center-manifold
gets smaller and smaller for large l these reduction methods become useless.
Thus, for the description of pattern forming systems on large domains
close to the threshold of instability we need a new approach. Since the
influence of far away boundaries is negligible in a certain sense for large
domains, similar to the PDEs in Part III, these pattern forming systems are
considered on unbounded domains to understand the intrinsic mechanism
for the occurrence of the pattern. Hence, throughout the rest of this chapter
all systems are considered on the whole real line or unbounded cylindrical
domains. We remark that the influence of the far away boundaries on the
dynamics have been estimated for instance in [Mie97a], see also [SS00].
318 10. Dynamics of pattern and the GL equation

The observation by Turing for (10.1) is nowadays accepted as the mech-

anism of pattern formation in animal furs. Moreover, this mechanism of
pattern formation is not restricted to biological systems. The same kind of
instability widely occurs in nature [Man92, Pis06, Mar07, CG09]. Ex-
amples are pattern in sand dunes, convection pattern in clouds or magma,
desertification pattern in savannas, and many more. Additionally, this kind
of instability occurs in laboratory experiments which have been created to
understand the occurrence of pattern formation. The Rayleigh-Bénard prob-
lem and the Couette-Taylor problem will be discussed as examples in §10.6
and in §10.7.
In the second half of the 1960s a number of scientists started to describe
the behavior of all these pattern forming systems close to the first instability
by one universal amplitude or modulation equation, cf. [dES71]. In the
simplest case, by a multiple scaling ansatz
v(x, t) = εA(εx, ε2 t)eikc x v1 (kc ) + c.c.
(using the notation following (10.1)), a GL equation
2
(10.3) ∂ T A = ν2 ∂ X A + ν0 A + ν3 A|A|3 ,
with T ≥ 0, X ∈ R, coefficients ν0 , ν2 , ν3 ∈ C, and A(X, T ) ∈ C has been
derived. For many years, only the associated GL equations have been an-
alyzed for the understanding of the pattern forming systems. In nonlinear
science it is still widely accepted that the derivation of a GL equation im-
plies that it is a good model for these systems, and that it is sufficient to
analyze the GL equation instead of the original system. In the late 1980s
mathematicians began to discuss to which extend this approximation is cor-
rect. The main purpose of this chapter is to give an introduction into the
mathematical theory developed around this question.
The history of the approximation results starts with [CE90] where the
(cubic) SH equation as original pattern forming system has been handled.
The proof has been simplified in [KSM92]. The case of quadratic nonlin-
earities has been treated first in [vH91], and the proof has been simplified in
[Sch94c] to the version presented here, and generalized to the Navier-Stokes
equations in [Sch94a]. See also [TBD+ 96]. There are attractivity results
[Eck93, Sch95a], showing that every small solution can be described af-
ter a certain time by the GL equation. Combining the approximation and
attractivity results allows to prove the upper semi-continuity of attractors
[MS95, Sch99c], shadowing by pseudo-orbits, and global existence results
for the pattern forming systems [Sch94b, Sch99b]. These results will be
described in detail in §10.8.
We focus on the case of one unbounded space direction. The exceptions
are §10.5.3, where Hopf bifurcations at the wave vector k = 0 are treated
10.2. The Swift-Hohenberg equation 319

for x ∈ Rd , and §10.6.3-§10.6.5, where we discuss amplitude equations for

pattern forming systems with x ∈ R2 × Σ, where Σ is a bounded cross
section.

10.2. The Swift-Hohenberg equation

The simplest equation where the GL equation can be derived is the so called
Swift-Hohenberg (SH) equation [SH77]

(10.4) ∂t u = −(1 + ∂x2 )2 u + αu − u3 ,

with u(x, t) ∈ R, x ∈ R, t ≥ 0, α ∈ R, and −(1 + ∂x2 )2 u = −∂x4 u − 2∂x2 u − u.

This problem has the spatially homogenous state u∗ ≡ 0, which becomes
unstable at α = 0 via an instability at the non-zero wave number kc = 1.
The linearization ∂t v = −(1 + ∂x2 )2 v + αv of the SH equation at the trivial
solution u∗ ≡ 0 possesses the solutions v(x, t) = eikx+λ(k)t where

(10.5) λ(k, α) = −(1 − k 2 )2 + α.

It is clear that u∗ ≡ 0 is asymptotically stable for α < 0. The maxima of

λ at k = ±1 are positive if α > 0. For the SH equation the ansatz for the
derivation of the GL equation is given by

(10.6) εψGL (x, t) = εA(X, T )eix + εA(X, T )e−ix

where 0 < α = ε2  1, X = εx, and T = ε2 t. The terms which do not cancel

after inserting the GL approximation into the SH equation are collected in
the residual
Res(u) = −∂t u − (1 + ∂x2 )2 u + ε2 u − u3 .
If Res(u) = 0, then u is an exact solution of the SH equation (10.4). With
the abbreviation E = eix we find
 
Res(εψGL ) = − E(ε3 ∂T A) − E(εA) − 2E εi2 A + 2ε2 i∂X A + ε3 ∂X 2
A
 
− E εi4 A + 4ε2 i3 ∂X A + 6ε3 i2 ∂X
2
A + 4ε4 i∂X
3
A + ε5 ∂ X
4
A
+ E(ε3 A) − E(3ε3 A2 A) − E3 (ε3 A3 ) + c.c.
 
= εE · 0 + ε2 E · 0 + ε3 E −∂T A + 4∂X 2
A + A − 3A|A|2
− ε3 E3 A3 − ε4 E(4i∂X
3
A) − ε5 E(∂X
4
A) + c.c.

The first line of the residual vanishes by choosing the envelope A of the
bifurcating spatially periodic pattern eix to satisfy the GL equation

(10.7) 2
∂T A = 4∂X A + A − 3A|A|3 .
320 10. Dynamics of pattern and the GL equation

10.2.1. Special solutions. Before we explain to which extend the dynam-

ics of the GL equation can be transferred to the original system, we look for
special solutions of the GL equation and formally approximate with these
GL solutions the dynamics of the SH equation. We consider three classes
of GL solutions, namely space-independent solutions, i.e., A = A(T ), time-
independent solutions, i.e., A = A(X), and traveling wave solutions, i.e.,
A = A(X − cT ). We consider real-valued solutions which form an invariant
subspace for the real GL equation (10.7).
Space-independent solutions. For space-independent solutions we get
the one-dimensional complex-valued ODE
d
A = A − 3A|A|2 .
dT
Real-valued solutions satisfy
d
A = A − 3A3 .
dT
This last ODE possesses three equilibria,
 the unstable origin A = 0 and
the two stable fixed points A = ± 1/3. A solution
 A = A(T ) with initial
condition A(0) > 0 satisfies limT →∞ A(T ) = 1/3 with some exponential
rate. Associated with this solution we find formally in the SH equation
u(x, t) =εA(T )eix + εA(T )e−ix + h.o.t. = 2εA(T ) cos x + h.o.t.

(10.8) → 2ε 1/3 cos x + h.o.t.
for T → ∞. Thus, we have formally found 2π-spatially periodic SH solutions
converging towards a spatially periodic equilibrium for T → ∞ as drawn in
Figure 10.1.

x
x

Figure 10.1. Convergence towards a spatially periodic equilibrium in

the SH equation.

This can be made rigorous by center manifold theory, and it can be

shown that in the subspace of 2π-spatially periodic SH solutions a super-
critical pitchfork bifurcation occurs at α = 0, cf. §13.2.2. Since the GL
equation is S 1 -invariant, with T → A(T ) also T → A(T )eiθ for all θ ∈ R is
a solution. Associated to this solution for the SH equation we find
u(x, t) =εA(T )eiθ eix + εA(T )e−iθ e−ix + h.o.t.

(10.9) =2εA(ε2 t) cos(x + θ) + h.o.t. → 2ε 1/3 cos(x + θ) + h.o.t.
10.2. The Swift-Hohenberg equation 321

for T → ∞. Hence, associated to the one-dimensional family of constant

equilibria for the GL equation there is a one-dimensional family of 2π-
spatially periodic equilibria for the SH equation in accordance with the
translation invariance of the SH equation.
Time-independent solutions. Next we consider real-valued stationary
solutions A = A(X) of the GL equation. They can be found as solutions of
the scalar second order ODE
2
0 = 4∂X A + A − 3A3 .
In order to analyze this equation we consider the phase portrait of the equiv-
alent first order system
(10.10) ∂X A = B, ∂X B = −A/4 + 3A3 /4,
cf. Exercise 2.10. We find√the three equilibria from above, namely (A, B) =
(0, 0) and (A, B) = (±1/ 3, 0). The last two equilibria are connected by
two heteroclinic solutions (A, B) = (A, B)het (ξ). Formally, these give ap-
proximations of equilibria of the SH equation of the form
u(x) = 2εAhet (εx) cos x + h.o.t.
√ √
2 ) for x → ∞ and with u(x) = −2ε cos x/ 3+
with u(x) = 2ε cos x/ 3+O(ε
√
O(ε2 ) = 2ε cos(x + π)/ 3 + O(ε2 ) for x → −∞. Therefore, these solutions
connect two members of the √ same family of translates of the spatially peri-
odic equilibrium 2ε cos x/ 3 which differ by a phase shift of π. The transi-
tion from one equilibrium at x → −∞ to the other at x → ∞ happens on a
spatial scale of size O(1/ε). These solutions are sketched in Figure 10.2.
A
u

1/ε

Figure 10.2. The spatially heteroclinic solution in the GL equation and

for the SH equation. There is a phase difference between the asymptoti-
cally approached spatially periodic equilibrium for x → ∞ and x → −∞.

In the phase portrait for (10.10) we also find periodic solutions (A, B) =
(A, B)per (X) surrounding the origin (A, B) = (0, 0). For the SH equation
we find formally the equilibria
u(x) = 2εAper (εx) cos x + h.o.t..
322 10. Dynamics of pattern and the GL equation

The envelope A = Aper (X) is spatially periodic with some period of order
O(1/ε) w.r.t. the fast spatial variable x and modulates the underlying pat-
tern cos x which has some period of order O(1). These solutions are sketched
in Figure 10.3.

A 1/ε

Figure 10.3. The spatially periodically modulated solution in the GL

equation and for the SH equation.

Again with X → A(X) also X → A(X)eiθ for all θ ∈ R is a solution of

(10.7). Since the GL equation is translational invariant, with X → A(X)
also X → A(X + X0 ) for all X0 ∈ R is a solution. Therefore, we find for the
SH equation the two parameter families of approximate solutions
u(x) = 2εAhet (X + X0 ) cos(x + θ) + h.o.t., and
(10.11)
u(x) = 2εAper (X + X0 ) cos(x + θ) + h.o.t..

Front solutions. Next we consider real-valued traveling waves A = A(ζ) =

A(X − cT ) of the GL equation. These satisfy
0 = 4∂ζ2 A + c∂ζ A + A − 3A3 .
Again we consider the phase portrait of the equivalent first order system
(10.12) ∂ζ A = B, ∂ζ B = −cB/4 − A/4 + 3A3 /4,
cf. Exercise 2.10. For all values of c = 0 we find heteroclinic connections
√
(A, B) = (A, B)front (ζ) between the fixed points (A, B) = (±1/ 3, 0) and
the origin (A, B) = (0,√0). Such solutions describe the spreading of the
stable phases A = ±1/ 3 into the unstable phase A = 0. The functions
ζ → Afront (ζ) are monotonic for |c| ≥ 4. For the SH equation the associated
solutions
(10.13) u(x, t) = 2εAfront (X − cT ) cos x + h.o.t.
are called modulating fronts. These solutions are sketched in Figure 10.4.
With the arguments from above we find again a two parameter family of
such approximate solutions, namely
u(x, t) = 2εAfront (X − cT + X0 ) cos(x + θ) + h.o.t.,
10.2. The Swift-Hohenberg equation 323

with X0 , θ ∈ R. Two GL fronts can be patched together to describe the

growths of the solutions to small positive spatially localized initial conditions
of A ≡ 0.
A

x
u

1/ε

Figure 10.4. The front solution in the GL equation and the modulating
front for the SH equation.

Figures 10.2-10.5 illustrate why the GL equation is called a modulation

equation. There are a number of interesting mathematical questions about
the approximation of solutions of the SH equation via the GL equation,
e.g., how good is the formal approximation, and can we give existence and
stability proofs for the formally constructed solutions? The first question
is considered in §10.2.2, and the second for spatially periodic stationary
solutions in §13.2.2. Their stability is discussed in §14.2.2. As a preview
and an illustration of the quality of the approximation we give a numerical
result in Figure 10.5.
Remark 10.2.1. For α > 0 we have a  continuum of wave numbers with
λ(k) > 0 for k ∈ (k− , k+ ), where k± = 1± α/2+O(α). Thus, 2π/k-periodic
solutions of the SH equation occur for all wave numbers k ∈ (k− , k+ ). There-
fore, the two parameter families (10.9), (10.11) and (10.13) are embedded
in larger families with the wave number as a third parameter.

10.2.2. Validity of the GL approximation cubic case. In this section

we explain how to justify the GL equation in case of cubic nonlinearities in
the original system. For expository reasons we restrict to the SH equation
(10.4) as original system. Our main purpose is to prove that solutions of
the SH equation (10.4) behave as predicted by the associated GL equation
(10.7). In the following many different constants are denoted by the same
symbol C if they can be chosen independently of the small perturbation
parameter 0 ≤ ε  1.
Formal smallness of the residual. As already said, the terms which do
not cancel after inserting the GL approximation into the SH equation are
collected in the residual
Res(u) = −∂t u − (1 + ∂x2 )2 u + ε2 u − u3 .
324 10. Dynamics of pattern and the GL equation

40

0.5

0 20

-0.5

-100
-50
0
Figure 10.5. Comparison of the true 50
(numerical)
100
0solution of the
SH equation with ε=0.5 and initial condition u0 (x)=A0 (εx) cos(x),
A0 (X) = 1/ cosh(X), with the (numerical) solution A (dashed line)
of the GL equation with initial condition A0 (X), indicating that
ψGL (x, t) = εA(εx, ε2 t) cos(x) gives a good approximation for all
times considered.

The ansatz (10.6) and the abbreviation E = eix give

 
Res(εψGL ) =ε3 E −∂T A + 4∂X 2
A + A − 3A|A|2
− ε3 E3 A3 − ε4 E(4i∂X
3
A) − ε5 E(∂X
4
A) + c.c.,
and by choosing A to satisfy the GL equation (10.7) the first line of the
residual vanishes. However, we still have terms of O(ε3 ), which turn out to
be too large for proving an approximation result.
By adding higher order terms to the approximation εψGL the residual
can be made arbitrarily small, i.e., for arbitrary but fixed n ∈ N there exists
an approximation εψn with formally εψn − εψGL = O(ε3 ) and Res(εψn ) =
O(εn ). Since εψn − εψGL = O(ε3 ) the approximation εψn makes the same
predictions as εψGL about the behavior of the solutions u of the original
system. We will show Res(εψn ) = O(εn ) for n = 4, 5. With these examples
the general situation will be clear.
In order to obtain
(10.14) Res(εψ4 ) = O(ε4 )
we define
 
εψ4 = εψGL + ε3 A3 (εx, ε2 t)E3 + c.c. .
We then have
Res(εψ4 ) = ε3 E3 (−A3 − 64A3 ) + O(ε4 ),
10.2. The Swift-Hohenberg equation 325

and by choosing A3 = − 64
1 3
A we obtain (10.14). In order to obtain
(10.15) Res(εψ5 ) = O(ε5 )
we define
εψ5 = εψ4 + (ε2 A12 (εx, ε2 t)E + ε4 A32 (εx, ε2 t)E3 + c.c.)
where A12 and A32 are new functions which are chosen below. We find
Res(εψ5 ) = ε4 E(−∂T A12 + 4∂X
2
A12 + A12 − 3A2 A12 − 6|A|2 A12 − 4i∂X
3
A)
+ε4 E3 (−64A32 − 3A2 A12 − 96∂X A3 ) + O(ε5 ).
By choosing A12 to satisfy the inhomogeneous linearized GL equation
(10.16) 2
∂T A12 = 4∂X A12 + A12 − 3A2 A12 − 6|A|2 A12 − 4i∂X
3
A,
and A32 to satisfy −64A32 − 3A2 A12 − 96∂X A3 = 0, we obtain (10.15). In
order to achieve Res(εψn ) = O(εn ) we choose
(m)
 α

(10.17) εψn = εα(m)+j Amj (X, T )Em
m=−N,...,N j=1

with N = n − 1, X = εx, T = ε2 t, and

m 0 1 2 3 ··· m ··· N
α(m) 2 1 2 3 ··· ||m| − 1| + 1 ··· N
(m) N −1 N −2 N −1 N −2 · · ·
α N + 1 − α(m) − 2δ|m|1 ··· 1
where δij = 1 for i = j, and 0 else, see Figure 10.6.

|
u(k)|

O(1)

O(ε)

O(ε2 )

−1 |O(ε)|
k

Figure 10.6. Fourier mode distribution of the GL approximation. Note

 k−1 ) is O(1) in L∞ , but O(ε) in L1 and O(ε) in L∞ in physical
that A( ε
space.

As before A11 satisfies the GL equation. The evolution of the A1j for
j ≥ 2 are determined by inhomogeneous linearized GL equations, and the
Amj for m = ±1 are determined by algebraic equations. The solutions Amj
of the resulting system exist as long as the solution A11 of the GL equation
326 10. Dynamics of pattern and the GL equation

exists. In case of an odd nonlinearity such as for (10.4) we can set Amj = 0
for m odd.
Estimates for the residual in Cb0 . The formal orders of the residual
can be turned into estimates in norms. We find for instance
 Res(εψGL )C 0 ≤ s1 + s2 + s3 ,
b

where
s1 = 3ε3 E 3 A(ε·)3 C 0 ≤ 3ε3 A3C 0 ,
b b

s2 = ε4 E(4i∂X
3
A(ε·))C 0 ≤ 4ε4 AC 3 ,
b b

s3 = ε5 E(∂X
4
A(ε·))C 0 ≤ ε5 AC 4 ,
b b

such that the following holds.

Lemma 10.2.2. Let A ∈ C([0, T0 ], Cb4 ) be a solution of the GL equation
(10.7). Then for all ε0 ∈ (0, 1] there exists a C > 0 such that for all
ε ∈ (0, ε0 ) we have
sup  Res(εψGL (t))C 0 ≤ Cε3 .
b
t∈[0,T0 /ε2 ]

Similarly, for every n ∈ N with n ≥ 4 there exists an approximation εψn

such that the following holds. Let A ∈ C([0, T0 ], CbθA ) with θA = 3(n − 3) + 1
be a solution of the GL equation (10.7). Then for all ε0 ∈ (0, 1] there exists
a C > 0 such that for all ε ∈ (0, ε0 ) we have
sup  Res(εψn (t))C 0 ≤ Cεn ,
b
t∈[0,T0 /ε2 ]

and
sup εψGL (t) − εψn (t)C 0 ≤ Cε3 .
b
t∈[0,T0 /ε2 ]

Remark 10.2.3. Looking in more detail at the approximation εψ3 we find

that most regularity is lost in (10.16) for A12 . Since ∂X
4 A (·, T )
12 Cb0 has to
be estimated, we need A12 (·, T ) ∈ Cb , and if we do not use the smoothing
4

properties of the linear part in (10.16), then we need A(·, T ) ∈ Cb7 . Since the
structure of the approximation equations is the same for the next orders, we
lose three derivatives in every step, such that the estimates are possible with
θA = 3(n − 3) + 1. However, by using the smoothing properties of the linear
part in (10.16) and by constructing the approximation in a more clever way,
θm can be chosen much smaller, see below.

Estimates for the residual in H θ . For many pattern forming systems

the Cbθ spaces are no suitable phase spaces, because often it is hard to
establish the existence of semigroups in these spaces. In this respect Sobolev
spaces turn out to be a better choice. The estimates for the residual are less
10.2. The Swift-Hohenberg equation 327

trivial than before due to the scaling properties of the L2 -norm. For εψGL
we find for θ > 1/2 similar as before
 Res(εψGL )H θ

≤ C ε3 A(ε·)2C θ A(ε·)H θ + ε4 A(ε·)H θ+3 + ε5 A(ε·)H θ+4 .
b

However, for ε ∈ (0, 1) we find


 1/2
A(ε·)L2 = |A(εx)|2 dx
R

 1/2
(10.18) = ε−1/2 |A(X)|2 dX ≤ ε−1/2 AL2 ,
R
such that finally
 Res(εψGL (t))H θ ≤ Cε5/2 .
Nevertheless, for the proof of an approximation theorem this loss of ε−1/2
is no problem, since as before the residual can be made arbitrarily small by
adding higher order terms to the approximation.
Lemma 10.2.4. For all n ∈ N with n ≥ 4, and θ ≥ 1, there exists an
approximation εψn such that the following holds. Let A ∈ C([0, T0 ], H θA )
with θA = 3(n − 3) + 1 + θ be a solution of the GL equation (10.7). Then
for all ε0 ∈ (0, 1] there exists a C > 0 such that for all ε ∈ (0, ε0 ) we have
sup  Res(εψn (t))H θ ≤ Cεn−1/2
t∈[0,T0 /ε2 ]

and
sup εψGL (t) − εψn (t)C 0 ≤ Cε3 .
b
t∈[0,T0 /ε2 ]

Estimates for the residual in Hul θ. In §8.3.1 we introduced the Hul θ-

spaces to use the advantages of both spaces Cb and H , namely that Cb0
0 θ

contains non-decaying functions for |x| → ∞ and the fact that Fourier trans-
form is a suitable tool in H θ . The residual is estimated in Hul
θ as follows.

For θ > 1/2 we find


 Res(εψGL )H θ ≤ C ε3 A(ε·)3H θ + ε4 A(ε·)H θ+3 + ε5 A(ε·)H θ+4 .
ul ul ul ul

The loss of ε−1/2 can be avoided by

 y+1 1/2
A(ε·)L2 = sup |A(εx)| dx
2
ul
y∈R y
$ 1/2 8
y+1
≤ sup 1 dx sup |A(εx)| ≤ AC 0 ≤ CAH 1
b ul
y∈R y x∈[y,y+1]
328 10. Dynamics of pattern and the GL equation

such that we finally obtain

 Res(εψGL (t))H θ ≤ Cε3 .
ul

Like before we have

Lemma 10.2.5. For all n ∈ N with n ≥ 4, and θ ≥ 1, there exists an
θA
approximation εψn such that the following holds. Let A ∈ C([0, T0 ], Hul )
with θA = 3(n − 3) + 1 + θ be a solution of the GL equation (10.7). Then
for all ε0 ∈ (0, 1] there exists a C > 0 such that for all ε ∈ (0, ε0 ) we have
sup  Res(εψn (t))H θ ≤ Cεn ,
ul
t∈[0,T0 /ε2 ]

and
sup εψGL (t) − εψn (t)C 0 ≤ Cε3 .
b
t∈[0,T0 /ε2 ]

The equations for the error. Estimates for the residual, even in norms,
are only a necessary condition for the GL equation to correctly predict the
behavior of the original systems. Such estimates are not sufficient. The
errors can accumulate in time and there are a number of counter-examples,
cf. [Sch95b, Sch05, SSZ15], §10.6.3, and §11.5.3, showing that modu-
lation equations, although derived in a formally correct way, make wrong
predictions about the behavior of the original system.
The error
εβ R = u − εψ,
i.e., the difference between the solution u and the approximation εψ = εψn ,
with β and n suitably chosen, is estimated with the help of Gronwall’s
inequality. It satisfies
∂t R = −(1 + ∂x2 )2 R + ε2 R − 3ε2 ψ 2 R − 3εβ+1 ψR2 − ε2β R3 + ε−β Res(εψ).
We use the variation of constant formula
 t
tΛ 2
(10.19) R(t) = e R(0) + ε e(t−τ )Λ F (τ ) dτ,
0

where (etΛ ) t≥0 is the semigroup generated by Λ = −(1 + ∂x2 )2 and

ε2 F (τ ) =ε2 R(τ ) − 3ε2 ψ(τ )2 R(τ ) − 3εβ+1 ψ(τ )R(τ )2
− ε2β R(τ )3 + ε−β Res(εψ(τ )).
We have to decide in which norm we want to estimate the error, and are
not completely free in this. If A does not decay to zero for x → ±∞,
then the residual will be infinitely large in H θ . There are essentially three
ingredients which have to be estimated, namely first the residual terms which
have already been estimated in the spaces Cb0 , H θ , and Hul
θ above. Secondly,
10.2. The Swift-Hohenberg equation 329

we have to estimate the other terms in F , and thirdly the linear semigroup
(eΛt )t≥0 .
In order to make clear that estimates for F are easy to obtain indepen-
dently of the chosen norm we will give these estimates in all three spaces.
θA
Therefore, we assume that A ∈ C([0, T0 ], Cb,unif ), or A ∈ C([0, T0 ], H θA ), or
θA
A ∈ C([0, T0 ], Hul ), is a solution of the GL equation with θA ≥ 0 sufficiently
large. Then we obtain
Lemma 10.2.6. For every θ ≥ 1 there is a C > 0 such that for all ε ∈ (0, 1]
we have
ε2 F X θ ≤ C(ε2 RX θ + εβ+1 R2X θ + ε2β R3X θ + ε2 ),
where X θ stands either for Cb0 , H θ , or Hul
θ depending on the chosen phase

space for A.
Proof. All spaces X θ are closed under multiplication if θ > 1/2. Since the

H θ -norm of ψ is of order O(ε−1/2 ) we have to estimate ψ in the Cbθ -norm,
where θ = 0 in case X θ = Cb0 and θ = θ in the other cases. Then the
estimate follows from
ψ 2 RX θ ≤ Cψ2 θRX θ ,
Cb

ψR X θ ≤ CψC θR2X θ ,

2
b

R X θ ≤
3
CR3X θ ,
−β
ε Res(εψ)X θ ≤ Cε , 2

where the last estimate is either a consequence of Lemma 10.2.2 or Lemma

10.2.4, or Lemma 10.2.5. Moreover, we have
ψGL C θ = x → (A(εx)eix + c.c.)C θ ≤ CAC θ
b b b

and similar estimates for the other terms in ψ. Sobolev’s embedding theorem
gives the additional bounds
AC θ ≤ CAH θ+1
 and AC θ ≤ CA 
θ+1 .
b b Hul

Therefore, we are done. 

The third step, the bound for the semigroup (etΛ )
t≥0 , depends in its
complexity on the chosen space. Uniform bounds in the spaces Cb0 and Hul θ

are hard to obtain. In order to keep this introductory section simple, we

θ to §10.4.
skip the proof in Cb0 and postpone the proof in Hul
Lemma 10.2.7. For every θ ≥ 0 there exists a C > 0 such that for all
ε ∈ (0, 1] the semigroup (etΛ )t≥0 generated by Λ = −(1 + ∂x2 )2 satisfies
sup etΛ H θ →H θ ≤ C.
t∈[0,T0 /ε2 ]
330 10. Dynamics of pattern and the GL equation

Proof. We use that Fourier transform is an isomorphism from H θ to Hθ0 ,

cf. Lemma 7.3.31, and obtain

etΛ uH θ ≤ Ce−(1−·k H 0 ≤ C sup |e−(1−k

2 )2 t 2 )2 t
u |
uH 0 ≤ CuH θ .
θ θ
k∈R


The associated estimate in 0
is less trivial and can be found in
Cb,unif
θ.
[KSM92, Lemma 2.3]. Our favorite choice is the space Hul

Lemma 10.2.8. For every θ ≥ 0 there exists a C > 0 such that for all
ε ∈ (0, 1] the semigroup (etΛ )t≥0 generated by Λ = −(1 + ∂x2 )2 satisfies

sup etΛ H θ →H θ ≤ C.
ul ul
t∈[0,T0 /ε2 ]

Proof. See page 343. 

Now we have all estimates which are necessary to prove the following
approximation result. In the following Theorem 10.2.9 and Remark 10.2.11
the space X θ stands either for H θ or Hul
θ . In case X θ = H θ we set β = n−5/2

and in the other cases β = n − 2.

Theorem 10.2.9. Let n ∈ N, θ ≥ 1, and A ∈ C([0, T0 ], X θA ) with θA =

3(n − 3) + 1 + θ be a solution of the GL equation (10.7). Then there exist
an ε0 > 0 and a C > 0 such that for all ε ∈ (0, ε0 ) there are solutions u of
the SH equation (10.4) which can be approximated by εψn such that

(10.20) sup u(t) − εψn (t)X θ ≤ Cεβ ,

t∈[0,T0 /ε2 ]

and thus

(10.21) sup u(t) − εψGL (t)X θ ≤ Cε3/2 ,

t∈[0,T0 /ε2 ]

where εψn is the full ansatz (10.17), and εψGL is the lowest order approxi-
mation (10.6).

Proof. We use Lemma 10.2.6 and Lemma 10.2.7, respectively Lemma 10.2.8,
to bound the mild solutions of the equations for the error in the variation
of constant formula (10.19). By Theorem 5.2.22 we have local existence and
uniqueness of the solutions in Cb0 , H θ , or Hul
θ for θ > 1/2. The solutions

exist until the norm of the solutions becomes infinite, and thus we only have
to derive a priori estimates to guarantee that the solutions exist and stay
10.2. The Swift-Hohenberg equation 331

bounded for t ∈ [0, T0 /ε2 ].

 t 
R(t)X θ ≤ Cε2 R(τ )X θ + εβ−1 R(τ )2X θ + ε2β−2 R(τ )3X θ + 1 dτ
0
 t
≤ Cε2 (R(τ )X θ + 2) dτ
0
 t
≤ 2CT0 + Cε 2
R(τ )X θ dτ,
0

which holds as long as

(10.22) εβ−1 R(τ )2X θ + ε2β−2 R(τ )3X θ ≤ 1.

Gronwall’s inequality then yields

2
R(t)X θ ≤ 2CT0 eCε t ≤ 2CT0 eCT0 =: M

for all t ∈ [0, T0 /ε2 ]. Choosing ε0 > 0 such that εβ−1

0 M 2 + ε2β−2
0 M3 ≤ 1
we have satisfied (10.22). This yields (10.20). The triangle inequality then
yields (10.21), i.e.,

sup u(t) − εψGL (t)X θ

t∈[0,T0 /ε2 ]
≤ sup u(t) − εψn (t)X θ + sup εψn (t) − εψGL (t)X θ
t∈[0,T0 /ε2 ] t∈[0,T0 /ε2 ]

≤Cεβ + Cε3/2

Remark 10.2.10. For clarity we formulate the last step of Theorem 10.2.9
in a different way. We define  t = sup{t : R(t)X θ ≤ M }, with M defined
as above. We are done, if we show t ≥ T0 /ε2 . Choosing ε0 > 0 in the above
way guarantees the validity of (10.22) for all t ≤ 
t and so the validity of
 t
R(t)X θ ≤ 2CT0 + Cε2 R(τ )X θ dτ
0

for all t ≤ 
t. As above, Gronwall’s inequality implies supt∈[0,T0 ε2 ] R(t)X θ ≤
M , and so  t ≥ T0 /ε2 .

Remark 10.2.11. The estimate (10.21) shows that already the original
ansatz yields a suitable approximation, i.e., for ε > 0 sufficiently small we
see the dynamics predicted by the GL equation in the original system. For
times t ∈ [0, T0 /ε2 ] the error of order O(εmin(3/2,β) ) in X θ and by Sobolev’s
embedding theorem also in Cb0 is much smaller than the solution and ap-
proximation which are both of order O(ε) in Cb0 .
332 10. Dynamics of pattern and the GL equation

10.3. The universality of the GL equation

Above we already claimed that the GL equation
2
(10.23) ∂ T A = ν2 ∂ X A + ν0 A + ν3 A|A|2 ,
with coefficients ν0 , ν2 , ν3 ∈ C, appears as a universal modulation equation
for many different systems. The underlying original system will condense
in the values of the coefficients ν0 , ν2 , and ν3 in (10.23). Here is another
simple example.
Example 10.3.1. The Kuramoto-Shivashinsky (KS) equation
(10.24) ∂t u = −(1 + ∂x2 )2 u + αu + ∂x (u2 ),
with u(x, t) ∈ R, x ∈ R, t ≥ 0, and α ∈ R appears as a model for flame
propagation [KY76, Siv77, Kur80]. The lowest order approximation
(10.25) εΨGL = εA1 (X, T )eix + εA−1 (X, T )e−ix ,
where 0 < α = ε2  1 is a small perturbation parameter, is no longer
sufficient for the derivation of the GL equation. The ansatz is now given by
(10.26) εΨ = εΨGL + ε2 Ψh
where
εΨh = +ε2 A2 (X, T )e2ix + ε2 A−2 (X, T )e−2ix + ε2 A0 (X, T ),
and A−j (X, T ) = Aj (X, T ). We obtain
2
∂T A1 =4∂X A1 + A1 + 2i(A0 A1 + A2 A−1 ),
0 = − 9A2 + 2iA1 A1 ,
0 = − A0 .
Eliminating A0 and A2 via the algebraic relations from the second and third
line in the equation for A1 shows that the envelope A1 has to satisfy in
lowest order the GL equation
4
(10.27) ∂T A = 4∂X2
A + A − A|A|3 .
9

In order to explain why the GL equation plays such an important role

and occurs as universal amplitude or modulation equation describing the
evolution of many pattern forming systems close to the first instability, we
first review the derivation of the GL equation from the SH equation (10.4)
from a different point of view. The following computations are of limited
practical use, i.e., the are not useful for obtaining the nonlinear coefficient
ν3 of the GL equation, see also Remark 10.3.2 below, but they give the
10.3. The universality of the GL equation 333

essential theoretical framework for understanding the universality of the GL

equation.
The Fourier transformed SH equation. It turns out that Fourier
transform is the key for understanding the universality. In Fourier space,
the linearized SH equation is given by
(k, t) = λ(k, α)
∂t u u(k, t), with λ(k, α) = −(1 − k 2 )2 + α.
For α = ε2 > 0 there are two intervals centered around k = ±1 on which
λ(k, ε2 ) > 0, namely I− (ε) ∪ I+ (ε) = (−k+ (ε), −k− (ε)) ∪ (k− (ε), k+ (ε)) with
√ 1
k± (ε) = 1 ± ε = 1 ± ε + O(ε2 ).
2
2
A wave eikx+λ(k,ε )t associated with a k ∈ I+ (ε) (or k ∈ I− (ε)) will grow
2
exponentially, whereas a wave eikx+λ(k,ε )t associated with k ∈ I± (ε) will be
damped with some exponential rate. In order to find a modulation equation,
which describes the evolution of these unstable modes in an ε-neighborhood
of k = ±kc = ±1, and which is independent of the small perturbation
parameter ε we introduce K by k = 1 + εK, such that
 
1 1
(10.28) k ∈ I+ (ε) ⇔ K ∈ − + O(ε), + O(ε) .
2 2
With
1
λ(1 + εK, ε2 ) =λ(1, 0) + ∂1 λ(1, 0)εK + ∂12 λ(1, 0)ε2 K 2
2
+∂2 λ(1, 0)ε + O(ε )
2 3

(10.29) =ε2 (1 − 4K 2 ) + O(ε3 )

we can now write a family of linearly unstable spatially periodic perturba-
tions of u∗ = 0 as
2 (1−4K 2 +O(ε))t 2 +O(ε))(ε2 t)
(10.30) ei(1+εK)x+ε = eiKεx+(1−4K eix .
This has the form of a “basic pattern” eix modulated by a slowly varying
amplitude
2 )T
(10.31) Alin (X, T ) = eiKX+(1−4K + O(ε), with X = εx and T = ε2 t.
Then, ignoring the O(ε) terms, we find
2
(10.32) ∂T Alin = 4∂X Alin + Alin
which is the linear part of the GL equation. It is natural to expect that the
nonlinear problem (10.4) also has solutions which are small amplitude mod-
ulations of eix . The nonlinear terms determine the scaling of the amplitude
of the bifurcating solutions and so, as above, we insert the ansatz
(10.33) εψGL (x, t) = εA(X, T )eix + εA(X, T )e−ix
334 10. Dynamics of pattern and the GL equation

into the SH equation (10.4). As illustrated by Example 10.3.1, even if we

have quadratic terms in the original system, the nonlinear term in the GL
equation must always be of the form ν3 A|A|2 . This is due to the fact that the
quadratic interaction of the unstable modes e±ix gives modes concentrated
at the wave numbers k = 0, ±2 and not at the unstable modes at k = ±1 in
contrast to ν3 A|A|2 . A term proportional to A|A|4 is also concentrated at
k = ±1, but is of higher order.
The instability mechanism. With this idea in mind we consider a
general dissipative PDE in a cylindrical domain R × Σ with coordinates
x ∈ R and y ∈ Σ, where Σ ⊂ Rd is an bounded domain. We assume that the
system does not depend explicitly on x ∈ R, and that there exists a trivial
solution U ∗ which depends on y ∈ Σ only. W.l.o.g. for our purposes we can
assume that U ∗ = 0. Hence, we consider
∂t U = LU + N2 (U, U ) + N3 (U, U, U ) + . . .
for an unknown function U . The operator L is linear and the terms Nj
are j-multilinear and w.l.o.g. symmetric in their arguments. They can
depend on y ∈ Σ and on a control parameter α ∈ R, but they do not
depend explicitly on x ∈ R. If for instance N2 is not symmetric, then
define N 2 (U, V ) = 1 (N2 (U, V ) + N2 (V, U )), which has the properties that
2
N2 (U, V ) = N2 (V, U ) and N
2 (U, U ) = N2 (U, U ).
Due to the translation invariance in x, the linearization ∂t U = LU
around U ∗ = 0 possesses solutions of the form
U (x, y, t) = ϕn (k, y)eikx+λn (k)t ,
with k ∈ R. Under suitable assumptions on L and the boundary conditions
on R × Σ, for fixed k the operator −Lk · = −e−ikx L(eikx ·), acting on func-
tions which live on the bounded cross section Σ, is elliptic. Therefore, Lk
possesses a countable number of eigenvalues λn (k) with associated eigen-
functions ϕn (k, ·). We order the eigenvalues for fixed k in such a way that
Re λj ≥ Re λj+1 .
We assume that U ∗ = 0 is stable for α < αc , i.e., for α < αc all curves
k → λn (k) are in the left complex half-plane, respectively, all curves k →
Reλn (k) are below the k-axis. We assume further that that for α = αc one
curve λ1 or a pair of complex conjugate curves touch the imaginary axis for
a wave number k = kc = 0. Since we consider a real-valued problem the
same happens at k = −kc . Generically, all other curves are strictly bounded
away from the imaginary axis. See Figure 10.7.
For a simplification of the presentation we assume for the rest of this
subsection that the linear operator L can be diagonalized, that only the
curve λ1 = λ1 (k, ε2 ) touches the k-axis at (k, α) = (kc , 0), and that this
10.3. The universality of the GL equation 335

Reλn (k) Reλn (k)

k kc k

Figure 10.7. The curves of eigenvalues k → Reλn (k) in the stable case
are shown in the left panel. In the right panel instability occurs due to
one curve taking positive values at the wave numbers ±kc .

curve is real-valued like for the SH equation. Thus, we have λ1 (kc , 0) = 0,

∂k λ1 (kc , 0) = 0, and ∂k2 λ1 (kc , 0) < 0.
Derivation of the GL equation in general systems. After an expan-
sion  ∞ 
U (x, y, t) = cn (k, t)ϕn (k, y)eikx dk
−∞ n∈N

in eigenfunctions the coefficients cn (k, t) satisfy

  ∞
∂t cn (k, t) = λn (k)cn (k, t) + s2,n,n1 ,n2 (k, k−l, l)cn1 (k−l)cn2 (l) dl
n1 ,n2 ∈N −∞
  ∞
+ s3,n,n1 ,n2 ,n3 (k, k−l, l−m, m)cn1 (k−l)cn2 (l−m)cn3 (m) dl dm
n1 ,n2 ∈N −∞
+ ...,

with complex-valued kernels s2,n,n1 ,n2 (k, k − l, l), s3,n,n1 ,n2 ,n3 (k, k − l, l −
m, m), etc. Like the eigenvalues and eigenvectors, the kernels additionally
depend on ε2 , e.g. s2,n,n1 ,n2 (k, k − l, l; ε2 ). We have explicit representations
such as
9 :
s2,n,n1 ,n2 (k, k − l, l) = ϕ∗n (k), e−ikx N2 (ϕn1 (k − l)ei(k−l)x , ϕn2 (l)eilx )
L2 (Σ)

where ϕ∗n (k) is the associated adjoint eigenfunction w.r.t. the scalar product

·, ·L2 (Σ) , with
ϕ∗1 , ϕ1 L2 (Σ) = 1.
The modes u of the Fourier transformed SH equation solely correspond
to the modes c1 . For general systems additionally to these modes we have
infinitely many modes c2 , c3 , . . . which are damped with some exponential
rate. Since the Fourier transform of x → εA(εx, ε2 t)eikc x is given by k →
336 10. Dynamics of pattern and the GL equation

 −1 (k − kc ), ε2 t) we make the ansatz

A(ε
1,−1 (ε−1 (k+kc ), ε2 t)+A
c1 (k, t) =A 1,1 (ε−1 (k−kc ), ε2 t)
1,−2 (ε−1 (k+2kc ), ε2 t)+εA
+εA 1,0 (ε−1 k, ε2 t)
1,2 (ε−1 (k−2kc ), ε2 t),
+εA
n,−2 (ε−1 (k+2kc ), ε2 t)+εA
cn (k, t) =εA n,0 (ε−1 k, ε2 t)
n,2 (ε−1 (k−2kc ), ε2 t),
+εA
for n ≥ 2, with complex valued functions A n,j , where An,−j = An,j in
physical space. With k − kc = εK, and for instance
s211n (kc +εK, kc +ε(K−L), εL; ε2 )−s211n (kc , kc , 0; 0) = O(ε(|K|+|L|),
we find at ε2 for the modes concentrated at kc that
∂T A 1,1 (K, T )+∂ 2 λ1 (kc , 0)K 2 A
1,1 (K, T ) =∂α λ1 (kc , 0)A 1,1 (K, T )/2
k
∞
  ∞
+2 s211n (kc , kc , 0; 0) 1,1 (K−κ, T )A
A n,0 (κ, T ) dκ
n=1 −∞

∞  ∞
(10.34) +2 s211n (kc , −kc , 2kc ; 0)  
A1,−1 (K−κ, T )An,2 (κ, T ) dκ
n=1 −∞
 ∞  ∞
+3s31111 (kc , kc , kc , −kc ; 0) 1,1 (K−κ1 , T )
A
−∞ −∞
1,1 (κ1 −κ2 , T )A
×A 1,−1 (κ2 , T ) dκ2 dκ1 ,
where we used the symmetry of the bilinear terms in their arguments. With
k−jkc = εK we find at ε in the n-th equation for the modes concentrated
at jkc for j = 0, 2 and n ≥ 1 that
0 =λn (0, 0)An,0 (K, T )
 ∞
+2s2n1−1 (0, −kc , kc ; 0) A 1,1 (κ, T ) dκ,
1,−1 (K−κ, T )A
−∞
n,2 (K, T )
0 =λn (2kc , 0)A
 ∞
+s2n11 (2kc , kc , kc ; 0) A 1,1 (κ, T ) dκ.
1,1 (K − κ, T )A
−∞
Since Reλn (0, 0) < 0 and Reλn (2kc , 0) < 0 these algebraic relations deter-
mine A n,0 and An,2 in terms of A 1,1 and A 1,−1 such that (10.34) becomes
the GL equation in Fourier space, namely
1,1 (K, T ) =ν0 A
∂T A 1,1 (K, T ) − ν2 K 2 A
1,1 (K, T )
 ∞ ∞
(10.35) +ν3 1,1 (K − κ, T )A
A 1,1 (κ − κ 1,−1 (
 , T )A κ, T ) d
κ dκ,
−∞ −∞
10.4. An abstract approximation result 337

with coefficients
1
(10.36) ν2 = ∂k2 λ1 (kc , 0), ν0 = ∂α λ1 (kc , 0),
2
and
∞

ν3 = − 4 s211n (kc , kc , 0; 0)s2n1−1(0, −kc , kc ; 0)/λn (0, 0)
n=1
∞
(10.37) −2 s211n (kc , −kc , 2kc ; 0)s2n11 (2kc , kc , kc ; 0)/λn (2kc , 0)
n=1
+3s31111 (kc , kc , kc , −kc ; 0).
This derivation of the Fourier transformed GL equation (10.35) explains the
universality of the GL equation (10.3) and the formulas (10.36) for ν2 and
ν0 . The universal behavior can also be seen when computing the higher
order approximations. We find linearized GL equations at c1 and k = kc
and algebraic equations for all other modes, cf. Exercise 10.2. We leave the
case of a complex critical curve of eigenvalues λ1 to Exercise 10.3.
Remark 10.3.2. The formula (10.37), using the complete expansion in
eigenfunctions, is conceptually useful but of little practical use. Instead, the
exponentially damped modes can be collected in one big vector, and the
value of ν3 can be found by restricting to 2π/kc -spatially periodic functions,
see for instance §10.5.1, and also Remark 10.5.2. This often allows to ob-
tain this coefficient from the literature about the associated center manifold
reduction.

10.4. An abstract approximation result

We provide an abstract approximation theorem which will allow us subse-
quently to establish the validity of the GL approximation for general pattern
forming systems, such as reaction-diffusion systems or classical hydrodynam-
ical stability problems. In order to make things more concrete and for the
motivation of the following steps we in parallel consider the KS equation
(10.38) ∂t u = Lε (∂x )u + ∂x (u2 ) with Lε (∂x ) = −(1 + ∂x2 )2 + ε2
as an example, cf. Example 10.3.1.
Outline of the idea. In order to establish an approximation theorem like
the one for the SH equation, for a given ansatz εΨ, we introduce the error
εβ R = u − εΨ which satisfies
∂t R = Lε (∂x )R + 2ε∂x (ΨR) + εβ ∂x (R2 ) + ε−β Res(εΨ),
where the residual Res(εΨ), defined by Res(u) = −∂t u + Lε (∂x )u + ∂x (u2 ),
contains all terms which do not cancel after inserting the ansatz εΨ into
338 10. Dynamics of pattern and the GL equation

(10.38). As before, to any given n we can find an approximation εΨ such

that
(10.39) Res(εΨ) = O(εn ) and εΨGL − εΨ = O(ε2 ).
Thus, β and n can be chosen so large that no difficulty will occur in getting
bounds for the terms εβ ∂x2 (R2 ) + ε−β Res(εΨ) for all t ∈ [0, T0 /ε2 ].
However, a direct estimate for all t ∈ [0, T0 /ε2 ] with the help of Gron-
wall’s inequality is not possible due to the term 2ε∂x (ΨR), which could lead
to a growth rate O(erεt ) for an r > 0. This becomes unbounded for t = T0 /ε2
and ε → 0.
In order to overcome this difficulty with the term 2ε∂x2 (ΨR) we intro-
duce so called mode filters Ec and Es . They are defined as multiplication
operators in Fourier space by
Ej u = F −1 (χj u
), j = ±c, s,
where χj : R → [0, 1] is C0∞ , with χ±c (k) = 1 ⇔ |k ∓ 1| ≤ 1/30, χ±c (k) =
0 ⇔ |k ∓ 1| ≥ 1/15 and χs = 1 − χc − χ−c .
Remark 10.4.1. Since we would like to apply Lemma 8.3.7 about multi-
θ -spaces we have chosen χ ∈ C ∞ . In H θ -spaces we can take a
pliers in Hul j 0
simple cut-off function χj : R → {0, 1}, with χ±c (k) = 1 ⇔ |k ∓ 1| ≤ 1/30,
and χs = 1 − χc − χ−c .

Clearly we have εEc Ψ = O(ε) and εEs Ψ = O(ε2 ). The idea is to split
the error into a critical and a non-critical part which are scaled differently,
i.e.,
εβ R = εβ Rc + εβ R−c + εβ+1 Rs ,
where, depending on the chosen space, the support of R±c in Fourier space
is contained in a set equal to or slightly larger than the support of E±c ,
and where the support of Rs in Fourier space is contained in a set equal to
or slightly larger than the support of Es . The parts of the error satisfy a
system of the form
∂t Rc =Lε (∂x )Rc + O(ε2 |Rc | + ε2 |Rs |) + O(ε2 ),
∂t Rs =Lε (∂x )Rs + O(|Rc | + ε|Rs |) + O(1).
This is because
Ec applied to the quadratic interaction of critical modes vanishes,
i.e., Ec (ε∂x2 ((Ec Ψ)Rc )) = 0. Since Rs is exponentially damped by the
semigroup generated by Lε (∂x ), it is easy to obtain an estimate Rs =
O(|Rc |) + O(1). Inserting this into the first equation yields
∂t Rc = Lε (∂x )Rc + O(ε2 |Rc |) + O(ε2 ).
10.4. An abstract approximation result 339

A direct estimate for all t ∈ [0, T0 /ε2 ] with the help of Gronwall’s inequality
is now possible and so we obtain Rc = O(1) and Rs = O(1) for all t ∈
[0, T0 /ε2 ].
In the following we will extract assumptions which will allow us to handle
not only the KS equation, but very general pattern forming systems.

10.4.1. The abstract pattern forming system. Our starting point is a

pattern forming system

(10.40) ∂t u = Λu + N (u)

which is written with the help of the variation constant formula as

 t
tΛ
(10.41) u(t) = e u(0) + e(t−τ )Λ N (u(τ )) dτ.
0

We assume the existence of a formal approximation εΨ of u which is given

for all t ∈ [0, T0 /ε2 ], where T0 > 0 is a constant independent of 0 < ε  1.
The error εβ R = u − εΨ, with β ≥ 3/2, made by the approximation εΨ
satisfies
 t
R(t) =etΛ R(0) + ε−β e(t−τ )Λ (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ ))) dτ
0
 t
+ ε−β e(t−τ )Λ Res(εΨ(τ )) dτ,
0

where the residual is defined by

Res(u) = −∂t u + Λu + N (u).

In order to prove that εΨ is a suitable approximation of u we need a number

of assumptions. The first one is quite natural:
(A0) X and Y are Banach spaces.

Then we introduce mode filters Ec and Es for extracting the critical and
stable modes. For these we assume
(A1) There exist linear operators Ec and Es which are continuous both
in X and Y. Moreover, they commute with the semigroup (etΛ )t≥0 , i.e.,

etΛ Ec = Ec etΛ and etΛ Es = Es etΛ .

Using these mode filters we split the error R in two parts which we scale
differently, i.e., we set R = Rc +εRs and define Rc and Rs to be the solutions
340 10. Dynamics of pattern and the GL equation

of

Rc (t) =etΛ Rc (0)

 t
−β
(10.42) +ε e(t−τ )Λ Ec (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ ))) dτ
0
 t
−β
+ε e(t−τ )Λ Ec Res(εΨ(τ )) dτ,
0
Λt
Rs (t) =e Rs (0)
 t
−(β+1)
(10.43) +ε e(t−τ )Λ Es (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ ))) dτ
0
 t
−(β+1)
+ε e(t−τ )Λ Es Res(εΨ(τ )) dτ.
0

It turns out that this system is a good interface for coupling the systems
which we have in mind to the abstract set-up. For the formulation of the
other assumptions we need additional mode filters Ech and Esh . This is due
to the fact that for functional analytic reasons Ec and Es cannot be chosen
as projections in Hulθ . We assume

(A2) There exist linear operators Ech and Esh which are continuous both
in X and Y and satisfy Ech Ec = Ec Ech = Ec and Esh Es = Es Esh = Es . The
mode filters applied to the semigroup give the following estimates. There
exist CΛ , σc ≥ 0, α ∈ [0, 1), and a σs > 0 such that for all t ≥ 0 and ε ≥ 0
we have
2
etΛ Ech Y→Y ≤CΛ eσc ε t ,
2
etΛ Ech X →Y ≤CΛ eσc ε t ,
etΛ Esh Y→Y ≤CΛ e−σs t ,
etΛ Esh X →Y ≤CΛ max(1, t−α )e−σs t .

(A3) The nonlinear terms obey the following estimates. There exist con-
stants C1,c and C1,s , monotonically growing functions C2,c (Mc , Ms ) and
C2,s (Mc , Ms ) and an ε0 > 0 such that for all ε ∈ (0, ε0 ) we have

ε−β Ec (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ )))X

≤ C1,c ε2 (Rc Y + Rs Y ) + C2,c (Mc , Ms ) min(ε3 , εβ )(Rc Y + Rs Y )2 ,
ε−(β+1) Es (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ )))X
≤ C1,s Rc Y + C1,s εRs Y + C2,s (Mc , Ms ) min(ε, εβ−1 )(Rc Y + Rs Y )2 ,

as long as Rc Y ≤ Mc and Rs Y ≤ Ms .

10.4. An abstract approximation result 341

(A4) The residual terms obey the following estimates. There exist con-
stants Cres and ε0 > 0 such that for all ε ∈ (0, ε0 ) we have
sup Ec Res(εΨ(τ ))Y ≤Cres εβ+2 ,
τ ∈[0,T0 /ε2 ]

sup Es Res(εΨ(τ ))Y ≤Cres εβ+1 .

τ ∈[0,T0 /ε2 ]

Remark 10.4.2. As already said, the choice in (A0) depends on what

kind of solutions we want to study, e.g., localized or not, and for which X
and Y we can check (A1)-(A4), and with what effort. Not for all systems
considered below the assumptions will be checked in detail, e.g., for some
hydrodynamical stability problems we will refer to the literature. In any
case, for checking the assumptions (A0)-(A4) one usually does not have to
start from scratch. Before mathematicians discussed the validity of the GL
approximation, they justified the Landau approximation in case of spatially
periodic boundary conditions with the help of the center manifold theorem.
In order to apply the center manifold theorem one has to check a number of
assumptions, which are relatively close to our assumptions (A0)-(A4), see
for instance the proof of Theorem 10.7.5. Moreover, the cubic coefficient in
the Ginzburg-Landau equation and in the reduced ODE on the associated
center manifold in case of spatially periodic boundary conditions coincide,
cf. §13.2.2.

The following abstract approximation theorem will be applied to various

systems below, where we use the triangle inequality as in (10.21) to formu-
late the results for the associated lowest order approximations εΨGL , where
εΨGL − εΨY ≤ Cεγ , with γ ≥ 3/2 depending on the chosen space Y.
Theorem 10.4.3. Assume the validity of (A0)-(A2), and let εΨ be an
approximation such that (A3)-(A4) hold. Then there exist a C > 0 and
an ε0 > 0 such that for all ε ∈ (0, ε0 ) we have solutions u of the pattern
forming system (10.40) with
sup u − εΨY ≤ Cεβ .
t∈[0,T0 /ε2 ]

Proof. Our starting point is (10.42)-(10.43). Taking the Y-norm of both

sides yields
Rc (t)Y ≤ etΛ Ech Y→Y Rc (0)Y
 t
+ e(t−τ )Λ Ech X →Y ε−β Ec (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ )))X dτ
0
 t
+ ε−β e(t−τ )Λ Ech Y→Y Ec Res(εΨ(τ ))Y dτ,
0
342 10. Dynamics of pattern and the GL equation

2
≤ CΛ eσc ε t Rc (0)Y
 t
2
+ CΛ eσc ε (t−τ ) (C1,c ε2 (Rc (τ )Y + Rs (τ )Y )
0
+ C2,c (Mc , Ms ) min(ε3 , εβ )(Rc (τ )Y + Rs (τ )Y )2 ) dτ
 t
+ ε−β
2
CΛ eσc ε (t−τ ) Cres εβ+2 dτ
0

and
Rs (t)Y ≤ etΛ Esh Y→Y Rs (0)Y
 t
+ e(t−τ )Λ Esh X →Y ε−(β+1) Es (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ )))X dτ
0
 t
−(β+1)
+ε e(t−τ )Λ Esh Y→Y Es Res(εΨ(τ ))Y dτ
0
≤ CΛ e−σs t Rs (0)Y
 t
+ CΛ max(1, (t − τ )−α )e−σs (t−τ ) (C1,s Rc (τ )Y + C1,s εRs (τ )Y
0
+ C2,c (Mc , Ms ) min(ε, εβ−1 )(Rc (τ )Y + Rs (τ )Y )2 ) dτ
 t
−(β+1)
+ε CΛ e−σs (t−τ ) Cres εβ+1 dτ,
0

as long as Rc Y ≤ Mc and Rs Y ≤ Ms . Introducing

Sc (t) = sup Rc (τ )Y and Ss (t) = sup Rc (τ )Y
τ ∈[0,t] τ ∈[0,t]

allows for all t ∈ [0, T0 /ε2 ] to estimate that

 t
Rc (t)Y ≤CΛ eσc T0 Sc (0) + CΛ eσc T0 (C1,c ε2 (Sc (τ ) + Ss (τ ))
0
+ C2,c (Mc , Ms ) min(ε3 , εβ )(Sc (τ ) + Ss (τ ))2 ) dτ
+ T0 CΛ eσc T0 Cres ,
Rs (t)Y ≤CΛ Ss (0) + CΛ Cσ (C1,s Sc (t) + C1,s εSs (t)
+ C2,s (Mc , Ms ) min(ε, εβ−1 )(Sc (t) + Ss (t))2 )
+ CΛ Cσ Cres ,
t
where Cσ = supt∈[0,T0 /ε2 ] 0 max(1, (t − τ )−α )e−σs (t−τ ) dτ < ∞ is indepen-
dent of 0 ≤ ε  1. Since the right-hand sides of the last estimates increase
monotonically w.r.t. t we can replace Rc (t)Y and Rs (t)Y on the left-
hand side by Sc (t) and Ss (t), respectively. From the second inequality we
then obtain: For all Mc and Ms there exists an ε1 > 0 such that for all
10.4. An abstract approximation result 343

ε ∈ (0, ε1 ) we have
(10.44) Ss (t) ≤ (CΛ Ss (0) + CΛ Cσ Cres + 1) + 2CΛ Cσ C1,s Sc (t),
as long as Sc (t) ≤ Mc and Ss (t) ≤ Ms , if 0 < ε1  1 is chosen so small that

(10.45) CΛ Cσ (C1,s εMs + C2,s (Mc , Ms ) min(ε, εβ−1 )(Mc + Ms )2 ) ≤ 1

for all ε ∈ (0, ε1 ). Inserting (10.44) into the inequality for Sc (t) yields
 t
Sc (t) ≤ β0 + ε 2
β1 Sc (τ ) dτ,
0

where
β0 =CΛ eσc T0 Sc (0)
+ T0 CΛ eσc T0 Cres + T0 CΛ eσc T0 C1,c (CΛ Ss (0) + CΛ Cσ Cres + 1) + 1,
β1 =CΛ eσc T0 C1,c (1 + 2CΛ Cσ C1,s ),
if we choose 0 < ε2  1 so small that for given Mc and Ms we have
(10.46) CΛ eσc T0 T0 C2,c (Mc , Ms ) min(ε3 , εβ )(Mc + M2 )2 ) ≤ ε2
for all ε ∈ (0, ε2 ). Using Gronwall’s inequality immediately yields

Sc (t) ≤ β0 eβ1 T0 =: Mc ,
and due to (10.44) we define
Ms := (CΛ Ss (0) + CΛ Cσ Cres + 1) + 2CΛ Cσ C1,s Mc .
To this Mc and Ms we define ε0 = min(ε1 , ε2 ) where ε1 > 0 is chosen so
small that for all ε ∈ (0, ε1 ) condition (10.45) is satisfied and ε2 > 0 so small
that for all ε ∈ (0, ε2 ) condition (10.46) is satisfied. 

With the help of the mode filters we are now able to give the
Proof of Lemma 10.2.8. For the proof of the estimate we will use Lemma
8.3.7. Unfortunately, a direct estimate of k → e−(1−k ) t C 2 would lead to
2 2

b
some unwanted growth rates in time. Hence, we introduce two mode filters
Ec and Es satisfying Ec + Es = 1 where Ec is defined by E c = χ1 + χ−1 ,
∞
with χ ∈ C0 (R, [0, 1]) an even cut-off function with χ(k) = 1 for |k| ≤ 1/10
and χ(k) = 0 for |k| ≥ 1/5, and where χj (k) = χ(k − j). Then we find

etΛ H θ →H θ ≤ Ec etΛ H θ →H θ + Es etΛ H θ →H θ

ul ul ul ul ul ul

and immediately
s (k)e−(1−k
Es etΛ H θ →H θ ≤ Ck → E
2 )2 t
C 2 ≤ Ce−σt
ul ul b
344 10. Dynamics of pattern and the GL equation

for a σ > 0 since the semigroup decays with some exponential rate for the
wave numbers in the support of Es . The estimate for the Ec part uses that
for θ > 1/2 we have
Ec uH θ ≤CEc uL2 ≤ C
χ1 uL2 ≤ C χ(ue−ik·x )L∞
χ1 uL∞ ≤ C
ul ul ul
−ik·x
≤CSε (
χ(ue χ(ue−ik·x ))H θ
))L∞ ≤ CSε (
ul
−ik·x
≤C
χ(ue )H θ ≤ CEc uH θ ≤ CEc uL2
ul ul ul

where the scaling operator Sε is defined by (Sε u)(x) = u(εx). Hence, for
functions with a fixed compact support in Fourier space we have the equiva-
χ(ue−ik·x ))H θ .
lence of the norms Ec uH θ , respectively Ec uL2 , and Sε (
ul ul ul
W.r.t. the last norm Lemma 8.3.7 requires us to estimate k → eλε (k)t C 2
b
with
λε (k) = −(1 − (εk − 1))2 )2 .
Since λε (k) ∼ −(εk)2 for small k the term eλε (k)t behaves as e−k T with
2

T = ε2 t and so its Cb2 -norm is uniformly bounded for t on the O(1/ε2 ) time
scale. 

10.4.2. The Kuramoto-Shivashinsky equation. For simplicity we will

restrict ourselves to the case of Sobolev spaces H θ . The case of uniformly
local Sobolev spaces Hulθ is given as Exercise 10.4; see also [Sch94c].

Since the nonlinear terms are given by N (u) = ∂x (u2 ) we make the
choice
(A0) We set X = H θ−1 and Y = H θ .
The semigroup (etΛ )t≥0 is defined as a multiplier in Fourier space via
tΛ F where e4
etΛ = F −1 e4 tΛ (k) = e(−(1−k2 )2 +ε2 )t . Therefore, we also introduce

the mode filters Ec and Es as multiplication operators in Fourier space,

namely
Ej u = F −1 χj F , j = ±c, s,
where χj : R → {0, 1}, with χ±c (k) = 1 ⇔ |k ∓ 1| ≤ 1/30, and χs =
1 − χc − χ−c . See Remark 10.4.1.
(A1) We have the continuity of the Ej in H θ due to
Ej uH θ ≤ C1 χj u
L2 ≤ C1 χj L∞ 
uL2 ≤ C1 
uL2 ≤ C2 uH θ .
θ θ θ

Moreover, the Ej commute with the semigroup (etΛ ) t≥0 due to

etΛ Ej = F −1 e4
tΛ F F −1 χ F = F −1 e4
j
tΛ χ F = F −1 χ e4
j j
tΛ F

= F −1 χj F F −1 e4
tΛ F = E etΛ
j

and
(e4 = (χj e4
2 )2 +ε2 )t 2 )2 +ε2 )t
tΛ χ )(k) = e(−(1−k χj (k) = χj (k)e(−(1−k tΛ )(k).
j
10.4. An abstract approximation result 345

With this setup it is not difficult to verify Assumption (A2).

(A2) We choose Ech = Ec and Esh = Es since Ec2 = Ec and Es2 = Es .
Then there exist positive constants CΛ , σc , and α ∈ [0, 1), and a strictly
positive constant σs such that for all t ≥ 0 and ε ≥ 0 we have

etΛ Ec uH θ ≤C1 e4 c L2 ≤ C1 sup |e

tΛ χ u (−(1−k2 )2 +ε2 )t
uL2
χc (k)|
θ θ
k∈R
ε2 t
≤CΛ eσc uH θ ,
etΛ Ec uH θ ≤C1 e4 c L2 ≤ C1 sup |e
tΛ χ u (−(1−k2 )2 +ε2 )t
χc (k)(1+k 2 )1/2 |
uL2
θ θ−1
k∈R
ε2 t
≤CΛ eσc uH θ−1 ,
etΛ Es uH θ ≤C1 e4 s L2 ≤ C1 sup |e
tΛ χ u (−(1−k2 )2 +ε2 )t
uL2
χs (k)|
θ θ
k∈R
≤CΛ e−σs t uH θ ,
etΛ Es uH θ ≤C1 e4 s L2 ≤ C1 sup |e
tΛ χ u (−(1−k2 )2 +ε2 )t
χs (k)(1 + k 2 )1/2 |
uL2
θ θ−1
k∈R

≤CΛ max(1, t−1/4 )e −σs t

uH θ−1 .

Before we come to the assumptions (A3) and (A4) we must have a

more detailed look at the mode filters applied to the approximation. We
have
Lemma 10.4.4. Let θA ≥ max(θ + 1, 3) and let A ∈ C([0, T0 ], H θA ) be a
solution of the associated GL equation (10.27). There exist Cψ and ε0 > 0
such that for all ε ∈ (0, ε0 ) we have that εΨc = Ec (εΨ) and ε2 Ψs = Es (εΨ)
satisfy
sup εΨc C θ ≤ Cψ ε and sup ε2 Ψs C θ ≤ Cψ ε2 .
b b
t∈[0,T0 /ε2 ] t∈[0,T0 /ε2 ]

Proof. The GL approximation for the solutions of the KS equation is of the

form
εΨ(x, t) = εA(εx, ε2 t)eix + c.c. + ε2 Ψh (x, t)
where supt∈[0,T0 /ε2 ] ε2 Ψh C θ ≤ Cψ ε2 . With
b


Es (A(ε·)E)C θ ≤CEs (A(ε·)E)  −1
b
L1 ≤ C(1 − χ1 )A(ε (· − 1))L1
θ θ

≤C sup |(1 − χ1 (k))(1 + k ) 2 θ/2

(1 + ((k − 1)/ε) ) 2 −s/2  1
|A L
θ
k∈R
 2
≤Cεθ A ≤ Cεs AH θ+1 ,
L θ+1

where E = eix , we are done. 

346 10. Dynamics of pattern and the GL equation

The fact that Ec (Ψc Rc ) = Ec (Rc2 ) = 0 due to disjoint supports in Fourier

space of Ec with Ψc Rc and Rc2 , and the boundedness of Ec and Es , yield the
following estimates for the nonlinear terms
N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ ))
= ∂x ((εΨc + ε2 Ψs + εβ Rc + εβ+1 Rs )2 ) − ∂x ((εΨc + ε2 Ψs )2 ).
(A3) We have
ε−β Ec (N (εΨ(τ ) + εβ R(τ )) − N (εΨ(τ )))H θ−1
≤ C(ε2 Ψc C θ Rs H θ + ε2 Ψs C θ Rc H θ + ε3 Ψs C θ Rs H θ
b b b

+ εβ+1 Rc H θ Rs H θ + εβ+2 Rs 2H θ )

and
ε−β−1 Es (N (εΨ(τ )+εβ R(τ )) − N (εΨ(τ )))H θ−1
≤ C(Ψc C θ Rs H θ +εΨc C θ Rs H θ +εΨs C θ Rc H θ +ε2 Ψs C θ Rs H θ
b b b b

+ε β−1
Rc 2H θ +εβ Rc H θ Rs H θ +εβ+1 Rs 2H θ ).
As for the SH equation, due to Ψc H θ = O(ε−1/2 ) and Ψs H θ = O(ε−1/2 )
it is essential at this point to estimate Ψc and Ψs in Cbθ and not in H θ .
The residual terms are of the form

N
Res(εΨ) = εβj sj (εx)eijx ,
j=−N

with coefficients βj satisfying β±1 = 4 and βj ≥ 3 for all other j ∈ Z\{−1, 1}.
Hence, it remains to prove
Ec εβj sj (ε·x )eij·x H θ ≤ Cε7/2 ,
but this follows exactly as the estimate for Es (A(ε·)E)C θ in the proof
b
of Lemma 10.4.4. With this remark it is an easy exercise to check that
the approximation from Example 10.3.1 satisfies the Assumption (A4) with
β = 3/2 and εΨ from (10.26).
(A4) The residual terms obey the following estimates. There exist con-
stants Cres and ε0 > 0 such that for all ε ∈ (0, ε0 ) we have
sup Ec Res(εΨ(τ ))H θ ≤Cres ε7/2 ,
τ ∈[0,T0 /ε2 ]

sup Es Res(εΨ(τ ))H θ ≤Cres ε5/2 .

τ ∈[0,T0 /ε2 ]

We delegate the missing details to the exercises. Here, we only remark that
due to the fact that we have to expand the curves of eigenvalues at the wave
number k = kc , cf. (10.29) and Exercise 10.6, we need that the solutions of
10.5. Reaction-Diffusion systems 347

the GL equation have to be three spatial derivatives more regular than the
solutions of the underlying pattern forming system.
Hence, the KS equation with the previous choices for the mode filters
and H θ as phase space satisfies the assumptions (A0)-(A4) and so we have
the following approximation theorem.
Theorem 10.4.5. Fix θ ≥ 1. Let A ∈ C([0, T0 ], H θ+3 ) be a solution of the
GL equation (10.27). Then there exist a C > 0 and an ε0 > 0 such that for
all ε ∈ (0, ε0 ) we have solutions u of the KS equation (10.38) with
sup u − εΨGL H θ ≤ Cε3/2
t∈[0,T0 /ε2 ]

where εΨGL is the lowest order approximation defined in (10.25).

10.5. Reaction-Diffusion systems

We apply the above theory to reaction-diffusion systems (9.1), i.e.,
(10.47) ∂t u = Dμ ∂x2 u + fμ (u),
that show a Turing instability as discussed in §9.3. In (10.47) we explicitly
denote the dependence on the bifurcation parameter μ. We show that the
assumptions (A0)-(A4) are satisfied if a simple spectral assumption and
some genericity conditions are valid. Moreover, we give the derivation of the
GL equation from a more applied point of view, compared to the derivation
in §10.3. Finally, we extend our theory to systems which show some Hopf-
bifurcation at the wave number k = 0, with the Brusselator as prominent
example.

10.5.1. Derivation of the GL equation. We consider a spatially ho-

mogeneous equilibrium u∗ of the reaction-diffusion system (10.47) which
undergoes a Turing instability. By a change of coordinates u = u∗ + v we
find
(10.48) ∂t v = Lμ (∂x )v + gμ (v),
where
Lμ (∂x ) =Dμ ∂x2 + ∂u fμ (u∗ ),
gμ (v) =fμ (u∗ + v) − fμ (u∗ ) − ∂u fμ (u∗ )v = O(v 2 ).
We assume the occurrence of a Turing instability for (10.48) at the fixed
 μ (k) : Cd → Cd
point v = 0, i.e., for the eigenvalues λj (k) for j = 1, . . . , d of L
defined by
 μ (k)·
Lμ (∂x )(eikx ·) = eikx L
we assume:
348 10. Dynamics of pattern and the GL equation

(SpecRD ) For μ < μc the homogeneous state u∗ is stable, i.e., we

have Reλj (k, μ) < 0 for all k ∈ R and j = 1, . . . , d. At μ = μc we have
Reλ1 (kc , μc ) = 0 for some kc > 0 and Reλj (k, μc ) < 0 for k = ±kc if j = 1
or for all k ∈ R if j = 1. Finally, for ε2 = μ − μc , we assume that
Reλ1 (k, μ) = ν0 ε2 − ν2 (k − kc )2 + O(ε3 + |k − kc |3 )
for all k in some neighborhood of kc , with ν0 , ν2 > 0.

Hence, Reλj (k, μ) < −σ < 0 for all eigenvalues and all k ∈ R except of
a small neighborhood of kc , and we have Reλ1 (kc , μ) = O(ε2 ) > 0.

Remark 10.5.1. Since (10.48) is real-valued we have λj (−k) = λj (k), and

due to the symmetry x → −x we have λj (−k) = λj (k), hence simple eigen-
values are real. Thus, in particular λ1 (k) ∈ R for k near kc . However, we
continue to use complex notations since the calculus stays the same if we
consider a system of the form (10.48) without the symmetry x → −x.

As before the curve λ1 near kc determines the linear coefficients in the

GL equation. In order to compute the coefficient in front of the cubic terms
we write the nonlinear terms as
g(v) = B(v, v) + C(v, v, v) + O(v 4 ),
with bilinear and symmetric B(u, v) and trilinear and symmetric C(u, v, w).
In order to derive the GL equation we make the ansatz
(10.49) v(x, t) =εΨ(x, t) = εΨGL (x, t) + ε2 Ψh (x, t),
with
εΨGL (x, t) =εA(X, T )eikc x φ1 + c.c.,
ε2
ε2 Ψh (x, t) = Φ0 (X, T ) + ε2 e2ikc x Φ2 (X, T ) + c.c.,
2
where X = εx, T = ε2 t, A(X, T ) ∈ C, Φ0 (X, T ), Φ2 (X, T ) ∈ Cd , and where
φ1 is the critical eigenvector, i.e.,
(10.50)  μc (kc )φ1 = λ1 (kc , μc )φ1 = 0.
L

Plugging (10.49) into (10.48) and sorting the terms w.r.t. coefficients of
εm En , where En = einkc x , gives
εE : 0 = 0,
ε2 E2 :  μc (2kc )Φ2 + A2 B(φ1 , φ1 ),
0=L
ε2 E0 :  μc (0)Φ0 + 2|A|2 B(φ1 , φ1 ),
0=L
10.5. Reaction-Diffusion systems 349

and at ε3 E1 we find
0 1
∂T A = ∂μ λ1 (kc , μc ) − 12 ∂k2 λ1 (kc , μc )∂X
2 A
(10.51) & '
+ 2B(Aφ1 , Φ0 ) + 2B(Aφ1 , Φ2 ) + 3C(Aφ1 , Aφ1 , Aφ1 ), φ∗1 .

For obtaining (10.51) we applied the L  μc (kc ) invariant projection onto φ1 ,

defined via φ∗1 , the adjoint eigenvector of L  μc (kc ) to the eigenvalue λ1 (kc , μc )
 μc (kc ) φ = λ1 (kc , μc )φ , and normalized as
φ∗ , φ1  = 1. See
satisfying L ∗ ∗ ∗
1 1 1
Remark 10.5.2.
Due to our spectral assumptions L  μc (0) and L  μc (2kc ) are invertible such
that we obtain
Φ2 = −A2 L  μc (2kc )−1 B(φ1 , φ1 ),
(10.52)
Φ0 = −2|A|2 L  μc (0)−1 B(φ1 , φ1 ).

Then plugging (10.52) into (10.51) and setting

 μc (2kc )−1 B(φ1 , φ1 )
φ2 = L and  μc (0)−1 B(φ1 , φ1 ),
φ0 = L
we obtain
(10.53) 2
∂ T A = ν2 ∂ X A + ν0 A + ν3 |A|2 A,
where
& '
(10.54) ν3 = −2B(φ1 , φ0 ) − 2B(φ1 , φ2 ) + 3C(φ1 , φ1 , φ1 ), φ∗1 .
 μc (kc ) are real-valued we also have φ1 ∈ Rd , and
Since λ1 (kc , μc ) and L
L(k ∗  c ) . In particular, the coefficients ν2 , ν0 , ν3 in (10.53) are all
 c ) = L(k T

real-valued.
Thus, under rather generic assumptions on (10.48) we find that the
envelope A of the critical modes satisfy a GL equation. For a given problem,
of course it remains to determine the coefficients ν2 , ν0 , ν3 .
Remark 10.5.2. In the literature, e.g., [Man92, CH93b, Mie02, Pis06,
Mer15] the following derivation of the GL equation can be found. Plugging
the extended ansatz
(10.55) u(x, t) = εΨ(x, t) + ε3 Φ3 (εx, εt)eikc x + c.c.,
into (10.48) yields at ε3 E1
0 1
 c )Φ3 = − ∂T Aφ1 + ν0 + ν2 ∂ 2 Aφ1
L(k X

(10.56) + 2B(Aφ1 , Φ0 ) + 2B(Aφ1 , Φ2 ) + 3C(Aφ1 , Aφ1 , Aφ1 ).

The GL equation (10.53) is obtained as the solvability condition if we want
 c ) is not invertible
to remove the terms at ε3 E1 from the residual. Since L(k
there exists a solution Φ3 to (10.56) if and only if the right-hand side is
350 10. Dynamics of pattern and the GL equation

 c ) which is precisely
orthogonal to the zero eigenvector of the adjoint of L(k
(10.53).

Example 10.5.3. (GL formalism for the Schnakenberg model) Fig-

ure 10.8 shows some numerical simulations for the Schnakenberg model
(9.23). It shows that the GL approximation describes the qualitative be-
havior pretty well even in case of non-small ε. In detail, we consider (9.23)
with (b, d) = (1.5, 15), which is close to a Turing instability, cf. Figure
c ≈ 0.63, λ1 (kc ) ≈ 0.063 =: ε , hence ε ≈ 0.25, and
9.9. We have k 2

0.97
φ1 (kc ) ≈ . Moreover,
−0.24
 
1 1
B(φ1 , φ1 ) = B(φ1 , φ1 ) = q , with q = 2bφ11 φ12 + φ211 ≈ −0.0563,
−1 b
       
−1.2 0 1 1.03
Φ2 = q , Φ0 = q , C(φ1 , φ1 , φ1 ) = φ21 φ2 , φ∗1 ≈ ,
0.13 1 −1 0.28

and thus altogether ν3 ≈ −0.42. The  spatially independent equilibria of the

GL equation are therefore Ae ≡ ± −1/ν3 ≈ 1.55, from which we get the
approximate solutions
   
u b
= + 2εAe cos(kc x + ϑ)φ1 + O(ε2 )
v 1/b
 
1.5 + 0.75 cos(kc x + ϑ)
≈ + O(ε2 ),
0.67 − 0.18 cos(kc x + ϑ)

which fit well with the numerical simulations for the Schnakenberg model
(9.23) shown in Figure 10.8.

Figure 10.8. Pattern formation for the Schnakenberg model (9.23),

(b, d) = (1.5, 15), initial data (u0 , v0 ) = (b+sech(4x), 1/b+0.01 sec(4∗x).
Left: evolution of u. Right, u(x, 100) and v(x, 100) (dashed).
10.5. Reaction-Diffusion systems 351

10.5.2. Validity of the GL approximation. The excellent agreement

between numerical solutions and predictions of the GL equation is not re-
stricted to the Schnakenberg model and this particular solution. It is the goal
of the rest of this section to check that for the important class of reaction-
diffusion systems the assumptions (A0)-(A4) of our abstract approximation
theorem are satisfied, and that thus the GL approximation makes correct
predictions if the simple spectral assumption (SpecRD ) is valid. Since the
nonlinear terms are given by N (v) = gμ (v) we make the choice:
θ (R, Rd ).
(A0) We set X = Y = Hul

The semigroup (etΛ )t≥0 is defined by etΛ = F −1 e4 tΛ F where e4tΛ (k) =


etLμ (k) .In order to extract the critical modes we define for all k ∈ [kc −
δ, kc + δ], with a δ > 0 sufficiently small, a projection Pc (k) on the critical
eigenfunction f1 (k) by
Pc (k) =
f1∗ (k), ·Rd f1 (k),
 μ (k)∗ .
where f1∗ (k) is the associated eigenfunction of the adjoint operator L
We introduce the mode filter Ec as multiplication operators in Fourier space
by
Ej = F −1 χj Pc F , j = ±c
where χj : R → [0, 1] is C0∞ , with χ±c (k) = 1 ⇔ |k ∓ kc | ≤ δ/4, χ±c (k) =
0 ⇔ |k ∓ kc | ≥ δ/2, and Es = 1 − Ec − E−c . Next we choose
Ejh = F −1 χhj Pc F , j = ±c,
where χhj : R → [0, 1] is C0∞ , with χh±c (k) = 1 ⇔ |k ∓ kc | ≤ δ/2, χh±c (k) =
0 ⇔ |k ∓ kc | ≥ δ. Finally, we set
Esh = I − F −1 χhs Pc F ,
where χhj : R → [0, 1] is C0∞ , with χh±c (k) = 1 ⇔ |k ∓ kc | ≤ δ/8, χh±c (k) =
0 ⇔ |k ∓ kc | ≥ δ/4. With this choice the validity of the assumptions (A1)
and (A2) follow almost line for line as for the KS equation. See Lemma
10.2.8 for such estimates in Hulθ -spaces.

θ is an algebra for θ > 1/2 and since

Since Hul
Ec B(Ψc , Rc ) = 0, Ec B(Rc , Rc ) = 0
for Ψc = Ech Ψc and Rc = Ech Rc the Assumption (A3) follows again almost
line for line as for the KS equation.
The residual terms are of the form

N
Res(εΨ) = εβj sj (εx, ε2 t)eijx ,
j=−N
352 10. Dynamics of pattern and the GL equation

with coefficients βj satisfying βj ≥ 3 for all j ∈ {−N, N }. Hence, it remains

to prove
sup Ec εβj sj (εx, ε2 t)eij·x H θ ≤ Cε4 ,
ul
t∈[0,T0 /ε2 ]

but this follows exactly as the estimate for the KS equation (10.38). There-
fore, (A4) is also valid for reaction-diffusion systems.
Theorem 10.5.4. Fix θ ≥ 1, assume that the reaction-diffusion system
(10.48) satisfies (SpecRD ), and let A ∈ C([0, T0 ], Hul
θ+3
) be a solution of the
GL equation (10.53). Then there exist a C > 0 and an ε0 > 0 such that for
all ε ∈ (0, ε0 ) we have solutions v of the reaction-diffusion system (10.48)
with
sup v − εΨGL H θ ≤ Cε2 ,
ul
t∈[0,T0 /ε2 ]

where εΨGL is the associated GL approximation defined in (10.49).

10.5.3. The Hopf bifurcation case. The GL equation can also be de-
rived in case when the bifurcating pattern is oscillatory in time without
oscillations in space, i.e., when a Hopf bifurcation occurs at the wave num-
ber k = 0, see Figure 9.8 d). In this case, the number of unbounded spatial
directions plays no role, and thus we consider x ∈ Rd . The most famous
example is the so called Brusselator, cf. (9.3) and [Kur84],
∂t v1 = d1 Δv1 + a − (b + 1)v1 + v12 v2 , ∂t v2 = d2 Δv2 + bv1 − v12 v2 ,
where a, b, d1 , d2 are non-negative constants, and v1 , v2 are functions of t ≥ 0
and x ∈ Rd . For this reaction-diffusion system there exists a unique uniform
steady state (v1∗ , v2∗ ) = (a, b/a). Introducing new coordinates (v1 , v2 ) =
(v1∗ + u1 , v2∗ + u2 ) gives the system
∂t u1 = d1 Δu1 + (b − 1)u1 + a2 u2 + f (u1 , u2 ),
(10.57)
∂t u2 = d2 Δu2 − bu1 − a2 u2 − f (u1 , u2 ),
where
f (u1 , u2 ) = (b/a)u21 + 2au1 u2 + u21 u2 .
We fix a and take b as a control parameter. The stability of u1 =u2 = 0
is determined by the linearization of (10.57). We find the eigenfunctions
(u1 (k), u2 (k))eik·x , where the associated eigenvalues λ satisfy
λ2 + α(|k|)λ + β(|k|) = 0,
d
with |k|2 = 2
j=1 kj and k = (k1 , . . . , kd ) ∈ Rd . The constants are given by

α(q) =1 + a2 − b + (d1 + d2 )q 2 ,
β(q) =a2 + (a2 d1 + (1 − b)d2 )q 2 + d1 d2 q 4 .
10.5. Reaction-Diffusion systems 353

The trivial solution becomes unstable via an oscillatory instability at the

critical wave
 vector k = 0√for a critical value bc = 1 + a2 of the control pa-
−1
rameter if d1 /d2 > a ( 1 + a2 − 1). Drawing the surfaces of eigenvalues
λ± we obtain near k = 0 the rotational symmetric Figure 9.8 d).
We are interested in the bifurcation scenario near the threshold of insta-
bility. Therefore, we introduce the small bifurcation parameter 0 < ε2  1
where ε2 = (b − bc )/bc . In order to derive the GL equation from (10.57) we
make the ansatz
(10.58) u = εΨGL + ε2 Ψh
where
 
i(Imλ+ |k=0,ε=0 )t 1
2
εΨGL = A(εx, ε t)e + c.c., ε2 Ψh = O(ε2 ),
−1 + ia−1
where the vector is the eigenvector corresponding to the eigenvalue λ+ = ia
for k = 0 and ε = 0. Following the computations of [Kur84] shows that the
amplitude A in lowest order satisfies a GL equation

p
(10.59) ∂ T A = ν2 2
∂X j
A + ν0 A + ν3 |A|2 A,
j=1

with T = ε2 t, Xj = εxj , A(X, T ) ∈ C, and the coefficients

ν0 =(1 + a2 )/2,
ν2 =(d1 + d2 − ia(d1 − d2 ))/2,
 
1 2 + a2 4 − 7a2 + 4a4
ν3 = − +i .
2 a2 3a3
The solutions of (10.57) which are described by the GL equation are slow
modulations in time and in space of the spatially homogeneous temporally
oscillating state u = (1, −1 + ia−1 )eiat .
The abstract set-up. In the following we consider an abstract reaction-
diffusion system
(10.60) ∂t u = D∂x2 u + fμ (u),
with diagonal diffusion matrix D = diag(d1 Δ, . . . , dd Δ), where dj > 0. For
this system we assume the existence of a stationary solution u∗ , w.l.o.g.
u∗ = 0, such that the linearization
Lμ (∂x ) = DΔ + ∂u fμ (0)
around u∗ satisfies the following spectral assumption.
(SpecH ) There exist ρ0 , ε0 > 0, such that for 0 < ε < ε0 and k ∈
Uρ0 (0) = {k ∈ Rd : |k| < ρ0 } the following holds: The matrices
−diag(d1 |k|2 , . . . , dd |k|2 ) + ∂u fμ (0)
354 10. Dynamics of pattern and the GL equation

possess two surfaces of eigenvalues λ± ∈ C ∞ (Uρ , C), with associated eigen-
0
± ∈ C ∞ (Uρ0 , Cq ), such that
functions ϕ
± (k, ε2 ) = ±iω + ε2 (λ0 ± iν0 ) + |k|2 (λ2 ± iν2 ) + O(ε3 + |k|3 ),
λ
with constants λ0 > 0, λ2 < 0, and ω > 0. Denote by Σ− (k, ε2 ) the set of all
 ε2 ). Then, there exists an ε-independent constant
other eigenvalues of Λ(k,
0 > 0 such that
σ  
;
−
sup Re Σ (k, ε ) < −
2
σ0 .
k

In general the abstract theorem 10.4.3 cannot be applied in this situa-

tion, even when restricting to x ∈ R, due to the fact that Assumption (A3),
with the crucial property that Ec N2 (Ψc , Rc ) = 0, is not satisfied in general.
The quadratic interaction of critical modes again gives critical modes in case
of a Hopf bifurcation at the wave vector k = 0. What helps is the oscillatory
character of these quadratic terms. It will allow us to remove them by a
normal form transformation similar to §3.3. This has been carried out in
detail in [Sch98b]. We come back to normal form transformations in §11.4.
After the transformation we have Ec N2 (Ψc , Rc ) = 0 and so an analogue of
Theorem 10.4.3 can be applied.
Theorem 10.5.5. Fix s > d/2 and assume that the reaction-diffusion sys-
tem (10.60) satisfies the spectral assumption (SpecH ). Let
A ∈ C([0, T0 ], Hul
θ+3
)
be a solution of the associated GL equation. Then there exist a C > 0 and
an ε0 > 0 such that for all ε ∈ (0, ε0 ) we have solutions u of the reaction-
diffusion system (10.60) with
sup u − εΨGL H θ ≤ Cε2 ,
ul
t∈[0,T0 /ε2 ]

where εΨGL is the associated GL approximation defined in (10.58).

10.6. Convection problems

Hydrodynamic stability problems have always inspired new methods in dy-
namical systems theory. All classical hydrodynamic stability problems are
experiments or Gedanken-experiments, which have been designed to under-
stand and to extract special phenomena in more complicated situations. Ex-
amples considered in this chapter are convection problems and the Couette-
Taylor problem. Further hydrodynamic stability problems such as Poiseuille
flow, or the flow down an inclined plane are discussed in Chapter 14 about
diffusive stability.
10.6. Convection problems 355

A convection cell is a characteristic fluid flow pattern which occurs for

instance when a fluid is heated from below and cooled from above. Because
of the heat loss at the top surface and the fact that the fluid is more dense
at low temperature, a rising fluid becomes denser than the fluid underneath.
Since it cannot descend through the rising fluid, it moves sidewards. At some
distance the gravitational force overcomes the rising force and the fluid be-
gins to descend. As it descends, it warms again and the cycle repeats itself.
This mechanism plays a big role in geophysics. Examples are atmospheric or
oceanic circulations, but also the movement of the tectonic plates on earth is
induced by convection in the mantle of the earth. The forces leading to con-
vection do not necessarily come from a heat-dependent density. Vibration
induced convection occurs in powders and granular materials in vibrating
containers, Marangoni convection is induced by a heat-dependent surface
tension, and electro convection is observed in nematic crystals.
In this section we apply our previous theory to convection in porous
media and the classical Bénard problem. Moreover, we comment on the
case of planar systems and systems with a time-periodic forcing.

10.6.1. Convection in porous media. For so called black smokers on

the ocean floor, convection takes place in a porous medium, namely in the
highly fractured rock between the air or the sea at the top and the magma
chambers at the bottom. Compared to classical hydrodynamical stability
problems the associated system of PDEs for convection in porous media is
easier since the velocity field of the fluid is determined by a constitutive law,
namely Darcy’s law, and does not have to be computed as a solution of the
Navier-Stokes equations.
In detail, we are interested in the velocity field u = (u1 , u2 ) and the
temperature field T of a fluid in a strip Ω = R × (0, 1) of a porous medium,
heated from below. If we denote the coordinates in the strip by (x, y) ∈
R × (0, 1), then we have to solve
(10.61) ∇ · u = 0,
(10.62) u = −∇p + RT e2 ,
(10.63) ∂t T + u · ∇T = ΔT,
under the boundary conditions T = 1, u2 = 0 at y = 0 and T = 0, u2 = 0
at y = 1. Here, ∇ = (∂x , ∂y )T , Δ = ∂x2 + ∂y2 , e2 = (0, 1)T , p denotes
a pressure field, and the so called Rayleigh number R is a dimensionless
parameter which is proportional for instance to the distance of the plates
and the temperature difference between the upper and lower plate.
For a detailed derivation of (10.61)-(10.63) see for instance [Fow97,
§14]. Equation (10.61) describes conservation of mass for an incompressible
356 10. Dynamics of pattern and the GL equation

fluid, while (10.62) is based on the balance of forces. It is a constitutive

law which relates the velocity field u, the pressure p and the temperature
T . It comes from Darcy’s law which relates the velocity field u, the pressure
p and the density ρ. Here the relation between the temperature T and the
density ρ is given by an affine law. Otherwise the small temperature induced
density differences are neglected and ρ is considered to be constant for all
(x, y) ∈ R × (0, 1). This is called the Boussinesq approximation. Equation
(10.63) describes the diffusion and transport of heat.
The trivial state and its stability. The trivial solution of (10.61)-
(10.63) is given by the purely conducting state
R
(10.64) u = 0, T = 1 − y, p = − (1 − y)2 .
2
Since (10.61)-(10.63) is supposed to be a model for convection we expect
that for large R, e.g., for large temperature difference between the upper
and lower plate, convection sets in, resulting in some pattern of convection
rolls. In the following we explain that this is indeed the case, and that it
can be described via a GL approximation.
Our first step is to find the dispersion relation for the linearized system.
We eliminate the pressure p by introducing the stream function ψ such that
u1 = ∂ y ψ and u2 = −∂x ψ,
and introduce the deviation Θ from the linear temperature profile by T =
1 − y + Θ. This yields
(10.65) Δψ = −R∂x Θ and ∂t Θ + ∂x ψ + (∂y ψ)∂x Θ − (∂x ψ)∂y Θ = ΔΘ.
The linearized system is given by
Δψ = −R∂x Θ, ∂t Θ + ∂x ψ = ΔΘ,
together with the boundary conditions Θ = ψ = 0 at y = 0, 1. Due to the
boundary conditions we make the ansatz
ψ = f sin(nπy)eikx+λt , Θ = g sin(nπy)eikx+λt ,
with n ∈ N, k ∈ R, and complex-valued coefficients f and g. This gives the
system of linear equations
−(π 2 n2 + k 2 )f = −ikRg, −(π 2 n2 + k 2 )g = ikf + λg.
We find
ikR Rk 2
(10.66) f= g and λ= − (n2 π 2 + k 2 ),
n π2 + k2
2 n2 π 2 + k 2
i.e., we have a family of curves k → λn (k) ∈ R of eigenvalues with n ∈ N
and k ∈ R. It is easy to see that λn+1 (k) ≤ λn (k) ∈ R for each fixed k ∈ R.
10.6. Convection problems 357

Moreover, λn (k) → −∞, for |k| → ∞ and fixed n, or for n → ∞ and fixed
k.
The trivial solution (Θ, ψ) = (0, 0) is stable if λ1 (k) < 0 for all k ∈ R.
Instability occurs when the curve λ1 touches the axis λ = 0 at a wave number
k = kc ∈ R for a parameter value R = Rc . This leads to the conditions
Rk 2 !
λ1 = 2 2
− (π 2 + k 2 ) = 0
π +k
and
R Rk 2 π2R !
∂k2 λ1 = − − 1 = − 1 = 0.
π 2 + k 2 (π 2 + k 2 )2 (π 2 + k 2 )2
From this we find πR1/2 = π 2 + k 2 and λ1 = R − 2πR1/2 . This shows that
λ1 = 0 for R = Rc = 4π 2 ≈ 39.48 at the critical wave number k = kc = π,
see Figure 10.9.

0
λ1(k)
−30

−60 λ2(k)
−6 −4 −2 0 2 4 k 6

Figure 10.9. The curves of eigenvalues k → λn (k) for n = 1, 2.

Remark 10.6.1. The problem can be extended periodically into the y-

direction and can be solved in the subspace of odd functions w.r.t. y. Hence
for a function

u(x, y) = (k, n)eikx dk sin(nπy)
u
n∈N R

the H θ -norm is equivalent to the norm

 1/2

|
u(k, n)| (1 + |k| + |n| )
2 2 2 θ
.
n∈N R

Remark 10.6.2. The value Rc for spectral instability coincides with the
critical value Rc,e for which energetic stability of the trivial solution can be
shown. For this we multiply (10.65) with ψ and integrate over Ω to obtain
 
− |∇ψ|2 dx dy = −R Θ∂x ψ dx dy,
Ω Ω

hence ∇ψL2 ≤ RΘL2 . This, Poincaré’s inequality ΘL2 ≤ 2π 1

∇ΘL2 ,
and the fact that the last two terms in the first line of the following formula
358 10. Dynamics of pattern and the GL equation

cancel after integration by parts, gives


1
∂t Θ2L2 = −∇Θ2L2 + −(∂x ψ)Θ − (∂y ψ)(∂x Θ)Θ + (∂x ψ)(∂y Θ)Θ dx dy
2 Ω
≤ −∇Θ2L2 + ∂x ψL2 ΘL2 ≤ (R − 4π 2 )Θ2L2 .
Thus, Gronwall’s inequality yields
2 −4π 2 )
(10.67) Θ(·, t)2L2 ≤ et(R Θ(·, 0)2L2 .
Hence, for R < Rc,e = 4π 2 solutions decay exponentially to zero in L2 , while
for R > Rc,e = 4π 2 the L2 -norm can grow exponentially.

The dynamical systems formulation. We have the spectral situation

necessary for the derivation of the GL equation. Therefore, we would like to
come back to the arguments of §10.3 where we explained the universality of
the occurrence of the GL equation as an amplitude or modulation equation.
In order to do so we first formulate (10.65) as a dynamical system, i.e., we
have to get rid of the problem that (10.65) contains variables and equations
without time derivatives.
(10.65) is a parabolic equation for Θ coupled to an elliptic equation for
ψ. The evolution for (10.65) is therefore determined by specifying initial
conditions Θ(x, y, 0) = Θ0 (x, y), while ψ is determined by the compatibilty
condition Δψ = −R∂x Θ, ψ|y=0,1 = 0. Using (10.66), and

Θ(x, y, t) = 
Θ(k, n, t)eikx dk sin(nπy),
n∈N R

we find
 ik
ψ(x, y) = R 
Θ(k, n, t)eikx dk sin(nπy),
k2 + n2 π 2
n∈N R

which defines a linear operator M via

4F Θ where 4(k, n) = ik
ψ = RM Θ = RF −1 M M .
k2 + n2 π 2
Thus, M is bounded from H θ to H θ+1 , cf. Exercise 10.7. The introduction
of M allows to rewrite (10.65) as a single scalar parabolic equation
(10.68) ∂t Θ = LΘ + N (Θ),
with
L = Δ − RM ∂x Θ and N (Θ) = −R(∂y (M Θ))∂x Θ + R(∂x (M Θ))∂y Θ.
However, there is no maximum principle for (10.68) as for, e.g., the KPP
equation or the Burgers equation, because (10.68) is nonlocal in physical
space.
10.6. Convection problems 359

Establishing local existence and uniqueness of solutions is straightfor-

ward in Sobolev spaces H θ . The linear operator L generates a C0 -semigroup
(etL )t≥0 in H θ , with the smoothing estimate
 tL     
e Θ0  θ+1 ≤C1   e
 
tL
Θ0 2 ≤ C 1 sup |e

tL(k,m)
(1+k 2
+m 2 1/2   
) | Θ0  2
H Lθ+1 Lθ
k∈R,m∈N

≤C(1+t−1/2 )eβt Θ0 H θ

 m) ≤ β − k 2 − m2 . Since M is smoothing we have
for a β ∈ R since L(k,
estimates such as
∂y (M Θ)∂x ΘL2 ≤ C∂y (M Θ))L∞ ∂x ΘL2 ≤ C∂y (M Θ))H θ ∂x ΘL2
≤ CΘH θ ΘH 1
for every θ > 1 by the Sobolev embedding H θ → L∞ for θ > d/2. Hence,
N is locally Lipschitz-continuous from H θ to H θ−1 for θ > 1 and using the
method from §5.2.3 we can construct local mild solutions via the variation
of constant formula
 t
tL
Θ(t) = e Θ0 + e(t−τ )L N (Θ(τ )) dτ.
0
Therefore, we have the following local existence and uniqueness theorem.
Theorem 10.6.3. Let θ > 1. For every Θ0 ∈ H θ there exists a t0 =
t0 (Θ0 H θ ) > 0 and a unique local mild solution Θ ∈ C([0, t0 ], H θ ) of
(10.68) with Θ|t=0 = Θ0 .

As a consequence we additionally have ψ = RM Θ ∈ C([0, t0 ], H θ+1 ).

Derivation of the GL equation. Our starting point is System (10.68)
which satisfies the spectral situation necessary for the derivation of the GL
equation. We introduce the small perturbation parameter 0 < ε2  1 by
R = Rc +ε2 . The solutions of the system can be expanded in eigenfunctions,
i.e., we set

Θ(x, y, t) = cn (k, t)eikx dk sin(nπy).
n∈N R
In contrast to the reaction-diffusion systems considered in §10.5 for fixed
k we have an infinite-dimensional problem. Nevertheless, the present sys-
tem is special in the sense that the eigenfunctions fn (k, y) = sin(nπy) do
not depend on k ∈ R. Moreover, due to the fact that the problem can
be extended periodically in the y-direction we also have some convolution
structure in the y-direction which allows us to compute an exact expression
for the coefficient in front of the cubic terms of the GL equation. We find
  ∞
∂t cn (k, t) =λn (k)cn (k, t) + s2,n,n1 ,n2 (k, k − l, l)cn1 (k − l)cn2 (l) dl
n1 ,n2 ∈N −∞
360 10. Dynamics of pattern and the GL equation

with eigenvalues
Rk 2
λn (k, ε2 ) = − (n2 π 2 + k 2 )
n2 π 2 + k 2
and symmetrized complex-valued kernels
1 i(k − l)
s2,n,n1 ,n2 (k, k − l, l) = Rπ (−n1 il + i(k − l)n2 )δn−n1 −n2
8 (k − l)2 + n21 π 2
1 il
+ Rπ 2 (−n2 i(k − l) + iln1 )δn−n1 −n2
8 l + n22 π 2
1
since 0 sin(nπy) cos(n1 πy) sin(n2 πy)dy = δn−n1 −n2 /4. Following the calcu-
lations of §10.5 we have to compute
ν2 =∂k2 λ1 (π; 0)/2 = 2,
1
ν0 =∂ε2 λ1 (π; 0) = ,
2
∞
ν3 = − 4 s211n (kc , kc , 0; 0)s2n1−1 (0, −kc , kc ; 0)/λn (0; 0)
n=1
∞
−2 s211n (kc , −kc , 2kc ; 0)s2n11 (2kc , kc , kc ; 0)/λn (2kc ; 0)
n=1
= − 4s2112 (kc , kc , 0; 0)s221−1 (0, −kc , kc ; 0)/λ2 (0; 0)
− 2s2112 (kc , −kc , 2kc ; 0)s2211 (2kc , kc , kc ; 0)/λ2 (2kc ; 0) = −4π 2 .
Thus, we finally find the GL equation
1
(10.69) 2
∂T A = 2∂X A + A − 4π 4 |A|2 A,
2
cf. Exercise 10.8. Via the solutions of the GL equation we describe slow
spatial and temporal modulations of small amplitude of the underlying con-
vection pattern via the approximation
 
ψ
(x, y, t) = εΨGL (x, y, t) + O(ε2 )
Θ  
(10.70)
2πi iπx
= εA(X, T ) e sin πy + c.c. + O(ε2 ),
1
where X = εx and T = ε2 t are the long spatial and long temporal scale.
An immediate observation is that the GL equation (10.69) possesses stable
steady states A = 2√12π2 eiφ which are constant in space, where φ ∈ [0, 2π) is
a free phase. For φ = 0 this formally yields the steady convection rolls
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
u1 0 4π 2 cos πy sin πx
⎝u2 ⎠ = ⎝ 0 ⎠ + √ ε ⎝−4π 2 sin πy cos πx⎠ + O(ε2 ).
(10.71)
2 2π 2
T 1−y sin πy cos πx
10.6. Convection problems 361

As usual, these formal calculations do not guarantee that (10.71) is a good

approximation of steady convection rolls for (10.61)-(10.63), nor that steady
rolls exist at all. A rigorous existence proof for the stationary convection
rolls follows with center manifold theory or the Lyapunov-Schmidt method.
See §13 or Exercise 10.9.
The approximation result. The error estimates for the GL approxima-
tion follow immediately in H θ -spaces, where all assumptions of our abstract
approximation theorem 10.4.3 are easy to check. Therefore, we have
Theorem 10.6.4. Fix θ > 1 and let A ∈ C([0, T0 ], H θ+3 ) be a solution
of the GL equation (10.69). Then there exist ε0 , C > 0 such that for all
ε ∈ (0, ε0 ) there are solutions of (10.61)-(10.63) with
(10.72) sup (ψ, Θ)(·, t) − εΨGL (·, t)(H θ )2 ≤ Cε3/2 .
t∈[0,T0 /ε2 ]

θ -spaces for this introductory example.

We refrain from working with Hul

10.6.2. Bénard’s problem. Next, we treat the two-dimensional Rayleigh-

Bénard problem between two fixed plates, i.e., in the strip Ω = R × (0, π).
The problem is very similar to the one of the last section except that no
longer a porous medium is considered. Therefore, the Navier-Stokes equa-
tions replace Darcy’s law. The so called Boussinesq approximation is used,
i.e., the density is considered to be a constant except in the buoyancy
π term.
The problem is supplemented with the mean flux condition 0 u1 dy = 0
and the boundary conditions
∂y u1 |y=0 = u2 |y=0 = ∂y u1 |y=π = u2 |y=π = 0, T |y=0 = T0 , T |y=π = T1 ,
where T0 ≥ T1 . These boundary conditions allow for a number of explicit cal-
culations. For the physically more realistic boundary condition, u1 |y=0,π = 0
instead of ∂y u1 |y=0,π = 0, the subsequent analysis can be carried out, too.
There is a trivial solution, the purely conducting state with an affine
temperature profile and no motion of the fluid, namely
u = 0, T = T0 +y(T1 −T0 )/π.
This state is stable if the temperature difference between the lower and
the upper plate is sufficiently small. If the temperature difference becomes
large, then it loses stability and convection rolls appear. For even larger
temperature differences the motion becomes more complicated and eventu-
ally turbulent. With Θ = T − T0 −y(T1 −T0 )/π we introduce the deviation
from the linear heat profile and consider
(10.73) ∂t u =Δu − ∇p − ρΘe2 − (u · ∇)u,
(10.74) ∂t Θ =κΔΘ + u2 − (u · ∇)Θ,
(10.75) 0 =∇ · u.
362 10. Dynamics of pattern and the GL equation

The equations contain two dimensionless parameters, namely the Rayleigh

number ρ = β0 (T0 − T1 )h3 /(π 3 ν 2 ) and κ = δ/ν, where δ stands for the heat
conductivity, ν for viscosity, β0 for the buoyancy parameter, and h for the
physical height of the fluid.
Linear stability analysis. In a first step we consider the linearization
around the pure heat conducting state. We find
∂t u =Δu − ∇p − ρΘe2 ,
∂t Θ =κΔΘ + u2 ,
0 =∇ · u,
which is solved by
⎛ ⎞ ⎛ ⎞
u1 u1,m cos(my)
⎜u2 ⎟ ⎜ 2,m sin(my) ⎟
⎜ ⎟ (x, y, t) = eλt eikx ⎜ u ⎟
⎝Θ⎠ ⎝Θ  m sin(my) ⎠
p pm cos(my)
for m ∈ {1, 2, 3, . . .} and where

2λm,± (k) = −(κ + 1)s ± (κ + 1)2 s2 + 4ρk 2 s−1 ,
and s = k 2 + m2 . The associated eigenfunctions in (u1 , u2 , Θ) are denoted
in the following by ϕm,± (k).
We are interested in the problem close to the threshold of instability,
fix κ and vary the control parameter ρ. Due to the vanishing mean flux
condition we have m ≥ 1. It turns out that the trivial solution is unstable
for ρ > ρc = 27κ/4. For ρ = ρc + ε2 the curve√of eigenvalues λ1,+ (k) is
positive in an ε-neighborhood around ±kc = ±1/ 2. See Exercise 10.10.
The GL approximation. Defining the small bifurcation parameter ε in
this way, allows us to define the GL approximation
(u, Θ) = εΨGL (x, y, t) + O(ε2 ) = εA(X, T )eikc x ϕ1,+ (kc ) + c.c. + O(ε2 ),
with X = εx and T = ε2 t. Inserting this into (10.73)-(10.75) we find that
the complex amplitude A, which modulates the spatially periodic pattern
eikc x ϕ1,+ (kc ) slowly in time and in space, has to satisfy the GL equation
4κ 2 2 1
(10.76) ∂T A = ∂X A + A− A|A|2 .
κ+1 9(κ + 1) 8(κ + 1)
For a derivation see Exercise 10.11. Since again the problem can be extended
periodically into the y-direction, for the calculation of the cubic coefficient
only finitely many modes play a role, and so this coefficient can be computed
explicitly. For Bénard’s problem our approximation theorem specializes to
Theorem 10.6.5. Fix θ ≥ 2 and let A ∈ C([0, T0 ], Hul
θ+3
(R, C)) be a solution
of the GL equation (10.76). Then there exists ε0 , C > 0 such that for all
10.6. Convection problems 363

ε ∈ (0, ε0 ) there are solutions (u, Θ) of Bénard’s problem (10.73)-(10.75)

with
sup (u, Θ)(·, ·, t) − εΨGL (·, ·, t)H θ ≤ Cε2 .
ul
t∈[0,T0 /ε2 ]

The proof follows by applying the abstract approximation theorem stated

in Theorem 10.4.3. How to check the assumptions (A0)-(A4) can be found
in [Sch94a] where a detailed proof of Theorem 10.6.5 has been given.

10.6.3. Remarks on planar pattern. In this section we consider pattern

forming systems with more than one unbounded space directions. There,
pattern such as rolls, hexagons, and squares occur. As a toy problem we
consider
(10.77) ∂t u = −(1 + Δ)2 u + α∂y2 u + ε2 u + β|∇u|2 + γu3 ,
where Δ = ∂x2 + ∂y2 , (x, y) ∈ R2 , t ≥ 0, and α, β ∈ R. The small bifurcation
parameter is denoted by 0 < ε2  1. The linearized problem is solved by
v(x, y, t) = eikx+ily+λt with
λ = λ(k, l, ε2 ) = −(1 − k 2 − l2 )2 − αl2 + ε2 .
The parameter α in our model allows to break the rotational symmetry.
The rotational symmetric situation. For α = 0 the surface of spectral
values is rotational symmetric. Like for the Rayleigh-Bénard problem the
Fourier modes belonging to the wave vectors in a ring, here |k 2 + l2 − 1| =
O(ε), have positive growth rates. Amplitude equations can be derived for the
various pattern. The case β = 0 is a phenomenological model for Rayleigh-
Bénard problem with a rigid top boundary and the case β = 0 for the
problem with a free top boundary. In the following we first restrict to the
case β = 0 and γ = −1.
i) For squares, the Fourier modes are concentrated in O(ε)-neighborhoods
of four orthogonal wave vectors. W.l.o.g. we choose (1, 0), (0, 1), (−1, 0),
and (0, −1). Modulations of square pattern are described via the ansatz
u(x, y, t) =εA10 (εx, εy)eix + εA01 (εx, εy)eiy + c.c. + O(ε2 ).
This ansatz leads to the system, see Exercise 10.12,
2
∂T A10 =4∂X A10 + A10 − 3A10 |A10 |2 − 6A10 |A01 |2 ,
∂T A01 =4∂Y2 A01 + A01 − 3A01 |A01 |2 − 6A01 |A10 |2 ,
which is degenerated in the sense that in the A10 equation there is no diffu-
sion in Y -direction and similar for the A01 equation. This reflects the fact
that a big part of the ring of unstable modes is not covered by our ansatz.
ii) For hexagons, the Fourier modes are concentrated in O(ε)-neighbor-
hoods of six wave vectors with angle π/3 in between. W.l.o.g. we choose
364 10. Dynamics of pattern and the GL equation

√
(±1, 0) and (±1/2, ± 3/2). Modulations of hexagons are described via the
ansatz
√
1 3
u(x, y, t) =εA1 (εx, εy)eix + εA2 (εx, εy)ei( 2 x+ 2
y)
√
1 3
+ εA3 (εx, εy)ei( 2 x− 2
y)
+ c.c. + O(ε2 )
leading to the system
∂T A1 =4∂X2
A1 + A1 − 3A1 |A1 |2 − 6A1 (|A2 |2 + |A3 |2 ),
√
3 1
∂T A2 =4( ∂X + ∂Y )2 A2 + A2 − 3A2 |A2 |2 − 6A2 (|A1 |2 + |A3 |2 ),
√2 2
3 1
∂T A3 =4( ∂X − ∂Y )2 A3 + A3 − 3A3 |A3 |2 − 6A3 (|A1 |2 + |A2 |2 ).
2 2
The degeneracy w.r.t. diffusion again reflects the fact that a big part of the
ring of unstable modes is not covered by the ansatz.
iii) For rolls there are two possibilities. If the Fourier modes are concen-
trated in O(ε)-neighborhoods of the two wave vectors (±1, 0), then the above
system for the squares, but with A01 = 0, is obtained. The second possibil-
ity is an O(ε)-concentration in k-direction and an O(ε1/2 )-concentration in
l-direction near (±1, 0). The ansatz is then given

u(x, y, t) = εA(εx, ε1/2 y)eix + εA(εx, ε1/2 y)e−ix + O(ε2 ),

leading to the so called Newell-Whitehead equation [NW69]
(10.78) ∂T A = −4(i∂X − ∂Y2 /2)2 A + A − 3A|A|2 .
This anisotropic scaling allows a better approximation of the ring of unstable
modes at the wave vector (1, 0), see Exercise 10.12. Nevertheless, the Newell-
Whitehead equation is a degenerated parabolic equation.
For β = 0 the validity of these approximations follows line for line as
for the SH equation in §10.2. In case β = 0 quadratic terms are present in
the problem and the essential assumption of the analysis in §10.4, namely
that the quadratic interaction of critical modes gives non-critical modes, is
no longer valid. In fact, it can be shown rigorously [Sch95b] that in case
β = 0 the Newell-Whitehead equation makes wrong predictions about the
dynamics of (10.77). See §10.6.4.
Although quadratic terms are present in Bénard’s problem, in case of
rigid boundaries the approximation property holds since the quadratic inter-
action of critical modes gives non-critical modes, due to reflection symmetry
of the problem in the bounded direction. We refer to [GS85] for a discussion
of the bifurcation scenarios for the Rayleigh-Bénard problem with rigid and
free top boundaries. However, the validity proof is only straightforward in
10.6. Convection problems 365

H s -spaces. In Hul
s -spaces new functional difficulties occur due to the projec-

tion on the divergence-free vector fields in case of more than one unbounded
space directions, cf. [CZ15]. For other approaches to describe planar pat-
tern via amplitude equations see also [Mel00, EINP03], and the further
reading at the end of this Chapter.
The anisotropic situation. Next we consider the case α > 0, i.e., we
break the rotational symmetry in (10.77). Then unstable modes occur in
O(ε)-neighborhoods of the wave vectors (±1, 0). With the ansatz
u(x, y, t) = εA(εx, εy)eix + εA(εx, εy)e−ix + O(ε2 )
a two-dimensional GL equation can be derived
(10.79) 2
∂T A = 4∂X A + α∂Y2 A + A − 3A|A|2 .
Our abstract approximation result, Theorem 10.4.3, applies in this situation
since now the quadratic interaction of the critical modes gives non-critical
modes.
The prototype of an anisotropic pattern forming system is electro-convec-
tion in nematic liquid crystals [OD08]. In this problem the pattern forma-
tion is due to some external time-periodic forcing. It can be brought back
with Floquet theory to a problem with an autonomous linear part. The rele-
vant nonlinear terms can be made autonomous with the help of normal form
transformations. Hence, the above theory applies to time-periodic systems
as well [SU07]. See §10.6.5.

10.6.4. A counter-example. We recall the example from [Sch95b],

which shows that in case α = γ = 0 and β = 1 the predictions made
by the Newell-Whitehead equation
(10.80) ∂T A = −4(i∂X − ∂Y2 /2)2 A + A − 4A|A|2 /9.
for (10.77) in general are wrong, although the Newell-Whitehead equation
is derived in a formally correct way. Although modulation equations are not
primarily derived for the handling of the spatially periodic situation, they
should be valid for this special case, too. We consider the Newell-Whitehead
equation (10.80) with the initial condition
√ √
3 i( √3y − x ) i− √3y − x )
(10.81) A(X, Y, 0) = + B(0)e 2 ε 2ε + B(0)e 2 ε 2ε ,
2
with B(0) = O(ε) ∈ √R. The solutions of (10.80) are then spatially pe-
riodic with periods 2π√3 ε and 2πε. In this space, (10.80) possesses a two-
dimensional center manifold tangential to the complex space spanned by ei0 .
The reduced system on the center manifold is given by
=A
∂T A  − 4/9A|
 A|
 2 + h.o.t.
366 10. Dynamics of pattern and the GL equation

Therefore, the Newell-Whitehead equation predicts on an O(1/ε2 )-time scale

w.r.t. t, that solutions u of (10.77) behave as u(x, y, t) = 32 eix +c.c.+O(ε3/2 ).
It turns out that this prediction is only correct on a much shorter time
scale. The initial condition (10.81) for (10.80) leads to a hexagonal pattern
for (10.77). Therefore, the solutions of (10.77) can be written as Fourier
series  √
u(x, y, t) = ujk (t)eijx+ik(x/2+ 3y/2) .
j,k∈Z
Equation (10.77) possesses a six-dimensional attracting center manifold with
coordinates u10 , u01 , u−11 , u−10 , u0−1 and u1−1 . The initial condition for
(10.80) gives solutions for (10.77) with a = u10 = u−10 ∈ R and b = u01 =
u−11 = u0−1 = u1−1 ∈ R. In this two-dimensional invariant subspace on the
center manifold the reduced equations are given by
∂t a = ε2 a + b2 − 4/9a3 − 2/3b2 a + h.o.t.,
∂t b = ε2 b + ba − 1/3a2 b − 7/9b3 + h.o.t..
The perfect roll solutions described by (10.80) are given by the stationary
solution at (a, b) ≈ (3ε/2, 0). The linearization at this fixed point is given
by ∂t
a = −2ε2 a + h.o.t., ∂tb = 3εb/2 + h.o.t..
There is a one-to-one relation between the regularity of the initial con-
dition A|T =0 of (10.80) and the order of B(0). If A|T =0 is chosen analytic
in a strip in the complex plane, then there there is an r > 0 such that
b(0) = εB(0) = O(εe−r/ε ), cf. §6.2.2. Hence, although there is an expo-
3ε
nential growth e 2 t in the b-equation, b stays less than O(ε2 ) on an O(1/ε2 )
3ε
time scale since εe−r/ε e 2 t  ε2 for t ≤ r/(3ε2 ). However, if only n-times
differentiability of A|T =0 is assumed, then b(0) = O(εn+1 ). The Newell-
Whitehead equation makes wrong predictions, if b is of order O(ε) on a time
scale shorter than O(1/ε2 ), since then the error is of the same order as the
solution. In a neighborhood of the stationary solution (a, b) ∼ (3ε/2, 0). we
have b(t) ≈ e3εt/2 b(0) and so b(t) = O(ε) for t ≈ (n + 1)| ln ε|/ε  1/ε2 .
Hence the Newell-Whitehead approximation in general makes wrong predic-
tions for initial conditions which are only n-times differentiable. By looking
at the phase portrait of the (a, b)-system we see that the same behavior
occurs in the nonlinear system, too.

10.6.5. Time-periodic systems. Here we explain how to transfer the

ideas from the autonomous to the time-periodic case. Prototypes of pattern-
forming systems with an external time-periodic forcing are electro-convection
in nematic liquid crystals and the so called Faraday problem. In the first
problem a time-periodic external electric field is applied since a time-indepen-
dent external field would destroy the experiment by electrolysis, and in the
second problem a container is shaken periodically. We restrict ourselves to
10.6. Convection problems 367

the underlying ideas. The precise conditions on the underlying systems are
similar to the previous ones and can be found in [SU07].
We consider a translationally invariant system
(10.82) ∂t V = M (t)V + N (t, V )
which depend on some parameter α ∈ R, on a cylindrical domain R × Σ,
where M (t)V stands for the linear and N (t, V ) for the nonlinear terms
satisfying M (t) = M (t + 2π/ω0 ) and N (t, V ) = N (t + 2π/ω0 , V ) for an
ω0 > 0. As explained above, electro-convection is an anisotropic problem.
For such problems the subsequent ideas also apply in domains R2 × Σ, and
yield systems of more complicated amplitude equations.
In order to analyze the stability of the trivial solution V =0 in (10.82),
we consider the linearized system
(10.83) ∂t V = M (t)V.
Due to the translational invariance of the problem the solutions are given
by Floquet-Fourier modes
(10.84) V = ϕm (k, z, t)eikx eλm (k)t ,
with k ∈ R, m ∈ N, z ∈ Σ, and ϕm periodic in t, i.e.
ϕm (·, ·, t) = ϕm (·, ·, t + 2π/ω0 ).
If V = 0 is asymptotically stable, then we have that Re λm (k) < 0 for all
m ∈ N and k ∈ R. We assume that V = 0 becomes unstable at α = αc ,
i.e., at α = αc there exists one curve of Floquet exponents λ1 satisfying
Reλ1 (kc ) = 0 for some kc > 0. We set ε2 = α−αc > 0, write λm = λm (k, ε2 ),
and assume that all Floquet exponents possess a real part strictly less than
−σ0 for a σ0 > 0, except of λ1 (k) for k in small neighborhoods of ±kc . Hence
we have the same assumptions as in the autonomous case, except that the
curve of critical eigenvalues has to be replaced by a curve of critical Floquet
exponents.
Like for the autonomous case, it turns out that for the mathematical
analysis, it is essential to consider the Fourier transformed system w.r.t. the
unbounded spatial variable. In Fourier space (10.82) is given by
(10.85) ∂t V (k, t) = M
4(k, t)V (k, t) + N
 (V )(k, t),

with k ∈ R and V (k, t) a vector-valued function of z. We now explain

how (10.85) can be brought in a form similar to the one of §10.3 about the
universality of the GL equation in the autonomous case.
Derivation of the GL equation. For the subsequent analysis it is
4(k, t) are simple near kc .
sufficient that the critical Floquet exponents λ1 of M
However, for simplicity we assume that the linear operator M4(k, t) possesses
368 10. Dynamics of pattern and the GL equation

for every k ∈ R and t ∈ [0, 2π/ω0 ) a Floquet Schauder basis (ϕj (k, t))j∈N of
L2 (Σ), with ϕj (k, t) = ϕj (k, t + 2π/ω0 ) satisfying
4(k, t)ϕj (k, t) − λj (k)ϕj (k, t).
∂t ϕj (k, t) = M
The functions ϕj are normalized by ϕj (k, 0)L2 =1. For defining pro-
jections onto the ϕj (k, t) we consider the adjoint problem −∂t V (k, t) =
4∗ (k, t)V (k, t). Consequently, also this problem has for every k ∈ R a Flo-
M
quet Schauder basis (ϕ∗j (k, t))j∈N of L2 (Σ), with ϕ∗j (k, t) = ϕ∗j (k, t + 2π/ω0 ),
solving
4∗ (k, t)ϕ∗ (k, t) − λj (k)ϕ∗ (k, t),
−∂t ϕ∗j (k, t) = M j j
and satisfying the orthogonality
(10.86)
ϕ∗i , ϕj L2 (Σ) = δij .
A solution V (k, t) of (10.85) is expanded in terms of the Floquet functions
ϕj (k, t), i.e.

(10.87) V (k, t) = 
aj (k, t)ϕj (k, t), with aj (k, t) ∈ C,
j∈N

such that
⎛ ⎞
 
∂t ⎝ aj (k, t)ϕj (k, t)⎠ =
 ((∂t
aj (k, t))ϕj (k, t) + 
aj (k, t)∂t ϕj (k, t))
j∈N j∈N

=  4 t)ϕj (k, t) + N
aj (k, t)M(k,  (V )(k, t).
j∈N

In order to find the equations for the coefficient functions 

aj (k, t) we apply
∗
the adjoint eigenfunction
ϕj (k, t), ·L2(Σ) and find
(10.88) ∂t  j (k)
aj (k, t) = λ  t)
aj (k, t) +
ϕ∗j (k, t), N(k,
for j ∈ N. We used (10.86) and
4 t)ϕi (k, t)
−
ϕ∗j (k, t), ∂tϕi (k, t)+
ϕ∗j (k, t), M(k,
j (k)ϕi (k, t) = λ
=
ϕ∗j (k, t), λ j (k)δij .
Our derivation of the GL equation is now based on (10.88). We make the
ansatz
a1 (x, t) =εA1 (X, T )eikc x + ε2 A2,1 (X, T )e2ikc x + ε2 A0,1 (X, T ) + c.c.,
aj (x, t) =ε2 A2,j (X, T )e2ikc x + ε2 A0,j (X, T ) + c.c.,
where j ∈ N\{1}, X = εx, and T = ε2 t. With this ansatz we derive formally
a GL equation with time-periodic coefficients. We write the nonlinearity of
(10.82) in the form
(10.89) N (V ) = B(t, V, V ) + C(t, V, V, V ) + O(V 4 ),
10.6. Convection problems 369

with bilinear and trilinear symmetric terms B and C and introduce the
abbreviations

j (t, k, k − m, m) = e−ikx B(t, ϕ1 (k − m, t)ei(k−m)x , ϕj (m, t)eimx ),

B
 k, k − l1 , l1 − l2 , l2 )
C(t,
= e−ikx C(t, ϕ1 (k − l1 , t)ei(k−l1 )x , ϕ1 (l1 − l2 , t)ei(l1 −l2 )x , ϕ1 (l2 , t)eil2 x ).

For ε2 e0ix in the j th equation we obtain

(10.90) 1 (t, 0, kc , −kc )|A1 |2 ,

λj (0)A0,j = −2
ϕ∗j , B

and for ε2 e2ikc x in the j th equation

(10.91) 1 (t, 2kc , kc , kc )A21 .

λj (2kc )A2,j = −
ϕ∗j , B

For ε3 eikc x in the equation for j = 1 we obtain

2
(10.92) ∂T A1 =ν0 A1 + ν2 ∂X A1

+ 2
ϕ∗1 , j (t, kc , kc , 0)A1 A0,j
B
j∈N\{1}

+ 2
ϕ∗1 , j (t, kc , −kc , 2kc )A−1 A2,j
B
j∈N\{1}

 kc , kc , kc , −kc )A1 |A1 |2 ,
+ 3
ϕ∗1 , C(t,

with ν0 = ∂ε2 λ1 (kc , 0) and ν2 = − 12 ∂k2 λ1 (kc , 0). In (10.92) we replace A0,j
by (10.90), A2,j by (10.91), and obtain the GL equation

(10.93) 2
∂T A1 = ν0 A1 + ν2 ∂X A1 + ν3 (t, ε)A1 |A1 |2 ,

with a time-periodic coefficient ν3 (t, ε), which depends smoothly on ε2 , and

hence has a limit ν3 (t, 0). Then averaging over the highly oscillating cubic
coefficient ν3 (T /ε2 , 0) shows that in lowest order it can replaced by its mean
value ν3 , and thus we finally have the autonomous GL equation

(10.94) 2
∂T A1 = ν0 A1 + ν2 ∂X A1 + ν3 A1 |A1 |2 .

Justification of the GL approximation. For the justification of this

approximation we only remark, that, since the estimates for the linear and
nonlinear terms are exactly the same as in the autonomous case, the proof
of the approximation result goes along the lines of the autonomous case in
§10.4. This has been worked out in [SU07].
370 10. Dynamics of pattern and the GL equation

10.7. The Couette-Taylor problem

This problem consists in finding the flow of a viscous incompressible fluid
filling the domain Ω = R × Σ between two concentric cylinders of infinite
length with radii r1 , r2 and which rotate with angular velocities ω1 , ω2 . Our
presentation follows the textbook [CI94]. We denote the bounded cross
section by Σ. The flow is governed by the Navier-Stokes equations in Ω
with no-slip boundary conditions. In this section we apply our previous
theory to this classical hydrodynamical stability problem and extend our
theory to Turing-Hopf bifurcations.

r2

ω2

ω1 r1

Figure 10.10. The Couette-Taylor problem.

10.7.1. The symmetries. Before we start with the linear stability analy-
sis we discuss the symmetries of the underlying PDEs. In cylindrical coor-
dinates (r, φ, z) the velocity field v = (vr , vφ , vz ) and the pressure p satisfy
vφ vφ2 1 2 vr
∂t vr +vr ∂r vr + ∂φ vr +vz ∂z vr − = − ∂r p+ν(Δvr − 2 ∂φ vφ − 2 ),
r r ρ r r
vφ vφ vr 1 2 vφ
∂t vφ +vr ∂r vφ + ∂φ vφ +vz ∂z vφ − = − ∂φ p+ν(Δvφ + 2 ∂φ vr − 2 ),
r r ρr r r
vφ 1
∂t vz +vr ∂r vz + ∂φ vz +vz ∂z vz = − ∂z p+νΔvz ,
r ρ
and
1 vr
∂r vr + ∂φ vφ + ∂z vz + = 0,
r r
where Δ = ∂r2 + 1r ∂r + r12 ∂φ2 + ∂z2 . The equations are invariant under the
transformations of the velocity field defined by
(10.95) (τa v)(r, φ, z) =(vr , vφ , vz )(r, φ, z + a),
(10.96) (Sv)(r, φ, z) =(vr , vφ , −vz )(r, φ, −z),
(10.97) (Rϕ v)(r, φ, z) =(vr , vφ , vz )(r, φ + ϕ, z).
10.7. The Couette-Taylor problem 371

The associated transformations for the pressure p are computed in Exer-

cise 10.13. After non-dimensionalization the following three dimensionless
parameters appear:
ω = ω2 /ω1 , η = r1 /r2 , R = r1 ω1 (r2 − r1 )/ν,
with ν the kinetic viscosity, and R is called the Reynolds number. The
non-dimensionalized equations which describe this physical system possess
an exact solution, which is called the Couette flow. It is a pure azimuthal
flow whose streamlines are concentric circles. In the cylindrical coordinates
it is given by
UCou =(0, (UCou )(φ) (r), 0), (UCou )(φ) (r) = Ar + B/r,
(10.98) ω − η2 η(1 − ω)
A= ,
B= ,
η(1 + η) (1 − η)(1 − η 2 )

with associated pressure pCou = R ((UCou )(φ) (r))2 /r dr. The Couette flow
UCou satisfies the symmetries (10.95)-(10.97).
Remark 10.7.1. The Couette-Taylor problem, written as an abstract dy-
d
namical system dt x = f (x) is an example of an equivariant system. The
transformations τa , S, and Rϕ define a group Γ acting on a suitable chosen
phase space M . The system is said to be Γ-equivariant if f (γx) = γf (x)
for all γ ∈ Γ and x ∈ M . An important consequence is that if t → x(t)
solves the dynamical system, then so does t → γx(t) for all γ ∈ Γ. Works
that discuss equivariant dynamical systems include [GSS88, CK91, CL00,
GS02, Hoy06].
We use the equivariance of the system in the analysis of the bifurca-
tion scenario in §13.2.2, and in the derivation of the associated amplitude
equations here. In the following we denote the coordinate in the unbounded
axial direction by the variable x. The variable y is used for the coordinates
in the bounded cross section Σ. Since we are interested in solutions which
bifurcate from the Couette flow we choose this solution as the new origin
of the Navier-Stokes equations. The deviation (U, q) from the Couette-flow
(UCou , pCou ) satisfies in Cartesian coordinates the system
∂t U = ΔU − R[(UCou · ∇)U + (U · ∇)UCou ] − ∇q
(10.99) −R(U · ∇)U,
∇·U = 0,
with Dirichlet-boundary conditions
U = 0 at r = η/(1 − η), r = 1/(1 − η).
We choose the vanishing mean flux condition

1
[U(x) ]Σ = U (x, y) dy = 0
|Σ| Σ (x)
372 10. Dynamics of pattern and the GL equation

for x ∈ R in order to solve this problem uniquely. Since ∂x [U(x) ]Σ = 0 due

to ∇ · U = 0, this can be achieved by adding to q an unbounded term αx.
Herein, U(x) denotes the velocity component along the cylinder.

10.7.2. Linear stability analysis. The Couette flow, U ≡ 0 in (10.99), is

linearly stable if all eigenvalues have a negative real part, and it is linearly
unstable if at least one eigenvalue has a positive real part. A value λ ∈ C is
an eigenvalue if we find non-trivial solutions of
ΔU − R[(UCou · ∇)U + (U · ∇)UCou ] − ∇q = λU,
(10.100)
∇ · U = 0, U |R×∂Σ = 0, [U(x) ]Σ = 0.
Since U is translational invariant in the unbounded x-direction we can use
Fourier series
(10.101)  n (y), with k ∈ R, y ∈ Σ, n ∈ N.
U = eikx Uk
 n are of the form U
Since U is 2π-periodic in φ, the eigenfunctions U  n (y) =
k k
n (r) with m = m(k, n) ∈ Z. For fixed k we obtain discrete eigen-
eimφ Uk,m
values with Re λn ≥ Re λn+1 depending on the three parameters ω, η, and
R. Since the geometry of the apparatus can not be changed in the experi-
ment we assume that η ∈ (0, 1) is fixed. According to [CI94] the following
assertions hold for instance for η ∼ 0.883.
In the (ω, R)-plane there are two essentially different parameter regions
called PR I and PR II in the following, see Figure 10.11. For R < Rc (ω)
the largest eigenvalue satisfies Re λ1 (k, R, η, ω) < 0.
secondary bifurcations R
or mode−interactions

PR II PR I 2 real eigenvalues
4 imaginary (considered parameter region)
Rc
eigenvalues

Figure 10.11. The bifurcation scenario for η ∼ 0.8.

PR I: Numerical computations show that when ω = ω2 /ω1 ≥ 0 or for

ω slightly negative, the instability at R = Rc (ω) occurs via a curve of real
eigenvalues crossing the k-axis at a wave number kc = 0. See Figure 10.12.
PR II: For strongly counter-rotating cylinders, i.e., ω sufficiently neg-
ative, the instability at R = Rc (ω) comes via a pair of pure imaginary
eigenvalues at a wave number kc = 0. All eigenvalues which occur for the
wave number k must also occur for the wave number −k due to the reflection
symmetry (10.96) of the problem. See Figure 10.13.
10.7. The Couette-Taylor problem 373

Re λ
ε2

kc k
λ1

−σ

all other eigenvalues

Figure 10.12. PR I: The real-valued curve λ1 close to the wave num-

bers ±kc and the rest of spectrum.

real parts (double curves) imaginary parts

k
− kc

− kc kc k

Figure 10.13. PR II:The complex-valued curves λ1 , λ2 close to the

wave numbers ±kc and the rest of spectrum.

10.7.3. The GL formalism in PR I. The starting point for the GL

formalism for the solution U of (10.99) in PR I is as usual the ansatz
(10.102) U (x, y, t) = εΨGL (x, y, t) + O(ε2 ),
with
 1 (y)eikc x + c.c.,
εΨGL (x, y, t) = εA(X, T )Ukc ,0
T = ε2 t and X = εx. As above the idea is to approximately describe
slow modulations in time and in space of the linearly most unstable pattern
 1 (y)eikc x in the Couette-Taylor problem for 0 < ε2 = R − Rc  1.
U kc ,0
Formally equating the coefficients of the εj1 eij2 kc x for j1 ∈ N, j2 ∈ Z, to zero
shows that the complex valued amplitude A = A(X, T ) ∈ C should satisfy
a GL equation
(10.103) 2
∂ T A = ν2 ∂ X A + ν0 A + ν3 |A|2 A
for which the complex-valued coefficients satisfy Reν2 , Reν0 , −Reν3 > 0.
For ν2 = 0 this equation is the lowest order approximation of the reduced
system of §13.2.2 on the center manifold in case of 2π/kc -spatial periodicity.
Therefore, the coefficients ν0 and ν3 are already known from the spatially
periodic case. From the reflection symmetry (10.96) it follows that the
coefficients ν2 , ν0 , and ν3 have to be real-valued. The detailed argument is
as follows. We find
 1 (y)eikc x ) = U
S(U  1 (y)e−ikc x
kc ,0 kc ,0
374 10. Dynamics of pattern and the GL equation

such that
 1 (y)eikc x ) + c.c. + O(ε2 )
S(εψ) =εA(−X, T ) S(U kc ,0

 1 (y)e−ikc x + c.c. + O(ε2 ).

=εA(−X, T )Ukc ,0
Therefore on the level of the amplitude function A the map S possesses a
representation S : A(X, T ) → A(−X, T ). Since the Couette-Taylor problem
is invariant under S also the associated GL equation has to be invariant
under S. If we apply S we find
2
∂ T A = ν2 ∂ X A + ν0 A + ν3 |A|2 A.
Since A satisfies the complex conjugate GL equation the coefficients neces-
sarily have to satisfy νj = νj ∈ R.
In order to write (10.99) as a dynamical system in an infinite dimensional
phase space we introduce a projection on the divergence free vector fields.
The folowing results are from [Sch99b].
Lemma 10.7.2. There exists a continuous projection Π0 : (L2ul (Ω))3 → H,
where
H = {U ∈ (L2ul (Ω))3 : ∇ · U = 0, U · n|R×∂Σ = 0, [U(x) ]Σ = 0}
is equipped with the norm of (L2ul (Ω))3 . This projection additionally satisfies
Π0 ∇q = 0.

With the help of the projection Π0 we define the unbounded linear opera-
tor ΛR and the nonlinearity N (R, ·), depending smoothly on the parameters
R, η, and ω, by
ΛR U =Π0 (ΔU − R[(UCou · ∇)U + (U · ∇)UCou ]),
N (R, U ) = − RΠ0 (U · ∇)U,
such that (10.99) writes as
(10.104) ∂t U = ΛR U + N (R, U ).
For the analysis of (10.104) we need the two spaces
Z = {U ∈ H : U ∈ (Hul
2
(Ω))3 , U |R×∂Σ = 0}
and
Z ∗ = {U ∈ H : U ∈ (Hul
1
(Ω))3 }.
It turns out that
Lemma 10.7.3. The operator ΛR is the generator of an analytic semigroup
eΛR t in H with domain of definition Z satisfying
eΛR t Z ∗ →Z ≤ Ct−3/4
for t ∈ (0, 1].
10.7. The Couette-Taylor problem 375

Lemma 10.7.4. The nonlinearity N (R, U ) is locally Lipschitz-continuous

from Z to Z ∗ .

With these preparations we can now formulate the approximation theo-

rem.
θA
Theorem 10.7.5. Let θA ≥ 5. Let A ∈ C([0, T0 ], Hul (R, C)) be a solution
of the GL equation (10.103). Then there exists ε0 , C > 0 such that for all
ε ∈ (0, ε0 ) there are solutions U of the Couette-Taylor problem (10.104) with
sup U (t) − εΨGL (t)Z ≤ Cε2 .
t∈[0,T0 /ε2 ]

A slightly modified approximation theorem has been combined with the

attractivity of the set of functions in GL form to show a number of other
θ for the weakly unstable
results such as the global existence of solutions in Hul
Couette-Taylor problem [Sch99b] or the upper semicontinuity of attractors
[Sch99c]. We will come back to this in §10.8.
The proof of Theorem 10.7.5 follows by applying the abstract approxi-
mation theorem stated in Theorem 10.4.3. How to check the assumptions
(A0)-(A4) can be found in [Sch99b] where a detailed proof of Lemma
10.7.2, Lemma 10.7.3, Lemma 10.7.4, and Theorem 10.7.5 has been given.
Following Remark 10.4.2, this essentially requires modifications of the re-
sults from [CI94], where 2π/k0 -spatially periodic boundary conditions are
considered and for instance estimates for the projection Π0 on the divergence
free vector fields and resolvent estimates are given in the spaces
Hk0 ={U ∈ (L2per (Ω2π/k0 ))3 : ∇ · U = 0, U · n|R×∂Σ = 0, [U(x) ]Σ = 0},
Zk0 ={U ∈ Hk0 : U ∈ (Hper
2
(Ω2π/k0 ))3 , U |R×∂Σ = 0},
and
Zk∗0 = {U ∈ Hk0 : U ∈ (Hper
1
(Ω2π/k0 ))3 },
where Ω2π/k0 = R/((2π/k0 )Z) × Σ. Specializing such estimates to functions
 (y) gives estimates which allow to apply the Multiplier Lemma 8.3.7
eikx U
and to obtain the associated estimates in Hul θ -spaces. Again, see [Sch94a]

or [Sch99b] for details.

10.7.4. PR II: The Turing-Hopf scenario. Panel b) is the last one in

Figure 9.8 where a GL equation can be derived. It appears in reflection
symmetric (x → −x) evolution problems when a trivial ground state loses
stability and bifurcates into a temporally oscillating, spatially periodic pat-
tern. A famous example of such a system is the Couette-Taylor problem
in PR II, i.e., in case of strongly counter-rotating cylinders. In PR II the
amplitude or modulation equations are given by a set of coupled singularly
perturbed GL equations. Again the symmetries (10.95)-(10.97) can be used
376 10. Dynamics of pattern and the GL equation

to show that in the associated amplitude equation only four and not eight
different coefficients occur. See Exercise 10.14.
A simple example of an original system with these properties is a system
of coupled KS equations
∂t u = −(1 + ∂x2 )2 u − ∂x u + ε2 u + ∂x (u2 + uv + v 2 ),
(10.105)
∂t v = −(1 + ∂x2 )2 v + ∂x v + ε2 v + ∂x (u2 + uv + v 2 ),
where t ≥ 0, x ∈ R, and (u(x, t), v(x, t)) ∈ R2 . As before 0 < ε2  1 is
used as a small bifurcation parameter. The system is invariant under the
transformation (x, u, v) → (−x, −v, −u).
The linearization at (u, v) ≡ (0, 0) possesses eigenfunctions of the form
(u, v)(x, t) = (ak eikx , bk eikx ), with coefficients ak , bk ∈ C, wave number k ∈
R, and associated eigenvalues
1,2 (k, ε2 ) = −(1 − k 2 )2 ∓ ik + ε2 .
λ
We observe that Reλ 1,2 (k, ε2 ) is positive for wave numbers k close to the
values ±kc = ±1. As above we expect that the bifurcating solutions are
slow modulations in time and in space of the bifurcating pattern eix due to
1,2 . Therefore, we make the ansatz
the form of λ
u = εA(ε(x − cg t), ε2 t)ei(x−ω0 t) + ε2 A2 (ε(x − cg t), ε2 t)e2i(x−ω0 t)
1
+ ε2 A00 (ε(x − cg t), ε2 t) + c.c.,
2
v = εB(ε(x + cg t), ε2 t)ei(x+ω0 t) + ε2 B2 (ε(x + cg t), ε2 t)e2i(x+ω0 t)
1
+ ε2 B00 (ε(x + cg t), ε2 t) + c.c.,
2
with cg , ω0 ∈ R. Inserting the ansatz into the original system and equating
the coefficients in front of εn eimx eijω0 t to zero gives cg = −1 and ω0 = 1.
Setting X1 = ε(x − cg t), X2 = ε(x + cg t), T = ε2 t and eliminating the
algebraic equations for m = 0, ±2 we finally obtain
2 A(X , T ) + A(X , T )
∂T A(X1 , T ) = 4∂X 1 1 1
− 9 (A(X1 , T )|A(X1 , T )|2 + a1 + a2 ),
6
(10.106) 2 B(X , T ) + B(X , T )
∂T B(X2 , T ) = 4∂X 2 2 2
− 9 (B(X2 , T )|B(X2 , T )|2 + b1 + b2 ),
6

where
a1 =A(X1 , T )B(X2 , T )B(X2 , T ),
< =
a2 =A(X1 , T ) B(X2 , T )A(X1 , T )e−2iαT /ε +A(X1 , T )B(X2 , T )e2iαT /ε
2 2

< 2 2
=
+B(X2 , T ) A(X1 , T )B(X2 , T )e4iαT /ε +B(X2 , T )B(X2 , T )e2iαT /ε ,
10.7. The Couette-Taylor problem 377

b1 =B(X2 , T )A(X1 , T )A(X1 , T ),

< =
b2 =A(X1 , T ) B(X2 , T )B(X2 , T )e−2iαT /ε +A(X1 , T )B(X2 , T )e−4iαT /ε
2 2

< =
+A(X1 , T ) B(X2 , T )B(X2 , T )e2iαT /ε +A(X1 , T )A(X1 , T )e−2iαT /ε .
2 2

A classical GL equation is only obtained if either A = 0 or B = 0. If this

is not the case the coupled system of GL equations depends singularly on
the small bifurcation parameter 0 < ε  1. Since a2 and b2 are highly
oscillatory w.r.t. time they can be eliminated by a normal form transform,
respectively averaging, cf. Exercise 10.5. Hence, a2 and b2 can be neglected
in lowest order. The terms collected in a1 and b1 also depend singularly on
the small bifurcation parameter 0 < ε  1 via X1 and X2 . Since a1 and b1
cannot be removed in general the modulation equations in the Turing-Hopf
case still depend on the small bifurcation parameter 0 < ε  1 and are
given by
2
∂T A(X1 , T ) =ν2 ∂X 1
A(X1 , T ) + ν0 A(X1 , T )
+ ν3 A(X1 , T )|A(X1 , T )|2 + ν4 A(X1 , T )|B(X2 , T )|2 ,
∂T B(X2 , T ) =ν2 ∂X2 B(X2 , T ) + ν0 B(X2 , T )
+ ν3 B(X2 , T )|B(X2 , T )|2 + ν4 B(X2 , T )|A(X1 , T )|2 ,
with coefficients νj ∈ C for j = 1, . . . , 4. An approximation theorem for this
system can be found for (10.105) in [Sch97], and for the Couette-Taylor
problem in [Sch99b].
Remark 10.7.6. In the equation for A which depends on the space variable
X1 = ε(x − αt) the function B depends on the fast variable X1 + 2αT /ε
and vice versa in the equation for B. The idea is that due to the high
velocity of order O(1/ε) of the counter-propagating waves they only influence
each other by their mean value. Therefore, in the literature, cf. [KDL90,
Kno92, DFKM96], a so called mean-field coupled system of GL equations
is considered
∂T A(X1 , T ) = 4∂X2 A(X , T ) + A(X , T )
1 1 1
− 9 (A(X1 , T )|A(X1 , T )| + A(X1 , T )|
BL (X2 , T )|2 ),
6 2
(10.107) 2 B(X , T ) + B(X , T )
∂T B(X2 , T ) = 4∂X 2 2 2
− 9 (B(X2 , T )|B(X2 , T )| + B(X2 , T )|
AL (X1 , T )|2 ),
6 2

where  L
1

F L (Y, T ) = F (Y + X, T ) dX
2L −L
for a fixed L > 0. In [Sch97, Theorem 3.1] it has been shown that the
non-averaged system (10.106) can be approximated by the averaged system
(10.107) up to an error O(εmin(3/2,2−ν) ) w.r.t. the Hul
θ -norm if L=O(1/εν )
378 10. Dynamics of pattern and the GL equation

with 0≤ν<1. As a consequence, the equations decouple in two classical

GL equations for spatially localized solutions, cf. [PW96] in case of NLS
approximations for dispersive wave systems, see also §11.7.2 and §12.1.

10.8. Attractors for pattern forming systems

Besides the approximation theorems showing that the GL equation makes
correct predictions about the dynamics of the associated original systems,
in this book we also show similar results for dispersive systems, which can
be described by the NLS equation and the KdV equation. We explain that
all three equations are universal modulation equations appearing for various
original systems under rather weak assumptions, and that they predict the
behavior of solutions starting in special sets of initial conditions for the
associated original systems. In contrast to the cases described by the NLS
equation and the KdV equation, an additional feature for dissipative systems
is that the set of initial conditions for which the GL equation makes correct
predictions is absorbing.
We now explain this attractivity property and some of its consequences,
namely the global existence and uniqueness of solutions in a neighborhood
of the unstable origin of the original system, the upper-semicontinuity of the
rescaled original system attractor towards the associated GL attractor, and
that every solution in the original system can be shadowed for all times by
some pseudo-orbit for the GL equation.
This has been carried out in detail for the KS equation in [Sch94b,
MS95] and for the Couette-Taylor problem in in [Sch99b, Sch99c]. For
instance, for 0 < ε  1 there exists a O(1)-neighborhood of the unstable
2 in which the global existence and uniqueness of solutions
Couette flow in Hul
of the 3D Couette-Taylor problem holds. We refrain from rewriting these
papers and restrict ourselves to the explanation of the underlying ideas by
considering the Swift-Hohenberg (SH) equation as the simplest dissipative
system showing a pattern forming instability.
The plan of this section is as follows. In §10.8.1 we prove the exis-
tence of the SH attractor. In order to do so we use weighted energy es-
timates to establish the existence of an absorbing set. However, weighted
energy estimates are not optimal for small values of the bifurcation parame-
ter 0 < ε  1, since they only yield O(ε1/2 ) estimates for the diameter of the
1 . In §10.8.2 we introduce pseudo-orbits of GL approxima-
attractor in Hul
tions, and use these in §10.8.4 to show a much stronger result, namely that
the diameter of the attractor in Hul1 is O(ε) for ε → 0. In §10.8.5 we prove

the upper-semicontinuity of the rescaled original system attractor towards

the associated GL attractor.
10.8. Attractors for pattern forming systems 379

Notation. As in §8.3 we choose the weight function ρ as

ρ(x) = 1/ cosh(x) or ρ(x) = 2/(2 + x2 ),
such that |ρ (x)| ≤ ρ(x) for all x ∈ R. We recall from Definition 8.3.21 that
distZρ (b, A) = inf a∈A b − aρ ,
distZu (b, A) = inf a∈A b − a = inf a∈A supy∈R Ty b − Ty aρ ,
where (Ty a)(x) = a(x − y) and aρ is the localized norm, for instance
aHρ1 = aρH 1 . Throughout this section we use
1
Z = Hul for the phase space of the SH equation, and
(10.108) 1
Y = Hul for the phase space of the GL equation.
Thus, this distinction is somewhat artificial here, but it helps to identify SH
variables vs. GL variables, and in other applications these spaces are really
different. For instance, for the Navier-Stokes equations as an original system
we may choose the space Z = Hul 2 (R × Σ) for some bounded cross-section Σ,

cf. [Sch94a]. Moreover, we use the notations Zρ and Yρ for the respective
localized spaces.

10.8.1. Existence of the Swift-Hohenberg attractor. We establish

1 , H 1 )-attractor in the sense of Definition 8.3.18 for
the existence of a (Hul ρ
the SH equation
(10.109) ∂t u = −(1 + ∂x2 )2 u + αu − u3 = Lu + αu − u3 .
The presentation follows in many parts the lines of [MS95]. Since no max-
imum principle is available for the SH equation we use weighted energy
estimates to establish the existence of an absorbing set.
 For b ∈ (0, 1] set
ρb (x) = ρ(bx). If wedifferentiate the 2
 weighted energy ρb u dx w.r.t. time
we obtain integrals ρb u4 dx and ρb u(1 + ∂x2 )2 u dx. The second integral
can be estimated as follows.
Lemma 10.8.1. a) For all u ∈ Hul
4 , the following estimate holds:
 
(10.110) − ρb u(1 + ∂x ) u dx ≤ (2b + 2 b ) ρb u2 dx.
2 2 2 3 4
R R
b) For each d > 0 there is a D > 0 such that
  
− ρu(1 + ∂x2 )2 u dx ≤ −d ρu 2 dx + D ρu2 dx.
R R R

Proof. a) By several integrations by part we obtain

 
− ρb u(1 + ∂x2 )2 u dx = − {ρb (u2 + 2uu + u 2 ) − ρ b u 2 + ρ b uu }dx.
R R
380 10. Dynamics of pattern and the GL equation

Using |ρ b | ≤ b2 ρb and
   
ρb u 2 dx = − (ρb u ) u dx = − ρ b ( 12 u2 ) dx − ρb u u dx
 1 2 
 
= b2
2 ρb u dx − ρb uu dx ≤ 2 ρb u2 dx + ρb |uu | dx
we conclude
 
> ?
− ρb u(1 + ∂x ) u dx ≤
2 2
ρb (b4 /2 − 1)u2 + 2(1 + b2 )|uu | − u 2 dx,
R R
which yields the desired result after maximizing the integrand w.r.t. u .
b) Applying the estimates of part a) with b = 1 we find
   
  d−1 2
ρ −u(1 + ∂x2 )2 u + du 2 dx ≤ ρ u + (4 + d)|uu | − u 2 dx.
2
Maximizing w.r.t. u gives the result with D = (2d − 2 + (4 + d)2 )/4. 
With these estimates we can construct a global semiflow and an absorb-
ing ball for (10.109).
Theorem 10.8.2. The SH equation (10.109) defines a global semiflow u(t) =
St (u0 ) where for each t > 0 the nonlinear map St maps bounded sets in
Z = Hul 1 into bounded sets in H 2 . Moreover, for all α ∈ R there is a
ul
constant Δ1 (α) such that
lim sup u(t)H 1 ≤ Δ1 (α).
ul
t→∞
More precisely, there is a constant C independent of α ∈ [0, 1] such that
Δ1 (α) ≤ Cα1/4 .

Proof. The local existence and uniqueness of solutions and continuous de-
pendence on the initial condition in Hul 1 follow as in §10.2.2. We consider

the variation of constant formula

 t
u(t) = etL u(0) + e(t−s)L F (u(s)) ds, where F (u) = αu − u3 .
0
1 into H 1 , and using Lemma 8.3.7 we have
F is a smooth map from Hul ul

(10.111) etL uH 1+j ≤ C max(1, t−j/4 )uH 1

ul ul

for all t > 0, with C > 0 independent of α.

To show global existence it is sufficient to bound the Hul 1 norm. The L2
ul
estimate follows from standard weighted energy estimates using part a) of
Lemma 10.8.1, namely
 
1 d 2
ρb u dx = ρb {−u(1 + ∂x2 )2 u + αu2 − u4 }dx
2 dt R R

≤ 4b2 u2ρb + {ρb u2 (α − u2 )}dx
R
10.8. Attractors for pattern forming systems 381


≤ (4b2 + α)u2ρb − ρb u4 dx ≤ c1 − c2 u2ρb ,
R
d2

with c1 = 2 R ρb dx and c2 = d − 4b2 − α, and where b ≤ 1 was used and
d ≥ 0 is arbitrary. Assuming c2 > 0 and applying Gronwall’s Lemma to the
differential inequality yields
c1
u(t)2ρb ≤ e−2c2 t u(0)2ρb + (1 − e−2c2 t ).
c2
Letting t → ∞ and choosing the optimal d gives

 
lim sup u(t)ρb ≤ C α + b
2 2
ρb dx = Δ0 (α, b).
t→∞ R

With R ρb dx = C/b we obtain the desired estimate for α ∈ [0, 1] when
optimizing w.r.t. b = O(α1/2 ) ∈ (0, 1].
The estimate in Hul 1 is derived via the variation of constant formula

 t
u(t) = e (t−τ )L
u(τ ) + e(t−r)L (αu(r) − u(r)3 )dr.
τ

Using the smoothing properties (10.111) and u2L∞ ≤ CuL2 uH 1 the
ul ul
nonlinear terms can be estimated as
erL u3 H 1 ≤erL L2 →H 1 u3 L2
ul ul ul ul
−1/4
≤Cr uL2 u2L∞ ≤ Cr−1/4 u2L2 uH 1 .
ul ul ul

For s ∈ {0, 1} we set es (t) = u(t)Huls and find the integral inequality

@  t
−1/4
(10.112) e1 (t) ≤C (t − τ ) e0 (τ ) + α (t − r)−1/4 e0 (r) dr
τ
 t A
+ (t − r)−1/4 e0 (r)2 e1 (r) dr ,
τ

where τ ∈ [t − 1, t]. We assume e0 (r) ≤ E on [t − 1, t] and with Lemma

7.3.19 we obtain
  p 1/p   1/p
τ +δ  t −κ  τ +δ −κ τ +δ
τ  τ (t − r) e1 (r) dr dt ≤ τ |t |dt τ |e1 (t)| dtp .
 τ +δ
We choose p = 3, and estimate Iδ = ( τ e31 dt)1/3 using (10.112) by
@ A
Iδ ≤ C Eδ 3/4 + αEδ 3/4 + E 2 δ 3/4 Iδ .

Thus, for sufficiently small δ ∈ (0, 1] the integral Iδ can be estimated in

terms of α, E, and δ, namely for
δ = min{1, (4CE 2 )−4/3 }
382 10. Dynamics of pattern and the GL equation

we have Iδ ≤ 2CE(1 + α)δ 3/4 . Inserting this result into (10.112) yields
@ A
e1 (τ + δ) ≤ C Eδ −1/4 + αEδ 3/4 + E 2 δ 5/8 Iδ .

This shows that u(t) is bounded in Hul1 and that the bound only depends on

the parameters and the L2ul bound E. For small E > 0 we have e1 ≤ CE.
Hence, solutions exist globally and for t → ∞ the bound E can be replaced
by Δ0 . 
Theorem 10.8.3. The SH equation possesses an absorbing set, and a global
1 , H 1 )-attractor Aε in the sense of Definition 8.3.18.
(Hul ρ

Proof. In order to apply Theorem 8.3.22 it remains to prove the continuity

of St in the localized space Hρ1 . This proof has been given in [MS95,
Theorem 3.9]. 

10.8.2. Shadowing by GL pseudo-orbits. The estimates from Theorem

10.8.2 are rather bad for 0 < α = ε2  1. We obtain Δ1 = O(ε1/2 ) and not
O(ε) which is the size of the bifurcating stationary solutions, and the size
of the solutions, which can be approximated by the GL approximation. In
this section we show that the result for small 0 < ε2  1 can be improved
to Δ1 = O(ε) by using a sequence of GL approximations. In order to do
so we have to apply the GL approximation to solutions larger than O(ε).
Therefore, we modify our approximation ansatz to
(10.113) u(x, t) = ψδ (A)(x, t) := δ(A(X, T )eix + A(X, T )e−ix )
with T = δ 2 t, X = δx, where δ > 0 is a small number, but independent
of 0 < ε  1. Then the amplitude A is a solution of the associated GL
equation
(10.114) 2
∂T A = 4∂X A − 3|A|2 A,
A+α
where α = ε2 /δ 2 . This does not lead to any conceptual differences, but has
the advantage that we are able to consider solutions of size larger than ε.
For the precise formulation of the results and the proofs we need a number
of preparations.
Preparations. The solution operator of (10.109) in the space Z = Hul 1

is denoted with St and for (10.114) in Y = Hul with GT . To deal with the
ε 1

slow spatial scale X = δx and the mode concentration at multiples of k = 1

we need scaling operators
Sδ : Z → Y, (Sδ u)(X) = u(X/δ),
and the multiplication operator
θ : Z → Z, (θu)(x) = e−ix u(x),
which is a translation operator in Fourier space.
10.8. Attractors for pattern forming systems 383

We define mode filters which extract the critical modes close to k =

kcrit = 1. In order to do so we define an even cut-off function φ0 ∈
C0∞ (R, [0, 1]) with φ0 (k) = 1 for k ∈ [−1/6, 1/6] and φ0 (k) = 0 for k ∈
[−1/3, 1/3]. To extract the modes close to k = −1, 0, 1 we define
ej (k) = φ0 (k − j) for j ∈ {−1, 0, 1}, ec = e1 + e−1 , and es = 1 − ec .
By Lemma 8.3.7 we associate to eσ , σ ∈ {s, c, 1, 0, −1}, an operator Eσ :
Z → Z which extracts the Fourier modes belonging to wave numbers in
the relevant intervals; Ec u ∈ Z contains the critical modes of u ∈ Z, and
Es u ∈ Z the stable modes of u.
The approximation ansatz (10.113), i.e., the map ψδ : Y → Z, can be
written as
ψδ (A) = δ(θ−1 Sδ−1 A + θ−1 Sδ−1 A).
and will be interpreted as a lift from the GL problem into the original
problem (10.109). For technical reasons we introduce a modified lift ψδ :
Y → Z defined as
(10.115) ψδ (A) = δ(θ−1 E0 Sδ−1 A + θ−1 E0 Sδ−1 A),
i.e., via a cut-off of Sδ−1 A in Fourier space.
To map u ∈ Z to an associated A ∈ Y we have to extract the modes
near k = 1, and subsequently rescale, and thus define
1
(10.116) Φδ : Z → Y, u → Sδ θE1 u.
δ
Obviously Φδ (u) contains all the information of uc = Ec u, since uc = E1 u +
E−1 u with E−1 u = E1 u. Thus, if we control Φδ (u) and us , then we control
all of u.
With these connections between the underlying spaces we want to com-
pare the dynamics of (10.109) and (10.114). This is done by lifting up the
semigroup GT from Y into Z, or by extracting the GL modes from St .
Approximation, attractivity, and shadowing. We state two results
which together yield a number of consequences for the dynamics of the orig-
inal system and the comparison of attractors. The first result is an approx-
imation theorem similar to, e.g., Theorem 10.2.9, or the abstract Approx-
imation Theorem 10.4.3 with its applications Theorem 10.5.4 or Theorem
10.6.5, but with two improvements. It states that if
A0 ∈ Y = Hul
1
, with u0 − ψδ (A0 )Z sufficiently small,
then ψδ [Gδ2 t (A0 )] ∈ Z is a good approximation of Stε (u0 ), where 0 ≤ t ≤ δ −2 .
Theorem 10.8.4. (Improved Approximation) For all R1 , T1 , d > 0
there exists C, δ0 >0 such that for all 0 < ε ≤ δ ≤ δ0 the following holds. If
384 10. Dynamics of pattern and the GL equation

A0 ∈ BY (R1 ) and u0 ∈ Z with u0 − ψδ (A0 )Z ≤ dδ 5/4 , then

(10.117) sup Stε (u0 ) − ψδ (Gδ2 t (A0 ))Z ≤Cδ 5/4
0≤t≤T1 /δ 2

(10.118) Φδ (STε1 /δ2 (u0 )) − GT1 (A0 )Y ≤Cδ 1/4 .
Thus, we relax the regularity requirements for A0 , and we allow for an
initial error u0 − ψδ (A0 ). These two improvements allow to connect the
approximation result with the second result, which is the attractivity of the
GL set, i.e., of the set of all functions u ∈ Z having in leading order the
form of ψδ (A) for some appropriate A ∈ Y .
Theorem 10.8.5. (Attractivity) For each r0 > 0 there exists constants
C, T0 , R1 , δ0 > 0 such that for all 0 < ε ≤ δ ≤ δ0 we have
(10.119) distZ (STε0 /δ2 (BZ (δr0 )), ψδ (BY (R1 ))) ≤Cδ 5/4 ,
(10.120) distY (Φδ (STε0 /δ2 (BZ (δr0 ))), BY (R1 )) ≤Cδ 1/4 ,
where BZ (r) = {u ∈ Z : uZ ≤ r} and distZ (A, B) = supa∈A inf b∈B a −
bZ .
Both theorems are slight generalizations (due to the additional scaling
parameter δ ≥ ε) of [Sch94b, Lemmas 10-12] and have first been formulated
in [MS95, Theorem 4.3 and Theorem 4.2]. We review their proofs in §10.8.3.
As a first consequence of the two results and of the existence of the globally
attracting set for the GL equation, cf. Corollary 8.3.25, we have the global
existence and uniqueness of solutions in a neighborhood of the unstable
origin of the pattern forming system.
Theorem 10.8.6. (Global existence and uniqueness) There exist T0 ,
T1 , δ0 > 0 such that for all 0 ≤ ε ≤ δ ≤ δ0 and R0 sufficiently large we have
S(T0 +T1 )/δ2 (BZ (δR0 )) ⊂ BZ (δR0 ). Therefore, solutions St (u0 ) with initial
conditions u0 ∈ BZ (δR0 ) stay bounded and exist globally in time.
Remark 10.8.7. For the SH equation this is not a new result. However,
for instance for the 3D Couette-Taylor problem, where no weighted a priori
estimates are available, Theorem 10.8.6 gives a nontrivial global existence
and uniqueness result, cf. [Sch99b]. For pattern forming systems global
existence in the sense of Theorem 10.8.6 can be shown, whenever analogs to
Theorem 10.8.4, Theorem 10.8.5, and Corollary 8.3.25 can be established.
See Figure 10.14.
Moreover, it is possible to show that all solutions u(t) = Stε (u0 ) can be
shadowed by the lift of a pseudo-orbit in the GL equation. Here the notion
of pseudo-orbits is similar to but slightly different from the one in §2.4.3. A
(T1 , κ)-pseudo-orbit is pieced together from true orbits of time span T1 with
jumps of maximal size κ in between:
10.8. Attractors for pattern forming systems 385

a) t = 0 b1) t = T1 /ε2 c1) t = (T1 + T0 )/ε2

Original system

GL equation

b2) T = 0 c2) T = T0

Figure 10.14. The GL formalism can be used to obtain a priori

estimates for the solutions of the pattern forming system. A neigh-
borhood (balls in a), b1), c1)) of the origin of the pattern forming
system is mapped by the attractivity, cf. Theorem 10.8.5, into a set
(ellipse in b1)) which can be described by the GL equation. The
GL equation possesses an exponentially attracting absorbing ball,
and hence the big ball in b2) is mapped into the smaller ball in c2).
Therefore, the original neighborhood (balls in a), b1), c1)) of the
pattern forming system is mapped after a time of order O(1/ε2 )
into itself (ellipse in c1)) due to the approximation property, cf.
Theorem 10.8.4. These a priori estimates combined with the local
existence and uniqueness gives the global existence and uniqueness
of solutions of the pattern forming system in a neighborhood of the
weakly unstable origin.

Definition 10.8.8. Let T1 > 0 and κ > 0. We call a function A = A(T ) a

(T1 , κ)-pseudo-orbit in Y for the GL equation (10.114) if for all n ∈ N we
have
A(nT1 + τ ) = Gτ (A(nT )) for all τ ∈ [0, T1 ), and
A((n + 1)T1 −0) − GT1 (A(nT1 ))Y ≤ κ,
where A(T −0) = limτ T A(τ ).
Theorem 10.8.9. (Shadowing by pseudo-orbits) For all T1 >0 there
exist δ0 , C, T0 >0 such that for all δ ∈ (0, δ0 ] the following is true. For all ε ∈
(0, δ] and all initial conditions u0 with u0 Z ≤ δ the solution u(t) = St (u0 )
exists for all time, and there is a (T1 , Cδ 1/4 )-pseudo-orbit A for (10.114)
which satisfies A(0)Y ≤ C and approximates u(t) as
u(t) − ψδ (A(δ 2 t − T0 ))Z ≤ Cδ 5/4 for all t ≥ T0 /δ 2 .
386 10. Dynamics of pattern and the GL equation

As already said, this will allow us to prove that the diameter of the
attractor of the SH equation in Hul1 is not only of size O(ε1/2 ), but of size

O(ε) for ε → 0. Moreover, it allows to show the upper-semicontinuity of the

rescaled SH attractor towards the GL attractor, cf. §10.8.5.
In the following Examples 10.8.10 and 10.8.11 we illustrate the idea
of pseudo-orbits (with δ = ε) numerically for our two canonical examples,
namely the SH equation (10.109), and the KS equation (10.24), for which
the above results can be shown as well [Sch94b]. However, these numerics
also illustrate some quantitative difference between cubic and quadratic sys-
tems: For the (cubic) SH equation, the next order correction to the ansatz
ψε (A) is O(ε3 ) and indeed “very small”, also for “moderate” ε > 0. There-
fore, pseudo-orbits with O(ε2 ) jumps (in u − ψε (A)) can be considered. In
contrast, for the (quadratic) KS equation, the next order correction to the
ansatz ψε (A) is O(ε2 ).
Example 10.8.10. Pseudo-orbits for the SH equation. We consider
(10.109), i.e.,
(10.121) ∂t u = −(1 + ∂x2 )2 u + ε2 u − u3
with associated GL equation
(10.122) 2
∂T A = 4∂X A + A − 3|A|2 A.
Numerically we need to work on a finite domain, and we choose x ∈ (0, l),
l = 100π with periodic boundary conditions, and similarly X = εx ∈ (0, L),
L = εl, with periodic boundary conditions. We choose ε = 0.5,
(X−L/2)2
(10.123) A0 (X) = e− 4 , u0 = εA0 eix + c.c.,
and solve (10.121) and (10.122) with a Fourier spectral method, see [Uec09].
From the solution u(·, t) we extract Φε (u) as in (10.116), i.e., with δ = ε.
According to Theorem 10.8.4, u and ψε (A) will stay O(ε5/4 ) close on a
1/ε2 time-scale. In fact, for the SH equation we can do better, and in the
simulations we track eu,∞ (t) := u(t) − ψε (A)(t)∞ and reset A to Φε (u) if
eu,∞ (t) > ε2 . Thus, we use a slightly different notion of pseudo-orbit than
in Theorem 10.8.9, where the jump-times Tj and sizes Cε1/4 are fixed on
the GL level. Here we only jump “when needed”, based on the L∞ -norm.
However, clearly a (T1 , Cε1/4 ) pseudo-orbit in the sense of Theorem 10.8.9
can be constructed just as easily. The results are given in Figure 10.15.

Example 10.8.11. Pseudo-orbits for the KS equation. We now repeat

Example 10.8.10 for (10.24), i.e.,
(10.124) ∂t u = −(1 + ∂x2 )2 u + ε2 u + ∂x (u)2 ,
10.8. Attractors for pattern forming systems 387

(a) initial condition and its Fourier modes (b) jumps in u and in A.
1
0.8
u0(x) |ût=0 |1/4
0.2 0.2
0.5 2A ( εx) 0.6
0 0.15
0.15
0.4 ||u−εψA||∞ ||A− φδ(u)||∞
0 0.1 0.1
0.2 0.05 0.05
−0.5
0
0 50 100 150 200 250 300 0 2 4 0 50 100 20 40 60 80

(c) snapshots of solutions, Fourier modes, amplitudes, and errors

t= 20 t= 20 t= 20 t = 20
u
ε ψ (A) 1.2 0.08
0.5
δ |û(k)|1/4 0.5
1 0.06
0.4
0.8
0.3 |A|
0 0.6 0.04 |A−φ (u)|
|Φδ(u)| δ
0.4 0.2
0.02
0.2 0.1
−0.5
0 50 100 150 0 2 4 0 50 100 150 0 50 100 150
t = 32.3 t = 32.3 t= 32.3 t = 32.3
u 0.25
0.5 ε ψ (A)
1.4
δ 1.2 |û(k)|1/4 0.5
0.2
1 0.4
0.8 0.15
0 0.3 |A| |A−φδ(u)|
0.6 |Φδ(u)| 0.1
0.2
0.4
0.1 0.05
0.2
−0.5
0 50 100 150 0 2 4 0 50 100 150 0 50 100 150
t = 100 t= 100 t = 100 −3 t = 100
x 10
0.5 |û(k)|1/4 0.577
1.5 3
u 0.576
|A−φδ(u)|
0 ε ψδ(A) 1 2
0.575
|A|
0.5 |Φδ(u)| 1
0.574
−0.5
0 50 100 150 0 2 4 0 50 100 150 0 50 100 150

Figure 10.15. A pseudo-orbit for the SH equation, ε = δ = 0.5, see

Example 10.8.10. (a) shows the initial condition and its Fourier
transform, where for clarity we do not plot | u(k)| but |
u(k)|1/4 .
(k) is not very strongly localized
Since ε = 0.5 is relatively large, u
around k = ±1. (b) shows the L∞ errors and jumps in u − ψε (A)
and A − Φε (u). Since the O(ε3 ) modes are quite small, it appears
that u − ψε (A) = 0 after resetting A, but this is in fact not
the case; see Example 10.8.11 for clearer illustration. Finally, (c)
shows, at selected time steps, the solutions (on half the domain for
u and ψε (A)), the Fourier transform, the extracted critical modes
Φε (u) and their comparison with the GL evolution. The rather
narrow initial wave packet widens, and the solution develops the
expected mode distribution; the error in A − Φε (u) is essentially
localized at the edges of the wave packet, and the error in u−ψε (A)
can hardly be seen up to t = 20, say. The first resetting occurs at
t ≈ 32.3, and is triggered by the edges of the wave packet. This
continues, with a second reset at t ≈ 74, until the wave packet has
filled the domain and the solution u converges to the stationary
2ε
roll uε given in lowest order by √ cos(x), which corresponds to
√ 3
A ≡ 1/ 3.

with associated GL equation

4
(10.125) 2
∂T A = 4∂X A + A − |A|2 A.
9
388 10. Dynamics of pattern and the GL equation

Again we work on x ∈ (0, l), l = 100π, X = εx ∈ (0, L), L = εl, and choose
the initial condition (10.123), but now we choose the smaller ε = 0.25.
Moreover, we reset A to Φε (u) if eu,∞ (t) > 2ε5/4 . The results are given in
Figure 10.16.

(a) initial condition and its Fourier modes (b) jumps in u and in A.
0.8 0.6
0.4 u (x)
0 |ût=0 |1/4
0.3
0.5
2A (ε x) 0.6
0.2 0
0.4
||u−εψA||∞
0.2 ||A−φ (u)||
0 0.4 0.3 δ ∞
0.2
−0.2 0.2 0.1
0.1
−0.4
0 100 200 300 0 2 4 50 100 150 200 50 100 150 200

(c) snapshots of solutions, Fourier modes, amplitudes, and errors

t= 20 t= 20 t= 20 t= 20
u 1
ε ψ (A) 1 0.15
0.4 δ |û(k)|1/4
0.8
0.8
0.2
0.6 0.1 |A−φ (u)|
0.6 |A| δ
0
0.4 0.4 |Φ (u)|
δ 0.05
−0.2
0.2 0.2
−0.4
0 50 100 150 0 2 4 0 20 40 60 0 20 40 60
t= 57.5 t= 57.5 t= 57.5 t= 57.5
u
1.4 1.4 0.6
ε ψδ(A) |û(k)|1/4
0.5 1.2 1.2
1 1
0.4
0.8 0.8 |A| |A−φ (u)|
0 δ
0.6 0.6 |Φ (u)|
δ 0.2
0.4 0.4
−0.5 0.2 0.2

0 50 100 150 0 2 4 0 20 40 60 0 20 40 60
t= 178 t= 178 t= 178 t= 178
u
ε ψ (A) 0.6
δ |û(k)| 1/4
0.5 1.5 1.4
|A−φ (u)|
0.4 δ
1 1.3 |A|
0
1.2 |Φ (u)|
δ 0.2
0.5
−0.5 1.1
0 50 100 150 0 2 4 0 20 40 60 0 20 40 60
t= 250 t= 250 t= 250 t= 250
1.52 0.05
|û(k)|1/4 |A| |A−φδ(u)|
0.5 1.5 |Φ (u)|
δ
u 1.5 0.04
0 ε ψδ(A) 1
1.48 0.03
0.5
−0.5
1.46
0.02
0 5 10 0 2 4 0 20 40 60 0 20 40 60

Figure 10.16. A pseudo-orbit for the KS equation, ε = δ = 0.25,

see Example 10.8.11. We mainly point out the differences to Figure
10.15. First of all, (b) clearly shows that after reset 0 = u −
ψε (A) = O(ε2 ) from the modes at k = 0 and k = 2, generated
by the quadratic terms, see the Fourier plots in (c). Moreover,
even though the criterion now is eu,∞ (t) > 2ε5/4 (as opposed to
eu,∞ (t) > ε2 in our SH example), more jumps are necessary than
in the SH equation; however, the jumps still occur after O(1/ε2 )
t-intervals, which can be checked by varying ε. Finally, as in the
SH example, the solution u converges to a stationary roll, here
uε = 3ε2 cos(x) + O(ε ), and A → 2 . Again, even after convergence
2 3

(t > 250, say), these O(ε ) terms can be seen more clearly in (b)
2

than the O(ε3 ) terms in the SH example in Figure 10.15(b).

10.8. Attractors for pattern forming systems 389

10.8.3. Ideas of the proofs. We want to prove that the set of initial
conditions for which the GL equation makes correct predictions is absorbing
in the sense of Theorem 10.8.5. In other words, given an arbitrary initial
conditions u0 ∈ BZ (δr0 ), we want to show that the associated solution
t → u(t) of the SH equation (10.109) near the Turing instability develops
in such a way that there exist a time 
t and functions B = B(X) ∈ Hul 1 and

R = R(x) ∈ Hul such that

1

u(x, 
t) = εB(εx) + c.c. + ε3 R(x).
Then we can use B as initial condition for the solution A ∈ C([0, T0 ], Hul
1 ) of

the associated GL equation (10.7), and apply the Approximation Theorem

10.8.4 to predict the dynamics of the solutions t → u(t) of (10.109) near the
Turing instability.
For convenience, in the following reformulation and proof of the attrac-
tivity result we return to the case δ = ε.
Theorem 10.8.12. Let u0 be an initial condition for (10.109) with
u0 H 1 ≤ R0 ε. Then there exist T1 > 0, ε0 > 0 such that for all ε ∈ (0, ε0 )
ul
the following holds. For t = T1 /ε2 the solution u(t, u0 ) of (10.109) can be
written as
u(T1 /ε2 , u0 ) = εw1 + εw−1 + ε3 ws
with w1 = E1h wc , w−1 = w1 , and ws = Ech ws . There exists constants C1 , C2
only depending on R0 such that S1/ε θwc H 1 ≤ C1 and ws H 1 ≤ C2 .
ul ul

Proof. The variation of constant formula applied to (10.109) yields

 t
u(t) = e u0 −
tΛ
e(t−τ )Λ u3 (τ )dτ.
0
Similar to the proof of Theorem 10.8.2, by a simple contraction argument
we have the existence of ε0 > 0, T1 = T1 (R0 ), C0 = C0 (R0 ), such that for all
ε ∈ (0, ε0 ) we have a unique solution u ∈ C([0, T1 /ε2 ], Hul
1 ) with u(0) = u
0
satisfying supt∈[0,T1 /ε2 ] u(t)H 1 ≤ C0 ε.
ul

 Then we define ws = ε−3 Es u(T1 /ε2 )

and use the exponential decay
 tΛ h  −σs t
e Es u 1 ≤ e uH 1 with a σs > 0 independent of 0 < ε  1 such
ul
Hul
that
 T1 /ε2
−2 Λ −2 −τ )Λ
ws H 1 =ε−3 Es eT1 ε u0 H 1 + ε−3 Es e(T1 ε u3 (τ )dτ H 1
ul ul ul
0
≤ε−3 e−σs T1 /ε2
R0 ε + ε−3 σs−1 (1 − e −σs T1 /ε2
)C03 ε3 ≤ C2 = O(1)
for ε → 0. Accordingly, we define wc = ε−1 Ec u(T1 /ε2 ) = s1 + s2 with
 T1 /ε2
−2 −2
s1 = ε−1 Ec eT1 ε Λ u0 , s2 = −ε−1 Ec e(T1 /ε −τ )Λ u3 (τ )dτ.
0
390 10. Dynamics of pattern and the GL equation

Using the estimate

 
 
(10.126) S1/ε θetΛ E1h u 1
≤ C max(1, (εt)−1/2 )E1h uH 1
Hul ul

we find
S1/ε θs1 H 1 ≤ Cε−1 E1h u0 H 1 ≤ CR0 = O(1)
ul ul

and similarly
 T1 /ε2
2 −τ )Λ
S1/ε θs2 H 1 ≤ S1/ε θE1h e(T1 /ε H 1 →H 1 dτ
ul ul ul
0
× sup ε−1 E1h (u3 (τ ))H 1
ul
t∈[0,T1 /ε2 ]

≤ε−1 T1 C03 ε3 = O(ε2 )

1/2

for ε → 0. 

Remark 10.8.13. It can be proved that what is indicated in Theorem

10.8.12 with O(ε2 ) has more structure, namely that for every N ∈ N the
function u|t=O(1/ε2 ) possesses a representation


N
(10.127) u|t=O(1/ε2 ) = εβ(m) Am (εx)eimx + O(εN +1 )
m=0

where the function β : Z → N is defined by β(0) = 2, β(±1) = 1, and

β(m) = |m| for |m| ≥ 2. The functions Am ∈ CbN satisfy Am C N ≤ C0 for
b
a constant C0 independent of ε. This is exactly the structure of a higher
order GL approximation. Its Fourier mode distribution is plotted in Figure
10.6. Such an attractivity property has first been stated in [dES71], and
first been proved in [Eck93] by checking the occurrence of the Fourier mode
|t=O(1/ε2 ) in case of initial conditions
distribution plotted in Figure 10.6 for u
|t=0 = O(ε) ∈ L . The result has been improved in [Sch95a].
u 1

It remains to prove the improved approximation result Theorem 10.8.4

with low regularity A0 ∈ Hul
1 , compared to A ∈ H 5 in Theorem 10.2.9. For
0 ul
simplicity we again set δ = ε and only explain the main idea, see [Sch94b]
for more details. We reconsider the rather simple proof of Theorem 10.2.9
in §10, where we made the ansatz u = εψ + ε3/2 R with
1 3 3
εψ = εA(εx, ε2 t)eix + ε A (εx, ε2 t)e3ix + c.c. .
64
With A a solution of the GL equation (10.7), we found the error R to satisfy

∂t R = ΛR − 3ε2 ψ 2 R − 3ε5/2 ψR2 − ε3 R3 − ε−3/2 Res(εψ).

10.8. Attractors for pattern forming systems 391

We are interested in mild solutions and consider the variation of constant

formula
 t
R(t) = etΛ R(0) + e(t−τ )Λ (−3ε2 ψ 2 R − 3ε5/2 ψR2 − ε3 R3 )(τ )dτ + sres (t),
0
t
with sres (t) =ε−3/20 e
(t−τ )Λ Res(εψ)(τ )dτ. In §10 we used that sup
τ ∈[0,T0 /ε2 ]
Res(εψ)(τ )H 1 ≤Cε 7/2 5
if A∈C([0, T0 ), Hul ), but this can be improved. By
ul
using the smoothing properties G(T )A0 H m+j ≤ C max(1, T −j/2 )A0 Hul
m
ul
2
of the linear semigroup for the GL equation, i.e., G(T ) = e4T ∂X , and smooth-
ing estimates and concentration properties for the linear semigroup etΛ of
the original system, such as (10.126), we obtain that for A0 ∈ Hul 1 we have

sup sres (t)H 1 ≤ C = O(1),

ul
τ ∈[0,T0 /ε2 ]

and this turns out to be sufficient to prove (10.117). With similar estimates
we can prove
S1/ε θE1h sres (T0 /ε2 )H 1 ≤ C = O(1),
ul
which in the end implies (10.118).

10.8.4. Improved estimates. In this section we demonstrate how the

pseudo-orbit method can be used to improve the estimate on the size of the
attractor from O(ε1/2 ) to O(ε). First, extending the analysis of §8.3.3 we
provide sharp bounds on the decay rates for the real (i.e., c2 , c3 ∈ R) GL
equation ∂T A = c2 ∂X2 A+α A − c3 |A|2 A. The main tool is the maximum
principle which is applicable for the equation for the radial part if polar
coordinates are introduced.
Lemma 10.8.14. Let A(T ) = GT (A0 ) be a solution of the real GL equation
with A0 ∈ Hul
1 . For every ν > 0 and T > 0 we have

 
2c3 α + ν 3/2
(10.128) A(T )L∞ ≤ (1 − e−νT ) + e−νT A0 L∞ .
ν 3c3
Moreover, for each τ ∈ [0, T ) we have
 
1 M (T −τ )
√
(10.129) A(T )H 1 ≤ C 1 + √ + Me T − τ A(τ )L∞ ,
ul T −τ

where M = α  + 3c3 max{ α /c3 , A(τ )L∞ } and C is a universal constant
, c3 , and A.
independent of α

Proof. Let A(T, X) = r(T, X)eiφ(T,X) , which for r yields the equation
2
∂ T r = c2 ∂ X r − (∂X φ)2 r + α
 r − c3 r 3 .
Using the maximum principle it is possible to compare r with the solution
a−c3 a3 . Hence, if A0 L∞ = a(0), then A(T )L∞ ≤ a(T ) for all
of ∂T a = α
392 10. Dynamics of pattern and the GL equation

T > 0. From α α +ν)/(3c3 ))3/2 −νathe desired L∞ estimate

a−c3 a3 ≤ 2c3 ((
follows. In particular, we have A(T )L∞ ≤ max{ α /c3 , A(0)L∞ }.
1 estimate is derived by using the variation of constant formula
The Hul
for V = ∂X A. From
 V + c2 ∂ X
∂T V = α 2
V − c3 (2|A|2 V + A2 V )
we find
 T 
V (T ) = G(T − τ )∂X A(τ ) + G(T − r) αV (r) − c3 (. . .) dr
τ
and hence
 T
C
V (T )L2 ≤ √ A(τ )L2 + C(α + 3c3 A(r)2L∞ )V (r)L2 dr.
l,u T −τ l,u
τ
l,u

Applying Gronwall’s inequality the result follows. 

With these preparations we estimate lim supT →∞ A(T )Y for pseudo-
orbits.
Lemma 10.8.15. There exists a constant C such that for all α , κ ∈ (0, 1],
all T1 ≥ 1, and all (T1 , κ)-pseudo-orbit A = A(T ) of (10.114) we have
√
(10.130) lim sup A(T )Y ≤ CΠ, with Π = α  + (κ/T1 )1/3 + κ.
T →∞

Proof. We first estimate the L∞ -norm of gn = A(nT1 )L∞ . Then (10.128)

and the jump condition give
 
2c3 α  + ν 3/2
gn+1 ≤ (1 − e−νT1 ) + e−νT1 gn + κ.
ν 3c3
Hence, we find
 
2c3 α  + ν 3/2 κ
lim sup gn ≤ + .
n→∞ ν 3c3 1 − e−νT1
Using the estimate 1/(1−e−νT1 ) ≤ 1+1/(νT1 ) and optimizing w.r.t. ν yields
lim supT →∞ A(T )L∞ ≤ CΠ. With (10.129) (where now M ≤ C and
τ = T − 1) we conclude lim supn→∞ A(nT1 + s)Y ≤ CΠ for all s ∈ [1, T1 ).
The missing intervals can be estimated by the uniform continuity of the
semigroup GT for T ∈ [0, 1]. 
Now Theorem 10.8.9 is applied iteratively with a decreasing sequence of
5/4
δj such that in each step the size of u(t) = ψδj (Aj (δj2 t − τj )) + O(δj ) is
decreased.
Theorem 10.8.16. There exist positive constants ε0 , δ1 , and C such that
for all ε ∈ (0, ε0 ] and every solution u(t) = St (u0 ) of (10.109) with u0 Z ≤
δ1 we have
lim sup u(t)Z ≤ Cε.
t→∞
10.8. Attractors for pattern forming systems 393

Proof. We take ε0 ≤ δ1 ≤ δ0 where in the following δ0 , T0 , and T1 are

the same values as in Theorem 10.8.9. By induction over j we construct a
monotonically decreasing sequence (δj )j=1,...,J where δJ = Cε with a C ≥ 1
independent of 0 ≤ ε  1.
Assume that we know u(tj )Z ≤ δj , which is true for j = 1 with t1 = 0.
1/4
We apply Theorem 10.8.9 with δ = δj ≤ δ0 to obtain a (T1 , Cδj )-pseudo-
orbit Aj such that
5/4
u(t − tj ) − ψδj (Aj (δj2 t − T0 ))Z ≤ Cδj .
j = ε2 /δj2 ≤ 1. Using the subsequent
Note that Aj satisfies (10.114) with α
estimate (10.132) and the estimate (10.130) we find a time tj+1 with
5/4
u(tj+1 )Z ≤Cδj Aj (δj2 (tj+1 + tj ) − T0 )Y + Cδj
   
1/4 1/3 5/4 13/12
(10.131) ≤Cδj ε/δj + δj + Cδj ≤ C1 ε + δj .

Thus, we are finished, as we have shown that it is possible to define δj such

13/12
that δj+1 ≤ C1 (ε + δj ). By decreasing δ1 and ε0 , if necessary, such that
1/12
C1 δ 1 ≤ 1/3 and C1 ε0 ≤ min{1, δ1 /3} we find δ2 ≤ 2δ1 /3 for all ε ∈ (0, ε0 ].
We let h(ε, δ) = C1 (ε + δ 13/12 ) and define δj+1 = h(ε, δj ). Then δ =
h(ε, δ) has a unique fixed point δε∗ in the interval (0, δ1 ). Obviously, δε∗ ∈
(C1 ε, 2C1 ε) and δ > h(ε, δ) for δ ∈ (δε∗ , δ1 ]. Thus, the sequence δj decays
monotonically and has the limit δε∗ . Hence, there is a finite J such that
δJ ≤ 2C1 ε, and we conclude that lim supt→∞ u(t)Z ≤ δJ+1 ≤ 2C1 ε. 
Remark 10.8.17. a) From §10.8.1 we know that δ1 in Theorem 10.8.16
can be taken as Cε1/2 . Using the construction in the proof we see that it
is possible to assume δj = Cεαj with αj+1 = min{ 13 12 αj , 1}. Hence, it is
sufficient to do J = 10 iteration steps. Therefore, the overall time needed
for a solution starting in BZ (δ) to reach the ball BZ (Cε) is of order 1/ε2 .
b) More generally, if for some pattern forming system a small invariant
set of size δ = O(1) can be established (e.g., Theorem 10.8.6 for the case of
the SH equation), then the pseudo-orbit approach allows to prove that in this
invariant set there is an attractor of diameter O(ε), cf. [MS95, Sch99b].

It remains to prove the following estimate.

Lemma 10.8.18. The linear operator ψδ : Y → Z satisfies for all δ ∈ (0, 1]
the estimate
(10.132) Es ψδ (A)Z ≤ Cδ 3/2 AY for all A ∈ Y.

Proof. For v = Es θ−1 E0 Sδ−1 A we have u = Es ψδ (A) = δ(v + v), and it

is sufficient to show that vZ ≤ Cδ 1/2 AY . We introduce the spaces
394 10. Dynamics of pattern and the GL equation

Y 0 = L2ul and Z 0 = L2ul . As multiplier operators commute we find v =

θ−1 F E0 Sδ−1 A = θ−1 E0 F Sδ−1 A where F : Z → Z is associated with the
multiplier f(k) = es (k + 1) = 1 − φ0 (k) − φ0 (k + 2). The scaled operator
Fδ = Sδ F Sδ−1 : Y → Y is again a multiplier operator with kernel fδ (K) =
f(δK). We now have v = θ−1 E0 Sδ−1 Fδ A and estimate
vZ ≤ θ−1 Z→Z E0 Z 0 →Z Sδ−1 Y 0 →Z 0 Fδ Y →Y 0 AY .
Obviously, θ−1  ≤ C and E0 Z 0 →Z ≤ C by Lemma 8.3.7 since (1 +
k 2 )1/2 φ0 (k) ∈ Cb2 (R, R). Moreover, we have Sδ−1 Y 0 →Z 0 ≤ Cδ −1/2 and

Fδ Y →Y 0 ≤ C(1 + K 2 )−1/2 fδ (K)C 2 (R,R) ≤ Cδ,

b

since fδ (K) = 0 for |K| ≤ 1/(6δ). These estimates give the desired result.


10.8.5. Upper semicontinuity of the rescaled SH attractor. In The-

orems 8.3.25 and 10.8.3 we have proved the existence of the (global) attrac-
tors AG ⊂ Y for GT and Aε ⊂ Z for Stε , respectively. Our aim is to compare
these attractors and the dynamics on them. A natural question is whether
the distance of Aε from ψε (AG ) in Z tends to zero faster than ε, which is
the diameter of the two sets. The two principles, attractivity and approx-
imation property, can be used to establish the upper-semicontinuity of the
rescaled SH attractor Φε Aε , i.e., it converges to the GL attractor AG , as the
bifurcation parameter goes to zero. Like for other problems of this kind,
lower semicontinuity cannot be expected in general, cf. Example 2.4.9 or
[HR90].
Theorem 10.8.19. For every σ > 0 there exist C, ε0 > 0 such that for all
ε ∈ (0, ε0 ] we have the estimates
distY (Φε Aε , AG ) ≤ σ and distZ (Es Aε , {0}) ≤ Cε5/4 .

Proof. From Theorem 10.8.16 we know that Aε is contained in BZ (εr0 ) for

some r0 > 0. Let v ∈ Aε . Since Aε is invariant under the flow Stε , there is a
u0 ∈ Aε such that v = ST0 /ε2 (u0 ), where T0 is chosen according to Theorem
10.8.5. Hence,
Es vZ = Es STε0 /ε2 (u0 )Z ≤ Es (STε0 /ε2 (u0 ) − ψε (A0 ))Z + Es ψε (A0 )Z
≤ Cε5/4 ,
where (10.119) and (10.132) are used. This shows the second estimate.
From (10.120) we find R1 > 0 such that distY (Φε Aε , BY (R1 )) ≤ Cε1/4 .
Since AG is an attractor, there exists, for given σ, a time T2 > 0 such
that distY (GT2 BY (R1 ), AG ) ≤ σ/2. Now let v ∈ Aε be arbitrary. By the
10.9. Further remarks 395

invariance of Aε there is a u0 ∈ Aε with v = STε2 /ε2 (u0 ). Applying the

approximation result (10.118) with T1 = T2 and (10.113) we find
distY (Φε v, GT2 (BY (R1 ))) = distY (Φε STε2 /ε2 (u0 ), GT2 (BY (R1 ))) ≤ Cε1/4 .
1/4
Choosing ε0 > 0 so small that Cε0 < σ/2, we complete the proof by
distY (Φε A , AG ) ≤distY (Φε Aε , GT2 (BY (R1 )))
ε

+ distY (GT2 (BY (R1 )), AG ) ≤ σ.

10.9. Further remarks

We close this chapter with a few remarks about other approaches which
have been used to justify the GL approximation, justification results for
other dissipative systems, and so called beyond-all-order asymptotics.
Remark 10.9.1. One direction has been to establish the GL reduction as
an exact reduction method with the help of a Lyapunov-Schmidt reduction,
cf. [Mie92, Mel98, Mel00]. Unfortunately, the reduction is only exact
for so called essential solutions. These are solutions which exist globally
forward and backward in time.
Remark 10.9.2. An attempt to generalize the center manifold theorem,
cf. §3.2 and §13.1, to the case of unbounded domains has been made in
[Sch96c]. The Fourier mode distribution of the attractive solutions shows
an exponential concentration around integer multiples of the critical wave
number kc . This allows to modify the original system in such a way that
by an exponentially small change w.r.t. the small perturbation parameter
0 < ε  1 a system can be created which possesses an O(1) spectral gap.
An exact center manifold reduction is possible for the modified system which
allows to approximate the original system on an O(1/ε3 ) time scale which is
much longer than the natural GL time scale O(1/ε2 ). However, in all these
attempts from this and the previous remark the reduced systems are only
valuable from a theoretical point of view, but no practical simplification.
Remark 10.9.3. A number of approximation theorems have been proven in
slightly modified situations. The degenerated situation that the coefficient
in front of the cubic terms of the GL equation vanishes at the bifurcation
point has been discussed in [She97, BS07]. Problems with a weak spatial
periodicity have been considered in, e.g., [SD98, SU01]. The time-periodic
situation has been treated in [SU07].
Convection problems, or more general pattern forming systems with a
free surface possess an additional curve of eigenvalues touching the Re λ = 0
line at the wave number k = 0 for all values of the bifurcation parameter.
The spectrum for such systems is of the form sketched in Figure 10.17.
396 10. Dynamics of pattern and the GL equation

Re λ
ε2

−kc kc k

Figure 10.17. Spectrum for the derivation of the GL equation in case of

some additional conservation law leading to a zero eigenvalue at k = 0.
The rest of the spectrum lies strictly below the line Reλ = 0. The
magnitude of the instability defines the small perturbation parameter
0 < ε
1.

If such systems become unstable via an instability at a non-zero wave

number kc , then a GL like system can be derived as the amplitude equation
of the system. The situation, occurs for instance in the Bénard-Marangoni
problem and for the flow down an inclined plane. In the classical case, cf.
Figure 9.8, the proof of the approximation theorem is based on the fact that
the quadratic interaction of the critical modes, i.e., the modes with positive
or zero growth rates, gives only non-critical modes, i.e., modes which are
damped with some exponential rates. This is no longer true for the Bénard-
Marangoni problem and the flow down an inclined plane. Nevertheless,
approximation results have been established in [HSZ11] in case of different
group velocities in k = 0 and k = kc and in [SZ13, Zim16] in case of
vanishing group velocities. See [DKSZ16] for an attractivity result.
Such systems can also become unstable via a long wave instability at the
wave number k = 0. Approximation results for the Cahn-Hilliard equation,
which is one of the modulation equations in this case, can be found in
[Sch99a]. See also [DHV04].

Remark 10.9.4. A lattice differential pattern forming system is given by

a discrete SH equation
∂t un = −un−2 + 2un−1 − 3un + 2un+1 − un+2 + ε2 un − u3n .
The linearization at u = 0 is solved by un = eikn+λt with
λ = −e−2ik + 2e−ik − 3 + 2eik − e2ik + ε2 = −2 cos(2k) + 4 cos(k) − 3 + ε2 .
By solving λ = 0 it is easy to see that the critical wave numbers are given
by ±kc = ±π/3. Again by looking at the Fourier transformed system we
obtain the abstract pattern forming system from §10.3. The approximation
proofs from §10 can transferred almost line for line from the PDE situation
to the lattice differential equation situation. We leave this as an exercise
to the reader. We only remark that one has to be careful when computing
10.9. Further remarks 397

higher order approximations not to forget terms as A5 which break the S 1 -

symmetry of the modulation equation.
Remark 10.9.5. The GL equation can also be derived for delay-differential
equations
d
u(t) = g(u(t), u(t − τ ))
dt
with large delay τ = 1/ε and 0 < ε  1. The linearization at u = 0 is given
by
d
u(t) = au(t) + bu(t − τ ),
dt
with coefficients a, b ∈ R. It is solved by u = eλt with λ satisfying
λ = a + be−λτ .
By adjusting a and b, an instability can be created and a GL equation can
be derived. This formal analysis has been justified via some approximation
result in [YLWM13].
Remark 10.9.6. One of the major open problems in the theory of modu-
lation equations is the justification of the stochastic GL equation,
2
∂ T A = ν1 ∂ X A + ν2 A + ν3 A|A|2 + ξGL ,
with coefficients νj ∈ C, and ξGL white noise in time and space. It occurs for
pattern forming systems with thermal fluctuations in the background such
as the stochastic SH equation
∂t u = −(1 + ∂x2 )2 u + αε2 u − u3 + ε3 ξSH ,
with coefficient α ∈ R, small parameter 0 < ε  1, and ξSH white noise
in time and space. Various results already have been established. A first
approximation result has been shown in [BMS01] in case of 2π-spatially pe-
riodic boundary conditions. In [BHP05] such a result has been established
in case of large domains of size O(1/ε) with spatially periodic boundary con-
ditions. Other kinds of stochastic forcing and stochastic convergence have
been considered in a number of papers, cf. [BM13, MBK14]. A promising
approach to tackle the full problem is given in [BR13].
Remark 10.9.7. There are a number of results, e.g., [CK09, DMCK11,
KC13], that go beyond the asymptotic expansion of solutions of pattern
forming systems in powers of ε. For instance, in [DMCK11] an expansion

u= N n
n=1 ε An +rN for stationary solutions of the cubic quintic SH equation

∂t u = ε2 u − (1 + ∂x2 )2 u + εsu3 − u5
398 10. Dynamics of pattern and the GL equation

is considered, where the An fulfill GL type amplitude equations as above,

and it is shown that by optimally choosing N the error rN can be made
exponentially small, i.e.,

rN L2 ≤ C1 ε−3 eC2 /ε .

In particular, this can be used to find stationary fronts between periodic

solutions and the zero solution, cf. Figure 10.4, and to find localized pattern,
i.e., solutions that are homoclinic to the origin but are close to periodic
solutions in an (arbitrary) large interval. These fronts and homoclinics,
which come in branches snaking around the so called Maxwell point, have
already been predicted in [Pom86] due to a locking between the GL envelope
front and the phase of the underlying periodic pattern. However, to see
this phase locking one needs precisely the “beyond all (algebraic) order”
asymptotics. See also [LSAC08, BKL+ 09, ALB+ 10, UW14] for related
results (partly numerical), including localized pattern and snaking in 2D.

Further Reading. There is abundant physical, chemical and biological lit-

erature describing and/or applying the amplitude equation (over bounded
domains) or modulation equation (over unbounded domains) formalism to
various pattern forming systems. As already noted, for reaction diffusion
problems a main textbook reference is [Mur89]. For hydrodynamical prob-
lems we recommend [Man92, CI94, Str04], and a small selection of phys-
ically oriented general pattern formation and amplitude and modulation
equations textbooks and reviews is [CH93b, DK98, Pis06].
Convection in porous media is discussed in [Fow97, Chapter 14] and
[Str08]. See also [Str04, AGGP12] for the handling of various convection
problems, [Lap10] for an comprehensive work on thermal convection, and
[Rad13] for so called double-diffusive convection; for all these systems GL
type modulation equations can be derived.
For reviews of and outlooks on the active field of the derivation and anal-
ysis of modulation equations for planar pattern, here only touched upon in
§10.6.3, see for instance [Pis06, Chapter 4], [CG09, Chapter 7], or [Mer15,
Chapter 6]. Additionally, see [GS02, Hoy06] for the crucial symmetry con-
siderations in this field, and [CK99, CK01] for steps to take this to the
three-dimensional case. Most of the above books and reviews also treat
formal consequences of the derived modulation equations, for instance the
various instabilities of planar patterns, such as Eckhaus, zig-zag, cross-roll,
and some more, that may be discussed on the level of amplitude equations;
as a starting point we again recommend [Hoy06, Chapters 8 and 9].
10.9. Further remarks 399

Exercises
10.1. Consider the Kuramoto-Shivashinsky-KdV equation
∂t u = −(1 + ∂x2 )2 u + ∂x u + ∂x3 u + ε2 u + ∂x (u2 ),
with x ∈ R, t ≥ 0, 0 < ε  1, and u(x, t) ∈ R. Make an ansatz
u(x, t) =εA1 (ε(x − ct), ε2 t)ei(x−ωt)
ε2
+ ε2 A2 (ε(x − ct), ε2 t)e2i(x−ωt) +
A0 (ε(x − ct), ε2 t) + c.c.
2
and derive equations for A0 , A1 , and A2 . Eliminate A0 and A2 to derive a GL
equation for A1 .
10.2. Compute the higher order GL approximations for the SH equations. How do
the higher order approximations for the abstract system in §10.3 look like?
10.3. Replace the assumption on page 334 on the critical curve of eigenvalues λ1 by
the assumptions that λ1 (kc , 0)=iω0 ∈iR, ∂k λ1 (kc , 0)= − icg , and ∂k2 Reλ1 (kc , 0)<0,
for a wave number k = kc > 0. Derive the GL equation for this situation.
10.4. Apply Theorem 10.4.3 to the KS equation by checking the assumptions (A0)-
θ
(A4) in Hul -spaces.
10.5. Prove that a2 and b2 in (10.106) can be eliminated from the terms of order
O(1) by a normal form transform, respectively averaging.
10.6. Let θ, θ0 ≥ 0 and let g(k) satisfy |g(k)| ≤ C|k − k0 |θ0 . Prove that
 −1 (· − k0 ))L2 ≤ Cεθ0 −1/2 A
g(·)ε−1 A(ε  L2 .
θ θ+θ 0

Prove the associated estimate in H θ - and in Hul

θ
-spaces.
10.7. Show that the linear operator M , cf. page 358, defined via
4FΘ, where M 4(k, n) = ik
M Θ = F −1 M , k ∈ R, n ∈ N,
k + n2 π 2
2

is bounded from H s to H s+1 .

10.8. Compute the GL equation (10.69) for the system (10.61)-(10.63) describing
convection in porous media. Hint: Use a classical perturbation ansatz.
10.9. Use the Lyapunov-Schmidt method to prove the existence of stationary spa-
tially periodic convection rolls for the system (10.61)-(10.63) describing convection
in porous media.
10.10. Derive the curves of eigenvalues

2λm,± (k) = −(κ + 1)s ± (κ + 1)2 s2 + 4ρk2 s−1 .
and the associated eigenfunctions for Bénard’s problem from §10.6.2. Show that√
the
trivial solution becomes unstable for ρc = 27κ/4 at the wave number ±kc = ±1/ 2.
10.11. Derive the GL equation (10.76) for Bénard’s problem from §10.6.2.
10.12. Expand λ(k, l) = −(1 − k2 − l2 )2 at the wave vector (1, 0), first with k =
1 + εK and l = εL w.r.t. ε. Do the same with k = 1 + εK and l = ε1/2 L.
400 10. Dynamics of pattern and the GL equation

10.13. Compute for (10.95)-(10.97) the associated transformations for the pressure.
10.14. Use the symmetries (10.95)-(10.97) to show that in the associated amplitude
equation (10.106) in PR II only four and not eight different coefficients occur. In
order to do so make a GL ansatz in PR II and compute representations of the
transformations τa , S, and Rϕ on the level of the amplitudes.
 Chapter 11

Wave packets and the

NLS equation

The transport of information through glass fibers by light is one of today’s

key technologies. Information is encoded digitally by ones and zeroes, in
one approach by sending a light pulse through the optical fiber or not.
From a physical point of view such a light pulse consists of an underlying
electromagnetic carrier wave moving with phase velocity cp and of a pulse-
like envelope moving with group velocity cg .

Figure 11.1. 0s and 1s are encoded physically by sending a light

pulse or not; thus, for instance, the above series of electromagnetic
waves encodes the sequence 101101.

The analysis of the evolution of the associated physical system is a non-

trivial task. It shows linear and nonlinear dispersion and (weak) dissipation.
As a result the pulses are smeared out which eventually causes an unwanted
loss of information. A numerical investigation of the problem leads to a very
large system due to the multiple scaling character of the problem. The wave
length of the underlying carrier wave is around 10−7 m. Resolving such
small oscillations in a fiber of 100 km =105 m gives in a one-dimensional

401
402 11. Wave packets and the NLS equation

uniform and not very accurate spatial discretization 1012 points, even if we
ignore the transverse directions and the temporal discretization. Hence, a
direct simulation of Maxwell’s equations which describe these electromag-
netic waves is very expensive, if not impossible. Therefore, before making
any numerical investigation, the system has to be analyzed and simpler,
numerically more suitable, models have to be derived.
It turns out that the multiple scaling character of the problem is not
only a curse, but also a blessing, since it allows to separate the dynamics
of the envelope from the dynamics of the carrier wave, such that by mul-
tiple scaling analysis the NLS equation can be derived for the description
of the slow modulations in time and space of the envelope of the spatially
and temporarily oscillating wave packet. Due to the immense reduction of
the dimension of the discretized problem by this procedure the NLS equa-
tion turned out to be a very successful model. Even though arguably its
most important application is in nonlinear optics, e.g., [Agr01], the NLS
equation has also been derived for water waves [Zak68, Osb10], for waves
in DNA [SH94b] and other discrete chains, for Bose-Einstein condensates
[Pel11], in plasma physics [Deb05, Chapter 10], and in many other fields
as a universal envelope or modulation equation, cf. also [Mil06, Chapter
10]. In this chapter we explain its justification by approximation theorems
for model problems. We explain its universal character, give an overview
about approximation results, and explain some applications.

11.1. Introduction
The Nonlinear Schrödinger (NLS) equation
(11.1) 2
∂T A = iν1 ∂X A + iν2 A |A|2 ,
with T ∈ R, X ∈ R, ν1 , ν2 ∈ R, and A(X, T ) ∈ C is a universal modulation
equation which can be derived via multiple scaling analysis in order to de-
scribe slow modulations in time and space of the envelope of a spatially and
temporarily oscillating wave packet. For instance, for the nonlinear wave
equation
(11.2) ∂t2 u = ∂x2 u − u − u3 , (x ∈ R, t ∈ R, u(x, t) ∈ R),
also called the cubic Klein-Gordon equation, the ansatz for the derivation
of the NLS equation is
 
(11.3) εψNLS = εA ε(x − cg t), ε2 t ei(k0 x+ω0 t) + c.c.,
where 0 < ε  1 is a small perturbation parameter, where cg is the group
velocity, and where the basic temporal and basic spatial wave number ω0
and k0 are related by the linear dispersion relation ω02 = k02 + 1. We obtain
11.1. Introduction 403

that the envelope A of the underlying carrier wave ei(k0 x+ω0 t) has to satisfy
in lowest order the NLS equation
(11.4) 2iω0 ∂T A = (1 − c2g )∂X
2
A − 3A |A|2 .
The dynamics of the NLS equation has been discussed in §8.1. The pulse
solutions found in §8.1.1 correspond to modulating pulse solutions in the
original system, cf. Figure 11.2.

cg
O(ε)

cp

O(ε−1 )

Figure 11.2. A modulating pulse described by the NLS equation.

The envelope advancing with group velocity cg in the laboratory
frame modulates the underlying carrier wave ei(k0 x+ω0 t) advancing
with group velocity cp . The envelope evolves approximately as a
solution of the NLS equation.

Here we explain mathematical results which justify this formal approx-

imation and show that the NLS equation makes correct predictions about
the behavior of the solutions in the original system. In case of no qua-
dratic terms in the original system the proof of error estimates turns out
to be rather easy. The estimates follow by a simple application of Gron-
wall’s inequality. This is the situation as it occurs in nonlinear optics due to
symmetries. A complete proof of the estimates in this situation is given in
§11.2. The proof of the approximation property in case of quadratic terms
is presented in §11.4. In this case there are serious difficulties due to the
fact that solutions of order O(ε) have to be bounded on the long O(1/ε2 )
time scale w.r.t. t, which corresponds to an O(1) time scale w.r.t. T = ε2 t.
However, if a non-resonance condition is satisfied, then by averaging or nor-
mal form techniques the quadratic terms can be eliminated and this case
can be brought back to the situation discussed in §11.2. In §11.5 we explain
how the theory from the previous sections can be extended to the situation
of additional resonances and to quasilinear systems. In §11.6 we transfer
the theory to problems with spatially periodic coefficients and explain why
standing light pulses can theoretically occur in photonic crystals. In this
chapter we concentrate on nonlinear wave equations as original systems.
404 11. Wave packets and the NLS equation

Nevertheless, the NLS equation is a universal modulation equation. This

is explained in §11.3 and an in §11.7 the connection to nonlinear optics is
given.
It turns out that the formal derivation of the NLS equation, its univer-
sality, and for cubic nonlinearities also its justification, has many similarities
with the derivation, justification and universality of the GL equation for the
simple model problems in §10.2 and §10.3. Nevertheless, we repeat most
details to better be able to explain the crucial differences, which mainly
lie in different suitable phase spaces, in the role of so called non-resonance
conditions, and in the lack of attractivity properties of the set of modulated
waves as in §10.8.

11.2. Justification in case of cubic nonlinearities

In this section we explain how to justify the NLS equation in case of cubic
nonlinearities in the original system. For expository reasons we restrict our-
selves to the cubic Klein-Gordon equation (11.2) as original system. Our
main purpose is to prove that its solutions behave as predicted by the as-
sociated NLS equation (11.4). The NLS approximation (11.3) is formally a
good approximation if the terms which do not cancel after inserting εψNLS
into (11.2) are small. They are collected in the residual
(11.5) Res(u) = −∂t2 u + ∂x2 u − u − u3 .
If Res(u) = 0, then u is an exact solution of (11.2). With the abbreviation
E = ei(k0 x+ω0 t) we find
 
Res(εψNLS ) =εE (ω02 − k02 − 1)A
+ ε2 E ((2ik0 − 2icg ω0 )∂X A)

2
+ ε E (−2iω0 ∂T A + (1 − cg )∂X A − 3A |A|
3 2 2

+ ε3 E3 (−A3 )
+ ε4 E(2cg ∂X ∂T A)
+ ε5 E(−∂T2 A) + c.c..
By choosing ω = ω0 and k = k0 to satisfy the linear dispersion relation
ω 2 = k 2 + 1,
by choosing cg to be the linear group velocity
d  k0
cg = ω = ,
dk k=k0 ,ω=ω0 ω0
and by choosing A to satisfy the NLS equation
(11.6) 2iω0 ∂T A = (1 − c2g )∂X
2
A − 3A |A|2 ,
11.2. Justification in case of cubic nonlinearities 405

the first three lines in the residual cancel. However, we still have Res(εψNLS ) =
O(ε3 ).
Formal smallness of the residual. It turns out that by adding higher
order terms to the approximation εψNLS the residual can be made arbi-
trarily small, i.e., for arbitrary, but fixed n ∈ N with n ≥ 3 there exists an
approximation εψn with εψn −εψNLS = O(ε3 ) and Res(εψn ) = O(εn ). Since
εψn − εψNLS = O(ε3 ) the approximation εψn makes the same predictions as
εψNLS about the behavior of the solutions u of the original system. We will
show Res(εψn ) = O(εn ) for n = 4, 5. With these two examples the general
situation can be understood.
In order to obtain
(11.7) Res(εψ4 ) = O(ε4 )
we define
   
εψ4 = εψNLS + ε3 A3 ε(x − cg t), ε2 t E3 + c.c. .
We find
 
Res(εψ4 ) = ε3 E3 −A3 − (9ω02 − 9k02 − 1)A3 + O(ε4 ).
Due to the non-resonance 9ω02 − 9k02 − 1 = 9(k02 + 1) − 9k02 − 1 = 8 = 0 we
can choose A3 = −(9ω02 − 9k02 − 1)−1 A3 in order to achieve (11.7). In order
to achieve
(11.8) Res(εψ5 ) = O(ε5 )
we define
 
εψ5 = εψ4 + ε2 A12 (ε(x − cg t), ε2 t)E + ε4 A32 (ε(x − cg t), ε2 t)E3 + c.c.
where A12 and A32 are new functions to be chosen below. We find
 
Res(εψ5 ) =ε4 E −2iω0 ∂T A12 + (1 − c2g )∂X
2
A12

+ ε4 E −3A2 A12 − 6 |A|2 A12 − 2cg ∂X ∂T A
+ ε4 E3 ((9ω02 − 9k02 − 1)A32 − 3A2 A12 ) + O(ε5 ) + c.c..
By choosing A12 to satisfy the linearized NLS equation
(11.9) −2iω0 ∂T A12 + (1 − c2g )∂X
2
A12 − 3A2 A12 − 6 |A|2 A12 − 2cg ∂X ∂T A = 0
and A32 to satisfy
(9ω02 − 9k02 − 1)A32 − 3A2 A12 = 0
we achieve (11.8). In order to achieve Res(εψn ) = O(εn ) we choose
(m)
 
α
εψn = εα(m)+(j−1) Amj (X, T )Em
m=−N,...,N j=1
406 11. Wave packets and the NLS equation

with N = n − 1, X = ε(x − cg t), T = ε2 t, and α, α

 chosen according to
m 0 1 2 3 ··· m ··· N
α(m) 2 1 2 3 ··· ||m| − 1| + 1 ··· N
(m) N −1 N −2 N −1 N −2 · · ·
α N + 1 − α(m) − 2δ|m|1 ··· 1
The mode distribution of the NLS approximation is similar to one for the
GL approximation which is sketched in Figure 10.6. As before A11 satisfies
the NLS equation, the A1j for j ≥ 2 linearized NLS equations, and the Amj
for m = ±1 algebraic equations, which are linear in Amj and can be solved
w.r.t. the Amj due to the validity of the non-resonance conditions
(11.10) (mω0 )2 − (mk0 )2 − 1 = (mω(k0 ))2 − ω(mk0 )2 = 0,
√
for m = −N, . . . , N , where ω(k) = 1 + k 2 . In case of an odd nonlinearity
such as for (11.2) we can set Amj = 0 for m even.
Estimates for the residual. The formal orders of the residual can be
improved to estimates in norms. We find for instance
 Res(εψNLS )C 0 ≤ s1 + s2 + s3
b

where
s1 = 2ε3 E 3 A3 C 0 ≤ 2ε3 A3C 0 ,
b b

s2 = 4ε4 Ecg ∂X ∂T AC 0 ≤ 4ε4 cg ∂T AC 1 ,

b b

s3 = 2ε5 E∂T2 AC 0 ≤ 2ε5 ∂T2 AC 0 .

b b

We can use the right-hand side of the NLS equation to estimate ∂T AC 1
b
and ∂T2 AC 0 . For instance we have
b

1 
∂T AC 1 ≤ (1 − c2g )∂X
2
AC 1 + 3A3C 1 < ∞
b 2ω0 b b

if A ∈ Cb3 . Similarly, we find ∂T2 AC 0 < ∞ if A ∈ Cb4 . This is completely

b
analogous to the GL case in §10.2.
Lemma 11.2.1. Let A ∈ C([0, T0 ], Cb4 ) be a solution of the NLS equation
and ε0 ∈ (0, 1]. There exists a C > 0 such that for all ε ∈ (0, ε0 ) we have
sup  Res(εψNLS (t))C 0 ≤ Cε3 .
b
t∈[0,T0 /ε2 ]

Similarly, for every n ∈ N with n ≥ 4 there exists an approximation εψn

such that the following holds. Let A ∈ C([0, T0 ], CbθA ) with θA = 3(n − 3) + 1
be a solution of the NLS equation. Then for all ε0 ∈ (0, 1] there exists a
C > 0 such that for all ε ∈ (0, ε0 ):
sup  Res(εψn (t))C 0 ≤ Cεn
b
t∈[0,T0 /ε2 ]
11.2. Justification in case of cubic nonlinearities 407

and
sup εψN LS (t) − εψn (t)C 0 ≤ Cε3 .
b
t∈[0,T0 /ε2 ]

For (11.2) we have supt∈[0,T0 /ε2 ] εψN LS (t)−εψβ (t)C 0 ≤ Cε3 . If we look
b
in more detail at the approximation εψ5 we recognize that most regularity
is lost in (11.9) for A12 . Since ∂T2 A12 (·, T )C 0 has to be estimated, we need
b
A12 (·, T ) ∈ Cb4 , and so we need ∂X ∂T A(·, T ) ∈ Cb4 , respectively A(·, T ) ∈ Cb7 .
Since the structure of the approximation equations is the same for the next
orders we lose three derivatives in each step such that the estimates are
possible with θA = 3(n − 3) + 1. In fact θA can be chosen much smaller by
a number of simple tricks, cf. §11.5.1.
The equations for the error. Estimates for the residual, even in norms,
are only a necessary condition for showing that the NLS equation makes
correct predictions about the behavior of the original systems. By no means
they are sufficient. The errors can sum up in time and there are a number
of counter-examples, cf. [Sch95b, SSZ15], showing that formally derived
modulation equations make wrong predictions about the behavior of the
original system.
The error εβ R = u − εψ, the difference between the solution u and the
approximation εψ = εψn , with β and n suitably chosen, is estimated with
the help of Gronwall’s inequality. It satisfies
(11.11) ∂t2 R = ∂x2 R − R − 3ε2 ψ 2 R − 3εβ+1 ψR2 − ε2β R3 − ε−β Res(εψ).
Although there is local existence and uniqueness for (11.11) in Cbm -spaces
by the method of characteristics, here some crucial differences to Chapter
10 arise, since these spaces and this method are not suitable for obtaining
estimates on the long time scale O(1/ε2 ).
Estimates for the residual in Sobolev spaces. Sobolev spaces turn
out to be more suitable for (11.11). Hence, we assume that
A ∈ C([0, T0 ], H θA ) is a solution of the NLS equation with θA ≥ 0 suffi-
ciently large.
As a first step we have to re-estimate the residual in Sobolev spaces,
taking into account the scaling properties of the L2 -norm. As in §10.2.2 we
find

 Res(εψNLS )H θ ≤C ε3 A(ε·)2C θ A(ε·)H θ
b

+ ε4 ∂X ∂T A(ε·)H θ + ε5 ∂T2 A(ε·)H θ
 
=O ε3 A(ε·)H θ+4 .
408 11. Wave packets and the NLS equation

However,
 1/2  1/2
−1/2
A(ε·)L2 = |A(εx)| dx
2
=ε |A(X)| dX
2
= ε−1/2 AL2
R R
such that finally

 Res(εψNLS (t))H θ = O ε5/2 AH θ+4 .

It is essential that we estimate A3 with A(ε·)2C θ A(ε·)H θ and not with

b
A(ε·)3H θ .
Nevertheless, ultimately this loss of ε−1/2 is no problem, since as before,
the residual can be made arbitrarily small by adding higher order terms to
the approximation.
Lemma 11.2.2. For all n ∈ N with n ≥ 4 and θ ≥ 1 the following holds.
Let A ∈ C([0, T0 ], H θA ) with θA = 3(n − 3) + 1 + θ be a solution of the NLS
equation. Then for all ε0 ∈ (0, 1] there exists a C > 0 such that for all
ε ∈ (0, ε0 ) there is an approximation εψn with
sup  Res(εψn (t))H θ ≤ Cεn−1/2
t∈[0,T0 /ε2 ]

and
sup εψN LS (t) − εψn (t)H θ ≤ Cε3/2 .
t∈[0,T0 /ε2 ]

The equations for the error in Fourier space. For (11.2) simple
energy estimates are possible, cf. Exercise 11.1. However, in order to have a
method which also works for more general systems we use semigroup theory
and write the equations for the error as first order system in Fourier space.
We set β = n − 5/2, choose εψ = εψn and find
 − 3ε2 ψ∗2 ∗ R
 = −ω 2 R
∂t2 R  − 3εβ+1 ψ ∗ R 
∗3 + ε−β Res(εψ),
∗2 − ε2β R
√
where ω(k) = k 2 + 1. This is conveniently written as a first order system
1 = iω R
∂t R 2 ,
1 + ε2 f,
2 = iω R
∂t R
where
1 ∗2   .

f = −3ψ ∗ R − 3εβ−1 ψ ∗ R ∗3 + ε−β−2 Res(εψ)
∗2 − ε2β−2 R
iω
This system is abbreviated in the following as
 t) = Λ(k)R(k,
∂t R(k,  t) + ε2 F(k, t),
with    
0 iω(k) 0
Λ(k) = , F (k, t) =  .
iω(k) 0 f (k, t)
11.2. Justification in case of cubic nonlinearities 409

We use the variation of constant formula

 t

R(k, t) = e tΛ(k) 
R(k, 0) + ε 2
e(t−τ )Λ(k) F(k, τ ) dτ
0
in order to estimate the solutions of this system.
Lemma 11.2.3. The semigroup (etΛ(k) )t≥0 is uniformly bounded in every
Hθ0 , cf. Definition 7.3.30 and Lemma 7.3.31, i.e., there exists a C > 0 such
that we have supt∈R etΛ H 0 →H 0 ≤ C.
θ θ

Proof. We have Λ(k) = SD(k)S −1 where

   
1 1 iω(k) 0
S= and D(k) =
1 −1 0 −iω(k)
such that etΛ(k) = SetD(k) S −1 . Hence,
H 0 ≤ sup etΛ(k) C2 →C2 
etΛ u uH 0
θ θ
k∈R
and further
sup etΛ(k) C2 →C2 ≤ SC2 →C2 sup etD(k) C2 →C2 · S −1 C2 →C2
k∈R k∈R
−1
≤ SC2 →C2 S C2 →C2 < ∞.

Lemma 11.2.4. For every θ ≥ 1 there is a C > 0 such that for all ε ∈ (0, 1]
we have

Hθ
 0 + εβ−1 R
F  0 ≤ C R Hθ
 2 0 + ε2β−2 R
 3 0 +1 .
Hθ Hθ

Proof. The estimate follows from

1 
   ∗3  0 ≤ CR
 3 0,
 u ≤  uH 0 , ε−β Res(εψ) Hθ0 ≤ C, R Hθ Hθ
ω Hθ0 θ
 
     2 1 R
 0 , ψ ∗ R ∗ R
 0 ≤ Cψ  1 R
 2 0,
ψ ∗ ψ ∗ R 0 ≤ Cψ L H θ H L θ H θ
Hθ θ θ

and
1 · − k    1 · 
 1 = 2
ψ  A
0  
+ h.o.t. 1 ≤ C  A

 + h.o.t.
Lθ
ε ε Lθ ε ε L1θ
≤ CAL1 + h.o.t. ≤ CAH 0 + h.o.t.
θ θ+1


Remark 11.2.5. Note that ψ  0 = O(ε−1/2 ) such that ψ has to be es-
Hθ
timated in the Lθ -norm, respectively ψ in the Cbθ -norm in order to get ε2
1

for the most dangerous term ε2 ψ ∗ ψ ∗ R.

 This power is necessary to obtain
estimates on the natural time scale O(1/ε2 ) w.r.t. t.
410 11. Wave packets and the NLS equation

Using the previous lemmas shows that

 t 

R(t)H 0 ≤ Cε 2
R(τ  )2 0 + ε2β−2 R(τ
 ) 0 + εβ−1 R(τ  )3 0 + 1 dτ
θ Hθ Hθ Hθ
0
 t   t
≤ Cε2  ) 0 + 2 dτ ≤ 2CT0 + Cε2
R(τ  ) 0 dτ,
R(τ
H θ H θ
0 0
which holds as long as
(11.12)  )2 0 + ε2β−2 R(τ
εβ−1 R(τ  )3 0 ≤ 1.
H θ H θ

Gronwall’s inequality then yields


R(t)
2
H 0 ≤ 2CT0 e ≤ 2CT0 eCT0 = M
Cε t
θ

for all t ∈ [0, T0 /ε2 ]. Choosing ε0 > 0 such that εβ−1

0 M 2 + ε2β−2
0 M 3 ≤ 1 we
have satisfied the condition (11.12) and so proved the following approxima-
tion result.
Theorem 11.2.6. For all n ∈ N with n ≥ 4 and θ ≥ 1 the following holds:
Let A ∈ C([0, T0 ], H θA ) with θA = 3(n − 3) + 1 + θ be a solution of the NLS
equation (11.6). Then there exists an ε0 > 0 and a C > 0 such that for all
ε ∈ (0, ε0 ) there are solutions u of the original system (11.2) which can be
approximated by εψn with
sup u(t) − εψn (t)H θ < Cεn−5/2
t∈[0,T0 /ε2 ]

and as a consequence
sup u(t) − εψN LS (t)H θ < Cε3/2 .
t∈[0,T0 /ε2 ]

Remark 11.2.7. We have local existence and uniqueness of the solutions of

the nonlinear wave equation (11.2), respectively the equations for the error
(11.11), in the spaces where we proved the error estimates: Fix θ ≥ 1 and let
(u0 , u1 ) ∈ H θ+1 × H θ . Then there exists a t0 > 0 such that (11.2) possesses
a unique solution u ∈ C([−t0 , t0 ], H θ+1 ) with u|t=0 = u0 and ∂t u|t=0 = u1 .
In order to construct solutions of (11.2) we use the formula
  t  x+(t−s)
1 1 x+t
u(x, t) = (u0 (x+t)+u0 (x−t))+ u1 (ξ) dξ+ f (y, s) dy ds
2 2 x−t 0 x−(t−s)

with f (x, t) = −u(x, t) − u(x, t)3 which is based on the solution formula for
the inhomogeneous wave equation. For t0 > 0 sufficiently small the right-
hand side F (u) is a contraction in the space C([−t0 , t0 ], H θ+1 ). Thus, there
exists a unique fixed point u∗ = F (u∗ ) which is a classical solution of (11.2)
if m ≥ 2.
11.3. The universality of the NLS equation 411

The solutions exist as long as the norm of the solutions stay bounded.
By using the error estimates as a priori estimates we can guarantee that
the solutions stay bounded for t ∈ [0, T0 /ε2 ] and so we can apply the local
existence and uniqueness again and again to guarantee the existence and
uniqueness of the solutions of the error equations which are obtained from
(11.2) by a smooth change of variables.

11.3. The universality of the NLS equation

As already said, the NLS approximation can be derived in various systems.
In order to explain why this is the case, and why the NLS equation plays such
an important role, we review the derivation of the NLS equation from (11.2)
from a different point of view. This derivation will explain why the NLS
equation occurs as a universal modulation equation describing the evolution
of modulated wave packets. The underlying system will condense in the
values of the coefficients ν1 and ν2 in (11.1).
The Fourier transformed system. As for the GL approximation in
§10.3 it turns out that Fourier transform is the key for the understanding
of the universality. Hence, we consider (11.2) in Fourier space. The Fourier
transform u satisfies

(11.13) (k, t) = −ω 2 (k)

∂t2 u ∗3 (k, t),
u(k, t) − u
√
where ω(k) = k 2 + 1. By introducing w(k)  = ( 1
u(k), iω(k) (k)) we
∂t u
rewrite (11.13) into the first order system

(11.14) ∂t w(k, 4(k)w(k,

 t) = M  (w)(k,
 t) + N  t),

where
   
4(k) = 0 iω(k)  (w)(k, 0
M , N  t) = −1 ∗3 .
iω(k) 0  (k, t)
iω(k) u

This system is diagonalized forfixed wave  number k. For (11.14) the asso-
1 1
ciated transformation S = √12 is independent of k and unitary,
1 −1
i.e., S−1 = S∗ . The transformed variable z = S∗ w
 satisfies the diagonalized
system

(11.15)  z + S∗ N
∂t z = Λ  (S
z ),


with Λ(k) = diag(iω(k), −iω(k)). It turns out that the NLS equation can
be derived whenever the original system can be transformed in a system of
this form.
412 11. Wave packets and the NLS equation

General dispersive wave systems. The nonlinear wave equation (11.2)

is an example of a dispersive wave system
∂t U = LU + N2 (U, U ) + N3 (U, U, U ) + . . .
for an unknown function U = U (x, t) with values in CN and x, t ∈ R. The
operator L is linear and skew symmetric, and the terms Nj are j-multilinear
and w.l.o.g. symmetric in their arguments, and do not depend explicitly on
x ∈ R. Due to this translation invariance w.r.t. x ∈ R the linearization
∂t U = LU around U ∗ = 0 possesses solutions of the form
U (x, t) = fn (k)eikx eiωn (k)t ,
with k ∈ R and n ∈ I, where I is some finite index set and fn (k) ∈ CN . The
case of U (·, t) : Ω × R → RN , where Ω = R × Σ, Σ ⊂ Rm a bounded cross
section with m ≥ 1 leads to a countable index set I = N or I = Z and to
fn (k) ∈ L2 (Σ, CN ). We sort the eigenvalues for fixed k as ωn ≤ ωn+1 . For
real-valued systems we assume ωn = −ω−n . After an expansion
 ∞
U (x, t) = cn (k, t)fn (k)eikx dk
−∞ n∈I

in eigenfunctions, the coefficients cn (k, t) satisfy

∂t cn (k, t)
  ∞
=iωn (k)cn (k, t) + s2,n,n1 ,n2 (k, k−l, l)cn1 (k−l)cn2 (l) dl
n1 ,n2 ∈I −∞
  ∞
+ s3,n,n1 ,n2 ,n3 (k, k−l, l−m, m)cn1 (k−l)cn2 (l−m)cn3 (m) dl dm + ....
n1 ,n2 ∈I −∞

with complex-valued kernels s2,n,n1 ,n2 (k, k−l, l), s3,n,n1 ,n2 ,n3 (k, k−l, l−m, m),
etc. We have for instance
s2,n,n1 ,n2 (k, k−l, l) =
fn∗ (k), e−ikx N2 [fn1 (k−l)ei(k−l)x , fn2 (l)eilx ]L2
where fn∗ (k) is the associated adjoint eigenfunction w.r.t. the scalar product

·, ·L2 .
Derivation of the NLS equation for (11.15). Taking the Fourier
transform of the ansatz in physical space leads to the ansatz
 
−1  k − k0 2
z(k, t) = εε A1 , ε t eiω(k0 )t eicg (k−k0 )te1
ε 
(11.16)
−1  k + k0 2
+εε A−1 , ε t e−iω(k0 )t eicg (k+k0 )te2
ε
for (11.15), where
   
1 0
e1 = and e2 = .
0 1
11.3. The universality of the NLS equation 413

The notation εε−1 refers to the amplitude scaling ε and the wave number
scaling ε−1 , but in the following we shorten this to εε−1 = 1. Since the
Fourier modes of the wave packet are concentrated in an O(ε) neighborhood
of the basic wave numbers ±k0 the evolution of the wave packet will be
strongly determined by the curves ±ω at ±k0 . At eiω(k0 )t eicg (k−k0 )te1 we
find
1 + iεcg K A
iω(k0 )A 1 + ε2 ∂T A
1

1 + iε∂k ω(k0 )K A
=iω(k0 )A 1 + i ε2 ∂ 2 ω(k0 )K 2 A
1
k
2
3i  1 ∗ A
−1 + O(ε3 ),
+ ε2 A1 ∗ A
4ω(k0 )
where k = k0 + εK, A 1 = A1 (K, T ), and where we used
    
 k − l − k0 iω(k0 )t icg (k−l−k0 )t  l − k0
A ,T e e A , T eiω(k0 )t eicg (l−k0 )t dl
ε ε
R
  
 k − 2k 0 
=ε A − m, T A(m, T )e2iω(k0 )t eicg (k−2k0 )t dm .
ε
R

At ε0 and ε1 we obtain the linear dispersion relation and the linear group
velocity. At ε2 we obtain a NLS equation. Undoing the transformation
w = Sz and u = w1 gives a multiple of the original approximation
  
(k, t) = √
u
1
A1 k − k0 , ε2 t eiω(k0 )t eicg (k−k0 )t
2 ε
  
 k + k0 2 −iω(k0 )t icg (k+k0 )t
+A−1 ,ε t e e .
ε

Derivation of the NLS equation in the general situation. By this

procedure it is clear that the NLS equation occurs as a modulation equation
for dispersive wave systems whenever the Fourier transform of the initial
condition is strongly concentrated at a wave number k0 = 0 and when the
concentration and the amplitude are of the correct order. It also occurs in
case k0 = 0 if ω(0) = 0, cf. Exercise 11.1.
W.l.o.g. let us assume that we derive the NLS equation for the curve ω1
and that we have a real-valued system. We make the ansatz
1,1 (ε−1 (k − k0 ), ε2 t)eiω(k0 )t eicg (k−k0 )t
c1 (k, t) =A
1,−2 (ε−1 (k + 2k0 ), ε2 t)e−2iω(k0 )t eicg (k+2k0 )t
+ εA
1,0 (ε−1 k, ε2 t)eicg kt
+ εA
1,2 (ε−1 (k − 2k0 ), ε2 t)e2iω(k0 )t eicg (k−2k0 )t ,
+ εA
414 11. Wave packets and the NLS equation

Figure 11.3. The derivation of the NLS equation is based on the

concentration of the Fourier modes at a certain wave number. The
figure shows the curves of eigenvalues and the concentration of the
Fourier modes. Hence, for the evolution of these modes only the
curves of eigenvalues close to these wave numbers play a role.

−1,−1 (ε−1 (k + k0 ), ε2 t)e−iω(k0 )t eicg (k+k0 )t )

c−1 (k, t) =A
−1,−2 (ε−1 (k + 2k0 ), ε2 t)e−2iω(k0 )t eicg (k+2k0 )t
+ εA
−1,0 (ε−1 k, ε2 t)eicg kt
+ εA
−1,2 (ε−1 (k − 2k0 ), ε2 t)e2iω(k0 )t eicg (k−2k0 )t ,
+ εA
n,−2 (ε−1 (k + 2k0 ), ε2 t)e−2iω(k0 )t eicg (k+2k0 )t
cn (k, t) =εA
n,0 (ε−1 k, ε2 t)eicg kt
+ εA
n,2 (ε−1 (k − 2k0 ), ε2 t)e2iω(k0 )t eicg (k−2k0 )t ,
+ εA

for |n| ≥ 2 and complex valued functions A n,j , where An,−j = An,j in
physical space. With k − k0 = εK we find at ε2 for the modes concentrated
at k0 that
1,1 (K, T )
∂T A
=i∂k2 ω1 (k0 )K 2 A1,1 (K, T )/2
  ∞
+2 s2,1,1,n (k0 , k0 , 0) A n,0 (κ, T ) dκ
1,1 (K − κ, T )A
n∈Z −∞
  ∞
+2 s2,1,1,n (k0 , −k0 , 2k0 ) A n,2 (κ, T ) dκ
1,−1 (K − κ, T )A
n∈Z −∞
 ∞  ∞
+ 3s3,1,1,1,1 (k0 , k0 , k0 , −k0 ) 1,1 (K − κ1 , T )
A
−∞ −∞
1,1 (κ1 − κ2 , T )A
×A 1,−1 (κ2 , T ) dκ2 dκ1 ,
11.3. The universality of the NLS equation 415

where we used the symmetry of the multilinear terms in their arguments,

cf. (11.21). With k − jk0 = εK we find at ε in the n-th equation for the
modes concentrated at jkc for j = 0, 2 that
0 =iωn (0)An,0 (K, T )
 ∞
+ 2s2n1−1 (0, −kc , kc ) A 1,1 (κ, T ) dκ,
1,−1 (K − κ, T )A
−∞
0 =i(ωn (2k0 ) − 2ω1 (k0 ))An,2 (K, T )
 ∞
+ s2n11 (2kc , kc , kc ) A 1,1 (κ, T ) dκ,
1,1 (K − κ, T )A
−∞
see (11.22)-(11.23) for an example. Under the non-resonance conditions
ωn (0) = 0, ωn (2k0 ) − 2ω1 (k0 ) = 0,
these two algebraic relations determine An,0 and An,2 in terms of A1,1 and
A1,−1 such that A1,1 finally satisfies the NLS equation in Fourier space,
namely
1,1 (K, T ) =−iν1 K 2 A
∂T A 1,1 (K, T )
 ∞ ∞
+ iν2 A 1,1 (κ−
1,1 (K−κ, T )A 1,−1 (
κ, T ) A κ, T ) d
κ dκ,
−∞ −∞
with coefficients
ν1 = − ∂k2 ω1 (k0 )/2,

ν2 =4 s2,1,1,n (k0 , k0 , 0)s2,n,1,−1 (0, −kc , kc )/ωn (0)
n∈Z

+2 s2,1,1,n (k0 , −k0 , 2k0 )s2,n,1,1 (2k0 , k0 , k0 )/ωn (2k0 )
n∈Z
− 3s3,1,1,1,1 (k0 , k0 , k0 , −k0 ).
Therefore, the NLS equation is the universal modulation equation describ-
ing slow modulations in time and space of a propagating wave packet in
dispersive systems
(11.17) ∂t 
cj (k, t) = iωj (k)
cj (k, t) + nonlinear terms,
where j is in some index set and where the nonlinear terms have some
convolution structure.
Remark 11.3.1. We have seen that for (11.2) the valdity of the non-
resonance condition (11.10) is necessary for the construction of higher order
approximations. For general dispersive wave systems (11.17) for the validity
of a counterpart to Lemma 11.2.1 the validity of the non-resonance condition
(11.18) (mω1 (k0 ))2 − ωj (mk0 )2 = 0,
is necessary for m= − N, . . ., N , where N =n − 1, and j in some index set.
416 11. Wave packets and the NLS equation

11.4. Quadratic nonlinearities

In this section we explain how to justify the NLS approximation in case of
quadratic nonlinearities. For expository reasons we restrict ourselves to
(11.19) ∂t2 u = ∂x2 u − u + u2 ,
with x ∈ R, t ∈ R, and u(x, t) ∈ R as original system. The ansatz for the
derivation of the NLS equation is then given by
 
εψNLS =εA1 ε(x − cg t), ε2 t ei(k0 x+ω0 t) + c.c.
 
(11.20) + ε2 A2 ε(x − cg t), ε2 t e2i(k0 x+ω0 t) + c.c.
 
+ ε2 A0 ε(x − cg t), ε2 t .
We find as before at εE the linear dispersion relation and at ε2 E the condi-
tion for the linear group velocity cg . At ε3 E we find
(11.21) 2iω0 ∂T A1 = (1 − c2g )∂X
2
A1 + 2A1 A0 + 2A2 A−1 .
The algebraic relations which are found at ε2 E 0 and ε2 E 2
(11.22) ε2 E 0 : 0 = −A0 + 2A1 A−1 ,
(11.23) 2
ε E :2
0 = −(−4ω02 + 4k02 + 1)A2 + A21 ,
can be solved w.r.t. A0 and A2 since
−4ω02 + 4k02 + 1 = −(2ω(k0 ))2 + ω(2k0 )2 = 0.
Inserting the solution for A0 and A2 into the equation for A1 finally yields
the NLS equation
(11.24) 2iω0 ∂T A1 = (1 − c2g )∂X
2
A1 + γA1 |A1 |2 ,
with
2
γ =4+ .
−4ω02
+ 4k02 + 1
Like in case of cubic nonlinearities the residual
Res(u) = −∂t2 u + ∂x2 u − u + u2
can be made arbitrarily small by adding higher order terms, i.e., we have
Lemma 11.4.1. For all n ∈ N with n ≥ 4 and θ ≥ 1 the following holds.
For A ∈ C([0, T0 ], H θA ) with θA = 3(n − 3) + 1 + θ and ε0 ∈ (0, 1] there
exists a C > 0 such that for all ε ∈ (0, ε0 ) there is an approximation εψn
with
sup  Res(εψn (t))H θ ≤ Cεn−1/2
t∈[0,T0 /ε2 ]
and
sup εψNLS (t) − εψn (t)C 0 ≤ Cε2 .
b
t∈[0,T0 /ε2 ]
11.4. Quadratic nonlinearities 417

In order to prove that the solution A1 of the NLS equation (11.24)

predicts the behavior of the solutions u of the original system correctly we
estimate as before the difference εβ R = u−εψ between the correct solution u
and its approximation, and as before we choose εψ = εψn with β = n − 5/2.
This difference satisfies
∂t2 R = ∂x2 R − R + 2εψR + εβ R2 + ε−β Res(εψ).
As a first order system in Fourier space this yields
1 = iω R
∂t R 2 ,
(11.25) 
     ∗2 −β
∂t R2 = iω R1 + iω 2εψ ∗ R1 + ε R1 + ε Res(εψ) ,
1 β 
√
where again ω(k) = k 2 + 1. The simple argument of the last section no
longer works because of the new O(ε)-term 2εψ ∗ R 1 . In principle this term
can give some exponential growth of order O(exp(εt)) which is not O(1)-
bounded on the time scale of order O(1/ε2 ). However, this term is oscillatory
in time and can be eliminated by averaging or a normal form transformation
such that it finally has an O(1)-influence on the size of the solutions. This
observation goes back to [Kal88].
Remark 11.4.2. Normal form transformations have already been consid-
ered in Part I of this book. For PDEs the idea is very similar. For the
abstract evolutionary system
∂t u = Au + NQ (u) + Nc (u),
where NQ stands for quadratic and Nc for the higher order terms, we seek a
near identity change of coordinates v = u − K(u) to eliminate the quadratic
terms NQ (u) and to transfer them into higher order terms. We find
∂t v =∂t u − K (u)∂t u
=Au + NQ (u) + Nc (u)−K (u)Au−K (u)NQ (u)−K (u)Nc (u)
=Av + AK(u)−K (u)Au + NQ (u) + Nc (u)−K (u)NQ (u)−K (u)Nc (u).
In order to eliminate the quadratic terms we choose K to satisfy
AK(u) − K (u)Au + NQ (u) = 0
such that after the transformation
∂t v = Av + Nc (u) − K (u)NQ (u) − K (u)Nc (u).
As an instructive example [Str89, Page 38] we consider
(11.26) i∂t u = Δu − ∇u · ∇u
and choose v = u − 12 u2 , i.e., K(u) = 12 u2 , cf. Example 11.4.3 for an
explanation how to find this transformation. We have
1
AK(u) − K (u)Au + NQ (u) = −iΔ( u2 ) − u(−iΔu) + i∇u · ∇u = 0
2
418 11. Wave packets and the NLS equation

such that
∂t v = Av − K (u)NQ (u) = −iΔu − ui∇u · ∇u
In order to obtain an evolutionary problem for v we have to invert the
transformation v = u − K(u). For small v this is possible by applying the
implicit function theorem.

We apply this idea to the equation for the error and eliminate the O(ε)-
term 2εψ ∗ R
1 with a normal form transformation. In order to do so we first
diagonalize (11.25) by introducing
     
R 
ψ 1 1 1
R=S −1 1 
, Ψ=S −1
, where S = √ .
R2 1  2 1 −1
iω(k) ∂t ψ(k))

We find
(11.27) ∂t R = ΛR + 2εB(Ψ, R) + εβ B(R, R) + ε−β RES(εΨ),
with Λ a symmetric linear map and B(·, ·) a bilinear map, which are given
in Fourier space by
   
 iω 0 −1 0
Λ = , RES(εΨ) = S ,
0 −iω Res(εψ)
 
 U, U  U
 ) = 1 S −1 B(S  , SU
 ), B(
 U
 , V ) = 0
B( iω 1 ∗ V1 ,
U
 = (U
where U 2 ). Then we make a near identity change of variables
1 , U

(11.28) w = R + εQ(Ψ, R)
with Q an autonomous bilinear map. This gives
∂t w =∂t R + εQ(∂t Ψ, R) + εQ(Ψ, ∂t R)
=ΛR + 2εB(Ψ, R) + εβ B(R, R) + ε−β RES(εΨ) + εQ(∂t Ψ, R)
+ εQ(Ψ, ΛR + 2εB(Ψ, R) + εβ B(R, R) + ε−β RES(εΨ))
=Λw − εΛQ(Ψ, R) + 2εB(Ψ, R) + εβ B(R, R) + ε−β RES(εΨ)
+ εQ(∂t Ψ, R) + εQ(Ψ, ΛR + 2εB(Ψ, R) + εβ B(R, R) + ε−β RES(εΨ)),
and so
∂t w = Λw + ε (−ΛQ(Ψ, R) + Q(∂t Ψ, R)
(11.29)
+Q(Ψ, ΛR) + 2B(Ψ, R)) + O(ε2 ).
In order to eliminate the dangerous term 2εB(Ψ, w) we have to find a bilinear
Q such that
(11.30) −ΛQ(Ψ, R) + Q(∂t Ψ, R) + Q(Ψ, Λw) + 2B(Ψ, R) = 0.
11.4. Quadratic nonlinearities 419

In this form this equation is hard to analyze. For its simplification we first
use that
   
√ εA1 ε(x − cg t), ε2 t ei(k0 x+ω0 t) + O(ε2 )
ε∂t Ψ = 2∂t
εA−1 ε(x − cg t), ε2 t e−i(k0 x+ω0 t) + O(ε2 )
   
√ iω0 εA1 ε(x − cg t), ε2 t  ei(k0 x+ω0 t) + ε2 G1
= 2
−iω0 εA−1 ε(x − cg t), ε2 t e−i(k0 x+ω0 t) + ε2 G−1
 1 = O(1). Hence, (11.30) is given in lowest order by
with G L θ

(11.31) −ΛQ(Ψ, R) + Q(Υ, R) + Q(Ψ, Λw) + 2B(Ψ, R) = 0,

 
(Λ(k0 )Ψ)1
where Υ = . In Fourier space we have for the j-th com-
(Λ(−k0 )Ψ)−1
ponent of B that
 
(B(  R))
 ψ,  j= bj (k, k − l, l)ψm (k − l)R
 n (l) dl,
mn
m,n=1,2 R

with bjmn = bjmn (k, k − l, l) some smooth kernel. Thus, we make the same
ansatz for the j-th component of Q, namely
 
  
(Q(ψ, R))j = qmn
j
(k, k − l, l)ψm (k − l)R
 n (l) dl,
m,n=1,2 R
j j
with qmn = qmn (k, k − l, l) some kernel which we have to compute. Inserting
these representations in (11.31) yields the relations
(11.32) i(ωj (k) − ω1 (k0 ) − ωn (l))j
q1n (k, k − l, l) = 2bj1n (k, k − l, l),
and
(11.33) i(ωj (k) − ω2 (−k0 ) − ωn (l))j
q2n (k, k − l, l) = 2bj2n (k, k − l, l),
with ω1,2 (k) = ±ω(k).
Example 11.4.3. For system (11.26), where the approximation is replaced
by a general function and where only one curve of eigenvalues is involved,
the associated relation would be
q(k, k − l, l) = b(k, k − l, l),
i(ω(k) − ω(k − l) − ω(l))
with ω(k) = −k 2 and b(k, k − l, l) = i(k − l)l. We find
ω(k) − ω(k − l) − ω(l) = −k 2 + (k − l)2 + l2 = −2(k − l)l
such that q(k, k − l, l) = − 12 .

In Fourier space, the approximation εψ has order one amplitude only

close to the wave numbers k0 in the first component and close to the wave
numbers −k0 in the second component. Therefore, only wave numbers |k −
l − k0 | ≤ δ for (11.32) and only wave numbers |k − l + k0 | ≤ δ for (11.33)
420 11. Wave packets and the NLS equation

have to be taken into account for a small δ > 0 independent of 0 < ε  1.

j
For the other wave numbers we set qmn = 0. Hence, in order to solve (11.32)
j
w.r.t. the qmn we need the non-resonance conditions
(11.34) inf inf |(ωj (k) − ω1 (k0 ) − ωn (l))| ≥ C > 0
j,n∈{1,2} k,l∈R,|k−l|≤δ

and
(11.35) inf inf |(ωj (k) − ω2 (−k0 ) − ωn (l))| ≥ C > 0
j,n∈{1,2} k,l∈R,|k−l|≤δ

for this δ > 0 fixed. The validity of the non-resonance conditions can be
checked graphically by looking for intersections of the curves k → ±ω(k)
and k → ω(k0 ) ± ω(k − k0 ), see Figure 11.4. Since the asymptotes k → ±k
and k → ω(k0 ) ± (k − k0 ) to these curves are separated, for no value of k0 a
quadratic resonance occurs.

Figure 11.4. The curves k → ±ω(k) and k → ω(k0 ) ± ω(k − k0 ) in case

k0 = 0 and k0 = 1. There is no intersection and so no quadratic reso-
nance.

Since
sup sup |bjmn (k, k − l, l)| ≤ C < ∞,
j,m,n∈{1,2} k,l∈R,|k−l|≤δ
(11.34) immediately implies
sup sup |j
qmn (k, k − l, l)| ≤ C < ∞.
j,m,n∈{1,2} k,l∈R,|k−l|≤δ

As a consequence we obtain
 1 R
(Q(ψ, R))H θ ≤ CψC θ RH θ ≤ Cψ  2.
b
L L θ θ

Thus, the transformation (11.28) can be inverted with the help of Neumann’s
series for ε > 0 sufficiently small. We denote the inverse by R = Rε (w) =
O(1). Therefore, (11.29) transforms into
(11.36) ∂t w = Λw + ε2 F
11.5. Extension of the theory 421

with
ε2 F =εβ B(Rε (w), Rε (w)) + ε−β RES(εΨ) + ε2 Q(G, Rε (w))
+ εQ(Ψ, 2εB(Ψ, Rε (w)) + εβ B(Rε (w), Rε (w)) + ε−β RES(εΨ)).
Since ε2 F obeys the same estimates as the one in Lemma 11.2.4 the rest of
the proof of the approximation property from §11.2 applies line for line to
(11.36).

11.5. Extension of the theory

There are various physically relevant systems, especially the water wave
problem, where the previous non-resonance conditions (11.38) and (11.39)
are not satisfied. These are systems with trivial resonances at the wave
number k = 0, additional non-trivial resonances, and quasilinear nonlinear-
ities. Here we give a short overview about the strategies which have been
developed in last years to overcome these difficulties. We start this section
with some technical improvements of the previous analysis.

11.5.1. Some technical improvements. By two simple technical

changes in the previous proofs the non-resonance conditions can be sim-
plified and less regularity for the initial conditions of the NLS equation is
needed.
Less regularity is needed. Since in Fourier space the approximation
is strongly concentrated at integer multiples of the basic wave number k0
we only make a small error if we cut-off the approximation outside small
neighborhoods of k0 Z.
The formal expansion of the curve of eigenvalues and of the kernels in
the multilinear maps can be estimated with the aid of the following lemma.
Lemma 11.5.1. Let θ, θ0 ≥ 0, θ∞ ∈ R, and let g(k) satisfy
|g(k)| ≤ C min(|k − k0 |θ0 , (1 + |k|)θ∞ ).
Then
g(·)ε−1 A(ε  2
 −1 (· − k0 )) 2 ≤ Cεθ0 −1/2 A .
L θ L max(θ+θ∞ ,θ0 )

Proof. This follows immediately from the fact that the left-hand side of this
inequality can be estimated by
 
 ε 

≤ sup g(k)(1+ ) − max(θ+θ∞ ,θ0 ) θ  −1 ·) 2
(1+|k−k0 |)  ε−1 A(ε
k∈R |k−k0 | Lmax(θ+θ ,θ )
∞ 0
 
 |εK| θ 0 (1+|k0 +εK|) θ+θ ∞ 
≤ sup C min( , 
) ε−1/2 A(·)
θ θ+θ ∞ L2max(θ+θ ,θ )
K∈R (1+|K|) 0 (1+|K|) ∞ 0

where the loss of ε−1/2 is due to the scaling properties of the L2 -norm. 
422 11. Wave packets and the NLS equation

This lemma can be applied for instance to

1
g(k) = ω(k) − ω0 − cg k − ω (k0 )k 2 = O(|k|3 )
2
such that the residual terms can also be estimated for the diagonalized
system in Fourier space.
Example 11.5.2. With the simple approach from §11.2 we need θA ≥ θ + 4
in order to estimate the error in H θ if we take the approximation ψ2 . The
reason is as follows. The highest loss of regularity comes from ∂T2 A3 H θ
which can be estimated by ∂T2 AH θ . This can be estimated further via
the right-hand side of the NLS equation by AH θ+4 . For the associated
diagonalized system we have to expand ω at k0 up to order three, i.e., we
use
1
|ω(k) − (ω(k0 ) + ω (k0 )(k − k0 ) + ω (k0 )(k − k0 )2 )χ(|k − k0 |)|
2
≤C min(|k − k0 |3 , |k|)
where χ is a C0∞ function with χ(k) ∈ [0, 1], χ(k) = 0 for |k| ≥ 2 and
χ(k) = 1 for |k| ≤ 1. Hence, the application of Lemma 11.5.1 allows to
reduce the above value of θA to a value satisfying θA ≥ max(θ + 1, 3) since
for all other terms less derivatives are needed.

Sometimes it is possible to reduce the value for θA even further by using

a cut-off function in Fourier space to mollify the original approximation. Let
again χ ∈ C0∞ with 0 ≤ χ(k) ≤ 1, χ(k) = 0 for |k| ≥ 2, and χ(k) = 1 for
 −1 ·) we work with χ(·)ε−1 A(ε
|k| ≤ 1. Instead of ε−1 A(ε  −1 ·). Then we have

(11.37) g(·)ε−1 A(ε  2

 −1 (·)) 2 ≤ Cεθ−1/2 A
L θ L θ

for g(k) = 1 − χ(k). Hence, the modified approximation and the original
approximation are O(εθ−1/2 ) close to each other in H θ if the ansatz functions
A are in H θ . Moreover, we have
Lemma 11.5.3. Let θ, θ0 ≥ 0, θ∞ ∈ R, and let g(k) satisfy
|g1 (k)| ≤ C min(|k − k0 |θ0 , (1 + |k|)θ∞ ).
Then
 −1 (· − k0 )) 2 ≤ Cεθ0 −1/2 A
g1 (·)χ(·)ε−1 A(ε  2 .
L θ L θ0

Proof. This follows by applying Lemma 11.5.1 to g(k) = g1 (k)χ(k). 

Therefore, less regularity θA is needed for estimating the residual terms
in the error equations. However, the original approximation and the cut-off
approximation have to be shown to be O(ε3/2 ) close. Therefore, in order to
have such an estimate in H θ we need at least A ∈ H θ+1 , cf. (11.37). For an
approximation result optimized with this respect see [MU16].
11.5. Extension of the theory 423

Weakening of the non-resonance condition. Checking the non-reso-

nance conditions (11.34) and (11.35) is non-trivial due to the fact that a
two-dimensional function has to be bounded away from zero. The following
version of Lemma 11.5.1 which avoids the loss of ε−1/2 due to the scaling
properties of the L1 -norm allows to reduce the checking of the non-resonance
condition to a one-dimensional problem
Lemma 11.5.4. Let θ, θ0 ≥ 0 and let g(k) satisfy |g(k)| ≤ C|k − k0 |θ0 .
Then
 −1 (· − k0 )) 1 ≤ Cεθ0 A
g(·)ε−1 A(ε  1 .
L θ L θ+θ0

In order to remove l from (11.34) and (11.35) we use the concentration

of the approximation at the wave numbers k = ±k0 and estimate with the
help of Lemma 11.5.4
 
 
 − l)w
 (ωn (l) − ωn (· − k0 ))ψ(· n (l) dl
  2
R Lθ
 
 
 
=  (ωn (· − (· − l)) − ωn (· − k0 ))ψ(· − l)w n (l) dl

R L2θ
 
 
 
≤ sup |ωn (k − (· − l)) − ωn (k − k0 )||ψ(· − l)| wn L2 ≤ Cεwn L2 .
 θ θ
k∈R L1θ

Hence, if we replace ωn (l) by ωn (k − k0 ) in (11.29) we produce additional

terms of order O(ε2 ) which can be included into the terms which do not
make trouble in obtaining error estimates on the O(1/ε2 ) time scale. Hence,
the non-resonance condition (11.34) can be weakened to
(11.38) inf inf |ωj (k) − ω1 (k0 ) − ωn (k − k0 )| ≥ C > 0
j,n∈{1,2} k∈R

and (11.35) can be weakened to

(11.39) inf inf |ωj (k) − ω2 (−k0 ) − ωn (k − k0 )| ≥ C > 0.
j,n∈{1,2} k∈R

11.5.2. Systems with a trivial resonance at k = 0. In physical systems

with conserved quantities very often the eigenvalue zero occurs at the wave
number k = 0, i.e., ωj (0) = 0 for a j in some index set. The most prominent
example is the so called water wave problem, cf. §12.2.1, for which the NLS
equation has been derived first [Zak68]. The fact ωj (0) = 0 will violate the
non-resonance condition (11.34) and will always lead to quadratic resonances
which at a first view will not allow to remove the quadratic terms. However,
due to the fact that the eigenvalue zero is created by a conserved quantity
also the nonlinear terms vanish at the wave number k = 0. The simplest
example for such a system is the so called Boussinesq equation
(11.40) ∂t2 u = ∂x2 u + ∂x2 ∂t2 u + ∂x2 (u2 ),
424 11. Wave packets and the NLS equation

with x ∈ R, t ∈ R, and u(x, t) ∈ R. This model occurs as a long wave

limit for the water wave problem, cf. Chapter 12. The linearized problem
possesses solutions eikx+iωt with dispersion relation
k2
(11.41) ω2 = .
1 + k2
Hence, the non-resonance condition (11.38) is not satisfied as can be seen
in Figure 11.5. The resonance at the wave number k = 0 is trivial since
the nonlinear terms vanish at the wave number k = 0, too. However, the
resonance at the wave number k = k0 is non-trivial. One possibility to get
rid of this second resonance is a k-dependent scaling of the error function.

Figure 11.5. Intersection of the curves k → ±ω(k) and k → ω(k0 ) ±

ω(k − k0 ), here with k0 = 1. There are two intersections at k = 0 and
k = k0 .

Before we explain this in more detail we derive the NLS equation for the
Boussinesq model (11.40). As before we make the ansatz
 
εψNLSbouss =εA1 ε(x − cg t), ε2 t ei(k0 x+ω0 t) + c.c.
 
(11.42) + ε2 A2 ε(x − cg t), ε2 t e2i(k0 x+ω0 t) + c.c.
 
+ ε2 A0 ε(x − cg t), ε2 t .

We find as before at εE the linear dispersion relation (11.41) and at ε2 E the

condition for the linear group velocity cg . At ε3 E we find
2iω0 (1 + k02 )∂T A1 = (1 − c2g − ω02 )∂X
2
A1 − 2k02 (A1 A0 + A2 A−1 ).

The algebraic relations which are found at ε4 E0 and ε2 E2

ε4 E0 : 0 = (1 − c2g )2 ∂X
2 2
A0 + 2∂X A1 A−1 ,
ε2 E2 : 0 = −(−4ω02 + 4k02 + 16ω02 k02 )A2 − 4k02 A21 ,
11.5. Extension of the theory 425

can be solved w.r.t. A0 since (1 − c2g )2 = 0 and w.r.t. A2 since −4ω02 + 4k02 +
16ω02 k02 = 0. Inserting the solutions for A0 and A2 into the equation for A1
finally yields the NLS equation
(11.43) 2iω0 (1 + k02 )∂T A1 = (1 − c2g − ω02 )∂X
2
A1 + γA1 |A1 |2 ,
with
4k02 k04
γ= + .
1 − c2g −ω02 + k02 + 4ω02 k02
The error εβ R = u − εΨ satisfies
∂t2 R = ∂x2 R + ∂x2 ∂t2 R + 2ε∂x2 (ΨR) + O(ε2 ).
Writing this as a first order system gives two equations of the form

   − m, t)R
∂t Rj (k, t) = iωj (k)Rj (k, t) + ερj (k) ψ(k j (m, t) dm + O(ε2 ),
R
with ωj (0) = ρj (0) = 0, but non-vanishing ωj (0) and ρ j (0). The approxi-
mation εψ is concentrated at the wave numbers ±k0 . Hence, a subsystem
is given by
j (k0 , t) =iωj (k0 )R
∂t R  0 , t)R
j (k0 , t) + ερj (k0 ) ψ(k j (0, t) + h.o.t.,

∂t R 
j (0, t) + ερj (0)ψ(−k
j (0, t) =iωj (0)R 
0 , t)Rj (k0 , t) + h.o.t..

For the second equation we have

ωj (0) − ωj (−k0 ) − ωj (k0 ) = 0 + ωj (k0 ) − ωj (k0 ) = 0
such that the non-resonance condition is not satisfied for k = 0. However,
we have ρj (0) = 0 such that the nonlinear terms vanish for this resonant
wave number, too. Hence, this resonance is called trivial. Since
ωj (k) − ωj (−k0 ) − ωj (k + k0 ) = O(|k|) and ρj (k) = O(|k|),
with the denominator in the normal form transform also the nominator
vanishes and therefore the quadratic nonlinear terms close to k = 0 can be
eliminated. For the first of these equations we have
ωj (−k0 ) − ωj (k0 ) − ωj (0) = ωj (k0 ) − 0 − ωj (k0 ) = 0
such that the non-resonance condition is not satisfied for k = −k0 and
similarly for k = k0 . However, we have ρj (k0 ) = 0 such that the nonlinear
terms do not vanish for this resonant wave number. Hence, this resonance
is called non-trivial. One way to get rid of this difficulty is to scale Rj (0)
with εβ+1 instead of εβ . Doing so we obtain
j (k0 , t) =iωj (k0 )R
∂t R  0 , t)R
j (k0 , t) + ε2 ρj (k0 )ψ(k j (0, t) + h.o.t.,

∂t R 
j (0, t) + ρj (0)ψ(−k
j (0, t) =iωj (0)R 
0 , t)Rj (k0 , t) + h.o.t..
426 11. Wave packets and the NLS equation

Hence, in the first equation the nonlinear terms do not make problems any-
more. However, in the second equation we have now terms of order O(1)
which can be eliminated in the full system with the argument from above.
The terms of order O(ε) resulting from this transformation in the second
equation can either be eliminated by another transformation or are of long
wave form, i.e., of a similar form as ε∂x (B(εx)R(x)) in physical space. Such
terms can be estimated by energy estimates to have an O(ε2 ) influence on
the dynamics, cf. Chapter 12.
This can be made rigorous for the full system by making the ansatz
 = u − εΨ
εβ ϑ(k)R
for the error function R, where ϑ(k) = min(ε+|k|/δ, 1) with δ > 0 sufficiently
small, but independent of 0 < ε  1. This has been carried out in [Sch98a]
with a correction explained in [DS06].
The above idea has been transferred to the water wave problem without
surface tension in case of finite depth in [DSW16]. However, the water
wave problem is a quasilinear problem. Below we will explain the additional
difficulties occurring for such systems. But also the so called FPU system
falls into this class, cf. Exercise 11.4. In [Sch10] it has been explained that
the proofs of the approximation theorems given for the PDE systems can be
transferred almost line for line to the FPU system by looking at the Fourier
transformed FPU system.

11.5.3. Stable and unstable non-trivial resonances. For the water

wave problem with small surface tension, i.e., for surface tension parameter
σ ∈ (0, 1/3) additional resonances are present in the system, i.e., there exist
spatial wave numbers k1 , k2 , k3 and associated temporal wave numbers ω1 ,
ω2 , ω3 satisfying
k1 + k2 + k3 = 0 and ω1 + ω2 + ω3 = 0,
cf. Exercise 11.6 and §12.2.1. The same happens for dispersive wave systems
with spatially periodic coefficients, or the poly-atomic FPU model. There
are at least two different approaches to get rid of these resonances.
Solutions of the NLS equation which are analytic in a strip in the com-
plex plane decay with some exponential rate in Fourier space. By nonlinear
interaction, the solutions of the original system have a Fourier mode dis-
tribution which is strongly localized at integer multiples of the basic wave
number k0 . In between, the solutions will be exponentially small, i.e., the
original system can be solved in a weighted L1 -space equipped with the norm


uL1w = |
u(k)|w(k) dk
R
11.5. Extension of the theory 427

with
−1 |k−mk |
(11.44) 1/w(k) = sup |e−αε 0
|.
m∈Z

Hence, the Fourier modes associated to the resonant wave numbers are ex-
ponentially small initially, i.e., of order O(exp(−rε−1 )) for an r > 0, in-
dependent of 0 < ε  1. The quadratic resonances will lead to growth
rates O(exp(εt)) for solutions of order O(ε). Hence, it takes a time of order
O(1/ε2 ) to have O(exp(−rε−1 ))O(exp(εt)) = O(1), i.e., it takes the time
scale of the NLS equation for the error to grow to the order of the NLS ap-
proximation. This idea can be used to prove error estimates on an O(1/ε2 )
time scale for the validity of the NLS approximation also in case of additional
non-trivial resonances, if the solutions of the NLS equation are analytic in
a strip in the complex plane, and if the set of wave numbers resonant to k0
is separated from the set of integer multiples of the basic wave number k0 .
This idea can be made rigorous by making the coefficient α time-dependent,
i.e., by choosing α(t) = α0 − β̌ε2 t in (11.44). This idea has been explained
in [Sch98c] and carried out in [DHSZ16].
The second approach is based on a more detailed analysis of the reso-
nances. Consider a basic wave number k1 = k0 , resonant to wave numbers
k2 and k3 . The ansatz
u(x, t) = εA1 (εt)ei(k1 x+ω1 t) ϕ1 + εA2 (εt)ei(k2 x+ω2 t) ϕ2 + εA3 (εt)ei(k3 x+ω3 t) ϕ3 ,
with vectors ϕj , then yields a three wave interaction (TWI) system
∂T A1 = iγ1 A2 A3 , ∂T A2 = iγ2 A1 A3 , ∂T A3 = iγ3 A1 A2 ,
with coefficients γj ∈ R, associated to the resonances, cf. [Sch05, §3.3]. In
[Sch05, Theorem 3.8] a NLS approximation theorem has been shown in case
that the subspace {A2 = A3 = 0} associated to the wave number k1 = k0 is
stable in the TWI system, cf. Figure 11.6. The proof is based on a mixture
of normal form transforms for the non-resonant wave numbers and energy
estimates for the resonant wave numbers. In [DS06] the ideas of [Sch98a]
and [Sch05] are brought together to handle Boussinesq equations which
model the water wave problem in case of small positive surface tension.
There is also a counter-example [Sch05, §4.1] showing that the NLS
equation fails to approximate solutions in the original system in case of an
unstable k0 -subspace in the associated TWI system and periodic boundary
conditions in the original system. This idea has been carried out in [SSZ15]
for the water wave problem with surface tension and periodic boundary
conditions showing that there exists a continuum of wave numbers and val-
ues of surface tension where the NLS approximation does not make correct
predictions.
428 11. Wave packets and the NLS equation

ImA2

ImA1

ImA3

Figure 11.6. The phase portrait of the TWI system in the invariant
subspace Re A1 = Re A2 = Re A3 = 0. The energy surface is an ellipsoid
since due to conservation of energy not all j have the same sign. The
axes are invariant subspaces associated to the wave numbers kj . There
are one unstable and two stable subspaces.

The situation on the whole real line for an unstable resonance is still
open. In this case the different group velocities ω (kj ) at the resonant
wave numbers kj no longer can be neglected. For a thorough discussion
see [Sch05, §4.2]. A recent attempt to understand this situation can be
found in [MN13].

11.5.4. Quasilinear quadratic nonlinearities. The first and very gen-

eral NLS approximation theorem was proved in [Kal88] for general quasi-
linear dispersive wave systems. However, the occurrence of quasilinear qua-
dratic terms has been excluded explicitly. A typical example is given by the
quasilinear wave equation
1
(11.45) ∂t2 u = ∂x2 u − u + ∂x2 (u2 ) + u3 ,
2
with x ∈ R, t ∈ R, and u(x, t) ∈ R. With the NLS ansatz (11.42) we obtain
the linear dispersion relation ω02 = k02 + 1, the group velocity cg = k0 /ω0 ,
and the NLS equation,
2
(11.46) iν1 ∂T A + ν2 ∂X A + ν3 A|A|2 = 0,
with
9αc − 2k04
(11.47) ν1 = 2ω0 , ν2 = (1 − c2 ), ν3 = ,
3
for the description of small spatio-temporal modulations of the underlying
carrier wave ei(k0 x+ω0 t) . As above, the idea is to use a normal form transform
to eliminate the quadratic terms. For (11.45) the quasilinear quadratic non-
linearity causes the required normal form transform to lose regularity such
11.6. Pulse dynamics in photonic crystals 429

that the original quasilinear system can no longer be handled after applying
the transformation (see §11.4). This can be seen as follows. Writing the er-
ror equations as first order system gives nonlinear terms with growth rates
proportional to |k| for |k| → ∞. This results in the normal form transform
in a nominator proportional to |k| and in a denominator proportional to
ω(k) − ω(k0 ) − ω(k − k0 ) = O(1)
for |k| → ∞. Thus, the normal form transform is of the form identity plus
a term which is small but loses one derivative. Hence, the normal form
transform can no longer be inverted with Neumann’s series.
So far there are only a few quasilinear systems where approximation
results for the NLS approximation could have been established. One example
is where the right-hand side of the quasilinear
√ dispersive wave system only
loses half a derivative, i.e., a factor of k in Fourier space, as a result of the
normal form transformation. In this case the elimination of the quadratic
terms is still possible and the transformed system can be handled with the
Cauchy-Kowalevskaya theorem [SW11]. The Lagrangian formulation of
the water wave problem in case of finite depth and zero surface tension
falls into this class, cf. [DSW16]. In case of zero surface tension and
infinite depth such a result has been established in [TW12, Tot15] by
finding a special transformation which allows to eliminate all quadratic terms
for this particular system without loss of regularity. Another example is
in the context of the KdV equation where the result can be obtained by
applying the Miura transformation [Sch11], cf. Exercise 11.5. In [CS13]
numerical evidence is given that the NLS approximation is valid for (11.45).
Very recently it turned out that the solutions of the transformed quasilinear
system can be estimated by more clever energy estimates [Dül16, CW16,
DH16].
Although relevant systems can be handled with the last approach a va-
lidity theory for general dispersive wave systems with quasilinear quadratic
terms is still an open problem.

11.6. Pulse dynamics in photonic crystals

One of the major goals of photonics is the construction of ’electronic’ devices
where the electrons are completely replaced by photons. Photonic crystals
turned out to be a suitable tool for the construction of such devices. They
consist of a dielectric material such as glass with a periodic structure with a
period comparable to the wave length of light. Due to the periodic structure
the linearized problem is no longer solved by Fourier modes, but by so called
Bloch modes. The curves of eigenvalues plotted as a function over the Bloch
wave numbers can now possess horizontal tangencies, i.e., vanishing group
430 11. Wave packets and the NLS equation

velocities. Thus, in principle, standing light pulses are possible. At the

horizontal tangencies very often spectral gaps occur, i.e., there are temporal
wave numbers for which the associated wave cannot travel through the pho-
tonic crystal. The first fact allows to use photonic crystals as optical storage
the second fact to use them for filtering. We will explain the possibility of
standing light pulses in photonic crystals by using the NLS approximation.
For simplicity we restrict ourselves again to a nonlinear wave equation
(11.48) ∂t2 u = χ1 ∂x2 u − χ2 u + χ3 u3 ,
but now with spatially periodic coefficients χj (x) = χj (x + L) for a L > 0,
x ∈ R, t ∈ R, and u(x, t) ∈ R. Moreover, we assume χ1 (x) ≥ γ1 > 0 and
χ2 (x) ≥ γ2 > 0.

11.6.1. The eigenvalue problem for photonic crystals. We start by

reviewing a number of well known results, cf. [Eas73, Sca99]. For nota-
tional simplicity assume here L = 2π. The linear problem
∂t2 u = χ1 ∂x2 u − χ2 u
possesses solutions u(x, t) = eiωt v(x) where v satisfies
−ω 2 v = χ1 ∂x2 v − χ2 v.
Uniformly bounded solutions are given by Bloch waves
v(x) = w(x)eix ,
with w(x) = w(x + 2π) satisfying the eigenvalue problem
L()w = −χ1 (∂x + i)2 w + χ2 w = ω 2 w.
With w(x) = w(x + 2π) also w(x)einx for n ∈ Z is 2π-periodic. Hence we
can restrict ourselves to  ∈ [− 12 , 12 ), the so called Brillouin zone.
Lemma 11.6.1. The properties of L() are as follows:
a) L() is self-adjoint in L2per (χ−1
1 dx).
b) L() is positive definite.
c) L() has discrete spectrum with ∞ the only accumulation point. All
eigenvalues are real, semi-simple, and non-negative.

Proof. a) follows from

 2π  2π
1 1
(L()w1 , w2 )L2 (χ−1 dx) = −χ1 ((∂x +i) w1 )w2 dx+
2
(χ2 w1 )w2 dx
1
0 χ1 0 χ1
 2π  2π
1
= ((∂x + i)w1 )(∂x + i)w2 dx + χ2 w1 w2 dx
0 0 χ 1
=(w1 , L()w2 )L2 (χ−1 dx) .
1
11.6. Pulse dynamics in photonic crystals 431

b) follows from
 2π
χ2 2
(L()w, w)L2 (χ−1 dx) = |(∂x + i)w|2 + |w| dx > 0, (w = 0).
1
0 χ1
2 (R/2πZ) →
c) follows from the fact that for fixed  the operator L() : Hper
L2per (R/2πZ) is elliptic, i.e., L() has a compact resolvent and so L() has
discrete spectrum with ∞ the only accumulation point. Due to the self-
adjointness all eigenvalues are real and semi-simple. Due to the positive
definiteness all eigenvalues are non-negative. 
Lemma 11.6.2. Except of intersection points the curves of eigenvalues  →
ωn () of L() are smooth.

Proof. This follows by a simple perturbation argument, cf. [Kat95]. 

11.6.2. The computation of the eigenvalues. Here and the subsequent

example we assume L = 1, hence  ∈ [−π, π). The eigenvalue problem
(11.49) χ1 ∂x2 u − χ2 u = −λu
is a second order scalar ODE with spatially periodic coefficients. We write
(11.49) as first order system
∂x u(x) = v,
(11.50)
∂x v(x) = −s(x)λu(x) + q(x)u(x),
where s(x) = 1/χ1 (x) and q(x) = χ2 (x)/χ1 (x). The fundamental matrix
of (11.50) is denoted by Φλ = Φλ (x, x0 ) where Φλ (x0 , x0 ) = I. Floquet’s
theorem 2.1.17 shows that
Φλ (x, x0 ) = Pλ (x, x0 )e(x−x0 )Mλ ,
with Pλ (x, x0 ) = Pλ (x + 1, x0 ) and a matrix Mλ , independent of x and x0 .
Note that Mλ is not unique since e2πin = 1 for n ∈ Z. The eigenvalues
of Mλ are the Floquet exponents. The two eigenvalues ρ− and ρ+ of the
monodromy matrix Cλ = eMλ are the Floquet multipliers. Since the trace
of the linear vector field on the right-hand side of (11.50) vanishes we have
the conservation of the phase volume for (11.50) leading to ρ+ ρ− = 1, cf.
the proof of Theorem 4.1.3. Hence, the Floquet multipliers can be computed
via the characteristic polynomial
ρ2 − D(λ)ρ + 1 = 0
and are given by
1 1
ρ± (λ) = D(λ) ± (D(λ))2 − 4,
2 2
where the trace of the monodromy matrix, D(λ) = trace Cλ , is called the
discriminant. We find that
432 11. Wave packets and the NLS equation

a) if |D(λ)| > 2 then the Floquet multipliers ρ± (λ) are real. As a

consequence the solutions have exponential growth or decay w.r.t. x.
b) if |D(λ)| < 2 then the Floquet multipliers ρ± (λ) are on the complex
unit circle. As a consequence the solutions are uniformly bounded w.r.t. x.
c) if |D(λ)| = 2 then the Floquet multipliers ρ± (λ) are 1 or −1. In this
case we have at most linear growth
Example 11.6.3. We consider the eigenvalue problem (11.49) with
s(x) = χ[0,6/13] + 16χ(6/13,7/13) + χ[7/13,1] (x mod 1)
and q(x) = μ ∈ R from [BCBLS11]. For this choice the ODE can be solved
and the monodromy matrix can be computed explicitly. We find for the
discriminant
   
25 16  9 8
(11.51) D(λ) = cos λ + μ − cos λ+μ .
8 13 8 13
See Exercise 11.8. The graph λ → D(λ) of the discriminant and the associ-
ated dispersion relation can be found in Figure 11.7 for μ = 0.

Figure 11.7. The left panel shows the multivalued inverse of the map
λ → D(λ) and the right panel the associated curves of eigenvalues k →
ω 2 (k) for Example 11.6.3.
11.6. Pulse dynamics in photonic crystals 433

For frequencies ω with ω 2 = λ which fall into a spectral gap the incoming
wave is damped in the photonic crystal with some exponential rate w.r.t.
the depth of penetration x. Hence, photonic crystals can be used as filters.
This is one of the reasons why the wings of a butterfly show their colorful
appearance.
Remark 11.6.4. There is a relation between the regularity of the coeffi-
cients χ1 and χ2 on one side and the size of the spectral gaps in
{ωn (l) : l ∈ (−1/2, 1/2], n ∈ Z \ {0}} ⊂ R
on the other side. For continuous or even smoother χ1 and χ2 the spectral
gaps become smaller as n increases. According to [Eas73, Nti76], the size
of the gaps which are found at ω ∼ n decays at least with 1/nθ+1 for n → ∞
if χ1 ∈ Cbθ and χ2 ∈ Cbθ−2 , i.e., the more regular χ1 and χ2 are, the faster
the gaps close.
Remark 11.6.5. Spectral gaps can be obtained from a spatially homo-
geneous situation by adding small spatially periodic perturbations to the
coefficients. As an example consider the eigenvalue problem
(∂x + i)2 w(x) + ω 2 (1 + 2ε cos(2x))w(x) = 0,
with w(x) = w(x + 2π),  ∈ [− 12 , 12 ), and 0 ≤ ε  1. For ε = 0 the problem
is given by
(∂x + i)2 w + ω 2 w = 0,
and can be solved by w(x) = einx with n ∈ Z and associated eigenvalues
ωn2 () = (n + )2 . Hence, at (, ω) = (0, 1) there is a crossing of the curves
of eigenvalues. All single eigenvalues ω vary smoothly w.r.t. small ε since
2ε cos(2x)· is a small perturbation of the operator (∂x + i)2 . Hence, the
smooth curves  → ωn () will only vary slightly w.r.t. ε. However, at the
crossing points the curves can split. As an example we consider the point
(, ω) = (0, 1). We use 2 cos 2x = e2ix + e−2ix , make the ansatz w(x) =

n∈2Z+1 cn e  2 . We obtain
inx , and set ω 2 = 1 + ω

(11.52)  2 )[c1 + ε(c3 + c−1 )] =0,

−(1 + )2 c1 + (1 + ω
(11.53)  2 )[c−1 + ε(c1 + c−3 )] =0,
−(−1 + )2 c−1 + (1 + ω
(11.54)  2 )[c3 + ε(c1 + c5 )] =0,
−(3 + )2 c3 + (1 + ω
..
.
 2 , ε are
The equations (11.54)-. . . can be solved w.r.t. c3 , c−3 , c5 , . . ., if , ω
small, i.e., there exist functions
 2 , ε, c−1 , c1 ) = O(|| + |
cj =cj (, ω ω 2 | + |ε|)O(|c1 | + |c−1 |)
434 11. Wave packets and the NLS equation

which are linear w.r.t. c1 and c−1 . Inserting this into (11.52)-(11.53) gives
 2 )c±1 + ε(1 + ω
0 =(∓2 − 2 + ω  2 )c∓1 + εO(|| + |
ω |2 + |ε|)O(|c1 | + |c−1 |),
or equivalently,
    
−2 − 2 + ω  2 + h.o.t. (1 + ω 2 )ε + h.o.t. c1 0
= .
(1 + ω 2 )ε + h.o.t. 2 − 2 + ω  2 + h.o.t. c−1 0
In order to have non-trivial solutions we need a vanishing determinant. We
find
ω 2 )2 − ε2 + h.o.t. = 0,
(
i.e., (ω 2 )1/2 = ±ε+h.o.t., i.e., there is a splitting of the eigenvalues. In higher
space dimensions spectral gaps cannot be obtained by a small perturbation
of the spatially homogeneous situation. For examples of spectral gaps in 2D
and 3D see [Kuc93, BFL+ 07].

11.6.3. Bloch transform. To derive and justify the NLS equation for
(11.48) we follow [BSTU06] and adapt the Fourier space approach for the
constant coefficient problem (11.2) from §11.2 to Bloch space. For Schwartz
functions u ∈ S, the Bloch transform is defined by

(11.55) (, x) = (T u)(, x) =
u ( + j),
eijx u
j∈Z

and its inverse by

 1/2
(11.56) u(x) = (T −1 u
)(x) = (, x) d.
eix u
−1/2

By construction we have
(11.57) (, x) = u
u (, x + 2π) and u
(, x) = u
( + 1, x)eix .
The Bloch transform turns out to be an isomorphism between H θ (R, C) and
L2 ((−1/2, 1/2], H θ ([0, 2π), C)), cf. [RS75b, Sca99], where
 1/2
1/2

uL2 ((−1/2,1/2],H θ ([0,2π),C)) = 
u(, ·)2H θ [0,2π] d .
−1/2

Multiplication u(x)v(x) in physical space corresponds in Bloch space to the

operation
 1/2
(11.58) u  v)(, x) =
( ( − m, x)
u v (m, x) dm,
−1/2

where (11.57) has to be used for | − m| > 1/2. However, if χ : R → R is

2π-periodic w.r.t. x, then
T (χu)(, x) = χ(x)(T u)(, x).
11.6. Pulse dynamics in photonic crystals 435

Applying the Bloch transform to (11.48) gives

(11.59)  ∂x )
(, x) = −L(,
∂t2 u u3 (, x),
u(, x) + χ3 (x)
 ∂x ) : H 2 ([0, 2π)) → L2 ([0, 2π)) are given by
where the Bloch operators L(,
 ∂x )
L(, u(, ·)(x) = −χ1 (x)(∂x + i)2 u
(, x) + χ2 (x)
u(, x).
According to Lemma 11.6.1 for fixed  these operators are self-adjoint and
positive definite in the space L2 (χ−1 1 dx). The induced norm  · L2 (χ−1 1 dx)
and the usual L -norm are equivalent since χ1 (x) ≥ γ1 > 0 for a constant
2

γ1 independent of x by assumption. Thus, for each fixed  there exists a

 ∂x ) with
Schauder basis (fj (, ·))j∈N of L2 ([0, 2π)) of eigenfunctions of L(,
strictly positive eigenvalues λj () > 0, i.e., L(, ∂x )fj (, ·) = λj ()fj (, ·).
We make the ansatz

(, x, t) =
u j (, t)fj (, x).
u
j∈N

 ∂x ) is self-adjoint in L2 (χ−1 dx), the eigenfunctions (fj (, ·))j∈N

Since L(, 1
can be chosen to form an orthonormal basis of L2 (χ−1 1 dx) for each fixed .
Hence,
j (, t) =
fj (, ·), u
u (, ·, t)L2 (χ−1 dx) ,
1
and therefore
j (, t) = −λj ()
∂t2 u uj (, t) +
fj (, ·), χ3 (·)
u3 (, ·, t)L2 (χ−1 dx)
1
  1/2  1/2 j
= − λj () uj (, t) + bj1 j2 j3 (,  − 1 , 1 − 2 , 2 )
j1 ,j2 ,j3 ∈N −1/2 −1/2

(11.60)
j1 ( − 1 , t)
×u uj2 (1 − 2 , t)
uj3 (2 , t) d2 d1 ,
where
bjj1 j2 j3 (,  − 1 , 1 − 2 , 2 )
=
fj (, ·), χ3 (·)fj1 ( − j1 , ·)fj2 (j1 − j2 , ·)fj3 (j2 −, ·)L2 (χ−1 dx) .
1

Since the nonlinear terms have some convolution structure we have a system
as in §11.3. Thus, we can proceed exactly as in §11.3 in order to derive a
NLS equation. However, due to the special structure of (11.60) we keep the
second order system and proceed as in §11.2 and make the ansatz
   
−1  · − 0 2 −1  · + 0 2
n0 (, t) = εε A1
u 1
, ε t E + εε A−1 , ε t E−1
ε ε

for a n0 ∈ N where Ej = ejiωn0 (0 )t eiωn0 (0 )(−j0 )t . In physical space the
ansatz corresponds to
(11.61) u(x, t) = εψA (x, t) = εA(ε(x+cg t), ε2 t)fn0 (0 , x)ei0 x eiωn0 (0 )t +c.c.,
436 11. Wave packets and the NLS equation

where again T = ε2 t, X = ε(x + cg t), and A(X, T ) ∈ C. We have a

cancelation at εε−1 E and ε2 ε−1 E, and at ε3 ε−1 E we find the NLS equation
1 = −(λ (0 ) − 2(ω (0 ))2 )κ2 A
2iωn0 (0 )∂T A 1 /2
n0 n0
(11.62)
+γ A1 ∗ A1 ∗ A−1 ,

where T = ε2 t, κ = ε−1 ( − 0 ), and γ = 3bnn00 n0 n0 (0 , 0, 20 , −0 ), i.e.,

 2π
3 χ3 (x)
(11.63) γ= |fn (0 , x)|4 dx ∈ R,
2ωn0 (0 ) 0 χ1 (x) 0

while A−1 = F −1 A −1 satisfies the complex conjugate equation. In order to

obtain (11.62), in the convolution term we use
 1/(2ε)  1/(2ε)
bnn00 n0 n0 (0 + εκ, ε(κ − κ1 ), 20 + ε(κ − κ1 ), −0 + εκ2 ))
−1/(2ε) −1/(2ε)

×A 1 (κ1 − κ2 )A
1 (κ − κ1 )A −1 (κ2 ) dκ2 dκ1
 ∞ ∞
→ bnn00 n0 n0 (0 , 0, 20 , −0 ) A 1 (κ1 − κ2 )A
1 (κ − κ1 )A −1 (κ2 ) dκ2 dκ1
−∞ −∞

and the symmetry of the kernel. The derivation from (11.62) is consistent
with the derivation from the associated first order system since for instance

−(λ n0 (0 ) − 2(ωn 0 (0 ))2 )/(4iωn0 ) = − ((ωn0 (0 )2 ) − 2(ωn 0 (0 ))2 )/(4iωn0 )

which equals iω (0 )/2.

11.6.4. An approximation result. The derivation of the NLS equation,

the fact that the NLS equation possesses standing pulses, and the possibility
of vanishing group velocities, in principle gives the possibility of standing
light pulses. In order to show that these standing light pulses exist at least
on an O(1/ε2 ) time interval we again prove an approximation result.
We consider the following situation. We assume that one of the curves
of eigenvalues, namely λn0 , has a horizontal tangency and that there is
no other curve in a neighborhood of this point in the (, λ)-plane. Hence,
λ = 2ωω = 0 in this point and so we have a vanishing group velocity for an
associated modulated wave packet. For simplicity we assume that λ 1 (0) = 0
with λ1 (0) > 0.
In order to bring together the NLS equation whose solutions in Fourier
space are given on the complete real line with the Bloch wave representation
of the nonlinear wave equation we introduce a cut-off operator χ ∈ C0∞ with
χ() ∈ [0, 1], χ() = 1 for  ∈ [−1/5, 1/5], and χ() = 0 for || ≥ 2/5, and an
extension operator P which extends a function with length of support less
11.6. Pulse dynamics in photonic crystals 437

than 1 to a function on the complete real axis with period 1. Then we make
the modified ansatz
 
1 · , ε2 t E1 )() + εε−1 P(χ(·)A
1 (, t) = εε−1 P(χ(·)A
u −1 · , ε2 t E−1 )()
ε ε
and find the NLS equation
(11.64) 2iω1 (0)∂T A 1 = −(λ 1 (0) − 2(ω1 (0))2 )κ2 A 1 ∗ A
1 /2 + γ A 1 ∗ A
−1 ,

where now γ = 3b1111 (0, 0, 0, 0) ∈ R, while A−1 = F −1 A

−1 satisfies again the
complex conjugate equation. In case of a cubic nonlinearity for the proof
of the approximation result it is not necessary to expand the problem in
eigenfunctions, and so we set
(, x, t) = u
u ⊥ (, x, t),
1 (, t)f1 (, x) + u
⊥ (, ·, t)L2 (χ−1 dx) = 0. Moreover, the separation will be
with
(f1 (, ·), u
1
made in such a way that the support of u 1 (, t) is contained in [−2/5, 2/5].
The two functions u 1 (, t) and u ⊥ (, x, t) are defined to satisfy
1 (, t) = − λ1 ()
∂t2 u u1 (, t) + Ec ()
f1 (, ·), χ3 (·)
u3 (, ·, t)L2(χ−1 dx) ,
1

∂t2 u  ∂x )
⊥ (, x, t) = − L(, u⊥ (, x, t) + χ3 (x)
u3 (, x)
− Ec ()
f1 (, ·), χ3 (·)
u3 (, ·, t)L2 (χ−1 dx) f1 (, x),
1

where Ec ∈ C0∞ , with Ec () ∈ [0, 1], with Ec () = 1 for  ∈ [−1/5, 1/5], and
χ() = 0 for || ≥ 2/5. Note that χ is defined on the Fourier wave numbers,
where Ec is defined on the Bloch wave numbers.
We add higher order terms to the ansatz to make the residual smaller,
i.e., we consider
 
1 (, t) =εε−1 P(χ(·)A
u 1 · , ε2 t E1 )() + εε−1 P(χ(·)A −1 · , ε2 t E−1 )()
ε ε
3 −1  · 2  3
+ ε ε P(χ(·)A3 , ε t E )()
ε · 

+ ε ε P(χ(·)A−3 , ε3 t E−3 )(),
3 −1
ε
⊥ 3 −1 ⊥  
 (, x) =ε ε u
u 1 ( , x, ε t)E + ε3 ε−1 u
2 1
⊥ 2
−1 ( , x, ε t)E
−1
ε ε
 3 −1 ⊥ 
+ ε3 ε−1 u⊥ 2 3
3 ( , x, ε t)E + ε ε u −3 ( , x, ε2 t)E−3 .
ε ε
As before we find A 1 as a solution of the NLS equation (11.64) and A −1 as
a solution of the complex conjugate equation. Moreover we choose
−9ω 2 (0)A 3 (κ, T )
1
3 (κ, T ) +
f1 (0, ·), χ3 (·)(f1 (0, ·))3  2 −1 A
= − λ1 (0)A ∗3
L (χ dx) 1 (κ, T ),
1

u⊥
−ω12 (0)1 (, x, t)
438 11. Wave packets and the NLS equation


= − L(εκ, u⊥  ∗2 
∂x )1 (κ, x, T ) + 3χ3 (x)(A1 f1 ) ∗ A−1 f−1 )(κ, x)
1 f1 )∗2 ∗ A
−Ec ()
f1 (εκ, ·), 3χ3 (·)(A −1 f−1 )(κ, ·, T ) 2 −1 f1 (εκ, x),
L (χ dx)
1

u⊥
−9ω12 (0)3 (, x, t)

= − L(εκ, u⊥   
∂x )3 (κ, x, T ) + χ3 (x)(A1 f1 ∗ A1 f1 ∗ A1 f1 )(κ, x)
1 f1 )∗3 (κ, ·, T ) 2 −1 f1 (εκ, x),
− Ec ()
f1 (εκ, ·), χ3 (·)(A L (χ dx)
1

and A−3 , u
⊥ ⊥
−1 , and u−3 as the solutions of the associated complex conjugate
equations. By this choice we eliminate all nonlinear terms w.r.t. A 1 and
A−1 in the part of the residual belonging to the u equation. This has the
⊥

advantage that we do not need more information about the spectrum and
the associated eigenfunctions, especially we avoid to estimate derivatives of
the eigenfunctions w.r.t. .
In order to solve the equations for A 3 , u
⊥ ⊥
1 , and u3 we need the non-
resonance conditions
−9ω12 (0) = − λ1 (0),

−ω12 (0) ∈spec(−L(εκ, ∂x ))|{f1 (εκ,·)}⊥ ,

−9ω12 (0) ∈spec(−L(εκ, ∂x ))|{f1 (εκ,·)}⊥ .
If these conditions are not satisfied for  = εκ for which Ec () = 1 we can
make the support of Ec smaller to satisfy the non-resonance conditions. This
is possible if
 ∂x ))|{f (0,·)}⊥ ,
−ω12 (0) ∈spec(−L(0, 1

 ∂x ))|
−9ω12 (0) ∈spec(−L(0, {f1 (0,·)}⊥ .

These formal calculations can be made rigorous, as in §11.2 it can be shown

that the residual is of order O(ε7/2 ) in H 1 in physical space.
The error estimates can be proved in physical space. Let εΨ be the
approximation in physical space satisfying
sup Res(εΨ)H 1 ≤ Cε7/2
t∈[0,T0 /ε2 ]

where
Res(u) = −∂t2 u + χ1 ∂x2 u − χ2 u + χ3 u3 .
The error ε3/2 R = u − εψ satisfies
∂t2 R = χ1 ∂x2 R − χ2 R + 3ε2 χ3 ψ 2 R + 3ε5/2 χ3 ψR2 + ε3 χ3 R3 + ε−3/2 Res(εΨ).
Multiplying
∞ this equation with χ−1
1 ∂t R and performing the integration
−∞ . . . dx yields
∂t E ≤ C1 ε2 E + C2 ε5/2 E 3/2 + C3 ε3 E 2 + C4 ε2 ,
11.6. Pulse dynamics in photonic crystals 439

with ε-independent constants C1 , . . . , C4 and where

 ∞
E= χ−1 2 2 −1 2
1 (∂t R) + (∂x R) + χ1 χ2 R dx.
−∞
A simple application of Gronwall’s inequality yields as in §11.2 the following
approximation theorem.
Theorem 11.6.6. Let the Fourier transform of A ∈ C([0, T0 ], H 4 ) be a
solution of the NLS equation (11.64). Then there exists an ε0 > 0 and a
C > 0 such that for all ε ∈ (0, ε0 ) there are solutions u of the original system
(11.48) which can be approximated by εΨ w.r.t. the H 1 -norm such that
sup u(t) − εΨ(t)H 1 < Cε3/2 .
t∈[0,T0 /ε2 ]

If the NLS equation (11.64) is a focussing one, then it possesses standing

time-periodic pulse solutions, see §8.1.1. This in combination with Theo-
rem 11.6.6 shows the existence of approximate standing light pulses on an
O(1/ε2 ) time scale. Since we have a finite speed of propagation for (11.48)
with the help of spatial dynamics and invariant manifold theory, cf. Chapter
13, the existence of the standing modulating pulses can even be shown on
O(1/εn ) time scales for fixed n > 2 if certain non-resonance conditions are
satisfied [LBCB+ 09]. If the coefficients χ1 and χ2 are chosen in a special
way, then standing modulating pulses can be shown [BCBLS11] to exist
for all t ∈ R.

11.6.5. Gap solitons. Another popular class of models for light propa-
gation in photonic crystals starts directly with a Nonlinear Schrödinger (or
Gross–Pitaevsky) equation in the form
(11.65) iEt = −ΔE + V (x)E + σ|E|2 E, V : R2 → R, x ∈ R2 , t ∈ R,
with E = E(x, t) ∈ C, σ = ±1, and where w.l.o.g. the potential V is 2π–
periodic in each coordinate. Here E is the “out–of–plane” electric field, i.e.,
perpendicular to the 2D periodicity structure, and the amplitude of V is
called the contrast of the material. An interesting problem then is to search
for so called gap solitons E(x, t) = φ(x)e−iωt , where φ ∈ C is a localized
solution of the stationary problem
(11.66) (−Δ + V (x) − ω)φ + σ|φ|2 φ = 0.
Localized means |φ(x)| → 0 exponentially as |x| → 0, and this implies
that ω has to lie in a gap of the essential spectrum of the operator L :=
−Δ + V (x), hence the name “gap soliton”. These have been discussed the
physics literature since the early 1990ties, see, e.g., [Ace00] for a review.
In 1D, where gaps open for small contrasts as discussed above, they are
typically described by so called coupled mode equations, cf. Exercise 11.9. In
440 11. Wave packets and the NLS equation

two and more space dimensions gaps are more difficult to open [Kuc93], and
in particular one needs a finite contrast. Rigorous proofs and asymptotics for
2D gap solitons, based on the reduction of (11.65) to systems of homogeneous
NLS equations, here also called coupled mode equations, can be found for
instance in [DPS09] (for the case of a separable potential) and in [DU09]
(for general V ).

11.7. Nonlinear optics

It is the purpose of this section to explain how the previous analysis is
related to the motivation given at the beginning of this Chapter, namely
the transport of information by light pulses through glass fibers. Moreover,
we explain why the rate of information transported through the fiber can be
increased by multiplexing, i.e., by taking simultaneously pulses with different
carrier waves.

11.7.1. Maxwell’s equations in glass fibers. Electromagnetic waves

are described by Maxwell’s equations. For x = (x1 , x2 , x3 ) ∈ R3 and t ∈ R
they are given by
(11.67) ∇ · E =ρ,
(11.68) ∇ × E = − ∂t B,
(11.69) ∇ × B =∂t E + j,
(11.70) ∇ · B =0,

where E = E(x, t) ∈ R3 is the electric field, B = B(x, t) ∈ R3 the magnetic

field, j = j(x, t) ∈ R3 an electric current density, and ρ = ρ(x, t) ∈ R an
electric charge density. There is conservation of charges, i.e., additionally
we have ∂t ρ + ∇ · j = 0. There are no magnetic monopoles, i.e., (11.67)-
(11.70) are not symmetric, since there is no magnetic current density and
no magnetic charge density. We put the vacuum velocity of light and the
vacuum electric permittivity to one.

Remark 11.7.1. For some purposes, alternative formulations of Maxwell’s

equations are useful. For time-independent solutions the system decouples
into electrostatics
∇ · E = ρ, ∇×E =0
and magnetostatics
∇ × B = j, ∇ · B = 0.
Hence, there exist a scalar potential φ with E = ∇φ satisfying Δφ = ρ and
a vector potential A with B = ∇ × A satisfying ∇ × ∇ × A = j.
11.7. Nonlinear optics 441

The integral formulation of Maxwell’s equations requires less regularity.

For a volume V with surface ∂V we find
  
E·n= ρ and B · n = 0.
∂V V ∂V

For a surface S with boundary ∂S we find

   

E, ds = − ∂t B and
B, ds = ∂t E + j.
∂S S ∂S S

Introducing the polarization. For given initial conditions E|t=0 and

B|t=0 and given external densities ρ = ρ(x, t) and j = j(x, t) the solutions
E = E(x, t) and B = B(x, t) of Maxwell’s equations can be computed. In
a medium, the electric field E and the magnetic field B affect the motion
of the charged particles in the medium leading to another electric and mag-
netic field, namely the polarization and the magnetization. Glass fibers and
photonic crystals are insulators, i.e., there are no free charge carriers. More-
over, there is no magnetization. However, the applied electric field yields to
a shift of the electron density, i.e., to a charge density distribution ρpol in
the atoms. Due to the conservation of charges there is an induced current
density jpol satisfying
∂t ρpol + ∇ · jpol = 0.
These induced densities on the other hand induce via (11.67) and (11.69)
an electric field P = P (x, t), which is called the polarization such that
ρpol = −∇ · P and ∇ · jpol = −∂t ρpol = ∂t ∇ · P,
leading to ∂t P = jpol . Thus, Maxwell’s equations in glass fibers and photonic
crystals are given by
(11.71) ∇ · E = − ∇ · P,
(11.72) ∇ × E = − ∂t B,
(11.73) ∇ × B =∂t E + ∂t P,
(11.74) ∇ · B =0.
This system can be simplified to a single equation. Applying ∇× to (11.72)
yields
∇ × (∇ × E) = −∂t (∇ × B).
Using
∇ × (∇ × E) = ∇(∇ · E) − ΔE
and (11.73) yields
(11.75) ∇(∇ · E) − ΔE = −∂t2 P − ∂t2 E.
442 11. Wave packets and the NLS equation

For linearly polarized light, e.g. E = u(x3 , t)ex1 and P = p(x3 , t)ex1 we find
∇ · P = 0, ∇ · E = 0, and so
(11.76) ∂x23 u = ∂t2 u + ∂t2 p,
where u(x3 , t) ∈ R and p(x3 , t) ∈ R.
Modeling the polarization. In order to have an evolutionary problem we
have to close (11.75) or (11.76) by a constitutive law P = P (E) or p = p(u),
respectively. There exist various models. Basically the law of motion of a
particle with coordinates x = x(t) of mass m and charge q is given by
d2 dx
m 2
x = q(E + × B).
dt dt
In the simplest model the polarization is modeled as an oscillator and the
influence of the magnetic field B is neglected. Thus we suppose that for an
atom placed in an electric field the center x = x(t) of the electron density
obeys the equation  2 
d 2
m x + ω0 x = qe E,
dt2
with qe the elementary charge, ω0 some normalized temporal wave number,
and m the mass of the electron. On the continuum level in some mean-field
limit for the electron position we find for the polarization
(11.77) ∂t2 P + ω02 P = dE,
with a constant d ∈ R. In case of damping by thermalization we have
 2 
d d 2
m x + γ x + ω0 x = qe E
dt2 dt
for a γ > 0. On the continuum level we find
(11.78) ∂t2 P + γ∂t P + ω02 P = dE.

For larger values of E the linear oscillator has to be replaced by a non-

linear oscillator. Due to the symmetries of the problem very often there
are no quadratic terms. As a nonlinear example we consider the oscillator
model  2 
d d 2 2
m x + γ x + ω0 x + r|x| x = qe E
dt2 dt
for an r ∈ R. On the continuum level we find
(11.79) ∂t2 P + γ∂t P + ω02 P + rP |P |2 = dE.
In general, matter consists of different atoms which combine to various
molecules. This is modeled by considering various kind of oscillators. Hence,
we finally come to a system
(11.80) ∇(∇ · E) − ΔE = − ∂t2 P − ∂t2 E,
11.7. Nonlinear optics 443


n
(11.81) P = Pj ,
j=0

(11.82) ∂t2 Pj + γj ∂t Pj + ωj2 Pj + rj |Pj | Pj =dj E,

2

where the dj > 0 are constants taking into account different masses, different
numbers of atoms, etc., of the various kinds of atoms. In order to solve
this system uniquely we need initial conditions for E|t=0 , ∂t E|t=0 , Pj |t=0 and
∂t Pj |t=0 for j = 0, . . . , n. We refer to [SU03b] for a mathematical analysis of
light pulse propagation in glass fibers with this modeling in case of damping,
i.e., γj > 0.
Remark 11.7.2. a) (11.80)-(11.82) models isotropic media. Anisotropic
media can be modeled for instance by choosing γj , ωj2 , dj , and rj as tensors.
b) In the linear case rj = 0 in (11.82), the constitutive laws can be solved
explicitly. For E = E0 eiωt we find P = P0 eiωt + P(t), with P(t) → 0 with
some exponential rate for t → ∞ and
P0 = α(ω)E0
where

n
dj
α(ω) = .
j=0
−ω 2 + iγj ω + ωj2
The numbers γj , dj , and ωj can be used to fit the constitutive law to ex-
perimental data. There is a remarkable good agreement with experimental
observations.
c) More generally, in the linear case there always is a Green’s function
χ1 such that P can be expressed in terms of E, i.e.,
 t
(11.83) P (t) = χ1 (t − τ )E(τ ) dτ.
−∞
In isotropic materials χ1 is scalar. In glass fibers the above constants and
also χ1 only depend on the transverse variables, in photonic crystals there
is a periodic dependence on the spatial variables. Although there is no
Green’s function for the nonlinear system, similar to the linear situation the
constitutive law for the polarization is very often modeled by
 t  t
(11.84) P (t) = χ1 (t − τ )E(τ ) dτ + χ3 (t − τ )|E(τ )|2 E(τ ) dτ,
−∞ −∞
i.e., again (11.81)-(11.82) is replaced by (11.84).
d) In nonlinear optics very often time and space are interchanged. Due
to the finite size of the fibers and due to the experimental data which can be
measured initial conditions are posed at one end of the fiber, namely at x =
0, and one is interested in the solution at the end of the fiber, namely at x =
444 11. Wave packets and the NLS equation

xe , i.e., x is considered as evolutionary variable, and t as unbounded variable.

From a mathematical point of view there is no difference if no dissipation is
considered. However, very often phenomenologically dissipation is added to
the NLS equation. It has been explained in [SU03b] that then it is highly
problematic if x and t is interchanged.

In a one-dimensional optical fiber without damping and one kind of

oscillators in the material we have the so called Maxwell-Lorentz system

(11.85) ∂x2 u =∂t2 u + ∂t2 p,

(11.86) ∂t2 p + ω02 p − r|p|2 p =du,

with coefficients ω0 , r, and d. The linearized problem

∂x2 u = ∂t2 u + ∂t2 p, ∂t2 p + ω02 p = du

possesses solutions

u(x, t) = uk ei(kx+ωt) , p(x, t) = pk ei(kx+ωt)

which yields
    
−k 2 + ω 2 ω2 uk 0
= .
−d ω02 − ω 2 pk 0
We have non-trivial solutions if the determinant vanishes, i.e., if

(−k 2 + ω 2 )(ω02 − ω 2 ) + dω 2 = 0.

We find four curves of solutions ω = ω1,2,3,4 (k) which are sketched in Figure
11.8. The Fourier transformed system can be written as first order system
and then diagonalized, leading to

∂t 
c1 (k, t) =iω1 (k)
c1 (k, t) + nonlinear terms ,
..
.
∂t 
c4 (k, t) =iω4 (k)
c4 (k, t) + nonlinear terms ,

such that the Maxwell-Lorentz system falls in the abstract class of systems
considered in §11.3 for which an NLS equation can be derived.

Remark 11.7.3. Other models for the polarization are in use. The choice
∂t2 p = −u − u3 leads to the Klein-Gordon model (11.2) for which we justified
the NLS approximation in §11.2. Hence, the pulse dynamics present in
the NLS equation is present in the models used for the description of the
propagation of light pulses in glass fibers, too. Experimental observations
confirm this approximation and modeling of reality.
11.7. Nonlinear optics 445

Figure 11.8. The curves of eigenvalues for the Maxwell-Lorentz

system

11.7.2. Multiplexing. So far we considered pulses modulating one car-

rier wave with some basic wave number k0 . Now we use two or more carrier
waves with different basic wave numbers k1 , . . . , kN . It turns out that pulses
belonging to different carrier waves do not interact in lowest order, and thus
the use of more than one carrier wave allows to increase the rate of informa-
tion through the fiber. This concept is called multiplexing. In the following
we explain the underlying ideas with the help of the NLS approximation.
Let us consider here again the nonlinear wave equation with cubic non-
linearity
∂t2 u = ∂x2 u − u − u3
as original system. For the situation of N different carrier waves we make
the ansatz

N
 
εψmultiNLS = εAj ε(x − cj t), ε2 t ei(kj x+ωj t) + c.c..
j=1

A system of coupled NLS equations is obtained, namely

2iωj ∂T Aj = (1 − c2j )∂X
2
j
Aj − 3Aj |Aj |2 − coupling termsj
for j = 1, . . . , N , where the jth coupling term is given by

6 Aj |An |2 ,
|n|=1,...,N,n=j

where we assumed for simplicity that there are no resonant wave numbers
kj1 , . . . , kj4 with kj1 +. . .+kj4 = 0 and ωj1 +. . .+ωj4 = 0. If this assumption
is not satisfied there will be additional coupling terms Aj2 Aj3 Aj4 in the
equation for A−j1 which however can be handled as explained below, too.
At a first view there seems to be a full coupling between all equations.
However, looking more closely at the coupling terms shows that they have
446 11. Wave packets and the NLS equation

different arguments if cj = cn . We have for example

   2
Aj |An |2 =Aj ε(x − cj t), ε2 t An ε(x − cn t), ε2 t 
  2
=Aj (Xj , T ) An Xj − ε−1 (cn − cj )T, T  .
Hence, two spatially localized functions interact only on an O(ε)-time in-
terval w.r.t. the T -time scale of the NLS equation if cj = cn . Thus, the
influence of the coupling terms on the dynamics of the NLS equations is
only O(ε). Hence, for spatially localized solutions the NLS equations de-
couple and the dynamics of the modulations of the carrier waves can be
computed for each carrier wave individually by solving
2iωj ∂T Aj = (1 − c2j )∂X
2
j
Aj − 3Aj |Aj |2 .
The argument that spatially localized wave packets with different group
velocities do not interact in lowest order has been made rigorous in case of the
above NLS approximation first in [PW96]. The idea has been generalized
in [BF06] for the interaction of wave packets in various original systems.
It has also been used in case of the interaction of counter-propagating long
waves [Kal89, SW00b, SW02], cf. §12.1. In case of dissipative systems
mean-field coupled GL equations take the role of the NLS equation [Sch97].
The interaction of pulse solutions to different carrier waves can be de-
scribed very precisely such that [PW96] and [BF06] can be improved
strongly. The more detailed description is

N
 
εψmultiNLS = εAj ε(x − cj t − εψj (x, t)), ε2 t ei(kj x+ωj t+εΩj (x,t)) + c.c.,
j=1

where the Aj satisfy decoupled equations and there are explicit formulae
[CBSU07] for the pulse shifts εψj and the phase shifts εΩj . The internal
dynamics of the wave packets (described via the Aj ) and the interaction
dynamics of the wave packets (described via ψj and Ωj ) can be separated
up to very high order. An almost complete description of the interaction of
general localized NLS described wave packets can be found for the nonlinear
wave equation in [CBCSU08] and in [CBS12] for general dispersive wave
systems. As a consequence the shift of the underlying carrier wave and
the shift of the envelope both can be shown to be of order O(ε) instead of
O(1) w.r.t. the original x-variable. See also [SUW11] for an application to
oscillator chains.
Further Reading. The description of waves in nonlinear optics by mod-
ulation equations is an active field of research. Three main current topics
are:
• ultra-short pulses [SW04, CdR15, PS13, New16], see also Ex-
ercise 11.10;
11.7. Nonlinear optics 447

• the (numerical) analysis of 3D photonic crystals, and in partic-

ular their design to achieve favorable band structures [BFL+ 07,
STE+ 10];
• gap solitons, cf. §11.6.5, and related phenomena. Besides gap soli-
tons, models of type (11.65) support other solution families, for
instance so called Nonlinear Bloch Modes and “Out–of-Gap” soli-
tons, see, e.g., [Yan10] for a comprehensive overview, or [DU16]
for recent rigorous results. See also [DD13] for related results for
(simplified) nonlinear Maxwell’s equations.
NLS equations also occur as singular limits of other dipersive PDE systems.
Examples are the Zakharov system [AA88] or the Klein-Gordon-Zakharov
system [MN05]. In [UGS+ 15], a so called shrinking limit of the NLS equa-
tion on fat graphs has been considered, and an analysis somewhat similar to
1D and 2D photonic crystals deals with wave propagation on infinite periodic
metric graphs, used as models for nanotubes or graphen, see, e.g. [GPS17].
Exercises
11.1. Consider the nonlinear wave equation ∂t2 u = ∂x2 u − u − u3 , with x, t, u(x, t) ∈
R. Make an ansatz u(x, t) = εA(εx, ε2 t)eit + c.c. to derive a NLS equation for A.
Estimate the residual and prove an approximation result by considering the energy

E= u2 + (∂x u)2 + (∂t u)2 dx.
R
11.2. Consider the nonlinear wave equation ∂t2 u = Δu − u − u3 , with
x, y, t, u(x, y, t) ∈ R. Make an ansatz
u(x, t) = εA(ε(x − c0 t), εy, ε2 t)ei(k0 x+ω0 t) + c.c.
to derive a 2D NLS equation for A.
11.3. Consider the lattice differential equation
∂t2 un = un+1 − 2un + un−1 − 3un − u2n ,
with un = un (t) ∈ R for n ∈ Z.
a) Find solutions un (t) = ei(kn−ωt) of the linearized problem. Compute the disper-
sion relation.
b) Check the validity of the non-resonance condition.
c) Make the ansatz un (t) = εA(εn, ε2 t)eiω0 t +c.c.+O(ε2 ) and derive a NLS equation
for the amplitude A = A(X, T ).
11.4. Consider the FPU system ∂t2 un = W  (un+1 ) − 2W  (un ) + W  (un−1 ), with
W an analyticfunction, i.e., W  (u) = au + bu2 + . . .. Derive an evolution equation
ikn
for u(k, t) = n∈Z un (t)e , cf. [Mil06]. Find differences and similarities between
this evolution equation and the Fourier transform of a nonlinear dispersive PDE.
11.5. Consider the KdV equation ∂t u − 6u∂x u + ∂x3 u = 0.
a) Apply the Miura transformation u = v 2 + ∂x v and derive the mKdV equation
∂t v − 6v 2 ∂x v + ∂x3 v = 0.
448 11. Wave packets and the NLS equation

b) Plug in the ansatz εψv (x, t) = εA(ε(x − ct), ε2 t)ei(kx+ωt) + c.c. into the mKdV
equation and derive a NLS equation for A.
c) Justify this approximation by energy estimates and conclude the validity of an
approximation theorem for the approximation of the KdV equation by the NLS
equation [Sch11].
11.6. The dispersion relation of the water wave problem with surface tension is
given by
ω 2 = (k + σk3 ) tanh(k),
with surface tension parameter σ ≥ 0. Show that for σ ∈ (0, 1/3) beside k = 0 and
k = k0 there are two additional resonances k1 and k2 with k0 + k1 + k2 = 0 and
ω0 + ω1 + ω2 = 0.
11.7. Consider the system for the resonant three wave interaction
Ȧ1 = iγ1 A2 A3 , Ȧ2 = iγ2 A1 A3 , Ȧ3 = iγ3 A1 A2 ,
for Aj (t) ∈ C and coefficients γj ∈ R.
a) Prove that this is a Hamiltonian system with
An An
qn = sgn(γn )  , pn =  , H = i(A1 A2 A3 + A1 A2 A3 ).
|γn | |γn |
b) Prove that the quantities
A1 A1 A2 A2 A1 A1 A3 A3 A1 A2 A3 A3
J1 = − , J2 = − , J3 = −
γ1 γ2 γ1 γ3 γ2 γ3
are conserved under the flow of the system.
c) Prove for the Poisson brackets that {H, J1 } = {H, J2 } = {J1 , J2 } = 0. Use this
to show the complete integrability of the Hamiltonian system.
11.8. Solve the eigenvalue problem ∂x u(x) = v, ∂x v(x) = −s(x)λu(x), with the 1-
periodic function s(x) = χ[0,6/13] + 16χ(6/13,7/13) + χ[7/13,1] (x mod 1). Compute the
monodromy matrix Cλ and the discriminant D(λ) = trace Cλ . Hint. On intervals
where s is constant, the problem can be solved explicitly. We look for u ∈ Cb1 .
11.9. Consider the nonlinear wave equation ∂t2 u = ∂x2 u + u + 2ε2 cos(2x)u − u3 ,
with spatially periodic perturbed coefficients, i.e., 0 ≤ ε  1. For ε > 0 a spectral
gap of order O(ε2 ) occurs, which is too small to derive an NLS equation. In this
case with the ansatz
u(x, t) = εa(ε2 x, ε2 t)ei(x−ω0 t) + εb(ε2 x, ε2 t)e−i(x+ω0 t) + c.c.
derive the coupled mode system, cf. [SU01]
−2iω0 ∂T a =2i∂x a + ib − 3a|a|2 − 6a|b|2 ,
−2iω0 ∂T b = − 2i∂x b + ia − 3b|b|2 − 6b|a|2 .
11.10. If the pulses become very narrow, then the so called short pulse equation
∂ξ ∂τ A = A + ∂ξ2 (A3 ),
11.7. Nonlinear optics 449

can be derived. Consider the quasilinear wave equation ∂t2 u = ∂x2 u + u + ∂x2 (u3 ),
and make the ansatz
x−t
u(t, x) = 2εA(τ, ξ), τ = εt, ξ =
2ε
to derive the short pulse equation, cf. [PS13].
11.11. Show that the inhomogeneous Maxwell equations
∂t E = ∇ × B − J, ∇ · E = ρ, ∂t B = −∇ × E, ∇ · B = 0,
can be transformed to the inhomogeneous wave equations
∂t2 u − Δu = ρ, ∂t2 A − ΔA = J,
where A is a vector potential of B, i.e., ∇ × A = B, and u is a scalar potential of
E, i.e., E + ∂t A = −∇u, see Remark 11.7.1.
Hint. Consider a suitable gauge transform, i.e., adding ∇λ to A and subtracting
−∂t λ from u.
 Chapter 12

Long waves and their

modulation equations

It is the purpose of this chapter to explain the role of the KdV equation as
a long wave modulation equation for various dispersive wave systems. Such
approximations have a long history in science, and they not only play a
role in the description of the water wave problem but also for other nonlin-
ear dispersive systems such as the equations of plasma physics or the FPU
model.
One of the simplest systems where the KdV equation can be derived is
the so called Boussinesq equation
(12.1) ∂t2 u = ∂x2 u − ∂x4 u + ∂x2 (u2 ),
with x ∈ R, t ∈ R, and u(x, t) ∈ R. The long wave ansatz
u(x, t) = ε2 v(ξ, τ ) = ε2 v(εx, εt),
where 0<ε1 is a perturbation parameter, yields ∂τ2 v=∂ξ2 v+O(ε2 ), hence
(12.2) ∂τ2 v = ∂ξ2 v
in lowest order. The linear wave equation (12.2) is solved by
v(ξ, τ ) = v+ (ξ − τ ) + v− (ξ + τ ),
with functions v+ and v− satisfying

v(ξ, 0) = v+ (ξ) + v− (ξ) and ∂τ v(ξ, 0) = −v+ (ξ) + v− (ξ).
Hence, the formal long wave approximation predicts a wave moving to the
left and a wave moving to the right. However, this approximation is formally
only valid on a time scale of order O(1/ε). Nevertheless the right and

451
452 12. Long waves and their modulation equations

left moving wave will be separated for spatially localized initial conditions.
Afterwards the wave packets will evolve individually. For each of the wave
packets a KdV equation can be derived, and these are formally valid on the
much longer time scale of order O(1/ε3 ). In detail, inserting the ansatz
(12.3) u(x, t) = ε2 A+ (ε(x − t), ε3 t) + ε2 A− (ε(x + t), ε3 t).
into the Boussinesq equation and equating the coefficient at ε6 to zero yields
the system of two decoupled KdV equations
(12.4) 2∂T A+ − ∂X
3
A+ + ∂X (A2+ ) = 0, −2∂T A− − ∂X
3
A− + ∂X (A2− ) = 0,
one for each of the wave packets. There are some coupling terms which
however turn out to be of higher order for spatially localized initial condi-
tions. Estimates that this formal KdV approximation and true solutions of
the Boussinesq model stay close together over the natural KdV time scale
are a non-trivial task since solutions of order O(ε2 ) have to be estimated on
an O(1/ε3 ) time scale. This question will be discussed in §12.1.
It turns out that the set of KdV equations can be derived whenever
the dispersion relation for harmonic waves ei(kx+ω(k)t) gives two curves of
eigenvalues k → ω± (k) with ω± (0) = 0, and the nonlinear terms also vanish
at the spatial wave number k = 0. This is the reason why the KdV equation
occurs as a universal modulation equation, as discussed in §12.2. There, we
introduce the water wave problem, the equations of plasma physics and the
FPU system and show that they belong to the class of systems for which the
KdV equation can be derived. There are many other modulation equations
which can be derived in the long wave limit. Some of them are independent of
the small perturbation parameter ε such as the inviscid Burgers equation or
the Whitham system and some depend on the small perturbation parameter
ε such as the regularized long wave equation or the Benjamin-Bona-Mahony
system. We will comment on the use of these equations and on their validity
in §12.3. In §12.4 we explain that in the so called long wave limit, where
a big number of these modulation equations have been derived, the above
splitting in two decoupled KdV equations is sufficient to describe the original
system.

12.1. An approximation result

The goal of this section is to prove an approximation result for a single KdV
equation. This is a non-trivial task due to the fact that solutions of order
O(ε2 ) have to be estimated on an O(1/ε3 ) time scale. Hence, one power of
ε is missing. There are at least two essentially different approaches.
In the first approach one works in a family of spaces of analytic functions
where the strip of analyticity shrinks with time. This shrinking corresponds
to some smoothing which allows to gain the missing power of ε by using
12.1. An approximation result 453

the derivative in front of the KdV nonlinearity. This approach has been
used in [KN86], but as pointed out in [CSS92] the error estimates for the
KdV approximation there have only been shown on an O(1/ε) time scale
instead of the O(1/ε3 ) KdV time scale. However, as pointed out in [Sch96b]
the approach of [KN86] can be extended to the correct O(1/ε3 ) KdV time
scale. This approach turns out to be very robust, but does not give optimal
results, cf. [CDS15] for another example.
In this book we will present the second approach which in general gives
better estimates and is conceptionally easier, but less robust. For the Boussi-
nesq equation (12.1) the KdV approximation can be justified with a simple
energy estimate. An energy is constructed in such a way that the influence
of the terms of order O(ε2 ) on the growth of the solutions can be estimated
to be of order O(ε3 ). The papers [Cra85, SW00b, SW02, Dül12] where
the KdV approximation has been justified for the water wave problem work
this way. The major difficulty in transferring this proof from the Boussinesq
equation to the water wave problem or other applications is to bring together
the energy estimates with the local existence and uniqueness theory for the
various systems. For the Boussinesq equation our approximation result is
as follows.
Theorem 12.1.1. Fix θ > 1/2 and let A+ = A ∈ C([0, T0 ], H θ+4 ) and
A− = 0 be solutions of the system of KdV equations (12.4). Then there
exist ε0 > 0 and C > 0 such that for all ε ∈ (0, ε0 ] we have solutions u of
(12.1) with
sup u(·, t) − ε2 A(ε(· − t), ε3 t)H θ ≤ Cε7/2 .
t∈[0,T0 /ε3 ]

Proof. For notational simplicity we only consider the case θ = 2. The

residual
Res(u) = −∂t2 u(x, t) + ∂x2 u(x, t) − ∂x4 u(x, t) + ∂x2 (u(x, t)2 )
quantifies how much a function u fails to satisfy the Boussinesq model (12.1).
For the approximation
ε2 Ψ(x, t, ε) = ε2 A(ε(x − ct), ε3 t)
we find
Res(ε2 Ψ) = − ε4 c2 ∂X
2
A + 2ε6 ∂T ∂X A − ε8 ∂T2 A
+ ε4 ∂ X
2
A − ε6 ∂ X
4
A + ε6 ∂ X
2
(A2 )
= − ε8 ∂T2 A
if we choose c2 = 1 and A to satisfy the first of the two KdV equations of
(12.4). Therefore, we have
454 12. Long waves and their modulation equations


Lemma 12.1.2. Fix θ ≥ 0 and let A+ = A ∈ C([0, T0 ], H θ+6 ) and A− = 0
be solutions of the system of KdV equations (12.4). Then there exist ε0 > 0,
Cres such that for all ε ∈ (0, ε0 ) we have
sup  Res(ε2 Ψ(·, t, ε))H θ ≤ Cres ε15/2
t∈[0,T0 /ε3 ]

and
sup ∂x−1 Res(ε2 Ψ(·, t, ε))H θ ≤ Cres ε13/2 .
t∈[0,T0 /ε3 ]

Proof. Using the KdV equation allows us to write

4∂T2 A = − 2∂T (∂X
3
A − ∂X (A2 )) = 2(∂X
3
∂T A − 2∂X (A∂T A))
3
=∂X 3
(∂X A − ∂X (A2 )) − 2∂X (A(∂X
3
A − ∂X (A2 ))).
 
This shows that A(·, T ) ∈ H θ+6 is necessary to estimate the residual in H θ .
−1
Moreover, we can apply ∂x−1 = ε−1 ∂X to the residual terms which however
loses a factor ε−1 . The remaining formal error of order O(ε7 ) is reduced
further by a factor ε−1/2 due to the scaling properties of the L2 -norm, cf.
(10.18). 
The difference ε7/2 R = u − ε2 Ψ satisfies
(12.5) ∂t2 R = ∂x2 R − ∂x4 R + 2ε2 ∂x2 (ΨR) + ε7/2 ∂x2 (R2 ) + ε−7/2 Res(ε2 Ψ).
In order to estimate R we construct an energy E with dt d
E = O(ε3 ). Thus,
we multiply the error equation (12.5) with ∂t R and integrate it w.r.t. x. We
find with integration by parts and τ = εt that
 
2 1 d
(∂t R)∂t R dx = (∂t R)2 dx,
2 dt
R R

1 d
(∂t R)∂x R dx = −
2
(∂x R)2 dx,
2 dt
R R
1 d
− (∂t R)∂x4 R dx = − (∂ 2 R)2 dx,
R 2 dt R x
  
(∂t R)∂x (ΨR) dx = − (∂t ∂x R)Ψ(∂x R) dx − ε (∂t ∂x R)(∂X Ψ)R dx
2
R R
  R
d
=− Ψ(∂x R) dx/2 + ε (∂τ Ψ)(∂x R)2 dx
2
dt R R
 
2 2
+ ε (∂t R)(∂X Ψ)∂x R dx + ε (∂t R)(∂X Ψ)R dx,
R R
 
(∂t R)∂x2 (R2 ) dx = − 2 (∂t ∂x R)R(∂x R) dx
R R
 
d
=− R(∂x R)2 dx + (∂x R)2 (∂t R) dx.
dt R R
12.1. An approximation result 455

All terms which can be written as a time derivative are included in E. The
others are estimated, namely
 
 
 (∂t R) Res(ε2 Ψ) dx ≤∂t RL2  Res(ε2 Ψ)L2 ,
 
R
 
 
 (∂τ Ψ)(∂x R)2 dx ≤∂τ ΨL∞ ∂x R2 2 ,
  L
 R 
 
 (∂t R)(∂X Ψ)∂x R dx ≤∂X ΨL∞ ∂t RL2 ∂x RL2 ,
 
R
 
 
 (∂t R)(∂ 2 Ψ)R dx ≤∂ 2 ΨL∞ ∂t RL2 RL2 ,
 X  X
R 
 
 (∂x R)2 (∂t R) dx ≤∂x RL∞ ∂t RL2 ∂x RL2
 
R
≤C∂x RH 1 ∂t RL2 ∂x RL2 ,

with C a constant coming from Sobolev’s embedding theorem. Next we

multiply the error equation with −∂t ∂x−2 R which is defined via its Fourier
transform w.r.t. x, i.e. ∂x−1 R = F −1 ((ik)−1 R) integrate it w.r.t. x, and find
 
−2 1 d
− (∂t ∂x R)∂t R dx =
2
(∂t ∂x−1 R)2 dx,
R 2 dt R
 
−2 1 d
− (∂t ∂x R)∂x R dx = −
2
R2 dx,
2 dt
R R
−2 1 d
(∂t ∂x R)∂x R dx = −
4
(∂x R)2 dx,
R 2 dt R
 
− (∂t ∂x−2 R)∂x2 (ΨR) dx = − (∂t R)ΨR dx
R R
 
d
=− ΨR dx/2 + ε (∂τ Ψ)R2 dx,
2
dt R R
  
−2 2 d
− (∂t ∂x R)∂x (R ) dx = − 2 (∂t R)R dx = −
2 2 2
R3 dx,
3 dt
 R  R R

− (∂t ∂x−2 R) Res(ε2 Ψ) dx = (∂t ∂x−1 R)∂x−1 Res(ε2 Ψ) dx.

R R

We can estimate
 
 
 (∂t ∂ −1 R)∂ −1 Res(ε2 Ψ) dx ≤∂t ∂ −1 RL2 ∂ −1 Res(ε2 Ψ)L2 ,
 x x  x x
R
 
 
 (∂τ Ψ)R2 dx ≤∂τ ΨL∞ R2 2 .
  L
R
456 12. Long waves and their modulation equations

We collect all total time derivatives in the energy


E = (∂t ∂x−1 R)2 + R2 + (∂t R)2 + 2(∂x R)2 + (∂x2 R)2
R
(12.6) + 2ε2 ΨR2 + 2ε2 Ψ(∂x R)2 + 4ε7/2 R3 /3 + 2ε7/2 R(∂x R)2 dx.
Since for ε > 0 sufficiently small the second line of E can be estimated by
the first line we find that for all M > 0 there exist C1 , ε1 > 0 such that for
all ε ∈ (0, ε1 ) we have
RH 2 ≤ C1 E 1/2
as long as E ≤ M . Therefore, E satisfies the inequality
d
(12.7) E ≤ Cε3 E + Cε7/2 E 3/2 + Cε3 E 1/2 ≤ 2Cε3 E + Cε7/2 E 3/2 + Cε3 ,
dt
with a constant C independent of ε ∈ (0, ε1 ). Under the assumption that
Cε1/2 E 1/2 ≤ 1 we obtain
d
E ≤ (2C + 1)ε3 E + Cε3 .
dt
Gronwall’s inequality immediately gives the bound
sup E(t) = CT0 e(2C+1)T0 =: M = O(1)
t∈[0,T0 /ε3 ]

and so supt∈[0,T0 /ε3 ] R(t)H 2 = O(1). Finally choosing ε2 > 0 so small

1/2
that Cε2 M 1/2 ≤ 1 gives the required estimate for all ε ∈ (0, ε0 ) with
ε0 = min(ε1 , ε2 ) > 0. Since the residual has only to be estimated in L2 , we
need A ∈ H 6 to estimate the error in H 2 . For general θ we set θ = θ − 2
such that θ + 4 = θ + 6 in Lemma 12.1.2. 

12.2. The universality of the KdV equation

As already said, the KdV approximation can be derived for various systems.
In order to explain why this is the case and why the KdV equation plays
such an important role, we review the derivation of the KdV equation from
a different point of view. This derivation will explain why the KdV equation
occurs as a universal modulation equation whenever the curves of eigenvalues
k → ω± (k) of the underlying dispersive wave system satisfy ω± (0) = 0, and
the nonlinear terms also vanish at the wave number k = 0. In [Bri13] it has
been shown that this is the case under very weak conditions. The underlying
system will condense in the values of the coefficients of the associated KdV
equation. We present with the water wave problem, with the FPU system,
and with the equations of plasma physics three dispersive systems for which
the KdV approximation plays a role.
12.2. The universality of the KdV equation 457

As in §10.3 and §11.3 it turns out that Fourier transform is the key for
the understanding of the universality. Hence, we consider the Boussinesq
equation (12.1) in Fourier space. The Fourier transform u satisfies
(12.8) (k, t) = −ω 2 (k)
∂t2 u u∗2 (k, t),
u(k, t) − ρ(k)
√
where ω(k) = k 2 + k 4 sign(k) and ρ(k) = k 2 . By introducing w(k)  =
1
(k)) we rewrite (12.8) into the first order system
u(k), ω(k) ∂t u
(

(12.9) ∂t w(k, 4(k)w(k,

 t) = M  (w)(k,
 t) + N  t),
where
   
0 iω(k) 0
4(k) =
M ,  (w)(k,
N  t) = ρ(k) ∗2 .
iω(k) 0 − iω(k)  (k, t)
u

This system is diagonalized for fixed wave number k. For (12.9) the associ-
ated transformation  
=√ 1 1 1
U
2 1 −1
 −1 = U
is independent of k and unitary, i.e., U  ∗ , cf. §11.1 where the same
transformation has been used a number of times. The transformed variable
 ∗w
z = U  satisfies the diagonalized system
(12.10) z + U
∂t z = Λ  ∗N
 (U
 z),

with Λ(k) = diag(iω(k), −iω(k)). In accordance with (12.3), for (12.10) we
make the ansatz
 
2 −1  k 3
z(k, t) = ε ε A+ , ε t eickte1
ε  
(12.11)
2 −1  k 3
+ε ε A− , ε t e−ickte2 ,
ε
where    
1 0
e1 = and e2 = .
0 1
Since the Fourier modes of the wave packets are concentrated in an O(ε)
neighborhood of the wave number k = 0, the evolution of the wave packets
will strongly be determined by the curves ±ω at k = 0. At eickte1 we find
1 + ε4 ∂T A
iε2 cK A 1

1 + i ε4 ∂ 3 ω(0)K 3 A
= iε2 ∂k ω(0)K A 1 − ε4 ∂k ρ(0) iK A
1 ∗ A
1 + O(ε6 ),
k
6 ω(0)
where k = εK. At ε2 we obtain the linear group velocity. At ε4 we obtain
a KdV equation.
458 12. Long waves and their modulation equations

By this procedure it is clear that the KdV equation occurs as modulation

equation of dispersive systems
(12.12) ∂t zj (k, t) = iωj (k)
zj (k, t) + nonlinear terms
for j in some index set whenever the Fourier transform of the initial condition
is strongly concentrated at k = 0, when the linear (ω1 (0) = ω−1 (0) = 0) and
nonlinear terms vanish at this wave number, and when the concentration
and the amplitude are of the correct order.

ω ω

k→
 k3 /6
k→ ω(k)

k k

Figure 12.1. The derivation of the KdV equation is based on the

concentration of the Fourier modes at the wave number k = 0. The
left panel shows the curves of eigenvalues. Hence, for the evolution
of these concentrated Fourier modes only the curves of eigenvalues
close to the wave number k = 0 plays a role. The right panel
therefore shows a blow-up of these curves in a frame moving with
the group velocity ω  (0) and an expansion of one curve at the wave
number k = 0.

12.2.1. The water wave problem. The most famous system where the
KdV equation can be derived is the so called water wave problem. Following
§12.1, there are no conceptional difficulties to justify this approximation,
but the realization of this approach for the water wave problem is rather
lengthy and involves a big number of estimates. Therefore, it is the only
goal of this section to explain that the water wave problem falls into the
class of systems for which the KdV equation can be derived. References to
the KdV approximation for the full water wave problem can be found at the
end of the chapter.
The water wave problem consists in finding the irrotational flow of an
inviscid incompressible fluid in an infinitely long canal of fixed finite depth
with an unknown free top surface subject to gravitational forces. The bot-
tom is impermeable and for expository reasons surface tension is neglected
in this subsection. The coordinates are denoted in horizontal direction by
x1 ∈ R and in the vertical bounded direction by x2 . The fluid fills the
domain Ω(t) in between the bottom {(x1 , −h) : x1 ∈ R} and the unknown
free top surface Γ(t). In the Eulerian formulation the free surface Γ(t) is
12.2. The universality of the KdV equation 459

parameterized as function η over the bottom, i.e., we have

Γ(t) = {(x1 , x2 ) : x2 = η(x1 , t)}.
The velocity field u = (u1 , u2 )(x, t) ∈ R2 and the pressure field p = p(x, t) ∈
R in Ω(t) are governed by Euler’s equations consisting of the balance of
forces
(12.13) ∂t u + (u · ∇)u = −∇p + g(0, −1),
and the incompressibility condition
(12.14) ∂x1 u1 + ∂x2 u2 = 0,
with the constant of gravitational acceleration g ∈ R. W.l.o.g. we assume
h = 1 in the following.

x2
Γ (t)

0
x1
Ω (t) u2
u1
−1

Figure 12.2. The water wave problem

It turns out that a solution with an irrotational initial condition stays

irrotational under the evolution of Euler’s equation, cf (6.5), and so we can
assume
(12.15) ∂x2 u1 − ∂x1 u2 = 0,
i.e., vanishing vorticity. Hence, we have a potential flow u = ∇φ with poten-
tial φ : Ω(t) → R. The impermeability of the bottom yields the boundary
condition u2 |x2 =−1 = 0. This system is completed with the boundary con-
ditions for the free surface Γ. We assume that (1, u1 , u2 ) is tangential to
(t, Γ(t)). Differentiating x2 = η(x1 , t) w.r.t. t, this is transferred into the
dynamic boundary condition u2 = (∂x1 η)u1 + ∂t η, respectively
(12.16) ∂t η = ∂x2 φ − (∂x1 φ)∂x1 η.
We finally assume that for fixed time t the pressure p is constant along the
free surface Γ(t) such that after inserting u = ∇φ into (12.13), this equation
transforms into the Bernoulli equation
1
(12.17) ∂t φ = ((∂x1 φ)2 + (∂x2 φ)2 ) − gη
2
on the free surface Γ(t), cf. Exercise 12.5. If surface tension is included,
then the last assumption has to be modified.
460 12. Long waves and their modulation equations

It turns out that the evolution of the problem is completely determined

by the elevation η = η(x1 , t) of the top surface and the horizontal velocity
component w = w(x1 , t) = ∂1 φ(x1 , η(x1 , t), t) at the top surface. Inserting
u = ∇φ into (12.14) yields Δφ = 0 which can be solved uniquely up to a
constant if the boundary conditions u2 |x2 =−1 = ∂x2 φ|x2 =−1 = 0 and w =
∂1 φ(x1 , η(x1 , t), t) are given. Hence all derivatives on the right-hand side of
(12.16) and (12.17) can be computed if η and w are known, and the water
wave problem can be written as an evolutionary system in η and w, alone.
The linear water wave problem. Next we explain that the water
wave problem belongs to the class of systems for which the KdV equation
can be derived and which has been described above. In order to do so we
start with the linear problem
(12.18) ∂t η = ∂x2 φ|x2 =0 , ∂t φ|x2 =0 = −η,
where φ solves Δφ = 0 in Ω0 = {(x1 , x2 ) : x1 ∈ R, x2 ∈ (−1, 0)} under the
boundary conditions ∂x2 φ|x2 =−1 = 0 and φ|x2 =0 = Φ.
Lemma 12.2.1. The operator Φ → ∂x2 φ|x2 =0 is a multiplication operator
in Fourier space defined by ∂x 
2 φ|x2 =0 (k) = k(tanh k)Φ(k).

Proof. The Fourier transform φ satisfies

 x2 ) + ∂ 2 φ(k,
−k 2 φ(k,  x2 ) = 0
x2
 x =−1 = 0. We obtain solutions
with the boundary conditions ∂x2 φ| 2

 x2 ) = ck cosh(k(1 + x2 )),
φ(k,
 x =0 (k) = ck cosh(k) and ∂x
and so Φ| 2 2 φx2 =0 (k) = kck sinh(k). Therefore,
we have
∂x  
2 φx2 =0 (k) = k(tanh k)Φ(k).

Remark 12.2.2. In a canal of finite depth h the factor tanh(hk) is obtained.

In the limit of infinite depth we obtain the Heaviside function H,  where
 
H(k) = 1 for k ≥ 0 and H(k) = −1 for k < 0.

Hence, (12.18) possesses solutions (Φ, η) = (Φk , ηk )ei(kx+ωt) with the

linear dispersion relation
(12.19) ω 2 = k tanh(k),
and we have two curves of eigenvalues iω1,2 satisfying iω1,2 (0) = 0. We
explained above that this is a necessary condition for the derivation of the
KdV equation.
 have a ∂x1 in front
In the equation for w all terms on the right hand side
by construction. Due to the conservation of mass, i.e. η(x1 , t) dx1 = const.
also all terms on the right-hand side in the equation for η must vanish for
12.2. The universality of the KdV equation 461

1.5
1
0.5
ω(k)
0
−ω(k)
−0.5
−1
−1.5
−3 −2 −1 0 1 2 3

Figure 12.3. The curves of eigenvalues k → ω1,2 (k) in case of zero

surface tension.

k = 0. Hence the water wave problem falls into the class of systems described
in §12.2.
The KdV approximation. In case of positive surface tension (12.17)
changes into
 
1 ∂x1 η
(12.20) ∂t φ = ((∂x1 φ)2 + (∂x2 φ)2 ) − gη + σ∂x1  ,
2 1 + (∂x1 η)2
where σ is the surface tension parameter. The linear dispersion relation
then modifies into
(12.21) ω 2 = (k + σk 3 ) tanh(k).
There are essentially three situations, namely σ ∈ [0, 1/3), σ = 1/3, and
σ > 1/3. Expanding the curves of eigenvalues k → ω1,2 (k) at the wave
number k = 0 gives
1 1
ω1 (k) = k + (σ − )k 3 + O(k 5 ).
2 3
Hence, the sign in front of the cubic terms changes at σ = 1/3. At the same
value the inflection points of the curves k → ω1,2 (k) disappear. See Figure
12.4.
ω ω

k k

Figure 12.4. The curves of eigenvalues k → ω1,2 (k) of the water wave
problem with surface tension with σ ∈ (0, 1/3) in the left panel and
σ > 1/3 in the right panel.
462 12. Long waves and their modulation equations

This is reflected in the associated KdV equations

1 3
2∂T A+ =(σ − )∂X A+ − ν2 ∂X (A2+ ),
3
1
2∂T A− =( − σ)∂X 3
A− + ν2 ∂X (A2− ),
3
with coefficient ν2 ∈ R, which are derived by the ansatz
     
η 1 1
(x, t) = ε A+ (ε(x − t), ε t)
2 3 2 3
+ ε A− (ε(x + t), ε t) .
w 1 −1
It turns out that ν2 > 0 and so the solitary waves change at σ = 1/3 from
waves of elevation to waves of depression.
At σ = 1/3 the coefficient in front of k 3 vanishes. The regime σ =
1/3 + νε2 , with ν ∈ R, independent of 0 < ε  1, is called the Kawahara
regime. With the modified ansatz
     
η 1 1
(x, t) = ε A+ (ε(x − t), ε t)
4 5 4 5
+ ε A− (ε(x + t), ε t)
w 1 −1
a system of two decoupled Kawahara equations
3 A − ν ∂ 5 A + ν ∂ (A2 ),
−2∂T A+ = ν∂X + 1 X + 2 X +
(12.22)
2∂T A− = ν∂X A− − ν1 ∂X A− + ν2 ∂X (A2− ),
3 5

with coefficients νj ∈ R can be derived. The Kawahara equation has been

derived first in [Kaw72], describes waves longer than the ones in the KdV
scaling, and possesses rather complicated solutions such as multi-pulse so-
lutions or solutions with a spatially chaotic behavior [BGT96]. Error esti-
mates for the Kawahara approximation can be found in [SW01].

12.2.2. The FPU system. The most famous conservative lattice differ-
ential equation where the KdV and NLS (see Exercise 11.4) equation can
be derived is the so called Fermi-Pasta-Ulam (FPU) system
(12.23) ∂t2 qj = W (qj+1 (t) − qj (t)) − W (qj (t) − qj−1 (t)) , j ∈ Z .
It was first studied numerically by Fermi, Pasta, and Ulam [FPU55] for
a finite set of oscillators in order to see how energy was spread through
the various modes of the system by the nonlinear coupling via the inter-
particle forces which are described by the potential function W : R → R.
They found that at low energy most trajectories did not “thermalize” as
expected, but rather exhibited a regular motion. This observation has been
explained in [ZK65], where Kruskal and Zabusky derived the KdV equation
as a formal approximation to the FPU system and in studying the KdV
equation numerically they found soliton dynamics. A rigorous proof that
long waves in the FPU system can be approximated via the KdV equation
has been given in [SW00a].
12.2. The universality of the KdV equation 463

It is the goal of this section to explain how lattice equations can be

brought into the abstract form from §11.3 and the beginning of §12.2 for the
derivation of these equations. The FPU system (12.23) is rewritten in terms
of the difference variables u(j, t) = qj+1 (t) − qj (t), so that (12.23) becomes
(12.24) ∂t2 u(j, t) = W (u(j + 1, t)) − 2W (u(j, t)) + W (u(j − 1, t)) , j ∈ Z .
With
(12.25) W (u) = a1 u + a2 u2 + a3 u3 + . . . , (a1 > 0, aj ∈ R for j ∈ N),
the linear problem is given by
(12.26) ∂t2 u(j, t) = a1 (u(j + 1, t) − 2u(j, t) + u(j − 1, t)).
It is solved by
(12.27) u(j, t) = ei(kj+ωt)
for all k ∈ R, and ω and k related through the linear dispersion relation
(12.28) ω 2 = −a1 (eik − 2 + e−ik ) = 2a1 (1 − cos k).
In order to simplify the notation in the following we assume that
(12.29) W (u) = u + u2 ,
i.e., we set a1 = a2 = 1, and except of the dangerous quadratic terms
we neglect all higher order terms. In order to relate the FPU model to
the abstract systems considered in §11.3 and at the beginning of §12.2 we
consider the Fourier transformed FPU model. Since
F (u(· + 1)2 − 2u(·)2 + u(· − 1)2 )(k)
1 
= (u(j + 1)2 − 2u(j)2 + u(j − 1)2 )e−ikj
2π
j∈Z
1  ik 1 
= (e − 2 + e−ik )u(j)2 e−ikj = −ω(k)2 u(j)2 e−ikj
2π 2π
j∈Z j∈Z
  π
1
= −ω(k)2 u(j)2 e−ikj = −ω(k)2 (k − l)
u u(l) dl,
2π −π
j∈Z

the FPU system in Fourier space is given by

 π
(12.30) (k, t)
∂t2 u = −ω (k)
2
u(k, t) − ω (k) 2
(k − m, t)
u u(m, t) dm
−π

where ω 2 has been defined in (12.28) and where in the convolution integrals
 has to be used. Hence, for the derivation of the
the 2π-periodicity of u
NLS and KdV equations and for the justification of the associated approx-
imations we can proceed as in §11.1 and as above. A proof that the NLS
approximation makes correct predictions for the FPU model can be found
464 12. Long waves and their modulation equations

in [Sch10]. The KdV and NLS approximation has been justified for the
poly-atomic FPU model in [CCPS12].

12.2.3. The equations of plasma physics. A plasma is modeled as an

ionized fluid consisting of positively charged ions and negatively charged
electrons which interact through the electro-magnetic field they generate.
The determining equations for the densities, velocities, and pressures of the
fluids and the electromagnetic field consist of the equations for the conser-
vation of mass and momentum for the ion fluid and the electron fluid and
Maxwell’s equation. Under a number of assumptions these equations are
simplified. The consideration of a ’cold’ ion-fluid allows to neglect the ion
pressure. Since the electrons are much lighter than the ions their inertia is
neglected. An isothermal constitutive law between the electron pressure and
density is chosen, and the time-dependent Maxwell’s equations are replaced
by electrostatics. In the one-dimensional situation the non-dimensionalized
equations are then given by an Euler-Poisson system
(12.31) ∂t n + ∂x (nu) = 0, ∂t u + u∂x u + ∂x ϕ = 0, −∂x2 ϕ + eϕ = n,
with ϕ the electric potential, n the ion density, and u the ion velocity. This
system possesses a trivial solution (n, u, ϕ) = (1, 0, 0). We introduce the
density fluctuation v = n − 1 and find
(12.32) ∂t v + ∂x ((1 + v)u) = 0, ∂t u + u∂x u + ∂x ϕ = 0,
and
−∂x2 ϕ + eϕ = 1 + v.
The linearized system is given by
(12.33) ∂t v + ∂x u = 0, ∂t u + ∂x ϕ = 0, −∂x2 ϕ + ϕ = v.
It is solved by the Fourier modes (v, u, ϕ) = ( , ϕ)e
v, u  ikx+iωt . We find with
 = v/(1 + k 2 ) that
ϕ
k2
ω2 = .
1 + k2
For small v the third equation of (12.32) can be solved w.r.t. ϕ, i.e., we
have
ϕ(v) = (−∂x2 + 1)−1 v + S(v),
with S(v)H s+2 ≤ Cv2H s for small v and s > 1/2. Hence, (12.32) can be
written as evolutionary system in u and v alone, namely
∂t v + ∂x ((1 + v)u) = 0, ∂t u + ∂x (u2 )/2 + ∂x ((−∂x2 + 1)−1 v + S(v)) = 0.
This representation shows that the right-hand side vanishes in Fourier space
at the wave number k = 0 and so this system falls into the class of sys-
tems which has been described above and for which a KdV equation can be
12.3. Whitham, Boussinesq, BBM, etc. 465

derived. With the ansatz

   
v 1
= ε A(ε(x − t), ε t)
2 3
u 1
we find the KdV equation
1 3
∂T A + ∂X A + A∂X A = 0.
2
An approximation result for this KdV approximation has been proven in
[GP14].

12.3. Whitham, Boussinesq, BBM, etc.

There are many other modulation equations which have been derived in the
long wave limit. It is the purpose of this section to comment on the use
of these equations and to discuss their validity. We start with modulation
equations which are independent of ε and then come to the zoo of modulation
equations which has been derived for the water wave problem by using lower
order relations to modify the higher order terms.
Modulation equations which are independent of ε. For (12.1) we
make the ansatz
(12.34) u(x, t) = εα A(X, T ),
where X = ε(x−t), T = ε1+α t, A(X, T ) ∈ R, and α > 0. For α > 2 the Airy
equation ∂T A = 12 ∂X3 A occurs. The KdV equation is recovered for α = 2,

and for α ∈ (0, 2) the inviscid Burgers equation

1
(12.35) ∂T A = − ∂X (A2 )
2
is obtained. There is another long wave limit which leads to an ε-independent
non-trivial amplitude equation. With the ansatz
(12.36) u(x, t) = U (X, T ),
where X = εx, T = εt, U (X, T ) ∈ R, and 0 < ε  1 a small perturbation
parameter, we obtain
(12.37) ∂T2 U = ∂X
2 2
U + ∂X (U 2 )
which can be written as a system of conservation laws
(12.38) ∂T U = ∂X V, ∂T V = ∂X U + ∂X (U 2 ).
This so called Whitham limit has been considered first in the description
of slow modulations in time and space of a periodic traveling wave in a
dispersive wave system [Whi99, CS98]. Since solutions of order O(1) are
considered an additional smallness condition in the approximation theorem
is necessary [DS09, BDS16]. The proof of approximation results for these
modulation equations goes almost line for line as in §12.1.
466 12. Long waves and their modulation equations

The zoo of ε-dependent long wave modulation equations. Other

modulation equations, such as the regularized long wave equation, the BBM
system, etc., have been derived in the long wave limit, too. There have been
many scientific discussions about the question which equation of this zoo of
long wave modulation equations is the most suitable one for the description
of the long wave limit. The Boussinesq equation (12.1) and modified versions
of it can be derived in the long wave limit from the water wave problem,
too. However, then it still contains the small perturbation parameter ε, i.e.,
it is of the form
(12.39) ∂τ2 v = ∂ξ2 v − ε2 ∂ξ4 v + ε2 ∂ξ2 (v 2 ),

with u = ε2 v, τ = εt, and ξ = εx. Very often the statement can be

found that it only can be derived under the assumption of unidirectional
waves. This is wrong. For instance in [Cra85, SW00b] it has been derived
without using this assumption for the so called Lagrangian formulation of
the water wave problem. For the Eulerian formulation of the water wave
problem we have η(x, t) ∼ ε2 v(εx, εt), while the expression for w is more
complicated. Instead of the water wave problem we take (12.39) as starting
point for the derivation of the above mentioned modulation equations. Using
∂τ2 v = ∂ξ2 v + O(ε2 ) yields the system

(12.40) ∂τ2 v = ∂ξ2 v − ε2 ∂τ2 ∂ξ2 v + ε2 ∂ξ2 (v 2 ) + O(ε4 )

which in the long wave limit is formally equivalent to (12.39) up to terms of
order O(ε4 ). Ignoring the terms of order O(ε4 ) however gives an ill-posed
system. In principle by this approach one gets a two parameter family of
equations, namely
∂τ2 v = ∂ξ2 v − (1 − α)ε2 ∂ξ4 v − αε2 ∂τ2 ∂ξ2 v + (1 − β)ε2 ∂ξ2 (v 2 ) + βε2 ∂τ2 (v 2 ),

with arbitrary α, β ∈ R. If our starting Boussinesq system contains +∂x4 u

instead of −∂x4 u then it is ill posed. However, the counterpart to (12.40)
then is well-posed.
This is not the end of the game of creating a zoo of other long wave
limit modulation equations. We write the Boussinesq equation (12.39) in
long wave form as a first order system
∂τ v = ∂ξ w, ∂τ w = ∂ξ v − ε2 ∂ξ3 v + ε2 ∂ξ (v 2 ).
This system is next diagonalized in lowest order, i.e., we introduce Y = v+w
and Z = v − w, and find
(12.41) ∂τ Y =∂ξ Y − ε2 ∂ξ3 (Y + Z)/2 + ε2 ∂ξ ((Y + Z)2 )/4,
(12.42) ∂τ Z = − ∂ξ Z + ε2 ∂ξ3 (Y + Z)/2 − ε2 ∂ξ ((Y + Z)2 )/4.
12.3. Whitham, Boussinesq, BBM, etc. 467

Now we can use

(12.43) ∂τ Y = ∂ξ Y + O(ε2 ), and ∂τ Z = −∂ξ Z + O(ε2 )
to manipulate the right-hand side. We will mainly leave this to the reader,
and only remark that with this procedure a large number of modulation
equations can and have been obtained in the literature.
In order to derive the KdV equation from (12.41)-(12.42) we set Z = 0
and obtain in the Y equation
(12.44) ∂τ Y = ∂ξ Y − ε2 ∂ξ3 Y /2 + ε2 ∂ξ (Y 2 )/4
By going into the moving frame Y (ξ, τ ) = A(ξ − τ, ε2 τ ) we arrive at the
standard form of the KdV equation
1 3 1
∂T A = ∂X A + ∂X (A2 )
2 4
which is now independent of the small perturbation parameter 0 < ε  1.
Starting from (12.44) by using again (12.43) allows to derive the so called
regularized long wave equation
1 1
(12.45) ∂τ Y = ∂ξ Y − ε2 ∂τ ∂ξ2 Y + ε2 ∂ξ (Y 2 ).
2 4
This equation is ill-posed, but for the ill-posed version of (12.1), namely
∂t2 u = ∂x2 u + ∂x4 u + ∂x2 (u2 ) one obtains a well-posed regularized long wave
equation, namely
1 1
∂τ Y = ∂ξ Y + ε2 ∂τ ∂ξ2 Y + ε2 ∂ξ (Y 2 ).
2 4
Since this equation can be written as
1
(12.46) ∂τ Y = (1 − ε2 ∂ξ2 )−1 (∂ξ Y + ε2 ∂ξ (Y 2 ))
4
the local existence and uniqueness of solutions can be established with the
Picard-Lindelöf theorem for every ε > 0. However, the equation still depends
on ε > 0 and the advantage of a simpler local existence and uniqueness
theory vanishes for ε → 0 since the Lipschitz-constant on the right-hand
side of (12.46) tends to infinity for ε → 0.
All bi-directional models can be approximated for localized initial condi-
tions by a system of two (decoupled) KdV equations and all uni-directional
models can be approximated by a single KdV equation which in contrast
to the other equations are independent of the small perturbation parameter
0 < ε  1. Even more is true. For the water wave problem the complete
zoo of long wave modulation equations is derived in the same limit, namely
for initial conditions of the form
η(x1 , 0) = ε2 Φ0 (εx1 ) and w(x1 , 0) = ε2 Φ1 (εx1 ).
468 12. Long waves and their modulation equations

We explain in the next subsection that at this point the analysis goes beyond
the formal calculations and makes a decision which of the equations should
be taken to approximate the long wave limit. The answer of [SW00b,
SW02] is that two ε-independent decoupled KdV equations are sufficient
to describe this limit. On the O(1/ε3 ) time scale for small ε > 0 no other
small amplitude dynamics can be found in the original system via the other
ε-dependent modulation equations.

12.4. The long wave limit

We consider the Boussinesq equation (12.1) with initial conditions
(12.47) u(x, 0) = ε2 Φ0 (εx), ∂t (∂x2 − ∂x4 )−1/2 u(x, 0) = ε2 Φ1 (εx).
It is the goal of this section to prove that for small ε > 0 and for spatially
decaying Φ0 , Φ1 ∈ H2θ ∩ H θ+4 the dynamics of (12.1) can be described by
two decoupled KdV equations describing wave packets moving to the left
and to the right. In order to do so we set
1
A+ (X, 0) = √ (Φ0 (X) + Φ1 (X)),
2ε2
1
A− (X, 0) = √ (Φ0 (X) − Φ1 (X)),
2ε2
and prove the following approximation theorem.
Theorem 12.4.1. Fix θ > 1/2, and let A+ ∈ C([0, T0 ], H2θ+4 ) and A− ∈
C([0, T0 ], H2θ+6 ) be solutions of the system of KdV equations (12.4). Then
there exist ε0 > 0 and C > 0 such that for all ε ∈ (0, ε0 ] we have
sup u(·, t) − (ε2 A+ (ε(· − t), ε3 t) + ε2 A− (ε(· + t), ε3 t))H θ ≤ Cε7/2
t∈[0,T0 /ε3 ]

for the solutions u of (12.1) with initial conditions (12.47).

Proof. For notational simplicity we only consider the case θ = 2. For the
approximation
ε2 Ψ(x, t, ε) = ε2 A+ (ε(x − ct), ε3 t) + ε2 A− (ε(x + ct), ε3 t)
we find using the previous calculations that
Res(ε2 Ψ) = −ε8 ∂T2 A+ − ε8 ∂T2 A− + 2ε6 ∂X
2
(A+ A− ).
As before we have
4∂T2 A+ =∂X
3 3
(∂X A+ − ∂X (A2+ )) − 2∂X (A+ (∂X
3
A+ − ∂X (A2+ ))).
12.4. The long wave limit 469

and similarly for A− . Next we have

A+ (· − cε−2 T, T )A− (· + cε−2 T, T )H 2
1
≤ sup | −2 −2
|A+ (·, T )H22 A− (·, T )H22
X∈R (1 + (X + cε T ) )(1 + (X − cε T ) )
2 2

C
≤ A+ (·, T )H22 A− (·, T )H22 .
1 + (ε−2 T )2
Therefore, we have with the previous arguments

Lemma 12.4.2. Fix θ ≥ 0, and let A+ ∈ C([0, T0 ], H2θ+6 ) and A− ∈

C([0, T0 ], H2θ+6 ) be solutions of the system of KdV equations (12.4). Then
there exist ε0 > 0 and Cres > 0 such that for all ε ∈ (0, ε0 ) we have
1
 Res(ε2 Ψ(·, t, ε))H θ ≤ Cres (ε15/2 + ε11/2 )
1 + (εt)2
and
1
∂x−1 Res(ε2 Ψ(·, t, ε))H θ ≤ Cres (ε13/2 + ε9/2 )
1 + (εt)2
for all t ∈ [0, T0 /ε3 ].

As before the difference ε7/2 R = u−ε2 Ψ satisfies (12.5) and as before we

estimate R with the help of the energy E which has been defined in (12.6).
Similarly, to the derivation of (12.7), we find now
d 1
(12.48) E ≤ 2Cε3 E + Cε7/2 E 3/2 + Cε3 + Cε ,
dt 1 + (εt)2
with a constant C independent of ε ∈ (0, ε1 ). Under the assumption that
Cε1/2 E 1/2 ≤ 1 we obtain
d 1
E ≤ (2C + 1)ε3 E + Cε3 + Cε .
dt 1 + (εt)2
∞ 1
Using −∞ ε 1+(εt)2 dt = π, Gronwall’s inequality immediately gives the
bound
sup E(t) = C(T0 + π)e(2C+1)T0 =: M = O(1).
t∈[0,T0 /ε3 ]
1/2
Finally choosing ε2 > 0 so small that Cε2 M 1/2 ≤ 1 gives the required
estimate for all ε ∈ (0, ε0 ) with ε0 = min(ε1 , ε2 ) > 0. 
Remark 12.4.3. a) We leave it to the reader to check that initial conditions
in H θ are sufficient, but then the relative error is o(1) for ε → 0.
b) The existence of solutions for the KdV equation in H2θ -spaces is guaran-
teed for initial conditions in H2θ ∩ H θ+4 , cf. [SW00b, Lemma 6.4].
470 12. Long waves and their modulation equations

Remark 12.4.4. The analogous result for the water wave problem gives an
almost complete description of its long-time behavior in the long wave limit,
cf. [SW00b, SW02]. On a time scale O(1/ε) the solutions split up into
two wave packets, one moving to the right and one to the left. These wave
packets evolve independently as solutions of the system of KdV equations.
Their long-time behavior can be computed explicitly with the help of the
inverse scattering transform, cf. Remark 8.2.10. Some solitons which are
ordered w.r.t. their height evolve out of a dispersive remainder, see Figure
12.5, cf. [EvH81]. Approximation results such as Theorem 12.1.1 imply
that the same behavior will be observed in (12.1) in the long wave limit.
The approximation results from [SW00b, SW02] show the same behavior
for the water wave problem.

<−−− −−−>

Figure 12.5. The long time behavior for the water wave problem in the
long wave limit for t = T0 /ε3 with T0 large. The solutions of the water
wave problem in the long wave limit splits up into two wave packets,
one moving to the right and one to the left, where each of these wave
packets evolves independently as a solution of a KdV equation. For large
T0 solitons evolve out of a dispersive remainder.

Further Reading. There are various formulations of the water wave prob-
lem which differ in the chosen parametrization of the top surface. For the
different formulations there are a number of local existence and unique-
ness theorems. For the Eulerian formulation in §12.2.1, local existence and
uniqueness results have been shown for instance in [Shi76, KN79] with
the Cauchy-Kowalevslaya approach, and in [Lan05] with a Nash-Moser ap-
proach. For all other formulations the proofs are adaptions of the local ex-
istence and uniqueness theory for quasilinear hyperbolic equations [Kat75].
For the Lagrangian formulation such results have been shown for instance in
[Nal74, Yos82, Yos83, Cra85, Wu97, Wu99, SW00b, Igu01, SW02].
For the arc-length description of Γ(t) a local existence and uniqueness the-
orem has been shown in [AM05]. The theorems can be distinguished w.r.t.
2D or 3D, finite or infinite depth, with or without surface tension, and reg-
ularity of the initial conditions. See also [ABZ14]. Recently, a number of
almost global and global existence and uniqueness results have been estab-
lished [AL08, Wu09, Wu11, GMS12, AD15]. We refer to [Lan13] as
recent textbook.
12.4. The long wave limit 471

The Camassa-Holm equation has been derived [CH93a] as a modula-

tion equation for the description of unidirectional surface water waves of an
irrotational, inviscid fluid in an infinitely long canal of fixed constant depth,
see also [FL96, Joh02, DGH03]. Similarly to the KdV equation, it is
obtained in the limit of small amplitude and in the limit of long waves, but
where now the limits are considered independently. The peakon equation
is a special form of the Camassa-Holm equation and has attracted a lot of
interest due to its complete integrability, and due to solitary wave solutions
with a discontinuous derivative at the peak, which look like breaking water
waves. See for instance [CH93b, ACHM94, ACF+ 99, ACF+ 01] and the
references therein. The question whether solutions of the water wave prob-
lem can really be approximated by solutions of the Camassa-Holm equation
has been addressed in a number of papers. In [KS05] an approximation
theorem has been established which says that in the KdV-limit solutions of
the Camassa-Holm equation remain close to the solutions of the associated
KdV equation. Therefore, on the KdV time scale no intrinsic Camassa-Holm
dynamics can be seen in the water wave problem since this dynamics hap-
pens on a much longer time scale than the KdV dynamics. In [KS05] also
a discussion about the non-validity of the peakon equation in this limit can
be found. A different asymptotic limit has been considered in [CL09] where
approximation and blow up results are established. The blow up results are
related to wave breaking in the full water wave problem. Higher order long
wave limit approximations have been established in [WW02, Wri06].
The KdV limit of the Vlasov-Poisson system has been considered in
[Han13]. Long wave limits in higher space dimensions are considered in
many papers. In [LLS13] the Zakharov-Kusnetzov equation has been de-
rived for the Euler-Poisson system. The validity and limitations of the KP
equation has been discussed in [GS01a].

Exercises
12.1. Prove that solutions u(x, t) = v(x + t, ε2 t) of the regularized long wave
equation
∂t u = ∂x u + ε2 ∂x3 u + ε2 ∂x (u2 )
can be approximated via the solutions v of the associated KdV equation on an
O(1/ε2 ) time scale.
12.2. Use simultanuosly
ε2 ∂ξ3 Y = ε2 ∂ξ2 ∂τ Y + O(ε4 ) and ε2 ∂ξ3 Z = ε2 ∂ξ2 ∂τ Z + O(ε4 )

in (12.41)-(12.42) with different prefactors αj , and bring these terms to the right-
hand side. Then use
∂τ Y − αε2 ∂ξ2 ∂τ Y = (1 − αε2 ∂ξ2 )∂τ Y
472 12. Long waves and their modulation equations

together with
 −1    
1 − α1 ε2 ∂ξ2 −α2 ε2 ∂ξ2 1 0 α1 ∂ξ2 α2 ∂ξ2
= + ε2 + O(ε4 ).
−α3 ε2 ∂ξ2 1 − α4 ε2 ∂ξ2 0 1 α3 ∂ξ2 α4 ∂ξ2
to manipulate the coefficients on the right-hand side further and derive an approx-
imation system which is no longer symmetric w.r.t. to the interchange of Y and
Z.
12.3. Discuss the form of the smooth curves ω1,2 = ω1,2 (k) defined by
ω 2 = (k2 + μk4 )/(1 + k2 )
for |k| → ∞ and |k| → 0. Compute ∂kj ω1 |k=0 for j = 0, . . . , 4. Is there a change of
sign?
12.4. Consider the two-dimensional Boussinesq model
∂t2 u = Δu + ∂t2 Δu + Δ(u2 )
with u = u(x, y, t) ∈ R, x, y, t ∈ R, and Δ = ∂x2 + ∂y2 for modeling three-
dimensional water waves. By making the ansatz u = ε2 A(ε(x − t), ε2 y, ε3 t) derive
the Kadomtsev-Petviashvili (KP) equation
3
∂X ∂T A = ∂X (ν1 ∂X A + ν2 A∂X A) + ν3 ∂Y2 A,
with X = ε(x − t), Y = ε2 y. Compute the values of the coefficients ν1 , ν2 , and
ν3 ∈ R. Remark. The solutions of the KP equation describe unidirectional waves
slowly modulated in the direction normal to the direction of propagation. For
this model problem it has been pointed out in [GS01b] that the approximation
property for the KP equation depends strongly on the chosen initial condition of
the KP equation.
12.5. Derive the Bernoulli equation (12.17) from the Euler equations (12.13).
12.6. Derive a KdV equation for ∂t2 u = ∂x2 u + ∂x2 ∂t2 u − μ(∂x4 u + ∂x4 ∂t2 )u + ∂x2 (u2 )
by making the ansatz u(x, t) = ε2 A(ε(x − c0 t), ε3 t).
3
12.7. Derive a Kawahara equation ∂T A = c1 ∂X 5
A + c 2 ∂X A + c3 ∂X (A2 ) with T ∈ R,
X ∈ R, coefficients cj ∈ R, and amplitude A(X, T ) ∈ R for
∂t2 u = ∂x2 u + ∂x2 ∂t2 u − μ(∂x4 u + ∂x4 ∂t2 u) + ∂x2 (u2 ),
in case 1 + μ = ε2 ν, by making the ansatz u(x, t) = ε4 A(ε(x − c0 t), ε5 t).
12.8. Find solitary wave solutions of the Boussinesq equation
∂t2 u − ∂x2 u + 3∂x2 (u2 ) − ∂x4 u = 0
in the form u(x, t) = a sech2 (b(x−ct)). Discuss the well-posedness of this Boussinesq
equation.
 Chapter 13

Center manifold
reduction and spatial
dynamics

In Chapters 10-12, our focus has been on the approximation of the dynam-
ics of complicated PDEs on the real line or cylindrical domains by simple
modulation equations. Often, by this approximation we described special
solutions such as bifurcating spatially periodic pattern, solitary waves, mod-
ulating front solutions, or modulating pulse solutions. It is not clear that
the corresponding exact solutions really exist in the original system, too.
The last decades saw big efforts and successes in the construction of these
special solutions using the Lyapunov-Schmidt reduction and the center man-
ifold reduction. There are a number of overview articles [Van89, VI92] and
textbooks [Car81, HI11] where especially center manifold theory and its
applications are well explained. We concentrate on the aspects of the theory
and its applications which have to do with modulation equations.

13.1. The center manifold theorem

The center manifold theorem has been introduced in §3.2 in the finite-
dimensional situation. Here we explain how this approach can be generalized
to the infinite-dimensional case. In the finite-dimensional case we considered

u̇c =Bc uc + gc (uc , us , uu ),

(13.1) u̇s =Bs us + gs (uc , us , uu ),
u̇u =Bu uu + gu (uc , us , uu ),

473
474 13. Center manifold reduction and spatial dynamics

with uc , us , uu some finite-dimensional vectors, Bc a matrix with eigenvalues

on the imaginary axis, Bs a matrix with eigenvalues with negative real part,
Bu a matrix with eigenvalues with positive real part, and gc , gs , gu : C r+1 -
maps without constant and linear terms. Then the center manifold theorem,
stated in Theorem 3.2.1, guarantees the existence of a neighborhood U of
uc = 0, and of a C r -map h : uc → h(uc ) defined in U , such that the manifold
Wc = {u = uc ⊕ h(uc ) : uc ∈ U, (us , uu ) = h(uc )}.
is invariant under the flow of (13.1). Wc is called a center manifold. The
reduced flow on this in general non-unique center manifold is determined by
(13.2) u̇c = Bc uc + gc (uc , hs (uc ), hu (uc )),
The function h = (hs , hu ) contains no constant and no linear terms w.r.t.
uc .
In the infinite-dimensional case the situation is as above except that us
and uu are now allowed to be infinite-dimensional vectors. The variable uc
will still be of finite dimension such that (13.2) is a finite-dimensional ODE.
Hence, the center manifold theorem allows to reduce an infinite-dimensional
problem to a finite-dimensional one.
However, a number of things which are obvious in the finite-dimensional
case are less clear in the infinite-dimensional case, and so we have to impose a
number of additional assumptions. In order to formulate these assumptions
we have to take a look at the existence proof. Usually the system
 t
tBc
uc (t) =e uc (0) + e(t−τ )Bc gc ((uc , us , uu )(τ )) dτ,
0
 t
(13.3) us (t) = e(t−τ )Bs gs ((uc , us , uu )(τ )) dτ,
−∞
 ∞
uu (t) = − e(t−τ )Bu gu ((uc , us , uu )(τ )) dτ,
t
which is obtained from the variation of constant formula, is used for the
construction of the center manifold.
Remark 13.1.1. To simplify our notation we do not distinguish between
(us , uc , uu ) = (0, uc , 0) and uc , and similarly for us and uu , i.e., in our
notation we can have u = us ⊕ uc ⊕ uu and u = us + uc + uu . From the
context the exact meaning will be clear.

For u ∈ 2,θ sufficiently small the right-hand side of (13.3) turns out to
be a contraction in a space of slowly exponentially growing functions. In the
finite-dimensional case etBc , etBs , etBu are well-defined and continuous from
Rd to Rd . In the infinite-dimensional case this has to be assumed. For our
purposes it is sufficient to assume the existence of β+ > 0 and β− < 0 such
13.1. The center manifold theorem 475

that (etBs )t≥0 and (etBu )t≤0 define C0 -semigroups in some space 2,θ which
satisfy
i) etBu 2,θ →2,θ ≤M eβ+ t , ∀ t ≤ 0,
ii) etBs 2,θ →2,θ ≤M eβ− t , ∀ t ≥ 0.
with a constant M . Moreover, the nonlinear terms gc , gs , and gu should be
C r+1 as functions from 2,θ to 2,θ . According to §5.1.2 in case of diagonal
matrices Bu and Bs the assumptions on the semigroups follow from the
position of the spectral values, for instance by assuming that
β+ = inf{Re λ : λ ∈ σ(Bu )} > 0, and β− = sup{Re λ : λ ∈ σ(Bs )} < 0.
Since Bc is finite-dimensional we have for all ε > 0 the existence of a M > 0
such that
(13.4) etBc 2,θ →2,θ ≤ M eε|t| , ∀ t ∈ R.
By assuming the existence of etBc , with this last estimate we can construct
a center manifold also in case that Bc is infinite-dimensional, cf. [GS01b,
GS05]. Similar to §5.1.2 the assumptions on the semigroup and nonlinearity
can be weakened.
The cut-off. Center manifolds in general only exist in a small neighbor-
hood of the fixed point, here u = 0. In order to construct them, we modify
the original system outside this neighborhood without changing the dynam-
ics close to the fixed point. To do so, we modify the nonlinearity g outside
a neighborhood of u = 0 by some cut-off function χ ∈ C ∞ ([0, ∞), R) with
the properties (i) 0 ≤ χ(r) ≤ 1 for all r ∈ [0, ∞), (ii) χ(r) = 1, if r ∈ [0, 1]
and (iii) χ(r) = 0, if r ≥ 2. We define gρ (u) = g(u)χ(ρ−1 u2,θ ). Since
gρ (u) = g(u) for u ∈ Bρ = {u ∈ 2,θ : u2,θ ≤ ρ} the vector fields of the
modified system and of the original system (13.1) coincide for all u ∈ Bρ .
In particular we have
Lemma 13.1.2. Let g ∈ C k (2,θ , 2,θ ) for a k ≥ 1 and gρ defined as above.
Then gρ ∈ C k (2,θ , 2,θ ) and
lim sup D
gρ (u)2,θ →2,θ = 0.
ρ→0 u∈2,θ

Proof. The proof is obvious, since g(u) = O(u22,θ ) for u → 0. 

Existence and invariance. Since we are only interested in the dynamics
close to u = 0, we consider in the following (13.1) with g replaced by g = gρ
with g satisfying g ∈ C k (2,θ , 2,θ ) for a k ≥ 1, g(0) = 0, and Dg(0) = 0.
The existence theorem for the center manifold is a follows.
Theorem 13.1.3. Let η ∈ (0, β) with β = min(−β− , β+ ). There exists a
δ0 > 0 such that for g with supu∈2,θ Dg(u)2,θ ≤ δ0 the center manifold,
476 13. Center manifold reduction and spatial dynamics

a smooth manifold in 2,θ which is invariant under (13.1) and tangential to

the center subspace, can be characterized as
Wc = {u0 ∈ 2,θ : sup e−η|t| u(t, u0 )2,θ < ∞}
t∈R
for a suitable chosen small η > 0. This set is unique. In detail, there
exists a neighborhood U ⊂ 2,θ of uc = 0 and a (k − 1)-times differentiable
h : uc → (us , uu ) such that Wc = {uc + h(uc ) : uc ∈ U }.

Remark 13.1.4. The uniqueness formulated in this theorem is not con-

tradicting the non-uniqueness shown in Example 3.2.4, since the uniquely
constructed manifold depends on the choice of the cut-off function.

The full proof of Theorem 13.1.3 is rather lengthy. We restrict ourselves

to a few points and refer to [Van89] and [VI92] for more details. We will
discuss the invariance, the existence, and the Lipschitz-continuity of the
center manifold, but leave out the proof of the (k − 1)-times differentiability.
Proof of Theorem 13.1.3. i) The invariance of Wc follows directly from
the definition. We have u(τ, u0 ) ∈ Wc , if u0 ∈ Wc , since
sup e−η|t| u(t, u(τ, u0 ))2,θ ≤ eη|τ | sup e−η|t+τ | u(t + τ, u0 ))2,θ < ∞.
t∈R t∈R

ii) For the existence proof we use the variation of constant formula (13.3)
which we abbreviate as
(13.5) u = Suc (0) + KG(u)
where
(Suc (0))(t) = etBc uc (0), G(u)(t) = g(u(t)),
and
 t
(Ky)(t) = e(t−τ )Bc yc (τ )) dτ
0
 t  ∞
+ e(t−τ )Bs ys (τ )) dτ − e(t−τ )Bu yu (τ )) dτ.
−∞ t
We prove the existence of a fixed point of (13.5) by showing that the right-
hand side is a contraction in the metric space
Yη = {u ∈ C 0 (R, 2,θ ) : yη = sup e−η|t| y(t)2,θ < ∞}.
t∈R

In order to do so we use a number of estimates. From (13.4) we obtain for

all ε > 0 a M > 0 such that
Suc (0)Yη = sup e−η|t| etBc uc (0)2,θ ≤ sup e−η|t| M (ε)eε|t| uc (0)2,θ
t∈R t∈R
≤ M (ε)uc (0)2,θ .
13.1. The center manifold theorem 477

Lemma 13.1.5. For g ∈ Cb0 (2,θ , 2,θ ) we have G : Yη → Yη . If g ∈

Cb1 (2,θ , 2,θ ) then for all η > 0 we get
G(y1 ) − G(y2 )Yη ≤ DgC 0 y1 − y2 Yη
b

for y1 , y2 ∈ Yη .

Proof. The first statement is obvious, since g is bounded. From the mean
value theorem it follows that
G(y1 ) − G(y2 )Yη ≤ sup e−η|t| g(y1 (t)) − g(y2 (t))2,θ
t∈R
−η|t|
≤ sup e DgC 0 y1 (t) − y2 (t)2,θ ≤ DgC 0 y1 − y2 Yη
b b
t∈R

for y1 , y2 ∈ Yη . 
Lemma 13.1.6. For each η ∈ (ε, β) the map K is a bounded linear operator
from Yη into Yη , i.e., there is a function γ : (ε, β) → R such that
KYη →Yη ≤ γ(η).
∞
Proof. We write the last two terms of K as ∞ κ(t−τ )y(τ ) dτ . For η ∈ (ε, β)
and y ∈ Yη it follows
e−η|t| (Ky)(t)2,θ
6  t   ∞ 7
−η|t|  (t−τ )Bc η|τ |  η|τ |
≤yη sup e  e e dτ  + κ(t − τ )e dτ
t∈R 0 −∞
6  t   ∞ 7
 (t−τ )Bc −η|t−τ |  η|t−τ |
≤yη sup  e e dτ  + κ(t − τ )e dτ
t∈R 0 −∞
6  ∞  0 
τ Bc −ητ
≤yη max e e dτ, e e dτ
τ Bc ητ
0 −∞
 ∞ 7
η|τ |
+ κ(τ )e dτ
−∞
0 1
≤yη M (ε) (η − ε)−1 + 2(β − η)−1
which implies the assertion. 
The last three estimates show that, for δ0 > 0 sufficiently small, the map
F (u) := Suc (0) + KG(u)
is a contraction in Yη for all η ∈ (ε, β) and g ∈ Cb1 (Rd ) with contraction
constant Kη DgC 0 < 1. Hence, F possesses a unique fixed point which
b
is a solution of (13.5). We write this solution of (13.5) as u = Ψ(Suc (0)).
iii) We define
ψ(uc (0)) := Ψ(Suc (0))(0) and set h = (0, ψs , ψu ).
478 13. Center manifold reduction and spatial dynamics

The continuity of Ψ implies the continuity of ψ and h. Since for g ∈ Cb1 the
map Ψ is Lipschitz-continuous the same is true for ψ and h. Thus, we have
proved the existence of a Lipschitz-continuous center manifold. 
Remark 13.1.7. a) To complete the proof of Theorem 13.1.3 it remains to
prove that h is in C r−1 . This turns out to be rather complicated, because G
is in general not differentiable from Yη into Yη . However, it turns out that
G is k-times continuously differentiable from Yη1 into Yη2 , if η1 > kη2 . For
the heuristics of this consider g(x) = xk and x(t) = eη|t| .
b) Stable and unstable manifolds can be constructed similarly,
cf. [Van89]. For instance, in case of a stable manifold a fixed point is
constructed for the map
F (u) := Sus (0) + Ks G(u)
in the space
Zη+ = {u ∈ C 0 (R+ , 2,θ ) : uη = sup eηt u(t)2,θ < ∞},
t≥0

where (Sus (0))(t) = etBs us (0), G(u)(t) = g(u(t)), and

 t  ∞  ∞
(Ks z)(t) = e (t−τ )Bs
zs (τ ) dτ − e (t−τ )Bc
zc (τ ) dτ − e(t−τ )Bu zu (τ ) dτ.
0 t t

13.2. Local bifurcation theory on bounded domains

In this section we explain the classical application of the center manifold
theorem to PDEs for two examples.

13.2.1. The Allen-Cahn equation. We reconsider the Allen-Cahn equa-

tion
(13.6) ∂t u = ∂x2 u + αu − u3 ,
with α ∈ R, u = u(x, t) ∈ R, t ≥ 0, and x ∈ [0, π] under the boundary
conditions
u(0, t) = u(π, t) = 0.
In §5.3 we already obtained the global bifurcation diagram sketched in the
left panel of Figure 5.5 by using its gradient structure and by analyzing the
set of stationary solutions with a simple phase plane analysis. Hence, we
already know that there is a sequence of supercritical pitchfork bifurcations
of fixed points for values α = m2 with m ∈ N.
Nevertheless, we will use the Allen-Cahn equation as first example for
the application of the center manifold theorem in the infinite-dimensional
situation. We use Theorem 13.1.3 to establish the existence of the super-
critical pitchfork bifurcations of fixed points with an alternative approach.
13.2. Local bifurcation theory on bounded domains 479

We start with the first bifurcation which occurs for α = 1. We introduce

a bifurcation parameter ε2 = α − 1. Making the ansatz
∞

u(x, t) = un (t) sin(nx)
n=1

allows to rewrite the Allen-Cahn equation (13.6) as

∂t un = λn un + gn (u1 , u2 , . . .), (n ∈ N),

with λn = −n2 + α and

2 π
(13.7) gn = − sin(nx)u(x, t)3 dx = O(|u1 |3 + |u2 |3 + . . .).
π 0
For α = 1 we have one central eigenvalue, namely λ1 . The other eigenvalues
are in the left half plane and so we have
uc = (u1 , 0, 0, . . .) and us = (0, u2 , u3 , . . .).

We already know from §5.3.1 that

gc = (g1 , 0, 0, . . .) and gs = (0, g2 , g3 , . . .)

are smooth maps from 2,θ to 2,θ if θ ≥ 1. Hence, all assumptions for the
application of the center manifold theorem as stated in Theorem 13.1.3 are
satisfied. As in the finite-dimensional case, we add the equation ε̇ = 0 in
order to apply the center manifold theorem not only for ε = 0, but also for
small ε = 0. As a consequence there exist smooth functions hj such that for
the solutions on the center manifold
(13.8) uj = hj (u1 , ε2 ) = O(u31 ).

Therefore, the reduced system on the center manifold is given by

∂t u1 = λ1 u1 + g1 (u1 , h2 (u1 , ε2 ), . . .).

Due to (13.8) we have

g1 (u1 , h2 (u1 , ε2 ), . . .) = g1 (u1 , 0, 0, , . . .) + O(u51 ).

From the representation (13.7) we find

∂t u1 = λ1 u1 − γu31 + O(u51 ),

with  π
2 3
γ= sin4 (x) dx = .
π 0 4
Hence, a supercritical pitchfork bifurcation occurs.
480 13. Center manifold reduction and spatial dynamics

13.2.2. Rolls in the Swift-Hohenberg equation. In §10 we introduced

the Swift-Hohenberg (SH) equation
(13.9) ∂t u = −(1 + ∂x2 )2 u + αu − u3 ,
with t ≥ 0, x ∈ R, u(x, t) ∈ R, and bifurcation parameter α ∈ R. In this
section we apply the center manifold theorem, stated in Theorem 13.1.3, to
prove rigorously the existence of spatially periodic equilibria for (13.9) which
bifurcate from the trivial fixed point u = 0. In order to do so we impose
periodic boundary conditions u(x, t) = u(x + 2π, t) on (13.9). This second
example shows that for Turing instabilities the flow on the center manifold
which is used for the rigorous construction of spatially periodic equilibria is
related to the flow of the associated Ginzburg-Landau equation.
For the linearized system ∂t v = −(1 + ∂x2 )2 v + αv with v(x, t) = v(x +
2π, t) we find solutions v(x, t) = eikx+λ(k)t with λk = λ(k) = −(1 − k 2 )2 + α,
but now with k ∈ Z instead of k ∈ R. We have λk < 0 for all k ∈ Z if
α < 0. For α = 0 we have two zero eigenvalues λ1 = λ−1 = 0. The rest of
the spectrum satisfies λk ≤ −1 for all k ∈ Z \ {−1, 1}. Thus, we can apply
the center manifold theorem to reduce the infinite-dimensional problem to
a two-dimensional center-manifold. In order to apply the center manifold
theorem also in case α = 0 we extend (13.9) by the equation α̇ = 0.
We introduce coordinates on the center manifold through
(13.10) u(x, t) = c1 (t)eix + c−1 (t)e−ix + h(c1 (t), c−1 (t), α),
where h is the reduction function which is of the form

h(c1 , c−1 ) = hk (c1 , c−1 , α)eikx .
k∈Z\{−1,1}

The reduced system on the center manifold can be written as

∂t c1 =λ1 c1 + f1 (c1 , c−1 , α), ∂t c−1 = λ−1 c−1 + f−1 (c1 , c−1 , α),

and we have cj (t) = c−j (t).

Remark 13.2.1. The ansatz (10.17) for the derivation of the GL equation
for the SH equation and (13.10) for the computation of the center manifold
reduction are very similar. By restricting the GL ansatz to solutions which
are constant in space, i.e., Aij (X, T ) = Aij (T ), and setting α = ε2 ,
∞
 ∞

c1 (t) = ε1+j A1j (ε2 t) and hk (c1 , c−1 , α) = εk+j Akj (ε2 t)
j=0 j=0

for k ∈ (2Z + 1) \ {−1, 1}, it is obvious that the coefficients in the center
manifold reduction can be obtained by the coefficients from the GL approx-
imation.
13.2. Local bifurcation theory on bounded domains 481

For the quadratic approximation of f1 and f−1 in terms of c1 , c−1 , α in

principle we have to compute 10 coefficients, for the cubic approximation
20 coefficients, etc.. Hence, it makes sense to use the symmetries of the
problem in order to reduce the complexity of f1 and f−1 . The SH equation
is invariant under u → −u, which yields via (13.10) to (c1 , c−1 ) → −(c1 , c−1 ).
It is translational invariant, i.e., u(·) → u(· + y), which corresponds to the
invariance of the reduced system under (c1 , c−1 ) → (c1 eiy , c−1 e−iy ). Since
c1 = c−1 it is sufficient to consider the equation for c1 alone, which must be
of the form
∂t c1 = λ1 c1 + c1 g(|c1 |2 , α)
due to these symmetries. The last symmetry is the reflection symmetry for
the SH equation, namely u(·) → u(−·), which corresponds to (c1 , c−1 ) →
(c−1 , c1 ). Hence, the function g must be real-valued. We introduce polar
coordinates c1 = reiφ and find
∂t r = λ1 r + rg(r2 , α), ∂t φ = 0.
Since the nonlinear term of the SH equation is given by αu − u3 we have
hk (c1 , c−1 ) = O(|α||c±1 | + |c±1 |3 ),
and as a consequence
f1 (c1 , c−1 ) = αc1 − 3c21 c−1 + O(|α|2 |c±1 | + |α||c±1 |3 + |c±1 |5 ).
Hence,
∂t r = αr − 3r3 + O(α2 r + αr3 + r5 ), ∂t φ = 0.
The stationary solutions satisfy
 α) = αr − 3r3 + O(α2 r + αr3 + r5 ) = 0
G(r,
To balance the lowest order terms we consider the scaled version

G(A, ε) = ε−3 G(εA, ε2 ) = A − 3A|A|2 + O(ε2 ) = 0.
√ √
Since i) G(±1/ 3, 0) = 0 and ii) ∂A G(±1/ 3, 0) = 0 we can√apply the
implicit function theorem to find solutions A± = A± (ε) = ±1/ 3 + O(ε2 )
with G(A± (ε), ε) = 0. Hence, we proved
Theorem 13.2.2. At α = 0 a supercritical pitchfork bifurcation of families
of 2π-spatially periodic equilibria occurs for the SH equation. They are of
the form
√
u(x, t) = ±2 α cos(x + x0 ) + O(α3/2 )
for α > 0 and arbitrary phase x0 ∈ R.

Obviously the same construction is possible for all spatial periods close
to 2π.
482 13. Center manifold reduction and spatial dynamics

Remark 13.2.3. Following Remark 13.2.1, the previous analysis applies

more or less for all pattern forming systems where a GL equation can be de-
rived. Examples which have been considered are reaction-diffusion systems
in §10.5, convection problems in §10.6, and the Couette-Taylor problem in
§10.7. The center manifold reduction for the Couette-Taylor problem has
been carried out for instance in [IA98]. The bifurcating spatially periodic
equilibria are called Taylor vortices and are of the form
 1 (z)eikc x + c.c. + h[q, ε](εA0 , εĀ0 , ε)
UTV [q, φ, ε](x, z) = εA0 [q, φ](X)U kc ,0

where 
ν0 − ν2 q 2 i(qX+φ)
A0 [q, φ](X) = − e
ν3
is an equilibrium of the associated GL equation (10.103). The critical wave
number kc of the most unstable pattern U  1 (z)eikc x and the bifurcation
kc ,0
parameter 0 ≤ ε2  1 have been introduced in §10.7. The function h from
the center manifold reduction satisfies h[q, ε](εA0 , εĀ0 , ε) = O(ε2 ). Since
our center manifold theorem, Theorem 13.1.3, has not been formulated to
apply in this situation, more abstract versions like [HI11, Theorem 2.9] have
to be used for the reduction.

13.3. Spatial dynamics for elliptic problems in a strip

For nonlinear problems the nowadays classical method of spatial dynamics
combined with center manifold theory goes back to Kirchgässner [Kir82].
He studied elliptic problems in unbounded cylindrical domains by consider-
ing the unbounded space variable as evolutionary variable. As an example
consider
(13.11) Δu = αu − u3
on the cylindrical domain R × (0, π) with Dirichlet boundary conditions
u|y=0,π = 0. In order to find uniformly bounded solutions we write the
problem as a dynamical system w.r.t. the unbounded space variable x ∈ R.
In order to apply our center manifold theorem 13.1.3 we make the ansatz
∞

u(x, t) = un (x) sin(ny)
n=1
and find
(13.12) (∂x2 − n2 )un (x) = αun + gn (u1 , u2 , . . .),
with
 π
2
(13.13) gn = − sin(nx)u(x, t)3 dx = O(|u1 |3 + |u2 |3 + . . .)
π 0
13.3. Spatial dynamics for elliptic problems in a strip 483

for n ≥ 1. This can be written as a first order system

(∂x − n)un (x) = vn (x), (∂x + n)vn (x) = αun + gn (u1 , u2 , . . .).
For α = −1 we have two central
√ eigenvalues for u1 . The other eigenvalues
are then given by λ±n = ± n + α for n ≥ 2, i.e., for all α ≈ −1 the other
2

eigenvalues form two real sequences going to ±∞. See Figure 13.1.
Im

Re

Figure 13.1. Spectrum of the linearized spatial dynamics formulation

for (13.12). There are two central eigenvalues leading to a Jordan block,
and infinitely many unstable and infinitely many stable eigenvalues up
to ±∞.

Hence, the initial value problem for (13.12) is ill-posed in every 2,θ -
space. Nevertheless, the center manifold theorem, Theorem 13.1.3, can be
applied. The equations for n ≥ 2 can be written as
u n = λn vn , vn = λn un + λ−1
n gn (u0 , u1 , . . .).
This system can be diagonalized by introducing cn = un + vn and c−n =
un − vn which satisfy
c n = λn cn + gn , c −n = −λn c−n − gn .
Since functions satisfying Dirichlet boundary conditions on (0, π) can be
extended to odd 2π-periodic functions w.r.t. y, and since odd functions to
the power three are odd again, we can consider (13.11) alternatively in the
invariant subspace of odd 2π-periodic functions w.r.t. y. As a consequence
the nonlinearity is a smooth map from 2,θ to 2,θ for every θ > 1/2. Hence,
all assumptions of Theorem 13.1.3 are satisfied and so there exist smooth
functions uj = hj (u1 , α) for j ≥ 2, such that on the center manifold the
reduced system is given by
3
u 1 = (α + 1)u1 + g1 (u1 , h2 (u1 , α), . . .) = αu1 − u31 + O(u51 ).
4
Ignoring the higher order terms yields
3
(13.14) u 1 = v1 , v1 = (α + 1)u1 − u31 .
4
For α + 1 > 0 small, this system possesses two homoclinic orbits at the
origin which correspond to spatially localized solutions of the original elliptic
problem (13.11).
484 13. Center manifold reduction and spatial dynamics

Remark 13.3.1. It remains to prove the persistence of the homoclinic or-

bits if the higher order terms O(u51 ) are not ignored. In general a homoclinic
orbit would break up. However, the original system is reversible, i.e., invari-
ant under (x, u) → (−x, u). The cut-off function in the center manifold
construction can be chosen in such a way that the reduction preserves this
property, cf. [HI11, §2.3.3]. Hence, the reduced system is invariant under
(x, u1 , v1 ) → (−x, u1 , −v1 ) and so the phase portrait must be reflection sym-
metric at the u1 -axis which implies the persistence of the homoclinic orbit
under higher order perturbations, cf. Figure 2.20.

13.4. Applications
The center manifold theorem in combination with spatial dynamics is nowa-
days a well established tool which has successfully been used in many appli-
cations. Examples are the construction of solitary surface water waves, of
standing light pulses in photonic crystals, or of breather solutions in lattice
differential equations. In this section we concentrate on the aspects which
have to do with modulation equations when the center manifold theorem
is applied, and explain how the spectrum of the linearized spatial dynam-
ics formulation is related to the spectrum of the linearized time-dependent
problem and how the reduced systems on the center manifold are related to
the associated modulation equations. We will concentrate on the construc-
tion of traveling wave solutions, modulating front solutions, and breather
solutions. We refrain from rewriting the literature and therefore skip all
functional analytic difficulties which have to be overcome in applying the
center manifold theorem and the discussion of the reduced system.

13.4.1. Solitary waves for the water problem. The center manifold
theorem in combination with spatial dynamics has especially been used in
the construction of solitary waves for the water wave problem as introduced
in §12.2.1, cf. [IK90, Ioo95]. The water wave problem can be written as
evolutionary system

(13.15) ∂t U = Λ(∂x , ∂y )U + O(U 2 ),

with x ∈ R, and y ∈ Σ, with Σ a bounded fixed cross section. We are

interested in waves of permanent form, i.e.,

U (x, y, t) = V (x − ct, y) = V (ξ, y).

Inserting this ansatz into the water wave problem yields

(13.16) −c∂ξ V = Λ(∂ξ , ∂y )V + O(V 2 ).

13.4. Applications 485

Written as an evolutionary system w.r.t. ξ we obtain the spatial dynamics

formulation
(13.17) ∂ξ W = M(∂y )W + O(W 2 ).
It is the purpose of the subsequent discussion to relate properties of the
temporal evolutionary system (13.15) and of spatial dynamics formulation
(13.17).
The spectrum. Solutions of the linearized spatial dynamics formulation
∂ξ w = M w are of the form
4 (y).
W (y, ξ) = eμξ W
The eigenvalues μ can be found from the eigenvalue problem
4 = M(∂y )W
μW 4.
Undoing the transformation which allowed to write (13.16) into (13.17)
yields to solutions
V (y, ξ) = eμξ V (y)
and to the eigenvalue problem
(13.18) −cμV (y) = Λ(μ, ∂y )V (y).
On the other hand the solutions of the linearized temporal dynamics formu-
lation (13.15) are of the form
 (y)
U (x, y, t) = eλt eikx U
leading to the eigenvalue problem
(13.19) λU  (y).
 (y) = Λ(ik, ∂y )U
In general we have a countable set of curves of eigenvalues k → λn (k) with
n ∈ N.
Comparing (13.18) with (13.19) shows that both coincide for the cen-
tral eigenvalues μ = ik of the spatial dynamics formulation. Hence, these
eigenvalues can be obtained by solving
(13.20) λn (k) = −cik,
where the wave speed c is a free parameter. For the water wave problem we
have 
λ± (k) = iω± (k) = ±isign(k) tanh(k)(k + σk 3 ).
In Figure 13.2 we analyze (13.20) in case of zero surface tension, in Figure
13.3 in case of surface tension σ > 1/3, and in Figure 13.4 in case of surface
tension 0 < σ < 1/3. We always find an intersection of k → −ck with ω+
and ω− at the wave number k = 0, but since the associated nonlinear terms
at the wave number k = 0 vanish identically, cf. the derivation of the KdV
equation, these two central eigenvalues play no role and can be eliminated.
486 13. Center manifold reduction and spatial dynamics

i) The reduction in case σ = 0. For |c| < 1 there are two intersections of
k → ω± (k) and k → −ck at wave numbers k with |k| > 0, which lead to two
central eigenvalues. Since the spatial dynamics formulation (13.17) of the
water wave problem can be written as an infinite-dimensional Hamiltonian
system it can be expected that an infinite-dimensional version of Lyapunov’s
subcenter theorem can be applied. Therefore, the existence of spatially
periodic traveling waves can be established, cf. [Str26]. If |c| approaches
1 these two central eigenvalues collide in zero and leave the imaginary axis
along the real axis. For |c| > 1 there exists a homoclinic orbit to the origin.
This scenario is analogous to the following example.

Figure 13.2. The curves of eigenvalues of the temporal dynamics for-

mulation (13.15) intersected with the curve k → −ck lead to the central
eigenvalues of the spatial dynamics formulation (13.17) of the water
wave problem without surface tension.

Example 13.4.1. We are interested in waves of permanent form for the

Boussinesq model
∂t2 u = ∂x2 u + ∂x2 ∂t2 u + ∂x2 (u2 ).
We make the ansatz u(x, t) = v(x − ct) = v(ξ) and find the fourth order
ODE
(13.21) c2 ∂ξ2 v = ∂ξ2 v + c2 ∂ξ4 v + ∂ξ2 (v 2 ).

The linearized problem is solved by v(ξ) = eλξ v where 0 = −c2 λ2 + λ2 +

λ4 = λ2 (λ2 + 1 − c2 ). Independently of the velocity c there are always two
eigenvalues 0, but since the nonlinear terms at the wave number k = 0 also
13.4. Applications 487

vanish identically, these two central eigenvalues play no role and can be
eliminated. In fact, (13.21) can be integrated twice, i.e.,
(13.22) c2 v = v + c2 ∂ξ2 v + v 2 .
The constants of integration vanish since we are interested in solitary waves.
We write (13.22) as the first order system
(13.23) ∂ξ v0 = v1 , ∂ξ v1 = c−2 ((c2 − 1)v0 − v02 ).
We have two fixed points (v0 , v1 ) = (0, 0) and (v0 , v1 ) = (c−2 (c2 − 1), 0). For
c2 < 1 the origin is a center surrounded by a family of periodic solutions.
For c2 > 1 the origin is a saddle, and by looking at the phase portrait we
find a homoclinic solution at the origin which corresponds to the solitary
wave we are interested in and which exists for velocities c2 > 1.

Remark 13.4.2. The ansatz

U (x, y, t) = ε2 A(ε(x − t), ε3 t)ϕ(y) + O(ε3 ).
for the derivation of the KdV equation for the water wave problem and the
ansatz
V (ξ, y) = V1 (ξ)ϕ(y) + h(V1 , ∂ξ V1 , c),
with ξ = x − ct, for the computation of the center manifold reduction are
very similar. By restricting the KdV ansatz to traveling wave solutions, i.e.,
A(X, T ) = A(X − cT ), and setting c − 1 = ε2 ,
V1 (ξ) = ε2 A(εξ) + O(ε3 ) and h(V1 , ∂ξ V1 , c) = O(ε3 ),
it is obvious that the coefficients in the center manifold reduction can be
obtained by the coefficients from the KdV approximation.

From Remark 13.4.2 and the fact that the associated KdV equation pos-
sesses solitary waves Asol and the reversibility of the water wave problem it
follows that the reduced system on the center manifold possesses a homo-
clinic orbit for every c, with |c| > 1 not to big. By making the previous
ideas rigorous the following result [Ioo98] can be shown.

Theorem 13.4.3. There exists a δ0 > 0 such that for δ ∈ (0, δ0 ) the wa-
ter wave problem without surface tension possesses a solitary wave solution
U (x, y, t) = V (x − ct, y) with lim|ξ|→∞ V (ξ, y) = 0 satisfying

sup |V (ξ, y) − δ 2 Asol (δξ)ϕ(y)| ≤ Cδ 3 ,

ξ∈R,y∈[0,1]
√ √
c = 1 + δ 2 / 2 + O(δ 4 ) or c = −1 − δ 2 / 2 + O(δ 4 ), where Asol is the rescaled
1-soliton solution of the associated KdV equation.
488 13. Center manifold reduction and spatial dynamics

ii) The reduction in case σ > 1/3. For |c| > 1 there are two intersection
points of k → ω± (k) and k → −ck at wave numbers k with |k| > 0 which
lead to two central eigenvalues. The existence of spatially periodic traveling
waves can be established [Zei71] by applying an infinite-dimensional version
of Lyapunov’s subcenter theorem. If |c| approaches 1 these two central
eigenvalues collide in zero and leave the imaginary axis along the real axis.
For |c| < 1 there exists a homoclinic orbit to the origin. This scenario is
exactly as in the case σ = 0. However, as we already know from the KdV
approximation the solitary wave for σ > 1/3 is a wave of depression and not
a wave of elevation.

Figure 13.3. The curves of eigenvalues of the temporal dynamics for-

mulation (13.15) intersected with the curve k → −ck lead to the central
eigenvalues of the spatial dynamics formulation (13.17) of the water
wave problem with surface tension σ > 1/3.

iii) The reduction in case 0 < σ < 1/3. For small values of |c| except
of the trivial intersection at k = 0 there are no further intersection points.
For a value |c| = cmin < 1 the curve k → −ck is tangent to k → ω(k) for a
k ≥ 0 which implies that two pairs of eigenvalues collide on the imaginary
axis. At this point the phase and group velocity of the associated NLS
approximation coincide. Thus, the modulating pulse solutions described
by the NLS approximation become traveling waves and can rigorously be
established with the present center manifold approach. For cmin < |c| < 1
there are four intersection points at wave numbers k with |k| > 0 which
lead to four central eigenvalues. The existence of spatially periodic traveling
waves can be established also in this case [Zei71]. As |c| approaches 1, two of
the four central eigenvalues collide in zero and leave the imaginary axis along
13.4. Applications 489

the real axis. The solitary waves in this case can be obtained via the KdV
approximation. However, the solitary waves do not decay to zero for |ξ| →
∞, but have some small oscillatory tails. The reason for this fact is that it
is very unlikely that a one-dimensional unstable manifold intersects with a
one-dimensional stable manifold in a four-dimensional space. However, the
three-dimensional center-unstable manifold and three-dimensional center-
stable manifold intersect due to the reversibility of the water wave problem
and so for |ξ| → ∞ the solutions in the intersection converge towards the
center manifold which is filled with periodic solutions.

15
ω(k)
ck
1
c2k
10
c3k

0
0 5 10 15 20

Figure 13.4. The curves of eigenvalues of the temporal dynamics for-

mulation (13.15) intersected with the curves k → −ck for different values
of c lead to the central eigenvalues of the spatial dynamics formulation
(13.17) of the water wave problem with surface tension 0 < σ < 1/3.

Remark 13.4.4. In case σ = 1/3 and c = 1 the Kawahara equation (12.22)

occurs as a modulation equation. The reduced system on the center manifold
of the spatial dynamics formulation possesses a Jordan block of size 4. Near
this point more complicated traveling wave solutions, such as multi-pulse
solutions, can be found, cf. [BGT96].

13.4.2. Modulating fronts for the Swift-Hohenberg equation. We

are interested in the construction of modulating front solutions with the
center manifold approach and spatial dynamics for the Swift-Hohenberg
(SH) equation. Such solution describe the spreading of the stable spatially
periodic equilibria uε constructed in §13.2.2 into the regions of the unstable
trivial solution u = 0, i.e., we are interested in solutions
u(x, t) = v(x − ct, x)
490 13. Center manifold reduction and spatial dynamics

1
Im

c Re

σ
1/3

Figure 13.5. Spectrum of the linearized spatial dynamics formulation

in the (σ, c) plane for (σ, c) ≈ (1/3, 1).

where v is 2π-periodic in its second argument and satisfies

lim v(ξ, p) = 0, lim v(ξ, p) = uε (p).
ξ→∞ ξ→−∞
 2
Then, v = v(ξ, p) satisfies −c∂ξ v = − 1 + (∂ξ + ∂p )2 v + αv − v 3 which
written as a first order system yields
∂ξ v0 = − ∂p v0 + v1 ,
∂ξ v1 = − ∂p v1 + v2 ,
(13.24)
∂ξ v2 = − ∂p v2 + v3 ,
∂ξ v3 = − ∂p v3 − 2v2 − v0 + c(−∂p v0 + v1 ) + αv0 − v03 .
With V = (v0 , v1 , v2 , v3 ) this system is abbreviated as
∂ξ V = −∂p V + Kc,α (∂p )V + N (V )
where Kc,α (∂p )V stands for the linear and N (V ) for the nonlinear terms.
Since the modulating front solutions bifurcate out of the trivial solution we
consider first the linearized problem
∂ξ V = −∂p V + Kc,α (∂p )V.
Undoing the above transformations, this system is equivalent to
 2
−c∂ξ v = − 1 + (∂ξ + ∂p )2 v + αv.
Since v(ξ, p) is 2π-periodic w.r.t. p we can use Fourier series and look for
solutions v(ξ, p) = eμξ+imp . This ansatz yields
 2
−cμ = − 1 + (μ + im)2 + α.
13.4. Applications 491

In order to use the center manifold theorem we have to compute the central
eigenvalues μ = ik ∈ iR which are solutions of
 2
−cik = − 1 + (ik + im)2 + α = λ(k + m, α).
For c = 0 and α = 0 there are infinitely many intersection points. Since the
eigenvalue curve λ possesses maxima in k+m = ±1, the solution k = ±1−m
is double for each of these ms, which leads to a Jordan block of size two in
the spatial dynamics formulation. Hence, for c = 0 and α = 0 there are
infinitely many eigenvalues on the imaginary axis. Thus, at a first view a
reduction to a finite-dimensional center manifold does not seem possible.
However, at a second view an interesting phenomenon occurs. In order to
see the phenomenon take one of the front solutions
A(X, T ) = Af (X − 
cT )
of the associated GL equation with
√
lim Af (ξ) = 0 and lim Af (ξ) = 1/ 3
ξ→∞ ξ→−∞

which yields an approximation of the modulating front solution

v(ξ, p) =v(x − ct, x) = 2εAf (X − 
cT ) cos(x) + O(ε2 )
=2εAf (εx − 
cε2 t) cos(x) + O(ε2 ) = 2εAf (εξ) cos(p) + O(ε2 ),
with velocity c = ε c = O(ε). Hence, we should simultaneously change
α = ε2 and c = ε c. Interestingly, by doing so the eigenvalues leave the
imaginary axis with different velocities for ε > 0. The four eigenvalues
from before stay in an O(ε)-neighborhood of the imaginary axis, whereas
the other eigenvalues are at least O(ε1/2 )-bounded away from the imaginary
axis. Thus, we have a spectral gap of size O(ε1/2 ) which allows us to prove
the existence of an invariant four-dimensional manifold of size O(ε1/2 ). Since
the bifurcating modulating front solutions in which we are interested are of
order O(ε), they are contained in this manifold. As above, in lowest order
the flow on the manifold is given via the associated flow of the GL equa-
tion. The approximate heteroclinic solutions persist due to the transversal
intersection of the four-dimensional unstable√manifold of the origin and the
two-dimensional stable manifold of A = 1/ 3. The detailed analysis can
be found in [CE86, EW91] for the SH equation and in [HCS99] for the
Couette-Taylor problem.

13.4.3. Breathers and generalized breathers. We close this chapter

with the construction of spatially localized time-periodic solutions, so called
breather solutions. It turns out that the occurrence of such solutions with
finite energy is rather rare. In general they only exist with small, but non-
vanishing tails. Such solutions are called generalized breather solutions. Be-
side the non-persistence of breathers for perturbations of the sine-Gordon
492 13. Center manifold reduction and spatial dynamics

equation we also discuss the existence of breathers for nonlinear wave equa-
tions with spatially periodic coefficients.
The non-persistence of breathers for perturbations of the sine-
Gordon equation. The sine-Gordon equation
(13.25) ∂t2 u = ∂x2 u − sin(u),
with t ∈ R, x ∈ R, and u(x, t) ∈ R, originally came up in differential geome-
try in the form ∂ξ ∂η u = sin(u), describing surfaces with a constant negative
curvature. It was found to govern the propagation of a dislocation in a
crystal whose periodicity is represented by sin u, it was posed as a tentative
model of an elementary particle, and it was shown to be an equivalent form
of the so called Thirring model, cf. [DJ89]. It turns out to be a completely
integrable Hamiltonian system.
Remark 13.4.5. The name sine-Gordon equation is a pun referring to the
Klein-Gordon equation
(13.26) ∂t2 u = ∂x2 u − m2 u,
which is just the linear wave equation with the additional term −m2 u, where
m is the rest mass of the particle. The Klein-Gordon equation is the lin-
earized version of the sine-Gordon equation and of the nonlinear Klein-
Gordon equation (11.2), which in physics is often called φ4 -model, and has
been derived in 1928 as a relativistic version of the Schrödinger equation
describing free particles, i.e., it is invariant under Lorentz transformations.
Hence it attempts to unite quantum mechanics and special relativity, but it
has some serious flaws; see [Law90] for a very basic introduction.

Traveling wave solutions u(x, t) = v(x − ct) = v(ξ) of (13.25) satisfy

(c2 − 1)∂ξ2 v = − sin(v).
For |c| > 1 this is the equation of the pendulum with the two heteroclinic
connections. These traveling wave solutions are called kink and antikink for
(13.25). For |c| < 1 the phase portrait is shifted by v → v + π. Hence, we
have heteroclinic solutions no longer between (−π, 0) and (π, 0) but between
(0, 0) and (2π, 0).
There is a family of explicitly known spatially localized time-periodic
solutions of the sine-Gordon equation, namely
 
(1 − λ2 )1/2 sin(λ(t − t0 ))
(13.27) u(x, t) = 4 arctan
λ cosh((1 − λ2 )1/2 )(x − x0 )

for arbitrary x0 , t0 , and λ, with 0 < |λ| < 1. Due to the dynamical behavior
of these solutions they are called breather solutions. Surprisingly, it turns
13.4. Applications 493

out that the sine-Gordon equation is the only nonlinear wave equation
(13.28) ∂t2 u = ∂x2 u − u + g(u),
with g : R → R a smooth, odd function which satisfies g(u) = O(u3 ) and
g (0) > 0 for which such breather solutions exist [Den93, BMW94]. In
the following we explain why this ’non-persistence of breathers’ result holds.
Moreover, we explain a number of positive results about the existence of
generalized breather solutions.
Remark 13.4.6. With the ansatz
u(x, t) = εA(εx, ε2 t)eit + c.c.
the NLS equation
2
2i∂T A = ∂X A − A|A|2
can be derived which possesses localized time-periodic solutions. From the
approximation results in §11.1 it is known that these solutions approximately
exist in the nonlinear wave equations, too, on a time interval of length
O(1/ε2 ). However, the sine-Gordon equation up to rescaling is the only
of the nonlinear wave equations for which these solutions exist as localized
time-periodic solution for all times. In fact, taking the limit λ = 1 − ε2 → 1
in (13.27) shows that sine-Gordon breathers can be approximated by the
localized time-periodic solutions of the associated NLS equation.

In order to construct time-periodic solutions for (13.28) we can use

Fourier series w.r.t. the time variable t, i.e., we write

u(x, t) = k (x)eiωkt .
u
k∈Z

(x) satisfy
The Fourier coefficients u
k (x) = ∂x2 u
−ω 2 k 2 u k (x) − u
k (x) + gk (u)(x),
where  2π/ω
ω
gk (u)(x) = e−iωkt g(u(x, t)) dt.
2π 0
The linearization of this spatial dynamics formulation possesses solutions of
k (x) = eλx with
the form u
λ2 = −ω 2 k 2 + 1.
From Remark 13.4.6 it follows that for small amplitude solutions we have
approximately ω = 1. Since we have an odd nonlinearity we only have to
consider k ∈ 2Z + 1. Hence, at the bifurcation point all eigenvalues are
on the imaginary axis. For ω a little bit smaller than 1 there is one stable
and one unstable eigenvalue and still infinitely many central eigenvalues.
Breather solutions lie in the intersection of the one-dimensional stable and
494 13. Center manifold reduction and spatial dynamics

one-dimensional unstable manifold, which is very unlikely to exist, and the

sine-Gordon equation up to rescaling turns out to be the only equation where
this intersection occurs.
Remark 13.4.7. For all other nonlinear wave equations only generalized
modulating pulse solutions can be expected. Such solutions do not decay
to zero for |x| → ∞, but have some small oscillatory tails of order O(εn ),
with n ∈ N arbitrary, but fixed, where 0 < ε  1 is the amplitude of the
generalized modulating pulse solution. They can be found in the intersection
of the infinite-dimensional center-unstable manifold and infinite-dimensional
center-stable manifold. Both have co-dimension one and they intersect due
to the reversibility of (13.28). For |x| → ∞ the solutions in the intersection
converge towards the center manifold. The boundedness for |x| → ∞ is not
at all obvious. In general the solutions on the center manifold can grow
slowly and so the existence of generalized modulating pulse solutions can
only be established for |x| ≤ 1/εn . Since for (13.28) there is a finite speed of
propagation such solutions also exist for all |t| ≤ 1/εn which is much larger
than the O(1/ε2 ) time scale of the NLS approximation. If the Hamiltonian
restricted to the center manifold is positive definite, then the solutions stay
bounded for all x ∈ R and hence for all t ∈ R. See [GS01b, GS05, GS08]
for detailed proofs.
Remark 13.4.8. Breathers for nonlinear wave equations
(13.29) s(x)∂t2 u(x, t) = ∂x2 u(x, t) − q(x)u(x, t) + r(x)u(x, t)3 ,
with x ∈ R, t ∈ R, u(x, t) ∈ R, and spatially periodic real-valued coefficient
functions s, q, and r, i.e.,
s(x) = s(x + 2π), q(x) = q(x + 2π), and r(x) = r(x + 2π),
can be constructed with center manifold theory and spatial dynamics, too.
The spatial dynamics formulation, with x ∈ R as new evolutionary variable,
is solved in the space of 2π/ω-time-periodic functions. The spatial dynamics
formulation is non-autonomous due to the x-dependence of the coefficients s,
q, and r. Breather solutions are homoclinic solutions of the spatial dynam-
ics formulation. We have seen that for spatially constant coefficients such
solutions do not exist since for the linearization around the origin except of
two eigenvalues all other eigenvalues are on the imaginary axis. However,
by making the problem x-dependent and adjusting s and q in a proper way
all the Floquet exponents can be moved off the imaginary axis and the ex-
istence of a two-dimensional center manifold with two homoclinic solutions
can be established. The equations on the center manifold are related to a
NLS equation which is the associated modulation equation for this problem,
cf. §11.6. The details have been worked out in [BCBLS11].
13.4. Applications 495

Further Reading. The previous theory is restricted to semilinear sys-

tems. Some theory for quasilinear systems can be found in [Mie88, Ren92,
Jam03, Sch15]. The rigorous construction of stationary and time-periodic
pattern in hydrodynamical stability problems in cylindrical domains can be
found in [DIM89, IM91]. Three-dimensional water waves are constructed
via center manifold reductions in [GHS02, GH03]. Recently, a lot of re-
search has been done for non-irrotational water waves, cf. [Con11]. For the
analysis of the reduced equations on the center manifold symmetries play
a big role. The mathematical theory behind this approach can be found in
[GS85, GSS88, CL00]. More generally, invariant manifold theory plays a
big role in the construction of more and more complicated bifurcating so-
lutions in pattern forming systems. Spiral waves have been constructed for
instance in [SSW99], or dislocations of pattern for instance in [HS12].
Exercises
13.1. Consider the Allen-Cahn equation ∂t u = ∂x2 u + αu − u3 , α∈R, u=u(x, t)∈R,
t ≥ 0, and x ∈ [0, π] under the boundary conditions u(0, t) = u(π, t) = 0. Make
a center manifold reduction for α = 4 and use the reduced equation to prove the
occurrence of a supercritical pitchfork bifurcation.
13.2. Consider Δu = αu − u2 for u = u(x, y) ∈ R on the cylindrical domain
(x, y) ∈ R × (0, π) with Neumann boundary conditions ∂y u|y=0,π = 0.
a) Write the elliptic problem as an evolutionary problem w.r.t. x as new time
variable.
b) Compute the spectrum of the linearization around u = 0 of this spatial dynamics
formulation for α close to zero.
c) Compute the reduction function h and the reduced ODE on the center manifold.
(Hint: Consider y-independent solutions).
d) Discuss the flow on the center manifold for α < 0, α = 0, and α > 0. Plot the
solutions x → u(x, y).
13.3. Discuss whether the solutions λ1,2,3,4 of 1 − c2 + (c2 − μ)λ2 − μc2 λ4 = 0 are
on the imaginary or real axis. Sketch the result in the (μ, c)-plane.
13.4. (Explicit formula for the kink solution) Verify that for arbitrary C and
λ, with |λ| < 1,
u(x, t) = 4 arctan(C exp((x − λt)/(1 − λ2 )1/2 ))
is a solution of the sine-Gordon equation ∂t2 u = ∂x2 u − sin u. Sketch the solution.
13.5. Show that the sine-Gordon equation ∂t2 u = ∂x2 u − sin u is invariant under the
transformation x  = γ(x − vt), 
t = γ(t − vx), where γ = 1/(1 − v 2 )1/2 and |v| < 1.
Hence, conclude that
 
(1 − λ2 )1/2 sin(γλ(t − vx − t0 ))
u(x, t) = 4 arctan
λ cosh(γ(1 − λ2 )1/2 )(x − vt − x0 )
solves the sine-Gordon equation. Sketch the evolution of the solution.
 Chapter 14

Diffusive stability

A basic concept in stability theory is the control of the nonlinear terms by

the linearized problem if the linearization dissipates energy with an expo-
nential rate. Such a behavior occurs when the spectrum lies in the left half
plane strictly bounded away from the imaginary axis, cf. Theorem 2.3.4 for
the ODE case, and Theorem 5.2.23 and Theorem 6.2.18 for examples in
the PDE case. However, for PDEs posed on unbounded spatial domains
there are various stability problems where we have essential spectrum up to
the imaginary axis. Especially, the bifurcating Turing pattern discussed in
Chapter 10 and 13 fall into this class. Therefore, it is the goal of this section
to explain how nonlinear stability can be established in this case, too. We re-
strict ourselves to problems which are related to some modulation equation,
as the real or complex GL equation.
For the problems we are interested in, the linear problem shows some
diffusive behavior and polynomial decay rates. As a consequence, not all
nonlinear terms can be controlled by the linearized problem. In order to see
the difference to the case of exponential decay rates we consider the stability
of u = 0 for the two toy problems

d
(14.1) u = −u + up ,
dt
(14.2) ∂t u = ∂x2 u + up .

For (14.1) the linearization ∂t u = −u around u = 0 is solved by u(t) =

e−t u(0). For the terms occurring in (14.1) we therefore find dt
d
u ∼ e−t , −u ∼
−t −pt
e , and u ∼ e . Hence for all p > 1 the nonlinear terms vanish much
p

faster than the linear ones, and so it can be expected that the nonlinear
problem behaves asymptotically as predicted by the linear one.

497
498 14. Diffusive stability

For the nonlinear diffusion equation (14.2), the linearized problem ∂t u =

∂x2 uin general shows no decay rate at all. However, for spatially local-
ized initial conditions polynomial decay can be observed. It turns out that
the solutions of the linearized problem with the slowest decay are given by
u(x, t) = ct−1/2 e−x /(4t) . For the terms occurring in (14.2) we then find the
2

asymptotic decay

∂t u ∼ t−3/2 , ∂x2 u ∼ t−3/2 and up ∼ t−p/2 ,

cf. Exercise 14.1. Hence only for p > 3 the nonlinear terms vanish faster than
the linear ones, and so only for p > 3 it can be expected that the nonlinear
problem behaves asymptotically as predicted by the linear one. Nonlineari-
ties which can be controlled with the help of the linearized problem in this
way are called irrelevant w.r.t. diffusion or simply irrelevant. Nonlinearities
exactly at the boundary, p = 3, are called critical. Surprisingly, it turns out
that for many interesting and rather complicated problems from physics the
nonlinear terms are irrelevant. It is has been observed at the beginning of
the 1990s that the concept of diffusive behavior and irrelevant nonlinearities
plays an important role in stability questions of pattern forming systems.
The plan of this chapter is as follows. In §14.1 we recall from §7.1 the
most important properties of the linear diffusion equation from a slightly dif-
ferent point of view, and introduce the concept of irrelevant nonlinearities
and a number of different methods to establish the irrelevance of nonlin-
earities. Then in §14.2.1 we show the occurrence of diffusive behavior and
irrelevant nonlinearities in the real GL equation. In §14.2.2 we explain how
this approach can be transferred to pattern forming systems in order to show
the diffusive stability of spatially periodic equilibria in the SH equation, the
Couette-Taylor problem or Bénard’s problem. The rest of this section is
devoted to critical nonlinearities. In §14.3.1 we discuss exponentially long
transient behavior in unstable Poiseuille flow, and in §14.3.2 we consider the
Burgers equation as limit equation. Examples are the inclined film problem
or self-similar mixing of phases in pattern forming systems where the group
velocity depends on the wave number. In §14.4 we introduce phase diffusion
equations, which are modulation equations occurring in diffusive stability
theory. Finally, in §14.5 we revisit similarities and differences between dif-
fusive and dispersive dynamics.

14.1. Linear and nonlinear diffusive behavior

14.1.1. Linear diffusive behavior. Solutions of the linear diffusion equa-
tion

(14.3) ∂t u = ∂x2 u,
14.1. Linear and nonlinear diffusive behavior 499

with u|t=0 = u0 , x ∈ R, t ≥ 0, and u(x, t) ∈ R, can be written explicitly as

 ∞
1 (x−y)2
(14.4) t∂x2
u(x, t) = (e u0 )(x) = √ e− 4t u0 (y) dy.
4πt −∞
Spatially constant functions stay constant in time, but for u0 ∈ Lq we obtain
from Lemma 7.3.19 that, with some constant C independent of t,
(14.5) u(t)Lp ≤ G(t)Lr u0 Lq ≤ Ct−1/(2r) u0 Lq ,
1 − x2
where 1/p = 1/q − 1/r for 1 ≤ q ≤ p ≤ ∞, and G(x, t) = √ e 4t . Thus,
4πt
spatially localized initial conditions give rise to solutions with polynomial
decay rates. Moreover, the solutions become flatter and flatter, since we
have for instance
∂xn uL∞ ≤ ∂xn G(t)L1 u0 L∞ ≤ Ct−n/2 u0 L∞ .
The decay always happens in the same self-similar way which can be seen
by looking at the Fourier transform of (14.3). The Fourier transformed
(k, t) = e−k t u
2
 = −k 2 u
diffusion equation ∂t u  is solved by u 0 (k). The self-
similar behavior can be extracted by considering the renormalized solution
⎛ ⎞
√ √ 
n
(k/ t, t) = e−k u0 (k/ t) = e−k ⎝ 0 (0) + o(t−n/2 )⎠ ,
t−j/2 k j u
2 2 (j)
u
j=0

0 is n-times differentiable. Since smoothness in Fourier space corresponds

if u
to decay rates in physical space, solutions to spatially localized initial condi-
tions vanish in a universal manner. Loosely speaking, if the initial conditions
decay like |x|−n , then

n−1
√
t−(j+1)/2 u
0 (0)Hj (x/ t) + O(t−n/2 )
(j)
u(x, t) =
j=0

for t → ∞, where Hj is a multiple of the jth Hermite polynomial. See

Exercise 14.2. Moreover, it is easy to see that
√
u(k/ t, t) − e−k u 0 (0)H θ ≤ Ct−1/2
2

2

for all θ ≥ 2. Using the fact that Fourier transform is an isomorphism from
Hmθ to H m , cf. Lemma 7.3.31, we obtain equivalently
θ
√ √ √
0 (0)e−x /4 H 2 ≤ Ct−1/2 .
2
(14.6)  t u(x t, t) − π u
θ

Most of the above ideas hold if we have a linear evolution operator eλ(k)t with
eigenvalues λ(k) ∼ −k 2 for k → 0. This is the reason why diffusive behavior
can be observed in a big variety of problems, as we will see. Textbooks about
self-similar solutions in various parabolic problems are [Bar96, SR10].
500 14. Diffusive stability

Remark 14.1.1. By introducing a logarithmic time scale the polynomial

decay rates can be transferred into exponential decay rates. This allows to
interpret the Hermite functions as eigenfunctions of a renormalized diffusion
operator. We follow the lines of [Way97] and introduce the new variable w
and the new coordinates ξ and τ by
√
(14.7) u(x, t) = t−1/2 w(x/ t, log t) = e−τ /2 w(ξ, τ ).
The transformed diffusion equation is then given by
(14.8) ∂τ w = ∂ξ2 w + (ξ/2)∂ξ w + w/2.

The spectral problem in Fourier space, for instance in H2θ , reads

(14.9)  − (k/2)∂k w
−k 2 w  = f,
 − λw

with f ∈ H2θ . The eigenfunctions ψs (k) = k s e−k to the real eigenvalues
2

λs = −s/2 are parameterized with s ∈ R. Since ∂kj ψs ∈ L2 is required

for j = 0, 1, . . . , θ, this leads to a set of θ eigenvalues {λj = −j/2 : j =
0, 1, . . . , θ − 1} and to essential spectrum {λ ∈ C : Reλ < −θ/2 − 1/4}. In
order to see this, one has to consider the singularity of ∂kj ψs at k = 0 for
large j.

14.1.2. Irrelevant nonlinearities. Following partly the lines of [MSU01]

we now explain for the nonlinear diffusion equation
(14.10) ∂t u = ∂x2 u + cup , u|t=0 = u0 ,
with t ≥ 0, x ∈ R, p ∈ N, and c ∈ R \ {0}, how the polynomial decay rates
for the linearized problem can be used to control the nonlinear terms.
For initial conditions u0 > 0 constant in space and c > 0 we always have
blow-up. However, for spatially localized initial conditions the situation is
more complicated. In case p = 2 and c > 0 we still have blow-up of solu-
tions to arbitrary small non-negative initial conditions with non-vanishing
integral, cf. [Wei81]. In case p = 3 the sign of c decides about stability
(c < 0) and instability (c > 0), but for p ≥ 4 the sign of c does not play a
role and small spatially localized perturbations vanish for t → ∞ with the
same polynomial decay rate as in the linear case.
There are at least three methods to prove the last statement, namely L1 -
L∞ estimates, the construction of Lyapunov functions, and the discrete and
continuous renormalization approach. We will present all three of them in
order to make clear that when L1 -L∞ estimates work also the renormaliza-
tion approach works and vice versa, although the statements seem different.
Moreover, we explain why the Lyapunov function approach in general gives
weaker results.
14.1. Linear and nonlinear diffusive behavior 501

14.1.3. L1 -L∞ estimates. This method relies on the Lq -Lp estimate (14.5),
the variation of constants formula, and suitable estimates for the nonlinear-
ity.
Lemma 14.1.2. Let p > 3. For all C > 0 there exists a δ > 0 such that
solutions u of ( 14.10) with u0 L1 + u0 L∞ ≤ δ satisfy

u(t)L1 ≤ C and u(t)L∞ ≤ C(1+t)− 2

1

for all t ≥ 0.

Proof. We introduce
a(t) = sup u(s)L1 , b(t) = sup (1+s)1/2 u(s)L∞ .
0≤s≤t 0≤s≤t

We use the variation of constant formula (2.6) estimate successively the

L1 -norm and the L∞ -norm of the terms on right-hand side. We have, cf.
Exercise 14.3,
2
(14.11) et∂x u0 L1 ≤ u0 L1
and
$
t∂x2 u0 L∞ , t≤1
(14.12) e u0 L∞ ≤ √ 1
4πt
u0 L1 , t>1

such that
2
(1+t)1/2 et∂x u0 L∞ ≤ u0 L1 + u0 L∞ .
For the nonlinear terms we use up L∞ ≤ upL∞ and up L1 ≤ uL p−1
∞ uL1 .

We obtain
 
 t (t−s)∂ 2 p  
(1+t) 
1/2
 0e
x u (s) ds
 ∞ ≤ (1+t)
1/2 t e(t−s)∂x2 
0 L1 →L∞ u L1 ds
p

 t/2 L
≤ ((1+t)1/2 0 (t−s)−1/2 (1+s)−(p−1)/2 ds
t
+(1+t)1/2 t/2 (t−s)−1/2 (1+s)−(p−1)/2 ds) · b(t)p−1 a(t)
 t/2
≤ ((1+t)1/2 0 (t−t/2)−1/2 (1+s)−(p−1)/2 ds
t
+(1+t)1/2 t/2 (t−s)−1/2 (1+t/2)−(p−1)/2 ds) · b(t)p−1 a(t)
≤ C1 b(t)p−1 a(t),
with a constant C1 independent of t for p > 3, and
 
 t (t−s)∂ 2 p   t (t−s)∂ 2
 e x u (s) ds ≤ 0 e
x 1
L →L1 u L1 ds
p
 0  1
t L
≤ 0 (1+s)−(p−1)/2 ds · b(t)p−1 a(t) ≤ C2 b(t)p−1 a(t),
with a constant C2 independent of t for p > 3. Thus,
a(t) ≤ a(0)+b(0)+|c|C2 b(t)p−1 a(t), b(t) ≤ a(0)+b(0)+|c|C1 b(t)p−1 a(t).
502 14. Diffusive stability

Introducing y(t) = a(t) + b(t) and adding the two inequalities yields
(14.13) y(t) ≤ 2y(0) + (C1 + C2 )y(t)p .
For y(0) < δ, with δ > 0 sufficiently small, the two curves y → y and
y → y(0) + (C1 + C2 )y p possess two intersection points y1 , y2 , with y1 < y2 .
Inequality (14.13) is valid for y ≤ y1 and y ≥ y2 . Since y(0) ≤ y1 and since
we have continuity of t → y(t) we must have y(t) ≤ y1 for all t ≥ 0. 
According to Remark 14.1.1 we cannot expect faster decay rates even
if the initial conditions vanish much faster than a general L1 -function for
|x| → ∞, and for L∞ -initial conditions there is no decay. For functions
with an intermediate spatial decay, e.g., L2 initial conditions, we obtain the
following. On a linear level we have via Lemma 7.3.19 that
u(t)L∞ = G(t) ∗ u0 L∞ ≤ G(t)L2 u0 L2 ,
x2
1
with G(x, t) = √4πt e− 4t . We find
   1/2  
1 − x2 2 1 − 2x2 √ 1/2
G(t)L2 = √ e 4t dx = e 4 dy t ∼ t−1/4 ,
R 4πt R 4πt

such that u(t)L∞ ≤ Ct−1/4 u0 L2 . By introducing

a(t) = sup u(s)L2 and b(t) = sup (1+s)1/4 u(s)L∞ ,
0≤s≤t 0≤s≤t

the following result easily can be proved almost line for line as above.
Corollary 14.1.3. Let p > 5. For all C > 0 there exists a δ > 0 such that
solutions u of (14.10) with u0 L2 + u0 L∞ ≤ δ satisfy
u(t)L2 ≤ C and u(t)L∞ ≤ C(1+t)− 4
1

for all t ≥ 0. Similarly, for u0 ∈ Lq ∩ L∞ small, p > 2q + 1, and q ∈ [1, ∞),

we have
−1
u(t)Lq ≤ C and u(t)L∞ ≤ C(1+t) 2q .

14.1.4. Lyapunov functions. The usage of Lyapunov functions is well

established in nonlinear stability problems if the nonlinearity has the correct
sign such as in ∂t u = ∂x2 u − u3 . For small initial conditions Lyapunov
functions can be used even if the nonlinearity has the wrong sign. For
exponentially stable equilibria this is obvious, but also for diffusive stability
problems the approach can be used, cf. [EW94, GM98]. We do not get
the optimal power p for irrelevant nonlinearities, but the method is also
applicable on unbounded domains Ω ⊂ Rd with Ω = Rd , cf. [ES98].
As an example we consider again
∂t u = ∂x2 u + cup , u|t=0 = u0 ,
14.1. Linear and nonlinear diffusive behavior 503

 2
in
 case Ω = R. We introduce
 the functionals I(u) = R u dx, J(u) =
2 2 2 2
R (∂x u) dx and K(u) = R (∂x u) dx. For initial conditions with an L -
∞
decay for |x| → ∞, due to the L -L estimate (14.5) we can only expect to
2

handle nonlinearities for p > 5.

With u2L∞ ≤ I 1/2 J 1/2 and J 2 ≤ I K we obtain
 
1 d
I = u∂t u dx = u∂x2 u + cup+1 dx
2 dt R R
 
p−1
= −(∂x u) + cu
2 p+1
dx ≤ −J + |c|u(x)L∞ u2 dx
R R
≤ − J + |c|I (p+3)/4 (p−1)/4
J = −J [1−|c|I (p+3)/4 (p−5)/4
J ],
and

1 d
J = (∂x u)(∂t ∂x u) dx
2 dt
R 
p−1
= (∂x u)(∂x u) + cpu (∂x u) dx ≤ − (∂x2 u)2 dx + |c|puL
3 p−1 2
∞ J
R R
≤ − K + |c|pI (p−1)/4 J (p+3)/4 ≤ −K [1−|c|pI (p+3)/4 J (p−5)/4 ].
Hence, we have I˙ ≤ 0 and J˙ ≤ 0 if I(u0 )(p+3)/4 J(u0 )(p−5)/4 > 0 is sufficiently
small. Thus, we have proved the following result.
Lemma 14.1.4. Let p ≥ 5. Then, there exists a δ0 > 0 such that for
all δ ∈ (0, δ0 ) and all solutions u of ( 14.10) with u0 H 1 ≤ δ we have
supt≥0 u(t)H 1 ≤ δ.

Note that the method also allows to handle the boundary case p = 5. In
higher space dimensions the critical exponent obtained with this approach
can be reduced to p = 1 + 4/d, cf. Exercise 14.4.

14.1.5. The continuous renormalization process. In addition to some

stability result, the renormalization process gives the asymptotics of the
decay towards zero. It relies on formula (14.6). By a fixed point argument
we prove not only for the linear √ but also
√ for the nonlinear system (14.10)
that the renormalized solution t u(t, x t) converges towards a multiple of
the Gaussian e−x /4 . There are two approaches, a continuous and a discrete
2

one. In the continuous approach the system satisfied by the renormalized

solution is considered directly, where additionally a logarithmic time scale is
used to transfer the polynomial into exponential decay rates. As in Remark
14.1.1 we introduce the new variable w and the new coordinates ξ and τ by
√
(14.14) u(x, t) = t−1/2 w(x/ t, log t) = e−τ /2 w(ξ, τ ).
The transformed equation is given by
(14.15) ∂τ w = w/2 + (ξ/2)∂ξ w + ∂ξ2 w + e(3−p)τ /2 wp ,
504 14. Diffusive stability

where we choose w|τ =0 ∈ H22 . Moreover, w.l.o.g. for our purposes we set
c = 1. In H22 we have a zero eigenvalue, and the rest of the spectrum
is left of the line {λ ∈ C : Reλ = −1/2}, cf. Remark 14.1.1. Since the
solutions of the linearized problem are uniformly bounded and since the
nonlinear terms vanish with an exponential rate there exist δ, C > 0 such
that supτ ≥0 w(τ )H22 < C for the solutions w of (14.15) if w|τ =0 H22 < δ
and p > 3. If we denote by w0 the part of w belonging to the eigenvalue 0
and with w1 the rest of w, then by integration of the variation of constant
formula w.r.t. time we can conclude that
3−p 1 3−p
w0 (·, τ ) = wlim e−(·) }τ
2 /4
+ O(e 2
τ
) and w1 (τ ) = O(emax{− 2 , 2 )
for τ → ∞, where wlim ∈ R is a constant only depending on the initial
conditions. This leads to the following convergence result, cf. [Way97].
Theorem 14.1.5. Let p > 3. There exist δ, C > 0 such that for all solutions
u of (14.10) with u|t=0 H22 ≤ δ, there exists a wlim ∈ R such that
√ √
 t u(· t, t) − wlim e−(·) /4 H22 ≤ C(1+t)− max{1/2,(3−p)/2} for all t ≥ 0.
2

14.1.6. The discrete renormalization approach. In the discrete ap-

proach, cf. [BK92, BKL94, Gal94], a sequence of problems is considered
which formally converges towards the linear diffusion equation. Define
(14.16) un (x, τ ) = Ln u(Ln x, L2n τ ),
with L > 1 fixed and n ∈ N. Then un satisfies
(14.17) ∂τ un = ∂x2 un + Ln(3−p) upn for τ ∈ [L−2 , 1]
and
(14.18) un (x, L−2 ) = Lun−1 (Lx, 1).
Formally, for p > 3 the influence of the nonlinear terms vanishes geometri-
cally for n → ∞. In the limit we obtain the linear diffusion equation. By
a fixed point argument we prove for the nonlinear problem the convergence
of the sequence of functions un |τ =1 = Ln u(Ln ·, L2n ) towards the limit of
the linear problem, namely a multiple of the Gaussian e−x /4 . We use this
2

approach to give a
Proof of Theorem 14.1.5. In Fourier space (14.10) is given by
(14.19)  = −k 2 u
∂t u ∗p ,
+u
where u∗p denotes the p-times convolution of a function u ∗p = u
, i.e., u  ∗. . .∗
. Corresponding to (14.16) we rescale u
u  to u −n
(L k, L2n τ ) for a
n (k, τ ) = u
fixed L > 1. Note that due to the scaling properties of Fourier transform
. We obtain
there is no Ln in front of u
(14.20) n = −k 2 u
∂τ u ∗p
n + Ln(3−p) un ,
14.1. Linear and nonlinear diffusive behavior 505

where a factor Ln(1−p) in front of the convolution terms comes from the
substitution in the integrals. In order to simplify notation we denote all
constants with the same symbol C if they can be chosen independently of
L. We solve (14.19) via (14.20) by the following renormalization procedure.
(RG)n→n+1 : Equation (14.20) is solved on the time interval [1/L2 , 1].
n+1 (k, 1/L2 ) = u
Then u n (k/L, 1) is taken as initial condition for n + 1.
For solving (14.20) we use the variation of constant formula
n (k, τ ) = e−k
2 (τ −1/L2 )
u n−1 (k/L, 1)
u
(14.21) τ
e−k ∗p
2 (τ −s)
+Ln(3−p) 1/L2 un (k, s) ds.

The size Rn := supτ ∈[1/L2 ,1]  n on the time in-

un (τ )H22 of the solutions u
terval [1/L2 , 1] can be controlled in terms of the initial conditions.
Lemma 14.1.6. There exist δ, C1 and L0 > 0 such that for all u n−1 |τ =1 ∈
H22 with 
un−1 |τ =1 H22 < δ and all L > L0 the solution un ∈ C([1/L2 , 1], H22 )
exists and satisfies
Rn ≤ C1 L5/2 
un−1 |τ =1 H22 .

Proof. Rescaling gives the estimate

v(·/L)H22 ≤ CL5/2 vH22 .
Thus, the linear terms in (14.21) can be estimated by
e−(·)
2 (τ −1/L2 )
sup n−1 (·/L, 1)H22
u
τ ∈[1/L2 ,1]

e−(·)
2 (τ −1/L2 )
≤ sup C 2 
un−1 (·/L, 1)H22 ≤ CL5/2 
un−1 (·, 1)H22 .
b
τ ∈[1/L2 ,1]

This estimate for the semigroup and Sobolev’s embedding theorem show for
the nonlinear terms
  τ 
 
 n(3−p) −k2 (τ −s) ∗p 
sup L e n (·, s) ds ≤ CLn(3−p) Rnp .
u
τ ∈[1/L2 ,1]  1/L 2  2
H2

Choosing δ > 0 sufficiently small and L0 sufficiently large it is easy to see

that the right-hand side of (14.21) defines a contraction in a ball of radius
O(L5/2 
un−1 |τ =1 H22 ) in the space C([1/L2 , 1], H22 ). Therefore, the solution
which is the fixed point of this contraction stays in H22 in a ball of the same
radius O(L5/2 un−1 |T =1 H22 ). 
In order to show the convergence of the points u n |τ =1 towards A∞ ψ in
H22 , with some A∞ ∈ R and the Gaussian ψ(k) = e−k , let us define the
2

projection Π n |k=0 . We are done, if we prove the convergence of

un = u
n |τ =1 − An ψ,
ρn = u
506 14. Diffusive stability

with An = Πun |τ =1 towards zero and the existence of the limit limn→∞ An .
By definition we have ρn |k=0 = 0. The variables An and ρn satisfy
 1
∗p
e−k (1−s) u
n(3−p) 2
An+1 =An + Π(L n+1 (k, s) ds),
1/L2
 1
−k2 (1−1/L2 )
e−k ∗p
2 (1−s)
ρn+1 (k) =e ρn (k/L) + L n(3−p)
un+1 (k, s) ds
1/L2

+ e−k
2 (1−1/L2 )
An ψ(k/L) − An+1 ψ(k).
For the terms on the right-hand side we get
 1
∗p
e−(·) (1−s) u
2
L n(3−p)
n+1 (·, s) dsH22 ≤ CRn L
p n(3−p)
,
1/L2
|Πu| ≤ CuH22 ,
e−(·)
2 (1−1/L2 )
ρn (·/L)H22 ≤ (C/L)
ρn H22 ,
e−(·)
2 (1−1/L2 )
An ψ(·/L) − An+1 ψ(·)H22 ≤ C|An+1 − An |.
Using these estimates we obtain
|An+1 − An | ≤CLn(3−p) Rnp ,
ρn+1 H22 ≤(C/L)
 ρn H22 + CLn(3−p) Rnp .

Using Lemma 14.1.6 and the abbreviations rn = 

ρn H22 gives

|An+1 − An | ≤CLn(3−p) (L5/2 )p (An + rn )p ,

rn ≤(C/L)rn + CLn(3−p) (L5/2 )p (An + rn )p .
We now choose L0 > 1 and n0 > 0 so big that for n > n0 we have
|An+1 − An | ≤L(n−n0 )(p−3) (An + rn )p ,
rn+1 ≤rn /10 + L(n−n0 )(p−3) (An + rn )p ,
if supn∈N (|An | + rn ) < δ, which was needed in the derivation of this es-
timate since we used Lemma 14.1.6. If supn∈N (|An | + rn ) < δ, then the
sequence of An converges geometrically to a limit A∞ < ∞ and we find that
limn→∞ rn = 0. Since the quantities |An | and rn increase only for at most
n0 steps we can stay smaller than δ > 0 from Lemma 14.1.6 if we choose
|A1 | and r1 of order O(L−m ), for an m > 0 sufficiently large.
Since the convergence goes as L− min(1,p−3)n , since Fourier transform is
an isomorphism in H22 , and since the convergence holds for all L ∈ [L0 , L20 ]
we obtain Theorem 14.1.5. 
As a consequence we have
14.2. Diffusive stability of spatially periodic equilibria 507

Corollary 14.1.7. The fixed point u∗ = 0 in (14.10) is (H22 , L∞ )-stable in

the sense of Definition 7.1.19.

Proof. Using the notation from the proof of Theorem 14.1.5 the statement
follows from
√
u(·, t)L∞ ≤Ct−1/2 e−(·) /(4t) L∞ + Ct−1 ρ(·/ t, t)L∞
2

=Ct−1/2 + Ct−1 ρ(·, t)L∞

≤Ct−1/2 + Ct−1 ρ(·, t)H22 = O(t−1/2 ).

Remark 14.1.8. a) The above methods work for more general nonlin-
earities, too. If we call the exponent p of the nonlinear term up the de-
gree of irrelevance, then up1 (∂x u)p2 (∂x2 u)p3 has the degree of irrelevance
p = p1 + 2p2 + 3p3 . For x ∈ R and the diffusion equation there are only
three relevant monomial nonlinearities, namely u2 , u3 , and u∂x u.
b) In higher space dimensions the solutions of the linear diffusion equa-
tion ∂t u = Δu, with x ∈ Rd , satisfy the estimate
u(t)L∞ ≤ Ct−d/2 uL1 .
With d increasing, the number of directions increases in which energy can
diffuse. From the above analysis it is easy to see that in R2 the only relevant
monomial is u2 . In Rd with d ≥ 3 all smooth nonlinearities are irrelevant
w.r.t. diffusion. See [Uec99] for an abstract theorem.

14.2. Diffusive stability of spatially periodic equilibria

Diffusive stability results for the GL equation have been shown first in
[CEE92, BK92]. We already discussed this topic in §8.3.6 in an infor-
mal way. It is the purpose of this section to make the formal arguments
of §8.3.6 rigorous. The previous arguments have to be extended in order
to handle additional exponentially damped terms. The presented approach
is a generalization of [MSU01] and is less involved than the ones given in
[CEE92, BK92]. The scheme developed subsequently is very general and
it will be explained afterwards how it can be used to prove for instance the
diffusive stability of spatially periodic equilibria in pattern forming systems.

14.2.1. Diffusive stability in the GL equation. The real GL equation

(14.22) 2
∂T A = ∂X A + A − A|A|2
with X ∈ R, T ≥ 0, and A(X, T ) ∈ C possesses a two-dimensional family of
stationary solutions

(14.23) A = Aper [q, φ0 ](X) = 1 − q 2 eiqX+iφ0 ,
508 14. Diffusive stability

parameterized by φ0 , q ∈ R. In §8.3.6 we discussed the linearization around

these equilibria and found diffusive modes and exponentially damped modes.
They can be separated easily by introducing polar coordinates
A(X, T ) = r(X, T )eiφ(X,T )
in the real GL equation, which then transforms into
2(∂X r)(∂X φ)
2
∂T r = ∂X r + r − (∂X φ)2 r − r3 , 2
∂T φ = ∂X φ+ .
r
For simplicity we restrict ourselves to (r, φ) = (1, 0) as the stationary pat-
tern. We introduce the deviation (s, φ) from the equilibrium by setting
r = 1 + s and find
∂T s = ∂X2 s − 2s − (∂ φ)2 − (∂ φ)2 s − 3s2 − s3 ,
X X
(14.24) 2(∂ X s)(∂ X φ)
∂T φ = ∂X2 φ+ .
1+s
As we already know from §8.3.6, the linearized system
2
∂T s = ∂X s − 2s, 2
∂T φ = ∂X φ,
possesses solutions (s, φ) = (sk , φk )eikx+μj (k)t , with μ1 (k) = −k 2 and μ2 (k) =
−k 2 − 2. Hence, the component s which corresponds to μ2 is exponentially
damped, and so we expect s to be slaved to φ, i.e., we expect s to behave in
lowest order as (∂X φ)2 . Since φ behaves diffusively, we expect decay rates
φ ∼ t−1/2 and s ∼ t−2 , and that φ formally satisfies
(14.25) 2
∂T φ = ∂X φ − ∂X ((∂X φ)2 ))(∂X φ) + h.o.t..
In §8.3.6 we argued that with these asymptotics the nonlinear terms on the
right-hand side have a degree of irrelevance of seven and hence are irrelevant
w.r.t. linear diffusion. Thus, with a slight improvement of the previous
methods the following L1 -L∞ diffusive stability result can be established.
The proof will be given below.
Theorem 14.2.1. The stationary solution A = 1 of (14.22) is diffusively
stable in the following sense. For all C > 0 there exists a δ > 0 such that
solutions A = 1 + v of (14.22) with v |T =0 L1 + 
v |T =0 L∞ ≤ δ satisfy
v (T )L∞ ≤ C,
 v (T )L1 ≤ C(1+T )−1/2

for all T ≥ 0.

Since the Fourier transform is continuous from L1 to L∞ we have as a

direct consequence the following result.
Corollary 14.2.2. For all C > 0 there exists a δ > 0 such that solutions
v |T =0 L1 + v|T =0 L1 ≤ δ satisfy
A = 1 + v of (14.22) with 
v(T )L∞ ≤ C(1+T )−1/2
for all T ≥ 0.
14.2. Diffusive stability of spatially periodic equilibria 509

Remark 14.2.3. In order to obtain an irrelevant nonlinearity with a degree

of irrelevance of seven we would have to handle (14.24) as a quasilinear
system, for instance by choosing s ∈ H 1 and φ ∈ H 2 . By handling it as a
semilinear one, for instance by choosing s ∈ H 1 and φ ∈ H 1 , we would come
from seven to six. This procedure then would not allow us to handle systems
with a degree of irrelevance of four due to the loss of one. In order to avoid
the more complicated analytic tools used for quasilinear systems we choose
s ∈ H 1 and φ ∈ H 2−ϑ with ϑ ∈ (0, 1) which finally leads only to a loss of
ϑ in the degree of irrelevance. Since ϑ > 0 can be chosen arbitrarily small,
the scheme used in the proof of Theorem 14.2.1 will allow us to handle all
systems with irrelevant nonlinearities.
Remark 14.2.4. Before we give a proof of Theorem 14.2.1 we consider
time-periodic solutions for the complex GL equation
(14.26) 2
∂T A = (1 + iα)∂X A + A − (1 + iβ)A|A|2 ,
where α, β ∈ R, X ∈ R, T ≥ 0, and A(X, T ) ∈ C, in order to make clear
that a more abstract consideration of the problem is advantageous. The
complex GL equation possesses wave trains

A(X, T ) = 1 − k 2 ei(kX−ω(k)T ) ,
where ω(k) = β + (α − β)k 2 . We introduce the amplitude and phase devia-
tions (s, φ) of the wave train to k = 0 via
A(X, T ) = (1 + s(X, T ))ei(−βt+φ(X,T )) ,
which satisfy
2
∂T s =∂X s − 2s − (∂X φ)2 − (∂X φ)2 s − 2α(∂X s)(∂X φ)
− α∂X
2
φ − α(∂X
2
φ)s − 3s2 − s3 ,
2 s
∂X 2(∂X s)(∂X φ)
2
∂T φ =∂X φ+α − α(∂X φ)2 + − 2βs − βs2 .
1+s 1+s
With the choice s ∈ H 2 and φ ∈ H 2 this system is a quasilinear system al-
though in the original coordinates the nonlinear terms contain no derivative
at all. Note that there is no choice of H s spaces which makes it a semilinear
system. Therefore, already for this simple problem it will be advantageous
to choose a more abstract approach as we do below.

Proof of Theorem 14.2.1. We go back to the GL equation (14.22). The

deviation v = A − 1 from the equilibrium A = 1 satisfies
(14.27) ∂T v = Lv + N (v)
where
2
Lv = ∂X v + v − (2v + v), N (v) = −(v 2 + 2vv + v 2 v).
510 14. Diffusive stability

The eigenvalue problem Lv = μv is solved by Fourier modes v(x, t) =

c,s (k)eikx with eigenvalues μc,s (k). In the Eckhaus stable domain we have
ϕ
constants σc > 0 and σs > 0 such that
Reμc (k) ≤ −σc k 2 and Reμs (k) ≤ −σs
for all k ∈ R, cf. §8.3.5. At this point it turns out to be advantageous to
work in Fourier space. Hence we consider
(14.28) v + N
∂T v = L  (
v ),
where v = F v, L = F LF −1 and N  (
v ) = F (N (F −1 v)). There exists a
k1 > 0 such that for all |k| ≤ k1 the two curves of eigenvalues are separated,
and so we define
pc (k)
v = χ(k)
ϕ ∗c (k), v,
where χ(k) = 1 for |k| ≤ k1 /2, and χ(k) = 0 for |k| > k1 /2, and where ϕ ∗c (k)
is the eigenvector associated to the adjoint eigenvalue problem normalized
by
ϕ∗c (k), ϕ
c (k) = 1. Moreover, define Ps (k) v = v − (
pc (k) c (k). We
v )ϕ
write v = vc ϕ c + vs and use the projections to separate (14.28) into two
parts, namely
 (
∂T vc = μc vc + pc N v ),  s vs + Ps N
∂T vs = LP  (
v ).
 s we have the estimates
For the semigroups generated by μc and LP
eμc T vc L1 ≤eμc T L∞ 
vc L1 ≤ 
vc L1 ,
eμc T vc L∞ ≤eμc T L∞ 
vc L∞ ≤ 
vc L∞ ,
vc L∞ ≤ CT −1/2 
eμc T vc L1 ≤eμc T L1  vc L∞ ,

eT LPs vs L1 ≤Ce−σs T /2 
vs L1 ,

eT LPs vs L∞ ≤Ce−σs T /2 
vs L∞ .

After showing the linear diffusive behavior we have to establish the irrel-
evance of the nonlinearity w.r.t. to this diffusive behavior. We stay a little
bit more abstract than necessary at this point and expand the nonlinear
terms into
 (
pc N vc , vc ) + B2,2 (
v ) =B2,1 ( vc , vs ) + B3 (
vc , vc , vc ) + gc (
vc , vs ),
 (
Ps N vc , vc ) + gs (
v ) =B2,3 ( vc , vs ),
where the B2,j are symmetric bilinear maps, where B3 is a symmetric trilin-
ear map, and where gc and gs stand for the remaining terms. The splitting
is motivated as follows. If vc decays like t−1/2 , then vs , which is slaved to
vc , decays at least like t−1 . Then gc = O(| vc |4 + |
vc |2 |
vs | + |
vs |2 ) decays like
−2
t and gs = O(| vc | + |
3 vc ||
vs | + |
vs | ) decays like t
2 −3/2 . They are therefore
both irrelevant w.r.t. diffusive behavior.
14.2. Diffusive stability of spatially periodic equilibria 511

In order to prove the irrelevance of the other terms we make a change

of coordinates which removes vs from the relevant terms. We set

(14.29) c ,
vc = w  −1 Es B2,3 (
vs = −L vc , vc ) + w
s

Note that L  −1 Es exists for the considered values of k and that w

s decays
like t −3/2 . We obtain
c = μc w
∂T w c + N1 (w
c ) + N2 (w c , w
s ),
(14.30)
∂T w  s Ps w
s = L s + N3 (wc , w
s ),
where

c ) =B2,1 (w
N1 ( w c , w
c ) + B2,2 (w  −1 Es B2,3 (w
c , −L c , w
c ))
c , w
+ B3 ( w c , w
c ),
c , w
N2 ( w s ) =gc (w  −1 Es B2,3 (w
c , −L c , w
c ) + w
s ) + B2,2 (w
c , w
s ),
c , w
N3 ( w s ) =gs (w  −1 Es B2,3 (
c , −L vc , vc ) + w  −1 Es B2,3 (w
s ) + ∂t (L c , w
c )).

By this transformation N2 decays like t−2 and N3 decays like t−3/2 such
that they are both irrelevant w.r.t. diffusive behavior. In order to prove the
irrelevance of N1 (wc ) we write it as

c ) = K2 (k, k − l, l)w
N1 ( w c (k − l)w
c (l) dl
R
 
+ K3 (k, k − l, l − m, m)wc (k − l)w
c (l − m)w
c (m) dm dl,
R R

with smooth kernels K2 (k, k − l, l) and K3 (k, k − l, l − m, m). A derivative

corresponds in Fourier space to a multiplication with a wave number k,
and hence terms are more irrelevant, if they are multiplied with a kernel
vanishing at k = l = 0. The faster the kernel vanishes, the more irrelevant
are the associated terms. In §8.3.6 we already computed formally that the
nonlinear terms are irrelevant w.r.t. diffusion. Since after the transformation
c -equation of the abstract system (14.30) coincides with (14.25)
(14.29) the w
up to terms with asymptotic behavior of O(t−3/2 ), we must have that the
quadratic terms have a decay of order O(t−2 ). Therefore, K2 has to vanish
quadratically and K3 at least linearly, i.e., we have the estimates

(14.31) |K1 (, −m, m)| ≤C|2 + (−m)2 + m2 |,

(14.32) |K2 (,  − m, m − k, k)| ≤C(|| + | − m| + |m − k| + |k|).

For the special choice of the stationary solution, A = 1, we have no quadratic

terms at all. However, in general (14.31) occurs, and so in the following we
will use the weaker estimate (14.31). See also §14.4.1 and §14.4.2.
512 14. Diffusive stability

We proceed as in §14.1.3 and consider the variation of constant formula

 T
wc (T ) =e w
μc T
c (0) + eμc (T −τ ) (N1 (w
c ) + N2 (w
c , w
s ))(τ ) dτ,
0
(14.33)  T
 
w s (0) +
s (T ) =eT Ls Ps w e(T −τ )Ls Ps N3 (w
c , w
s )(τ ) dτ
0
for (14.27). In the following we use the abbreviations
ac (T ) = sup w
c (τ )L∞ ,
0≤τ ≤T

bc (T ) = sup (1+τ )1/2 w

c (τ )L1 ,
0≤τ ≤T

cc,ν (T ) = sup (1+τ )ν/2 |k|ν w

c (τ )L∞ ,
0≤τ ≤T

dc,ν (T ) = sup (1+τ )(ν+1)/2 |k|ν w

c (τ )L1 .
0≤τ ≤T

with cc = cc,1 , dc = dc,1 , and

as (T ) = sup (1+τ )ν/2 w
s (τ )L∞ ,
0≤τ ≤T

bs (T ) = sup (1+τ )(ν+1)/2 w

s (τ )L1 ,
0≤τ ≤T

with ν < 2 fixed arbitrarily close to 2, and denote many different con-
stants with the same symbol C if they can be chosen independently of
ac (T ), . . . , bs (T ), and T .
From the previous representations and Lemma 7.3.19 we find
c )L1 ≤C(|k|w
N1 (w c 2L1 + w
c L1 |k|2 w
c L1 + w
c 2L1 |k|w
c L1 ),
c , w
N2 (w s )L1 ≤C(w
c 4L1 + w
c L1 w
s L1 + w
s 2L1 ),
c , w
N3 (w s )L1 ≤C(w
c 3L1 + w
c L1 w
s L1 + w
s 2L1 + w
c L1 ∂T w
c L1 ),
and
c )L∞ ≤C(|k|w
N1 (w c L1 |k|w
c L∞ + w
c L1 |k|2 w
c L∞
+ w
c 2L1 |k|w
c L∞ ),
c , w
N2 (w s )L∞ ≤C(w
c 3L1 w
c L∞ + w
c L∞ w
s L1 + w
s L1 w
s L∞ ),
c , w
N3 (w s )L∞ ≤C(w
c 2L1 w
c L∞ + w
c L∞ w
s L1 + w
s L1 w
s L∞
+ w
c L∞ ∂T w
c L1 ),
c L1 on the right-hand side will be replaced via
where ∂T w
c L1 ≤ C(|k|2 w
∂T w c L1 + N1 (w
c )L1 + N2 (w
c , w
s )L1 ).
c has compact support in Fourier space, |k|2 w
Since w c can be estimated in
terms of |k| w
ν c for every ν ∈ [0, 2).
14.2. Diffusive stability of spatially periodic equilibria 513

The proof of Theorem 14.2.1 proceeds in three steps, first in A) estimates

c equation, in B) estimates for the terms in the w
for the terms in the w s
equation, which are brought together in C).
A1) We estimate
 T   T
 
 e μc (T −τ ) 
c )(τ ) dτ 
N1 (w ≤ eμc (T −τ ) L∞ →L∞ N1 (w c )(τ )L∞ dτ

0 L∞ 0
 T
≤C (1+τ )−3/2 dτ · (cc (T )dc (T ) + cc,μ (T )bc (T ) + b2c (T )cc (T ))
0
≤C(cc (T )dc (T ) + cc,μ (T )bc (T ) + b2c (T )cc (T )).

Similarly we find
 T 
 
 e μ c (T −τ )
N2 (w s )(τ ) dτ 
c , w ≤ C(b3c (T )ac (T )+ac (T )bs (T )+as (T )bs (T )).
 
0 L∞

A2) Next we estimate

 

1/2 
T 
(1+T )  e μc (T −τ )
c )(τ ) dτ 
N1 (w 
0 L1
 T
≤(1+T )1/2 eμc (T −τ ) L∞ →L1 N1 (w
c )(τ )L∞ dτ
0
 T
≤C(1+T ) 1/2
(T − τ )−1/2 (1+τ )−3/2 dτ
0
× (cc (T )dc (T ) + cc,μ (T )bc (T ) + b2c (T )cc (T ))
 T /2
≤C(1+T ) 1/2
(T /2)−1/2 (1+τ )−3/2 dτ
0
× (cc (T )dc (T ) + cc,μ (T )bc (T ) + b2c (T )cc (T ))
 T
+ C(1+T ) 1/2
(T − τ )−1/2 (1+T /2)−3/2 dτ
T /2

× (cc (T )dc (T ) + cc,μ (T )bc (T ) + b2c (T )cc (T ))

≤C(cc (T )dc (T ) + cc,μ (T )bc (T ) + b2c (T )cc (T )).

Similarly we find
  
 T 
(1+T )1/2 e μc (T −τ )
s )(τ ) dτ 
c , w
N2 (w
 
0 L1
≤ C(b3c (T )ac (T ) + ac (T )bs (T ) + as (T )bs (T )).
514 14. Diffusive stability

A3) We estimate
 T 

ν/2 

(1+T )  e μc (T −τ ) ν
c )(τ ) dτ 
|k| N1 (w  ∞
0 L
 T
≤(1+T )ν/2 eμc (T −τ ) |k|ν L∞ →L∞ N1 (w c )(τ )L∞ dτ
0
 T
≤(1+T ) ν/2
C (T − τ )−ν/2 (1+τ )−3/2 dτ · (cc (T )dc (T ) + b2c (T )cc (T ))
0
≤C(cc (T )dc (T ) + b2c (T )cc (T )).
 T  T /2  T
again by separating 0 = 0 + T /2 . Similarly we find
 T 

ν/2 

(1+T )  e μc (T −τ )
|k| N2 (w
ν
s )(τ ) dτ 
c , w 
0 L∞
≤ C(b3c (T )ac (T ) + ac (T )bs (T ) + as (T )bs (T )).

A4) The last estimate for the diffusive part is

 T 

(ν+1)/2 

(1+T )  e μc (T −τ )
c )(τ ) dτ 
|k| N1 (w
ν
 1
0 L
 T −1
≤(1+T )(ν+1)/2 eμc (T −τ ) |k|ν L∞ →L1 N1 (w c )(τ )L∞ dτ
0
 T
+ (1+T ) (ν+1)/2
eμc (T −τ ) |k|ν L1 →L1 N1 (w
c )(τ )L1 dτ
T −1
 T −1
≤(1+T )(ν+1)/2 C (T −τ )−(ν+1)/2 (1+τ )−3/2 dτ
0
× (cc (T )dc (T ) + b2c (T )cc (T ))
 T
+ (1+T ) (ν+1)/2
C (T −τ )−ν/2 (1+τ )−2 dτ · (d2c (T ) + b2c (T )dc (T ))
T −1
≤s1 + C(cc (T )dc (T ) + b2c (T )cc (T )).
 T −1  T /2  T −1
We split 0 . . . = 0 . . . + T /2 . . ., respectively s1 = s2 + s3 , and find
 T /2
s2 ≤(1+T ) (ν+1)/2
(T /2)−(ν+1)/2 (1+τ )−3/2 dτ · (d2c (T ) + b2c (T )dc (T )).
0
Moreover,
 T −1
s3 ≤(1+T ) (ν+1)/2
(T − τ )−(ν+1)/2 (1+T /2)−3/2 dτ
T /2

× (d2c (T ) + b2c (T )dc (T ))

14.2. Diffusive stability of spatially periodic equilibria 515

such that finally

s1 ≤ C(d2c (T ) + b2c (T )dc (T )).
Similarly we find
 
(ν+1)/2 
 T 
(1+T )  e μc (T −τ )
|k| N2 (w
ν
s )(τ ) dτ 
c , w 
0 L1
≤ C(d2c (T ) + b2c (T )dc (T ) + b4c (T ) + bc (T )bs (T ) + b2s (T )).
B1) For the linearly exponentially damped part we first estimate
 T 

ν/2   s Ps 
(1+T )  e (T −τ )L
N3 (w s )(τ ) dτ 
c , w 
0 L∞
 T

≤(1+T )ν/2 e(T −τ )Ls Ps L∞ →L∞ N3 (w
c , w
s )(τ )L∞ dτ
0
 T
≤(1+T )ν/2 e−σs (T −τ ) (1+τ )−ν/2 dτ
0
× C(b2c (T )ac (T )+ac (T )bs (T )+as (T )bs (T )
+ac (T )(dc,ν +d2c (T )+b2c (T )dc (T )+b4c (T )+bc (T )bs (T )+b2s (T )))
≤C(b2c (T )ac (T )+ac (T )bs (T )+as (T )bs (T )
+ac (T )(dc,ν (T )+d2c (T )+b2c (T )dc (T )+b4c (T )+bc (T )bs (T )+b2s (T )))
due to the uniform boundedness of
 T
(1+T ) ν/2
e−σs (T −τ ) (1+τ )−ν/2 dτ
0
 T /2
≤ (1+T ) ν/2
e−σs T (1+τ )−ν/2 dτ
0
 T
+(1+T ) ν/2
e−σs (T −τ ) (1+T /2)−ν/2 dτ.
T /2

B2) Secondly, we estimate

 T 

(ν+1)/2   s Ps 
(1+T )  e (T −τ )L
N3 (w s )(τ ) dτ 
c , w  1
0 L
 T

≤(1+T )(ν+1)/2 e(T −τ )Ls Ps L1 →L1 N3 (w c , w
s )(τ )L1 dτ
0
 T
≤(1+T )(ν+1)/2 e−σs (T −τ ) (1+τ )−(ν+1)/2 dτ
0
× C(b3c (T )+bc (T )bs (T )+b2s (T )
+bc (T )(dc,ν +d2c (T )+b2c (T )dc (T )+b4c (T )+bc (T )bs (T )+b2s (T )))
≤C(b3c (T )+bc (T )bs (T )+b2s (T )
+bc (T )(dc,ν +d2c (T )+b2c (T )dc (T )+b4c (T )+bc (T )bs (T )+b2s (T )))
516 14. Diffusive stability

due to the uniform boundedness of

 T
(1+T )(ν+1)/2 e−σs (T −τ ) (1+τ )−(ν+1)/2 dτ
0
 T /2
≤(1+T )(ν+1)/2 e−σs T (1+τ )−(ν+1)/2 dτ
0
 T
+ (1+T ) (ν+1)/2
e−σs (T −τ ) (1+T /2)−(ν+1)/2 dτ.
T /2

C) We set
r(T ) = ac (T ) + bc (T ) + cc (T ) + dc (T ) + cc,ν (T ) + dc,ν (T ) + as (T ) + bs (T ).
Summing up all estimates yields an inequality
r(T ) ≤ r(0) + f (r(T ))
where f is at least quadratic in its argument. Comparing the curves r → r
and r → δ + r2 it is easy to see that r cannot go beyond 2δ if δ > 0 is
sufficiently small. Hence, if r(0) < δ, with δ > 0 sufficiently small, we have
the existence of a C > 0 such that r(T ) ≤ C for all T ≥ 0. 

We will apply the scheme of the proof a second time, namely for showing
the diffusive stability of the time-periodic solution A(X, T ) = e−iβt of the
complex GL equation (14.26). In order to do so we introduce B(X, T ) =
A(X, T )eiβt which satisfies
2
∂T B = (1 + iα)∂X B + (1 + iβ)B − (1 + iβ)B|B|2 .
This equation possesses the equilibrium B = 1 for which we would like to
prove diffusive stability. The deviation v = B − 1 satisfies
(14.34) ∂T v = Lv + N (v)
where
2
Lv =(1 + iα)∂X v + (1 + iβ)v − (1 + iβ)(2v + v),
N (v) = − (1 + iβ)(v 2 + 2vv + v 2 v).
The eigenvalue problem Lv = μv is solved by Fourier modes v(x, t) =
v1,2 (k)eikx with eigenvalues μ1,2 (k). In the Eckhaus stable domain we have
constants σc > 0 and σs > 0 such that
Reμ1 (k) ≤ −σc k 2 and Reμ2 (k) ≤ −σs
for all k ∈ R. System (14.34) has exactly the same properties as System
(14.27). Therefore, we can follow the proof of Theorem 14.2.1 starting at
(14.28) line for line. The formal calculations for the irrelevance of the nonlin-
ear terms can be found in the subsequent §14.4.2. With the same arguments
14.2. Diffusive stability of spatially periodic equilibria 517

as above it can be concluded that the kernels K2 and K3 satisfy (14.31)-

(14.32).
Theorem 14.2.5. The time-periodic solution A = e−iβt of (14.26) is diffu-
sively stable in the following sense. For all C > 0 there exists a δ > 0 such
that for solutions A = e−iβt + v of (14.22) with v |T =0 L1 + 
v |T =0 L∞ ≤ δ
we have
v (T )L∞ ≤ C,
 v (T )L1 ≤ C(1+T )−1/2

for all T ≥ 0.

Remark 14.2.6. Obviously, with the presented method the diffusive stabil-
ity of all Eckhaus-stable equilibria of (14.22) and of all Eckhaus-stable time-
periodic solutions of (14.26) can be established. We already explained that
whenever the L1 -L∞ approach works the renormalization approach works,
too. See [BK92] for a detailed proof for the real GL equation. The renor-
malization approach gives more informations than the L1 -L∞ approach, but
needs more localization of the initial conditions.
Theorem 14.2.7. There exists a δ > 0 such that for all p ∈ (0, 1/2) there
exists a C1 > 0 such that for all (φ0 , s0 ) ∈ H22 × H22 with φH22 + sH22 ≤ δ
there exists a unique global solution (φ, s) ∈ C([1, ∞), H22 × H22 ) of (14.24)
to initial conditions (φ0 , s0 ), and a φlim ∈ R such that
(14.35)
T 1/2 φ(T 1/2 ·, T )−φlim e−· H22 + T 1/2 s(T 1/2 ·, T )H22 ≤ C1 T −1/2+p ,
2 /4

and thus in particular

φlim 2
(14.36) φ(·, T ) − √ e−· /4T ∞ + s(·, T )∞ ≤ C1 T −(1−p) .
T
14.2.2. Diffusive stability of rolls in pattern forming systems. In
the last section we have seen the diffusive stability of A = 1 in the real and
of A = e−iβt in the complex GL equation. In Chapter 10 we proved that
the GL equation makes correct predictions on an O(1/ε2 )-time scale about
the dynamics of various pattern forming systems such as the SH equation or
the Couette-Taylor problem. However, since we only have an exponentially
growing error bound for this approximation the polynomial decay rates ob-
served for the GL equation are insufficient to conclude diffusive stability of
the spatially periodic pattern in the original system. Thus, a direct proof of
diffusive stability is necessary. It will be based on a linear stability analysis
via Bloch waves and the computation of the effective nonlinearity such that
finally the approach of §14.2.1 can be transferred.
We refrain from greatest generality and restrict ourselves to the SH equa-
tion (10.4). The subsequent analysis is by no means restricted to bifurcating
518 14. Diffusive stability

solutions or equilibria. It works for all spectrally stable spatially periodic

equilibria or traveling waves.
In order to keep the notation on a reasonable level we consider the
stability of the spatially periodic equilibrium
2
cos(x) + O(ε2 )
uper,ε [x] = ε
3
√ cf. §13.2.2. It corresponds to the diffusively stable
for the SH equation,
equilibrium A = 1/ 3 in the associated GL equation
2
∂T A = 4∂X A + A − 3A|A|2 ,
cf. §10.2. The linearization of (14.37) around a roll uper,ε yields a linear
operator with periodic coefficents. Thus, the Fourier transform has to be
replaced by the Bloch transform and the splitting into a diffusive part and
an exponentially damped remainder becomes more complicated. Finally,
it turns out that also the nonlinear terms are best studied in Bloch space
where the effective nonlinearity can be computed. The presentation is not
optimized w.r.t. to the small perturbation parameter 0 < ε  1. In fact, the
set of initial conditions for which the associated solutions converge towards
the equilibrium is of the same size as the bifurcating equilibrium w.r.t. 0 <
ε  1. See [Sch96a, Sch98d].
In the SH equation we take uper,ε as new origin, i.e., we set u = uper,ε + v
and obtain
(14.37) ∂t v = Lv + N (v)
with
Lv = −(1 + ∂x2 )2 v + ε2 v − 3u2per,ε v, N (v) = −3uper,ε v 2 − v 3 .

Linear diffusive behavior. In a first step we analyze the linearized prob-

lem ∂t v = Lv. In addition to the spectral properties we gain a useful rep-
resentation of the semigroup which allows us to analyze the asymptotic
behavior of the perturbations. The linearized problem is solved by Bloch
waves, cf. §11.6.3,
v(x, t) = eμt eix v(, x), with v(, x) = v(, x + 2π)
satisfying
(14.38) L v(, x) = μ
v (, x)
where
L v(, x) := −(1 + (i + ∂x )2 )2 v(, x) + ε2 v(, x) − 3uper,ε (x)2 v(, x).
For fixed  ∈ [−1/2, 1/2) the operator L is self-adjoint in L2 (T). Therefore,
all eigenvalues are real-valued. The operator L : H 4 (T) → L2 (T) has a
14.2. Diffusive stability of spatially periodic equilibria 519

compact resolvent (L − μI)−1 for some μ ∈ C. From this, for fixed , it
follows the existence of a discrete set of eigenvalues
{μj () ∈ C : j ∈ N, μj ≥ μj+1 → −∞ for j → ∞}
and a corresponding set of eigenfunctions {fj () : j ∈ N}. We normalize fj
such that fj (, ·)L2 (T) = 1.
For ε = 0 we have uper,ε = 0 and so the eigenvalue problem can be solved
explicitly. The well known curve λ(k) = −(1−k 2 )2 is cut into pieces of length
one. These pieces are plotted via k = m+, with m ∈ Z and  ∈ [−1/2, 1/2),
as functions w.r.t. , i.e., we consider the spatial homogenous situation
artificially as spatially periodic. Spectral perturbation theory for fixed , cf.
[Kat95] or Lemma 11.6.2, yields the smoothness of the curves of eigenvalues
with the intersection points as exception. Hence, it is easy to see that for
fixed 1 = O(1) > 0 for ε → 0 there exist ε0 > 0 and σ0 (1 ) = O(1) > 0
for ε → 0, such that for all ε ∈ (0, ε0 ) and all || > 1 all eigenvalues satisfy
μj () < −σ0 . Except of the two curves μ1 , μ2 touching zero this is also true
for all other eigenvalues, i.e., supj≥3 sup∈[−1/2,1/2) μj () < −σ0 .
In order to understand what happens with the two curves μ1 , μ2 for
ε > 0 we proceed as follows. The eigenvalue problem
v = L v − μ
G(, ε, μ) v=0
always has the trivial solution v = 0 for all , μ, and ε. Solutions can bifur-
cate from this trivial branch if (∂vG(, ε, 0))−1 does not exist. We know al-
ready that ∂w G(0, 0, 0) = L0 |ε=0 is not invertible and has a two-dimensional
kernel spanned by U1 = sin x and U2 = cos x. Thus, we apply the Lyapunov-
Schmidt reduction method to compute the bifurcating solutions. Let P be
the orthogonal projection on this kernel and let v = aU1 + bU2 + V with
(1−P )V = V. The solution of the hyperbolic part (1−P )G(, ε, μ, a, b, V) = 0
is denoted by V = V(, ε, μ, a, b). Inserting this into P G(, ε, μ, a, b, V) = 0
gives the bifurcation equation
 
a
G1 = 0,
b
where
   
ρ−μ −iδ O(2 + |μ|) O(|| + |μ|)
G1 = + O(ε ) 4
,
iδ ρ + c(ε) − μ O(|| + |μ|) O(2 + |μ|)
with
ρ = −42 − 4 , δ = −43 , c(ε) = −2ε2 + O(ε4 ).
Computing μ1,2 (, ε) in such a way that the determinant of G1 vanishes,
gives
520 14. Diffusive stability

Lemma 14.2.8. The curves of eigenvalues μ1,2 have the expansions

μ1 () = −c1 (ε)2 + O(4 ), μ2 () = −2ε2 − c1 (ε)2 + O(4 + ε4 ),
where c1 (ε) = 4 + O(ε2 ).

Irrelevance of the nonlinear terms. In order to prove the irrelevance of

the nonlinear terms we proceed as in §14.2.1 except that Fourier transform
is replaced by Bloch transform. Hence we consider
(14.39) v + N
∂t v = L  (
v ),
 = BLB −1 , and N
where v = Bv, L  (
v ) = B(N (B −1 v)). There exists an 1 > 0
such that for all || ≤ 1 the first curve of eigenvalues is separated from the
rest. For fixed  the operator L is self-adjoint in L2 (T) and therefore
v () = χ()

pc () v (, ·), f1 (, ·)L2 (T) f1 (, ·),

where χ() = 1 for || ≤ 1 /2 and χ() = 0 for || > 1 /2, allows to separate
the diffusive modes from the exponentially damped modes. Moreover, define
v = v − (
Ps () v)fc ().
pc ()
We write v = vc fc + vs and use the projections to separate (14.39) into
two parts, namely
 (
∂t vc = μc vc + pc N v ),  s vs + Ps N
∂t vs = LP  (
v ).
 s we have the estimates
For the semigroups generated by μc and LP
eμc t vc L1 ≤eμc t L∞ 
vc L1 ≤ 
vc L1 ,
eμc t vc L∞ ≤eμc t L∞ 
vc L∞ ≤ 
vc L∞ ,
vc L∞ ≤ Ct−1/2 
eμc t vc L1 ≤eμc t L1  vc L∞ ,

eLPs t vs L1 ≤Ce−σs t/2 
vs L1 ,

eLPs t vs L∞ ≤Ce−σs t/2 
vs L∞ ,
with a constant σs > 0, where we used the abbreviation
L1 = L1 ([−1/2, 1/2), H 1 (T)) and L∞ = L∞ ([−1/2, 1/2), H 1 (T)).
By the choice H 1 (T) the nonlinear terms are well-defined subsequently. Af-
ter showing the linear diffusive behavior we have to establish the irrelevance
of the nonlinearity w.r.t. to this diffusive behavior. With the same motiva-
tion as before we expand the nonlinear terms into
 (
pc N vc , vc ) + B2,2 (
v ) =B2,1 ( vc , vs ) + B3 (
vc , vc , vc ) + gc (
vc , vs ),
 (
Ps N vc , vc ) + gs (
v ) =B2,3 ( vc , vs )
14.2. Diffusive stability of spatially periodic equilibria 521

where the B2,j are symmetric bilinear maps, B3 symmetric trilinear maps,
and where gc and gs stand for the remaining terms. In order to prove the
irrelevance all these terms we make a change of coordinates. We set
(14.40) c ,
vc = w  −1 Es B2,3 (
vs = −L vc , vc ) + w
s
and obtain
c = μc w
∂t w c + N1 (w
c ) + N2 (w c , w
s ),
(14.41) 
s = Ls Ps w
∂t w s + N3 (wc , w
s ),
where
c ) =B2,1 (w
N1 ( w c , w
c ) + B2,2 (w  −1 Es B2,3 (w
c , −L c , w
c )) + B3 (w
c , w
c , w
c ),
c , w
N2 ( w s ) =gc (w  −1 Es B2,3 (w
c , −L c , w
c ) + w
s ) + B2,2 (w
c , w
s ),
c , w
N3 ( w s ) =gs (w  −1 Es B2,3 (
c , −L vc , vc ) + w  −1 Es B2,3 (w
s ) + ∂t (L c , w
c )).
By this transformation N2 decays like t−2 and N3 decays like t−3/2 such that
they are both irrelevant w.r.t. diffusive behavior.
In order to prove the irrelevance of N1 (w c ) we write it as

c ) = K2 (,  − 1 , 1 )
N1 ( w vc ( − 1 )
vc (1 ) d1
R
 
+ K3 (,  − 1 , 1 − 2 , 2 )
vc ( − 1 )
vc (1 − 2 )
vc (2 ) d2 d1 ,
R R
with smooth kernels K2 (,  − 1 , 1 ) and K3 (,  − 1 , 1 − 2 , 2 ). As be-
fore, the faster the kernels vanish near (k1 , . . . , kj ) = (0, . . . , 0), the more
irrelevant are the associated terms. We have
(14.42) K2 (0, 0, 0) = 0, K3 (0, 0, 0, 0) = 0
by considering the SH equation in the space of 2π-spatially periodic func-
tions. There exists a one-dimensional center manifold associated to family
of equilibria {uper,ε (· + x0 ) : x0 ∈ R}, cf. §13.2.2. The flow on the center
manifold is given by
(14.43) c (0) = μ1 (0)w
∂t w c (0) + β2 w
c (0)2 + β3 w
c (0)3 + . . . ,
with coefficients βj ∈ R. Since the flow on the center manifold is trivial,
we must have ∂t wc (0, t) = 0 and so β2 = β3 = 0. Since after the transform
(14.40) the right-hand side of (14.41) in the space of 2π-spatially periodic
functions for ws = 0 coincides with the right-hand side of (14.43) up to
terms of order O(|wc |4 ) we must have
β2 = K2 (0, 0, 0), β3 = K3 (0, 0, 0, 0)
and so (14.42). So it remains to prove that the first derivatives of K2 van-
ish at (0, 0, 0), too. This property can be computed explicitly for the SH
522 14. Diffusive stability

equation using that the curve of eigenfunctions f1 () has the following useful
property.
Lemma 14.2.9. The eigenfunction f1 can be expanded as f1 () = g0 +ig1 +
O(2 ), where g0 , g1 ∈ H 1 (T) are real-valued functions.

Proof. We have g0 = ∂x uper,ε /∂x uper,ε (x)L2 (0,π) . The operator L is of the
form L = A+iB +O(2 ), where A and B are real -independent operators.
The eigenvalues μ1 satisfy μ1 () = O(2 ). Let h1 ∈ L2 (T) also be real-valued.
Inserting the ansatz f1 () = g0 + ig1 + h1 + O(2 ) into L f1 () = μ1 ()f1 ()
gives for the terms of O() the condition iAg1 + Ah1 + iBg0 = 0. Thus,
h1 ∈ kerA, i.e., h1  g0 = β∂x uper,ε for a β ∈ R. Since we have normalized
f1 we have to choose h1 = 0. 
Using the above representation of the projection P1 () shows

K2 (,  − m, m) = 3 f1 (, x)f1 ( − m, x)f1 (m, x)
u0 (x) dx.
R
Using Lemma 14.2.9 shows that
 6
K2 (, −m, m) = 3 g0 (x)g0 (x)g0 (x)
u0 (x)
R
−ilg1 (x)g0 (x)g0 (x)
u0 (x)+g0 (x)i(−m)g1 (x)g0 (x)
u0 (x)
7
+g0 (x)g0 (x)img1 (x)u0 (x) + O(l2 +(−m)2 +m2 ) dx.

Note that u0 (x) is an even function, and so g0 (x) = ∂x uper /∂x uper L2 (0,2π)
is an odd function. Thus, the integral over the zero order terms goes over an
odd function and vanishes. Since additionally the first order terms cancel,
we have shown
|K2 (, −m, m)| ≤C|2 + (−m)2 + m2 |,
|K3 (,  − m, m − k, k)| ≤C(|| + | − m| + |m − k| + |k|).
Therefore, we have a system with exactly the same properties as for the GL
equation, and so we get
Theorem 14.2.10. The stationary solution u = uper,ε of (10.4) is diffu-
sively stable in the following sense. For all C > 0 there exists a δ > 0 such
that for solutions u = uper,ε + v of (14.22) with v |t=0 L1 + 
v |t=0 L∞ ≤ δ
we have
v (t)L∞ ≤ C,
 v (t)L1 ≤ C(1+t)−1/2

for all t ≥ 0.

Since the Bloch transform is continuous from L1 to L∞ and due to

H 1 (T) ⊂ L∞ (T) we have
14.3. The critical case 523

Corollary 14.2.11. For all C > 0 there exists a δ > 0 such that solutions
u = uper,ε + v of (14.22) with 
v |t=0 L1 + v|t=0 L1 ≤ δ satisfy
v(t)L∞ (R,R) ≤ C(1+t)−1/2
for all t ≥ 0.

Some remarks. We already explained that whenever the L1 -L∞ approach

works, then the renormalization approach also works. With the discrete
renormalization approach introduced in §14.1.6, in [Sch96a] the following
diffusive stability result has been shown.
Theorem 14.2.12. There exists an ε0 > 0 such that the following holds.
For all ε ∈ (0, ε0 ) there exists a C1 (ε) > 0, with C1 (ε) = O(1) as ε → 0, such
that for initial conditions uper,ε + v|t=0 of the SH equation (10.4), with v|t=0
sufficiently small in H22 , there exists an A∞ ∈ R and for all p ∈ (0, 1/2) a
C > 0 such that
2
1 − x
v(x, t) − √ A∞ e 4C1 (ε)t ∂x uper,ε [x]L∞ ≤ Ctp−1 for t → ∞.
t
The associated nonlinear diffusive stability result for spectrally stable
Taylor vortices UTV [q, φ, ε] can be found in [Sch98d].
In [SS99a] the authors consider the situation where the homogeneous
steady state at infinity for traveling pulses in reaction-diffusion systems be-
comes unstable and bifurcates to a spatially periodic Turing pattern. The
spectral situation for the bifurcating modulating pulse is as for the spatially
periodic equilibria except of some zero eigenvalue coming from the trans-
lational invariance of the original traveling wave. In [GSU04] it has been
shown that the previous diffusive dynamics persists under the perturbation
by the mode associated with the zero eigenvalue, and that therefore such
modulating traveling waves are nonlinearly stable w.r.t. spatially localized
perturbations.
The main difficulties for the two-dimensional case appear on the linear
level, see, e.g., [Mie97b, Mie97c]. Once linear diffusive stability is estab-
lished, the nonlinear problem becomes simpler in higher space dimensions
due to the faster decay rate t−1 , see, e.g., [Uec99].

14.3. The critical case

After the consideration of problems with nonlinearities which are irrelevant
w.r.t. diffusion we consider now problems with nonlinearities which are
critical w.r.t. diffusion. For systems with an effective +u3 nonlinearity it
takes an O(e−1/δ ) time scale before the +u3 wins against diffusion, with δ
the effective size of the initial condition. Systems with an effective u∂x u
524 14. Diffusive stability

nonlinearity show the same decay rates as diffusion but the limiting profile
is no longer a Gaussian.

14.3.1. Exponentially long transients in unstable Poiseuille flow.

There exists a number of ODEs and PDEs where an instability manifests
only after an exponentially long time in terms of the norm of the pertur-
bation. Famous examples are the theorem of Nekhorosov, cf. Theorem
4.4.8, the long time existence of solutions on a time interval of length of
order O(exp(1/δ)) for nonlinear wave equations [JK84] with initial data of
order O(δ), and the exponentially slow evolution of interfaces in bistable
nonlinear diffusion equations with δ being the order of the diffusion coef-
ficient [CP89, CP90]. For nonlinear diffusion problems with an effective
+u3 nonlinearity the same phenomenon occurs. Poiseuille flow at criticality
falls into this class. There is an exponentially long transient decay before
positive growth rates occur [SU03a].
The starting point of our explanations is again the nonlinear diffusion
equation
(14.44) ∂t u = ∂x2 u + up , u|t=0 = u0 , t ≥ 0, x ∈ R, u = u(x, t) ∈ R,
but now at the threshold of irrelevant nonlinearities, namely p = 3. The
result is as follows. Take an initial condition in L1 ∩ L∞ of norm less than δ
with δ > 0 sufficiently small. Then the associated solution will be less than
4δ for t ∈ [0, exp(q/δ 2 )] with a constant q > 0 independent of δ. This means
that in terms of the norm of the initial conditions it takes an exponentially
long time for the solution to leave a small neighborhood of the origin. This
instability is almost not observable since on this very long time interval
the solution decays with a rate O(t−1/2 ) in L∞ , i.e., for t = exp(q/δ 2 ) in
L∞ the solution has a norm of order O(exp(−q/δ 2 )) in L∞ . Before the
explosion takes place the solution becomes flatter and flatter until the mass,
respectively the L1 -norm, is sufficiently big to start the explosion. The
phenomenon is robust under adding higher order terms, i.e., it also holds for
∂t u = ∂x2 u + u3 + H(u, ∂x u, ∂x2 u),
where |H(u, ∂x u, ∂x2 u)| ≤ C(|u|4 + |∂x u|2 + |u||∂x2 u|), with a constant C, i.e.,
H is an irrelevant nonlinearity w.r.t. linear diffusion. Moreover, adding the
critical term u∂x u, will not affect the transient decay rates and the above
phenomenon, cf. §14.3.2, Thus, whenever the lowest order nonlinear terms
are critical in the above sense we can expect the above long time stability,
where after a time O(exp(1/δ γ )), for a γ > 0, an explosion may occur.
We now give the precise statement and the proof for the nonlinear dif-
fusion equation at criticality.
14.3. The critical case 525

Theorem 14.3.1. Consider (14.44) with p = 3. There exist q > 0 and

δ0 > 0 such that for all δ ∈ (0, δ0 ) the following holds. For an initial
condition u0 with u0 L1 + u0 L∞ ≤ δ the associated solution u = u(t) of
(14.44) with u(0) = u0 exists for all t ∈ [0, exp(q/δ 2 )] and satisfies
sup u(t)L1 ∩L∞ ≤ 4δ.
t∈[0,exp(q/δ 2 )]
 −(x−y)2 /(4t) u (y) dy of
Proof. We use the explicit solution u(x, t) = √1
2πt Re 0
t∂ 2
the linear diffusion equation, abbreviated as u(t) = e x u0 , and recall that
for p > q we obtain
u(t)Lp ≤ Ct−1/(2r) u0 Lq , where 1/p = 1/q − 1/r, 1 ≤ p ≤ q ≤ ∞,
for some constant C independent of time. Next we consider the variation of
constant formula
 t
2 2
u(t) = et∂x u0 + e(t−s)∂x u3 (s) ds
0
for (14.44). With u3  L∞ ≤ u3L∞ and u3L1 ≤ u2L∞ uL1 , the abbrevi-
ations
a(t) = sup u(s)L1 and b(t) = sup (1+s)1/2 u(s)L∞
0≤s≤t 0≤s≤t

and the estimate (14.5) we obtain similar to before

 t 

1/2 

(1+t)  e u (s) ds
(t−s)∂x2 3
 ∞
0 L
 t
2
≤ (1+t)1/2 e(t−s)∂x L1 →L∞ u3 (s)L1 ds
0
 t
≤ (1+t)1/2 (t−s)−1/2 (1+s)−1 ds · b(t)2 a(t) ≤ C1 ln(1 + t)b(t)2 a(t),
0
with a constant C1 independent of t. Furthermore, we have
 t   t
 
 e(t−s)∂x2 u3 (s) ds 2
  1 ≤ e(t−s)∂x L1 →L1 u3 (s)L1 ds
0 L 0
 t
≤ C2 (1+s)−1 ds · b(t)2 a(t) ≤ C2 ln(1 + t)b(t)2 a(t),
0
with a constant C2 independent of t such that finally
a(t) ≤ a(0) + b(0) + C1 ln(1 + t)b(t)2 a(t),
b(t) ≤ a(0) + b(0) + C2 ln(1 + t)b(t)2 a(t).
a(t) and b(t) by a(t) = δ
We introduce  a(t) and b(t) = δb(t) which satisfy
 a(0) + C1 δ 2 ln(1 + t)b(t)2
a(t) ≤ a(t),
a(0) + C2 δ 2 ln(1 + t)b(t)2 
b(t) ≤b(0) +  a(t).
526 14. Diffusive stability

a(0) + b(0) ≤ 1 we have 

If  a(t) + b(t) ≤ 4 as long as for instance
max{C1 , C2 }δ 2 ln(1 + t) ≤ 1/64,
i.e., for all t ∈ [0, exp(q/δ 2 )] for a positive constant q. 

We have additionally proved

Proposition 14.3.2. Under the assumptions of Theorem 14.3.1 we have a
constant C > 0 independent of δ ∈ (0, δ0 ) such that
u(·, t)L∞ ≤ Cδ(1 + t)−1/2
for all t ∈ [0, exp(q/δ 2 )], i.e.,
 exp(−q/(2δ 2 )),
u(·, exp(q/δ 2 ))L∞ ≤ Cδ
 > 0 independent of δ ∈ (0, δ0 ).
with another constant C

The analogue to Theorem 14.3.1 and Proposition 14.3.2 has been proved
in [SU03a] for Poiseuille flow. This makes the measurement of the critical
Reynolds number Rc a delicate experiment. Initial perturbations may be
amplified on a short transient time scale due to the non-normality of the
linearization around the laminar flow, see, e.g., [SH94a]. After the short
time transient growth the solutions seem to decay for a very long time, i.e.,
at criticality for an exponentially long time. The better the experiment is
performed the more and more this observation plays a role and the longer
it takes to observe the final growth of localized perturbations. For more
details see Figure 14.1.

14.3.2. The Burgers equation as a limit. For diffusion in 1D there

are three critical nonlinear monomials, namely −u3 which leads to diffusive
stability, +u3 which leads to instability and which we discussed in detail in
§14.3.1, and ∂x (u2 ) which is the topic of this section. In contrast to ±u3
which yields a logarithmic correction to the growth rate, ∂x (u2 ) will only
change the limit profile. In §7.4 we already considered the Burgers equation,
given by
1
∂t u = ∂x2 u − ∂x (u2 ), t ≥ 0, x ∈ R, u = u(x, t) ∈ R.
2
By the Cole-Hopf transformation we found that solutions u to spatially
localized initial conditions in the Burgers equation satisfy
√ √ 2z e−x /4
2

tu(x t, t) → − √ =: fz∗ (x),

4π 1 + zerf(x)
√
for t → ∞ with a rate O(1/ t), and thus the renormalized solutions con-
verge towards a non-Gaussian limit. In the following we explain why we see
14.3. The critical case 527

stable
Re λ
unstable

kc k

Rc R
rest of spectrum

Figure 14.1. The left panel shows a simplified schematic bifurcation

diagram for Poiseuille flow. A direct transition to turbulent flow with
large amplitudes is observed experimentally based on the fact that, in
contrast to the first bifurcation in the other hydrodynamical stability
problems which we considered so far, the basic laminar flow becomes
unstable at a critical Reynolds number Rc via a subcritical bifurcation.
The right panel shows the curve of critical eigenvalues and the rest of
the spectrum for Poiseuille flow at the threshold of instability. From
the form of k → Reλ1 (k) close to wave number kc , namely Reλ1 (k) =
−C1 (k − kc )2 + O((k − kc )3 ) for a constant C1 > 0, it is clear that
the linearized system exhibits some diffusive behavior. However, the
nonlinear terms are critical, but not irrelevant w.r.t. diffusion. This can
be seen easily by making a GL ansatz which shows that at criticality,
i.e., R = Rc , the complex-valued amplitude A(X, T ) satisfies in lowest
2
order a GL equation ∂T A = (C1 + iC2 )∂X A + γA|A|2 , but now with
Reγ > 0 due to the subcritical bifurcation.

the same asymptotic decay rates in a perturbed Burgers equation

1
(14.45) ∂t u = ∂x2 u − ∂x (u2 ) + H(u, ∂x u), u|t=0 = u0 ,
2
where H is a smooth function R2 → R only containing irrelevant nonlinear-
ities.
Lemma 14.3.3. For all C > 0 there exists a δ > 0 such that solutions u of
(14.45) with u0 L1 + u0 L∞ ≤ δ satisfy
u(t)L1 ≤ C and u(t)L∞ ≤ C(1+t)−1/2
for all t ≥ 0.

Proof. For notational simplicity we restrict ourselves to the example

H(u, ∂x u) = up with p > 3. We consider the variation of constant formula
  t
t∂x2 1 t (t−s)∂x2 2
(14.46) u(t) = e u0 − e 2
∂x (u (s)) ds + e(t−s)∂x up (s) ds
2 0 0
t 2
for (14.45). We follow Lemma 14.1.2 where estimates of 0 e(t−s)∂x up (s) ds
in terms of
a(t) = sup u(s)L1 , b(t) = sup (1+s)1/2 u(s)L∞
0≤s≤t 0≤s≤t
528 14. Diffusive stability

can be found. The second term on the right-hand side of (14.46) is estimated
as follows. We have
 t 

1/2 

(1+t)  e (t−s)∂x2
∂x (u (s)) ds
2
 ∞
0 L
 t/2
2
≤ (1+t)1/2 e(t−s)∂x ∂x L1 →L∞ u2 L1 ds
0
 t
2
+ (1+t)1/2 e(t−s)∂x ∂x L∞ →L∞ u2 L∞ ds
t/2
 t/2
≤ ((1+t)1/2 (t−s)−1 (1+s)−1/2 ds · b(t)a(t)
0
 t
+(1+t)1/2 (t−s)−1/2 (1+s)−(p−1)/2 ds) · b(t)2
t/2
 t/2
≤ ((1+t)1/2 (t−t/2)−1 (1+s)−1/2 ds · b(t)a(t)
0
 t
+(1+t) 1/2
(t−s)−1/2 (1+t/2)−1 ds) · b(t)2
t/2

≤ C1 (b(t)a(t) + b(t)2 ),
with a constant C1 independent of t, and
 
 t (t−s)∂ 2   t (t−s)∂ 2
 e x ∂ (u2 (s)) ds
 1 ≤ 0 e
x∂  1
x L →L1 u L1 ds
2
 0 x
t L
≤ 0 (t − s)−1/2 (1+s)−1/2 ds · b(t)a(t) ≤ C2 b(t)a(t),
with a constant C2 independent of t such that finally
a(t) ≤a(0) + b(0) + C2 (b(t)a(t) + b(t)p−1 a(t))
b(t) ≤a(0) + b(0) + C1 (b(t)a(t) + b(t)2 + b(t)p−1 a(t)).
If a(0) + b(0) < δ, with δ > 0 sufficiently small, we have the existence of a
C > 0 such that a(t), b(t) ≤ C for all t ≥ 0. 
Also here, whenever the L1 -L∞ approach works, the renormalization
approach works too. In [BKL94] the following diffusive stability result has
been shown with the discrete renormalization approach from §14.1.6.
Theorem 14.3.4. Let b ∈ (0, 1/2). Then there exist C, δ > 0 such that for
all initial conditions u0 of (14.45) with u0 H22 ≤ δ there is a z > −1 such
that the solution u of (14.45) with u|t=1 = u0 satisfies
√ √
 tu(x t, t) − fz∗ (x)H22 ≤ Ct−1/2+b ,
hence in particular
u(x, t) − t−1/2 fz∗ (t−1/2 x)L∞ ≤ Ct−1+b .
14.4. Phase diffusion equations 529

Remark 14.3.5. An interesting application of the above theory is the in-

clined film problem, which is concerned with the flow of a viscous liquid down
an inclined plane. It occurs in cooling processes, when wiping a blackboard,
in coating processes, such as the painting of a wall, and it also appears on
the windshield of a car on a rainy day. It is related to the other free surface
problem which we considered in this book, namely the water wave problem
which is studied in §12.2.1. However, for this problem viscous effects had
been neglected completely.
In detail, we consider a film of viscous liquid flowing down an inclined
plane with inclination angle θ and assume that for |x| → ∞ the film is flat,
i.e., h(x, t) → h0 for |x| → ∞. See Figure 14.2 for a two-dimensional sketch.
For such flat films there always exists the trivial so called Nusselt solution

y, v g

h(t, x)

x, u

Figure 14.2. The inclined film problem; a fluid of height y =

h(x, t) runs down a plate with inclination angle θ subject to con-
stant gravitational force g.

with a laminar flow and a flat surface. For sufficiently small Reynolds num-
ber this flow turns out to be stable with a self-similar decay as in Theorem
14.3.4. Using the discrete renormalization approach from §14.1.6, this has
been proved for the so-called integrated boundary layer (IBL) approximation
of the inclined film problem in [Uec03], for the free surface Navier-Stokes
problem in [Uec07], and for the IBL over wavy bottoms in [HSU12], which
adapts the Bloch wave analysis from §14.2.2 to quasilinear systems.

14.4. Phase diffusion equations

In the previous chapters we derived modulation equations, such as the GL
equation, the KdV equation and the NLS equation for the description of the
dynamics near a spatially homogeneous background state. If the background
state is spatially periodic new modulation equations occur. In dissipative
530 14. Diffusive stability

systems these are phase diffusion equations for the description of wave num-
ber modulations of a family of spatially periodic equilibria and depending
on the scaling the Burgers equation or a conservation law for the description
of wave number modulations of a family of traveling waves where the group
velocity depends in a non-trivial way on the wave number. For conservative
systems Whitham’s equations for the description of wave number and am-
plitude modulations of a family of traveling waves appear, cf. [Whi99]. We
explain the derivation of these equations for the real and complex GL equa-
tion. Moreover, we explain that due to the reconstruction of the solutions
from the wave number, approximation results can only be expected locally,
but not globally, in space. These equations give a feeling about the diffu-
sive dynamics which can be expected in more complicated pattern forming
systems.

14.4.1. The phase diffusion equation for the real GL equation. The
real GL equation
(14.47) 2
∂T A = ∂X A + A − A|A|2 ,
with X ∈ R, T ≥ 0, and A(X, T ) ∈ C possesses the two-dimensional family
of stationary solutions

(14.48) A = Aper [q, φ0 ](X) = 1 − q 2 eiqX+iφ0 ,
which is parameterized by φ0 , q ∈ R. In §8.3.6 we discussed the linearization
around these equilibria and found diffusive modes and exponentially damped
modes. In order to derive a nonlinear effective equation for the diffusive
modes alone, namely the so called phase diffusion equation, we introduce
polar coordinates
A(X, T ) = r(X, T )eiφ(X,T )
in the real GL equation and obtain
∂T r = ∂X2 r + r − (∂ φ)2 r − r 3 ,
X
(14.49) 2 2(∂X r)(∂X φ)
∂T φ = ∂X φ + r .
We introduce the deviation (s, φ) from the equilibrium (r, φ) = (1, 0) by
setting r = 1 + s and find
∂T s = ∂X 2 s − 2s − (∂ φ)2 − (∂ φ)2 s − 3s2 − s3 ,
X X
(14.50) 2 2(∂X s)(∂X φ)
∂T φ = ∂X φ + 1+s .
We replace the equation for the phase φ by an equation for the local wave
number ψ = ∂X φ and obtain
2 s − 2s − ψ 2 − ψ 2 s − 3s2 − s3 ,
∂T s = ∂X
 
(14.51) 2 (∂X s)ψ
∂T ψ = ∂X ψ + 2∂X 1+s .
14.4. Phase diffusion equations 531

To derive the phase-diffusion equation we make a long wave ansatz

ψ(X, T ) = ψ̌(δX, δ 2 T ) and s(X, T ) = δš(δX, δ 2 T ),

with 0 < δ  1 a small perturbation parameter. This yields

δ 2 ∂τ š = δ 2 ∂ξ2 š − 2š − ψ̌ 2 − ψ̌ 2 š − 3š2 − š3 ,
 
(14.52) 2 ψ̌∂ξ š
∂τ ψ̌ = ∂ξ ψ̌ + 2∂ξ 1 + š ,

with τ = δ 2 T and ξ = δX. After neglecting the terms of order O(δ 2 ) for
small ψ̌, the implicit function theorem yields that the first equation

(14.53) 0 = −2š − ψ̌ 2 − ψ̌ 2 š − 3š2 − š3

possesses a unique solution š = s∗ (ψ̌) with s∗ (ψ̌) = −ψ̌ 2 /2 + O( ψ̌ 3 ). Due

to the formula for the equilibria, we have in fact that 1 + s∗ = 1 − ψ̌ 2 .
Inserting this into the equation for ψ̌ leads to the phase diffusion equation
 
ψ̌∂ (s∗ (ψ̌))
ξ
(14.54) ∂τ ψ̌ = ∂ξ2 ψ̌ + 2∂ξ =: ∂ξ2 h(ψ̌),
1 + s∗ (ψ̌)

where h : R → R is smooth. Indeed, h is the solution to the equation

2q ∗ 1 − 3q 2
h (q) = 1 + s (q) = .
1 + s∗ (q) 1 − q2
This can be verified as follows. For a basic equilibrium r = rq , φ = qX + φ0 ,
q ∈ (−1, 1) the corresponding phase diffusion equation is given by

(14.55) ∂τ ψ̌ = ∂ξ2
h(ψ̌, q)

where h(ψ̌, q) = h(ψ̌ + q). For each q, let k → μ1,2 (k, q) denote the smooth
curves of eigenvalues corresponding to the  Fourier wave numbers k for the
linearization of (14.49) around (r, φ) = ( 1 − q 2 , qX). In particular, let

μ1 (k, q) = −k 2 − (1 − q 2 ) + (1 − q 2 )2 + 4q 2 k 2

denote the critical curve for which μ1 (0, q) = 0. Then the phase-diffusion
equation (14.54) must give at lowest order a linear diffusion equation with
diffusion coefficient

h (q) = −∂k2 μ1 (0, q)/2 = (1 − 3q 2 )/(1 − q 2 ).

An approximation result that the phase diffusion equation (14.54) makes

correct predictions about the dynamics of the original real GL equation can
be found in [MS04b, Theorem 3.1].
532 14. Diffusive stability

Theorem 14.4.1. Let 2 ≤ m ≤ n − 2. Then there exists a Cψ∗ > 0 such that
the following is true. Let ψ ∗ be a solution of the phase diffusion equation
(14.54) with
sup ψ ∗ (τ )Hul ∗
n ≤ C ,
ψ
τ  ∈[0,τ0 ]

and let s∗ be defined by the solution s∗ = s∗ (ψ ∗ ) of (14.53). Then there

exist δ0 > 0 and C2 > 0 such that for all δ ∈ (0, δ0 ) we have solutions (š, ψ̌)
of the (š, ψ̌)-system (14.52) such that
sup (š, ψ̌)(τ ) − (s∗ (ψ ∗ (τ )), ψ ∗ (τ ))Hul
m ×H m ≤ C2 δ .
ul
2
τ ∈[0,τ0 ]

Remark 14.4.2. Although the approximation result is uniformly in space

for the (s, ψ)-variables, in the original A variable it is only locally in space
due to the reconstruction of φ from ψ by spatial integration, see [MS04b,
Theorem 4.1]. Since this is a serious restriction we recall the arguments.
Our starting point is the relation
 X 
2
A(X, T ) = (1 + š(δX, δ T )) exp i ψ̌(δX , δ 2 T ) dX
0

which defines the solution A of the real GL equation (14.47) in terms of

solutions (š, ψ̌) of (14.52). These solutions are approximated by
 X 
∗ 2
Aapp (X, T ) = (1 + s (δX, δ T )) exp i ψ ∗ (δX , δ 2 T ) dX .
0

Then we obtain
|A(X, T ) − Aapp (X, T )|
  X 

≤(1 + š(δX, δ T )) exp i
2
ψ̌(δX , δ 2 T ) dX
0
 X 
∗ 
− (1 + s (δX, δ T )) exp i
2
ψ ∗ (δX , δ 2 T ) dX 
0
  X 

≤(1 + š(δX, δ T )) exp i
2
ψ̌(δX , δ 2 T ) dX
0
 X 

− (1 + š(δX, δ T )) exp i
2
ψ ∗ (δX , δ 2 T ) dX 
0
  X 

+ (1 + š(δX, δ 2 T )) exp i ψ ∗ (δX , δ 2 T ) dX
0
 X 

− (1 + s∗ (δX, δ 2 T )) exp i ψ ∗ (δX , δ 2 T ) dX 
0
14.4. Phase diffusion equations 533

  X 

≤|1 + š(δX, δ T ))| exp i
2
ψ̌(δX , δ 2 T ) dX
0
 X 

− exp i ψ ∗ (δX , δ 2 T ) dX 
0
∗
+ |s (δX, δ T )) − š(δX, δ 2 T ))|
2

 X 
 
≤C  |ψ̌(δX , δ 2 T ) − ψ ∗ (δX , δ 2 T )| dX  + Cδ 2
0
 X 
 
≤ Cδ 2 dX  + Cδ 2 ≤ Cδ 2 (1 + |X|)
0
using the approximation result of Theorem 14.4.1.
By adding higher order terms in (s∗ , ψ ∗ ), the approximation can be
improved to hold uniformly on space intervals with length of order O(δ −p )
with p ≥ 0 arbitrary.

14.4.2. The Burgers equation for the complex GL equation. We

derive the Burgers equation for the complex GL equation
(14.56) 2
∂T A = (1 + iα)∂X A + A − (1 + iβ)A|A|2 ,
where α, β ∈ R, X ∈ R, T ≥ 0, and A(X, T ) ∈ C. It possesses wave trains

A(X, T ) = 1 − k 2 ei(kX−ωnl (k)T )
where ωnl (k) = β + (α − β)k 2 . We introduce the amplitude and phase
deviations (s, φ) of the wave train to k = 0 via
A(X, T ) = (1 + s(X, T ))ei(−βt+φ(X,T ))
which satisfy
2
∂T s =∂X s − 2s − (∂X φ)2 − (∂X φ)2 s − 2α(∂X s)(∂X φ)
− α∂X
2
φ − α(∂X
2
φ)s − 3s2 − s3 ,
2 s
∂X 2(∂X s)(∂X φ)
2
∂T φ =∂X φ+α − α(∂X φ)2 + − 2βs − βs2 .
1+s 1+s
Next, we replace the equation for the phase φ by an equation for the local
wave number ψ = ∂X φ and obtain
2
∂T s =∂X s − 2s − ψ 2 − ψ 2 s − 2α(∂X s)ψ − α∂X ψ − α(∂X ψ)s − 3s2 − s3 ,
 2 
α∂X s 2(∂X s)ψ
2
∂T ψ =∂X ψ + ∂X − αψ +2
− 2βs − βs .
2
1+s 1+s
To derive the Burgers equation, we assume that the wave number varies
slowly in time and space and seek solutions of the form
s(X, T ) = δ 2 š(δX, δ 2 T ), ψ(X, T ) = δ ψ̌(δX, δ 2 T ; δ).
534 14. Diffusive stability

Substituting this ansatz into (14.56), we get

δ 2 ∂τ š =δ 2 ∂ξ2 š − 2š − ψ̌ 2 − δ 2 ψ̌ 2 š − 2δ 2 α(∂ξ š)ψ̌
(14.57) − α∂ξ ψ̌ − δ 2 α(∂ξ ψ̌)š − 3δ 2 š2 − δ 4 š3 ,
∂τ ψ̌ =∂ξ2 ψ̌ − ∂ξ (2βš + αψ̌ 2 )
 
α∂ξ2 š 2(∂ξ š)ψ̌
+ δ ∂ξ2
+ − βš ,
2
1 + δ 2 š 1 + δ 2 š
where ξ = δX and τ = δ 2 T . Neglecting terms of order O(δ 2 ) and higher
gives the equations
(14.58) 0 = − 2š − ψ̌ 2 − α∂ξ ψ̌,
∂τ ψ̌0 =∂ξ2 ψ̌ + ∂ξ (−αψ̌ 2 − 2βš).
Eliminating š via the first equation shows that ψ̌(ξ, τ ) has to satisfy in lowest
order the viscous Burgers equation
(14.59) ∂τ ψ̌ = (1 + αβ)∂ξ2 ψ̌ + (β − α)∂ξ (ψ̌ 2 ).

An approximation result that the Burgers equation (14.59) makes correct

predictions about the dynamics of the original complex GL equation can be
found in [DSSS09, Theorem 3.4]. As above the approximation result is
uniformly in space for the (s, ψ)-variables, but in the original A variable it
is only locally in space due to the reconstruction of φ from ψ by spatial
integration.
Remark 14.4.3. For the description of long wave perturbations of ampli-
tude O(1) a conservation law can be derived. In order to do so we make the
long wave ansatz
ψ = ψ̌(δX, δT ) and s = š(δX, δT ),
and obtain
δ∂τ š =δ 2 ∂ξ2 š − 2š − ψ̌ 2 − ψ̌ 2 š − 2δα(∂ξ š)ψ̌
− δα∂ξ ψ̌ − δα(∂ξ ψ̌)š − 3š2 − š3 ,
 
δ 2 α∂ξ2 š 2δ(∂ξ š)ψ̌
∂τ ψ̌ =δ∂ξ2 ψ̌ + ∂ξ (−2βš − αψ̌ 2 − βš2 ) + ∂ξ + ,
1 + š 1 + š
where τ = δT , ξ = δX. Neglecting terms of order O(δ) and higher gives
0 = − 2š − ψ̌ 2 − ψ̌ 2 š − 3š2 − š3 ,
∂τ ψ̌ =∂ξ (−2βš − αψ̌ 2 − βš2 ).
For small ψ̌ the first equation can be solved uniquely by the implicit function
theorem, so that there exists a smooth even function s∗ : R → R such that
14.5. Dispersive dynamics 535

š = s∗ (ψ̌). Inserting this into the second equation gives the conservation
law
(14.60) ∂τ ψ̌ = ∂ξ (−2βs∗ (ψ̌) − αψ̌ 2 − βs∗ (ψ̌)2 ) = ∂ξ h(ψ̌)
where h : R → R is smooth. Estimates that (14.60) makes correct pre-
dictions about the dynamics of the complex GL equation can be found in
[MS04a].

14.5. Dispersive dynamics

As explained in the introduction of this chapter, the method of polyno-
mial decay rates and asymptotically irrelevant nonlinearities plays an even
more prominent role in dispersive systems, cf. [Str89, Rac92, LP09].
Here we explain similarities and differences between diffusive and disper-
sive dynamics. We restrict to one particular example, namely the nonlinear
Klein-Gordon equation
(14.61) ∂t2 u = Δu − u − up ,
with x ∈ Rd , t ≥ 0, p ∈ N, and u(x, t) ∈ R, which we write as the first order
system
(14.62) ∂t U = iΛU + N (U ),
where
    
/= 0 iω(k) u  1 0
iΛU , N (U ) = − .
iω(k) 0 v iω(k) u∗p
We have the local existence and uniqueness of solutions in H s , if s > d/2,
since the skew symmetric Λ is the generator a unitary group (eitΛ )t∈R in
every H s , and since N (U ) is locally Lipschitz continuous from H s to H s ,
cf. §11.2. As before we are interested in the global existence and uniqueness
of solutions. In contrast to the previously considered diffusive systems for
general dispersive systems there is no L∞ -L∞ estimate for the semigroup
generated by the linearized system. Hence the scheme presented in §14.1.3
has to be modified. Instead of the two spaces L1 and L∞ we will work in
three spaces, namely H s2 = W s2 ,2 , W s1 ,1 , W s∞ ,∞ , with s1 , s2 , s∞ suitably
chosen below. As before the W s∞ ,∞ -norm will decay, the H s2 -norm stays
bounded, but for dispersive systems the W s1 ,1 -norm will grow for t → ∞.
For (14.61) the modulus of the possible group velocities is bounded from
above by 1 and so a L∞ -L∞ estimate can be established for (14.61). How-
ever, we will not use this estimate in order to present a method which works
for general dispersive systems, too.
Hence, for dispersive systems we consider
b(t) = sup (1 + τ )d/2 U (τ )W s∞ ,∞ and c(t) = sup U (τ )H s2 .
τ ∈(0,t) τ ∈(0,t)
536 14. Diffusive stability

For the unitary group eitΛ we obtain

eitΛ U0 H s2 ≤ U0 H s2
and
(1 + t)d/2 eitΛ U0 W s∞ ,∞

C(1 + t)d/2 eitΛ U0 H s2 ≤ CU0 H s2 , for t ∈ [0, 1],

≤
C(1 + t)d/2 t−d/2 U0 W s1 ,1 , for t ≥ 1,
where for t ∈ [0, 1] we used Sobolev’s embedding H s2 ⊂ W s∞ ,∞ for s2 ≥
s∞ + d/2 and for t ≥ 1 we used the L1 -L∞ estimate from [Sha85] which
holds for s1 > s∞ + d + 2.
As a consequence, we have to improve the previous estimates for the
nonlinear terms. Using the product rule shows that ∂xs (up ) creates terms of
the form
up−1 ∂xs u, up−2 (∂x u)∂xs−1 u, . . . , up−2 (∂xs/2 u)∂xs/2 u.
All these terms have in common that there is only one factor with more
than s/2 derivatives. As a consequence we have the estimates
p−1
up H s ≤CuW [s/2]+1,∞ uH ,
s

p−1
up H s ≤CuH [s/2]+1 uH ,
s

p−2
up W s,1 ≤CuW [s/2]+1,∞ uH [s/2]+1 uH .
s

For s∞ ≥ ' s22 ( + 1 we obtain for G(t, τ ) = ei(t−τ )Λ N (u(τ )) that

 t   t
 
 G(t, τ ) dτ  ≤ G(t, τ )H s2 dτ
 
0 H s2 0
 t  t
p−1
≤ N (u(τ ))H s2 dτ ≤ Cu(τ )W s∞ ,∞ u(τ )H s2 dτ
0 0
 t
− d2 (p−1)
≤ C(1 + τ ) dτ · b(t)p−1 c(t) ≤ C1 b(t)p−1 c(t),
0

if p > 1 + and for t ≥ 1 that

2
d,
 t
(1 + t) 
d/2
G(t, τ ) dτ W s∞ ,∞
0
 t/2  t
≤(1 + t) d/2
G(t, τ ) W s∞ ,∞ dτ + (1 + t) d/2
G(t, τ )W s∞ ,∞ dτ
0 t/2
 t/2
≤(1 + t) d/2
ei(t−τ )Λ W s1 ,1 →W s∞ ,∞ N (u(τ ))W s1 ,1 dτ
0
 t
+ (1 + t) d/2
ei(t−τ )Λ N (u(τ ))H s2 dτ,
t/2
14.5. Dispersive dynamics 537

 t/2
≤(1 + t) d/2
C(t − τ )−(d/2) u(τ )W
p−2
s∞ ,∞ u(τ )H s2 dτ
2
0
 t
p−1
s∞ ,∞ u(τ )H s dτ
d/2
+ (1 + τ ) Cu(τ )W
t/2
  − d
t/2
t 2
(1 + τ )− 2 (p−2) dτ b(t)p−2 c(t)2
d
≤(1 + t) d/2
C
0 2
 t   d
d/2 t − 2 (p−1)
+ (1 + t) C 1+ dτ b(t)p−1 c(t)
t/2 2
 
≤C2 b(t)p−2 c(t)2 + b(t)p−1 c(t) ,
where we used Sobolev’s embedding theorem for s2 > s∞ + d2 and for the
existence of the integrals we assume d2 (p − 2) > 1 and d2 (p − 1) ≥ d2 + 1. For
t ∈ (0, 1] we have
 t   1
 
(1+t)d/2  e i(t−τ )Λ
N (u(τ )) dτ 
 s ,∞ ≤ 2 d/2
C ei(t−τ )Λ N (u(τ ))H s2 dτ
0 W ∞ 0
 1
≤2d/2 C N (u(τ ))H s2 dτ ≤ C3 b(t)p−1 c(t)
0
Therefore we find
c(t) ≤c(0) + C1 b(t)p−1 c(t),
b(t) ≤C(c(0) + b(0)) + (C2 + C3 )(b(t)p−1 c(t) + b(t)p−2 c(t)).
Hence y(t) = b(t) + c(t) satisfies y(t) ≤ Cy(0) + Cy(t)p . It follows as above
that supt≥0 y(t) ≤ 2Cy(0).
Theorem 14.5.1. Consider (14.61) with p > 2 + d2 , s2 > s∞ + d2 , s1 >
s∞ + d + 2, and s∞ ≥ [ s22 ] + 1. Then there exists δ > 0 and C > 0 such that
the following holds. If
U0 H s2 + U0 W s∞ ,∞ + U0 W s1 ,1 ≤ δ,
then the associated solution U with U |t=0 = U0 exists globally and satisfies
u(t)W s∞ ,∞ ≤ C(1 + t)− 2 .
d
u(t)H s2 ≤ C and
Remark 14.5.2. Although (14.61) can be solved forward and backward in
time, Theorem 14.5.1 has a direction in time. The reason is the growth
of the W s1 ,1 -norm as time evolves. Since the W s1 ,1 -norm appears in the
assumptions of Theorem 14.5.1 the application of the Theorem cannot be
iterated.
Remark 14.5.3. The condition p > 1+2/d was used in the estimates for the
H s2 -norm. The stronger condition p > 2+2/d was only used in the estimates
for the W s∞ ,∞ -norm. By a normal form transform [Sha85] the nonlinearity
538 14. Diffusive stability

up can be transformed into a nonlinearity of order O(up+1 ). This transform

leads to a loss of regularity and cannot be inverted. However, these problems
can be avoided by considering (14.61) twice. The original equation is used to
estimate the H s2 -norm. In the second equation we perform the normal form
transform and use it to estimate the W s∞ ,∞ -norm. A loss of r derivatives is
not a problem as long as s2 > s∞ + r. After the transform only the weaker
condition p > 1+2/d is necessary. The facts that the normal form transform
has not be inverted and a loss of regularity can be accepted, also allows to
handle quasilinear systems, where for the H s2 -estimate an energy estimate
instead of the variation of constant formula is used. The fact that the loss
of regularity coming from the normal form transform is not a problem, is
different from §11.5.4.

Further Reading. For the GL equation, the nonlinear diffusive stability of

the spatially periodic equilibria has also been established in [Kap94a]. In
[EWW97] the continuous renormalization approach has been used to give
another proof of the stability of the bifurcating spatially periodic equilibria
for the Swift-Hohenberg (SH) problem. The nonlinear stability of front solu-
tions in the GL equation, i.e., of solutions which connect an Eckhaus-stable
solution for x → −∞ with the trivial solution for x → ∞ has been shown in
[EW94, Gal94, BK95]. The associated results for the modulating fronts
in the SH equation, respectively Couette-Taylor problem has been estab-
lished in [ES00, ES02]. Diffusive mixing results for pattern with different
phases, respectively wave numbers, for x → ±∞ has been established in
[BKL94, GM98, SSSU12, IS16]. A new suitable coordinate system for
diffusive stability results of spatially periodic solutions has been constructed
in [GS11]. The previous ideas have been transferred and extended to sys-
tems of conservation laws or systems with an additional conservation law in
a number of papers. See [JZ10, JZ11, JNRZ14] for an overview. In all
these papers about diffusive mixing of phases or wave numbers a more elab-
orate analysis is used to obtain more details about the effective nonlinear
terms than we obtained in §14.2.
The method of polynomial decay rates and asymptotically irrelevant
nonlinearities is also used in other situations, cf. [JGK93, Kap94b], es-
pecially, in dispersive systems it plays a very prominent role, cf. [Str89,
Rac92, LP09]. The method of space-time resonances, cf. [Sha10], is a
recent development which allows to control nonlinear terms by the linear
evolution even in case when other advanced tools such as the normal forms
method [Sha85] and the vector field method [CK90] fail.
14.5. Dispersive dynamics 539

Exercises
14.1. For u(x, t) = ct−1/2 e−x
2
/(4t)
and t → ∞, prove the estimates
−3/2
sup |∂t u| ∼ t , sup |∂x2 u| ∼ t−3/2 , and sup |up | ∼ t−p/2 .
x∈R x∈R x∈R

14.2. Compute the first four Hermite polynomials via the inverse Fourier transform
j = (ik)j e−k for j = 0, 1, 2, 3, cf. Exercise 7.19.
2
of u
14.3. Prove the estimates (14.11) and (14.12).
14.4. Transfer the Lyapunov approach from §14.1.4 to 2D and 3D spatial domains.
14.5. Consider ∂t u = −∂x4 u, with x ∈ R, t ≥ 0, u(x, t) ≥ 0. Make an ansatz
1 x
u(x, t) = t1/4 v( t1/4 ) for finding self-similar solutions. Show that v satifies
1 1
−v (4) + ξv  + v = 0.
4 4
Show that v(k) = e−k satisfies the Fourier transformed equation. Discuss the
4

 −k4 ikξ
smoothness
 and decay of the function G(ξ) = c R e e dk where c ∈ R is chosen
such that G(ξ)dξ = 1.
Bibliography

[AA88] H. Added and S. Added, Equations of Langmuir turbulence and nonlinear

Schrödinger equation: smoothness and approximation, J. Funct. Anal. 79
(1988), no. 1, 183–210. MR950090 (89h:35273)
[AA11] A. Ambrosetti and D. Arcoya, An introduction to nonlinear functional anal-
ysis and elliptic problems, Progress in Nonlinear Differential Equations and
their Applications, 82, Birkhäuser Boston Inc., Boston, MA, 2011.
[ABZ14] T. Alazard, N. Burq, and C. Zuily, On the Cauchy problem for gravity water
waves, Invent. Math. 198 (2014), no. 1, 71–163.
[Ace00] A. B. Aceves, Optical gap solitons: Past, present, and future; theory and
experiments, Chaos 10 (2000), no. 3, 584–589.
[ACF+ 99] M. S. Alber, R. Camassa, Y. N. Fedorov, D. D. Holm, and J. E. Marsden,
On billiard solutions of nonlinear PDEs, Phys. Lett. A 264 (1999), no. 2-3,
171–178.
[ACF+ 01] , The complex geometry of weak piecewise smooth solutions of inte-
grable nonlinear PDE’s of shallow water and Dym type, Comm. Math. Phys.
221 (2001), no. 1, 197–227.
[ACHM94] M. S. Alber, R. Camassa, D. D. Holm, and J. E. Marsden, The geometry
of peaked solitons and billiard solutions of a class of integrable PDEs, Lett.
Math. Phys. 32 (1994), no. 2, 137–151.
[AD15] T. Alazard and J.-M. Delort, Global solutions and asymptotic behavior for two
dimensional gravity water waves, Ann. Sci. Éc. Norm. Supér. (4) 48 (2015),
no. 5, 1149–1238.
[ADN59] S. Agmon, A. Douglis, and L. Nirenberg, Estimates near the boundary for
solutions of elliptic partial differential equations satisfying general boundary
conditions. I, Comm. Pure Appl. Math. 12 (1959), 623–727.
[ADN64] , Estimates near the boundary for solutions of elliptic partial differen-
tial equations satisfying general boundary conditions. II, Comm. Pure Appl.
Math. 17 (1964), 35–92.

541
542 Bibliography

[AGGP12] V. K. Andreev, Y. A. Gaponenko, O. N. Goncharova, and V. V. Pukhnachev,

Mathematical models of convection, De Gruyter Studies in Mathematical
Physics, vol. 5, De Gruyter, Berlin, 2012.
[Agr01] G. P. Agrawal, Solitons and nonlinear wave equations, Academic Press, 2001.
[Air45] G. B. Airy, Tides and waves, Enzyclopedia Metropolitana 5 (1845), 241–396.
[AJMR95] L. Arnold, C. Jones, K. Mischaikow, and G. Raugel, Dynamical systems, Lec-
ture Notes in Mathematics, vol. 1609, Springer-Verlag, Berlin, 1995, Lectures
given at the Second C.I.M.E. Session held in Montecatini Terme, June 13-22,
1994, Edited by R. Johnson.
[AK02] I. S. Aranson and L. Kramer, The world of the complex Ginzburg-Landau
equation, Rev. Modern Phys. 74 (2002), no. 1, 99–143.
[AKN06] V. I. Arnold, V. V. Kozlov, and A. I. Neishtadt, Mathematical aspects of clas-
sical and celestial mechanics, third ed., Encyclopaedia of Mathematical Sci-
ences, vol. 3, Springer-Verlag, Berlin, 2006, [Dynamical systems. III], Trans-
lated from the Russian original by E. Khukhro.
[AL08] B. Alvarez-Samaniego and D. Lannes, Large time existence for 3D water
waves and asymptotics, Invent. Math. 171 (2008), no. 3, 485–541.
[ALB+ 10] D. Avitabile, D. J. B. Lloyd, J. Burke, E. Knobloch, and B. Sandstede, To
snake or not to snake in the planar Swift-Hohenberg equation, SIAM J. Appl.
Dyn. Syst. 9 (2010), no. 3, 704–733.
[Alt16] H. W. Alt, Linear functional analysis. An application-oriented introduction,
Springer, 2016.
[AM05] D. M. Ambrose and N. Masmoudi, The zero surface tension limit of two-
dimensional water waves, Comm. Pure Appl. Math. 58 (2005), no. 10, 1287–
1315.
[Ama95] H. Amann, Linear and quasilinear parabolic problems. Vol. I. abstract linear
theory, Monographs in Mathematics, vol. 89, Birkhäuser Boston Inc., Boston,
MA, 1995.
[Arn78] V. I. Arnold, Mathematical methods of classical mechanics, Graduate Texts
in Mathematics. 60. Springer-Verlag, 1978.
[Arn88] V. I. Arnold, Geometrical methods in the theory of ordinary differential equa-
tions, second ed., Grundlehren der Mathematischen Wissenschaften [Funda-
mental Principles of Mathematical Sciences], vol. 250, Springer-Verlag, New
York, 1988, Translated from the Russian by Joseph Szücs.
[AS81] M. J. Ablowitz and H. Segur, Solitons and the inverse scattering transform,
SIAM, Philadelphia, Pa., 1981.
[Avi11] A. Avila, Dynamics of renormalization operators, Proceedings of the inter-
national congress of mathematicians (ICM 2010), Hyderabad, India, August
19–27, 2010. Vol. I: Plenary lectures and ceremonies, Hackensack, NJ: World
Scientific; New Delhi: Hindustan Book Agency, 2011, pp. 154–175.
[Bar96] G. I. Barenblatt, Scaling, self-similarity, and intermediate asymptotics, Cam-
bridge: Cambridge University Press, 1996.
[BB98] H. Brezis and F. Browder, Partial differential equations in the 20th century,
Advances in Mathematics 135 (1998), 76–144.
[BBH94] F. Bethuel, H. Brezis, and F. Hélein, Ginzburg-Landau vortices, Progress
in Nonlinear Differential Equations and their Applications, 13, Birkhäuser
Boston Inc., Boston, MA, 1994.
Bibliography 543

[BBTW15] N. Bellomo, A. Bellouquid, Y. Tao, and M. Winkler, Toward a mathematical

theory of Keller-Segel models of pattern formation in biological tissues, Math.
Models Methods Appl. Sci. 25 (2015), no. 9, 1663–1763.
[BCBLS11] C. Blank, M. Chirilus-Bruckner, V. Lescarret, and G. Schneider, Breather
solutions in periodic media, Comm. Math. Phys. 302 (2011), no. 3, 815–841.
[BCD+ 90] M. Bartuccelli, P. Constantin, C. R. Doering, J. D. Gibbon, and Magnus Gis-
selfält, On the possibility of soft and hard turbulence in the complex Ginzburg-
Landau equation, Physica D 44 (1990), no. 3, 421–444.
[BDS16] R. Bauer, W.-P. Düll, and G. Schneider, The KdV, Burgers, and Whitham
limit for a spatially periodic Boussinesq model, arXiv:1608.05632 (2016).
[Ben72] T. B. Benjamin, The stability of solitary waves, Proc. Roy. Soc. (London) Ser.
A 328 (1972), 153–183.
[BEX14] W. A. Brock, G. Engström, and A. Xepapadeas, Spatial climate-economic
models in the design of optimal climate policies across locations, European
Economic Review 69 (2014), 78–103.
[BF06] A. Babin and A. Figotin, Linear superposition in nonlinear wave dynamics,
Rev. Math. Phys. 18 (2006), no. 9, 971–1053.
[BFL+ 07] K. Busch, G. Von Freymann, S. Linden, S. F. Mingaleev, L. Tkeshelashvili,
and M. Wegener, Periodic nanostructures for photonics, Physics reports 444
(2007), no. 3, 101–202.
[BGT96] B. Buffoni, M. D. Groves, and J. F. Toland, A plethora of solitary gravity-
capillary water waves with nearly critical Bond and Froude numbers, Philos.
Trans. R. Soc. Lond., Ser. A 354 (1996), no. 1707, 575–607.
[BHP05] D. Blömker, M. Hairer, and G. A. Pavliotis, Modulation equations: Stochastic
bifurcation in large domains, Commun. Math. Phys. 258 (2005), no. 2, 479–
512.
[BK92] J. Bricmont and A. Kupiainen, Renormalization group and the Ginzburg-
Landau equation, Comm. Math. Phys. 150 (1992), 193–208.
[BK95] , Renormalizing partial differential equations, Constructive physics.
Results in field theory, statistical mechanics and condensed matter physics.
Proceedings of the conference, held at Ecole Polytechnique, Palaiseau, France,
25-27 July, 1994, Berlin: Springer-Verlag, 1995, pp. 83–115.
[BK00] J. Billingham and A. C. King, Wave motion, Cambridge University Press,
Cambridge, 2000.
[BKL94] J. Bricmont, A. Kupiainen, and G. Lin, Renormalization group and asymp-
totics of solutions of nonlinear parabolic equations, Comm. Pure Appl. Math.
6 (1994), 893–922.
[BKL+ 09] M. Beck, J. Knobloch, D. J. B. Lloyd, B. Sandstede, and T. Wagenknecht,
Snakes, ladders, and isolas of localized patterns., SIAM J. Math. Anal. 41
(2009), no. 3, 936–972.
[BM13] D. Blömker and W. W. Mohammed, Amplitude equations for SPDEs with
cubic nonlinearities, Stochastics 85 (2013), no. 2, 181–215.
[BMS01] D. Blömker, S. Maier-Paape, and G. Schneider, The stochastic Landau equa-
tion as an amplitude equation, Discrete Contin. Dyn. Syst., Ser. B 1 (2001),
no. 4, 527–541.
[BMW94] B. Birnir, H. P. McKean, and A. Weinstein, The rigidity of sine-Gordon
breathers, Commun. Pure Appl. Math. 47 (1994), no. 8, 1043–1051.
544 Bibliography

[BN85] N. Bekki and K. Nozaki, Formations of spatial patterns and holes in the
generalized Ginzburg-Landau equation, Physics Letters A 110 (1985), no. 3,
133–135.
[BNSZ14] M. Beck, T. T. Nguyen, B. Sandstede, and K. Zumbrun, Nonlinear stability
of source defects in the complex Ginzburg-Landau equation, Nonlinearity 27
(2014), no. 4, 739–786.
[Bou77] M. B. Boussinesq, Essai sur la théorie des eaux courantes, Inst. France (séries
2) 23 (1877), 1–680.
[Bou99] J. Bourgain, Global solutions of nonlinear Schrödinger equations, American
Mathematical Society, Providence, RI, 1999.
[BR13] D. Blömker and M. Romito, Local existence and uniqueness for a two-
dimensional surface growth equation with space-time white noise, Stochastic
Anal. Appl. 31 (2013), no. 6, 1049–1076.
[Bra83] M. Bramson, Convergence of solutions of the Kolmogorov equation to travel-
ling waves, Mem. Amer. Math. Soc. 44 (1983), no. 285.
[Bri13] Th. J. Bridges, A universal form for the emergence of the Korteweg-de Vries
equation, Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. 469 (2013),
no. 2153, 20120707, 15.
[BS07] J. Bitzer and G. Schneider, Approximation and attractivity properties of the
degenerated Ginzburg-Landau equation, J. Math. Anal. Appl. 331 (2007),
no. 2, 743–778.
[BSTU06] K. Busch, G. Schneider, L. Tkeshelashvili, and H. Uecker, Justification of the
Nonlinear Schrödinger equation in spatially periodic media, ZAMP 57 (2006),
1–35.
[BV90] A. V. Babin and M. I. Vishik, Attractors of partial differential equations in
an unbounded domain, Proceedings of the Royal Society of Edinburgh 116A
(1990), 221–243.
[BV92] , Attractors of evolution equations, North Holland, 1992.
[BX08] W. Brock and A. Xepapadeas, Diffusion-induced instability and pattern for-
mation in infinite horizon recursive optimal control, Journal of Economic
Dynamics and Control 32 (2008), no. 9, 2745–2787.
[BX10] , Pattern formation, spatial externalities and regulation in coupled
economic–ecological systems, Journal of Environmental Economics and Man-
agement 59 (2010), no. 2, 149–164.
[Car66] L. Carleson, On convergence and growth of partial sumas of Fourier series,
Acta Math. 116 (1966), 135–157.
[Car81] J. Carr, Applications of centre manifold theory, Applied Mathematical Sci-
ences, Vol. 35. New York Heidelberg Berlin: Springer Verlag. XII. 142 p. DM
32.00; $ 14.60 (1981)., 1981.
[CBCSU08] M. Chirilus-Bruckner, C. Chong, G. Schneider, and H. Uecker, Separation
of internal and interaction dynamics for NLS-described wave packets with
different carrier waves, J. Math. Anal. Appl. 347 (2008), no. 1, 304–314.
[CBS12] M. Chirilus-Bruckner and G. Schneider, Detection of standing pulses in peri-
odic media by pulse interaction, J. Differential Equations 253 (2012), no. 7,
2161–2190.
[CBSU07] M. Chirilus-Bruckner, G. Schneider, and H. Uecker, On the interaction of
NLS-described modulating pulses with different carrier waves, Math. Methods
Appl. Sci. 30 (2007), no. 15, 1965–1978.
Bibliography 545

[CCPS12] M. Chirilus-Bruckner, C. Chong, O. Prill, and G. Schneider, Rigorous de-

scription of macroscopic wave packets in infinite periodic chains of coupled
oscillators by modulation equations, Discrete Contin. Dyn. Syst., Ser. S 5
(2012), no. 5, 879–901.
[CdR15] G. M. Coclite and L. di Ruvo, Well-posedness results for the short pulse
equation, Z. Angew. Math. Phys. 66 (2015), no. 4, 1529–1557.
[CDS15] M. Chirilus-Bruckner, W.-P. Düll, and G. Schneider, NLS approximation of
time oscillatory long waves for equations with quasilinear quadratic terms,
Math. Nachr. 288 (2015), no. 2-3, 158–166.
[CE80] P. Collet and J.-P. Eckmann, Iterated maps on the interval as dynamical
systems, Progress in Physics, Birkhaeuser, 1980.
[CE86] P. Collet and J.-P Eckmann, The existence of dendritic fronts, Commun.
Math. Phys. 107 (1986), no. 1, 39–92.
[CE90] P. Collet and J.-P. Eckmann, The time dependent amplitude equation for the
Swift-Hohenberg problem, Comm. Math. Phys. 132 (1990), 139–152.
[CEE92] P. Collet, J.-P. Eckmann, and H. Epstein, Diffusive repair for the Ginzburg-
Landau equation, Helv Phys Acta 65 (1992), 56–92.
[CG09] M. Cross and H. Greenside, Pattern Formation and Dynamics in Nonequilib-
rium Systems, Cambridge University Press, 2009.
[CH82] S. N. Chow and J. K. Hale, Methods of bifurcation theory, Springer-Verlag,
New York, 1982.
[CH89] R. Courant and D. Hilbert, Methods of mathematical physics. Vol. II, Wiley
Classics Library, John Wiley & Sons Inc., New York, 1989.
[CH93a] R. Camassa and D. D. Holm, An integrable shallow water equation with peaked
solitons, Phys. Rev. Lett. 71 (1993), no. 11, 1661.
[CH93b] M. C. Cross and P. C. Hohenberg, Pattern formation outside equilibrium,
Rev. Mod. Phys. 65 (1993), 854–1190.
[Chi06] C. Chicone, Ordinary differential equations with applications, second ed.,
Texts in Applied Mathematics, vol. 34, Springer, New York, 2006.
[CI94] P. Chossat and G. Iooss, The Couette-Taylor problem, New York: Springer-
Verlag, 1994.
[CI75] N. Chafee and E. F. Infante, A bifurcation problem for a nonlinear partial
differential equation of parabolic type, Applicable Anal. 4 (1974/75), 17–37.
[CK90] D. Christodoulou and S. Klainerman, Asymptotic properties of linear field
equations in Minkowski space, Commun. Pure Appl. Math. 43 (1990), no. 2,
137–199.
[CK91] J. D. Crawford and E. Knobloch, Symmetry and symmetry-breaking bifurca-
tions in fluid dynamics, Annual review of fluid mechanics, Vol. 23, Annual
Reviews, Palo Alto, CA, 1991, pp. 341–387.
[CK99] T. K. Callahan and E. Knobloch, Pattern formation in three-dimensional
reaction-diffusion systems, Physica D 132 (1999), 339–362.
[CK01] , Long wavelength instabilities of three-dimensional patterns, Phys.
Rev. E 64 (2001), 036214.
[CK09] S. J. Chapman and G. Kozyreff, Exponential asymptotics of localised patterns
and snaking bifurcation diagrams, Physica D 238 (2009), no. 3, 319–354.
[CL00] P. Chossat and R. Lauterbach, Methods in equivariant bifurcations and dy-
namical systems, Advanced Series in Nonlinear Dynamics, vol. 15, World
Scientific Publishing Co., Inc., River Edge, NJ, 2000.
546 Bibliography

[CL09] A. Constantin and D. Lannes, The hydrodynamical relevance of the Camassa-

Holm and Degasperis-Procesi equations, Arch. Ration. Mech. Anal. 192
(2009), no. 1, 165–186.
[Cod61] E. A. Coddington, An introduction to ordinary differential equations,
Prentice-Hall Mathematics Series, Prentice-Hall Inc., Englewood Cliffs, N.J.,
1961.
[Con01] P. Constantin, Some open problems and research directions in the mathe-
matical study of fluid dynamics, Mathematics unlimited-2001 and beyond
(Engquist et al., ed.), Springer, 2001, pp. 353–360.
[Con11] A. Constantin, Nonlinear water waves with applications to wave-current inter-
actions and tsunamis., Philadelphia, PA: Society for Industrial and Applied
Mathematics (SIAM), 2011.
[CP89] J. Carr and R. Pego, Metastable patterns in solutions of ut = ε2 uxx − f (u),
Commun. Pure Appl. Math. 42 (1989), no. 5, 523–576.
[CP90] , Invariant manifolds for metastable patterns in ut = ε2 uxx − f (u),
Proc. R. Soc. Edinb., Sect. A 116 (1990), no. 1-2, 133–160.
[Cra85] W. Craig, An existence theory for water waves and the Boussinesq and
Korteweg-de Vries scaling limits, Comm. Partial Differential Equations 10
(1985), no. 8, 787–1003.
[Cri95] D. G. Crighton, Applications of KdV, Acta Appl. Math. 39 (1995), no. 1-3,
39–67.
[CS98] C. Cercignani and D. H. Sattinger, Scaling limits and models in physical
processes, DMV Seminar, vol. 28, Birkhäuser Verlag, Basel, 1998.
[CS13] C. Chong and G. Schneider, Numerical evidence for the validity of the NLS
approximation in systems with a quasilinear quadratic nonlinearity, ZAMM
Z. Angew. Math. Mech. 93 (2013), no. 9, 688–696.
[CSS92] W. Craig, C. Sulem, and P.-L. Sulem, Nonlinear modulation of gravity waves:
a rigorous approach, Nonlinearity 5 (1992), no. 2, 497–522.
[CW16] P. Cummings and C. E. Wayne, Modified energy functionals and the nls ap-
proximation, arXiv:1606.00028v1, 2016.
[CZ15] V. Chepyzhov and S. Zelik, Infinite energy solutions for dissipative Euler
equations in R2 , J. Math. Fluid Mech. 17 (2015), no. 3, 513–532.
[DD13] T. Dohnal and W. Dörfler, Coupled mode equation modeling for out-of-plane
gap solitons in 2D photonic crystals, Multiscale Model. Simul. 11 (2013),
no. 1, 162–191.
[Dea16] E. J. Doedel et al, AUTO: Continuation and bifurcation software for ordinary
differential equations, cmvl.cs.concordia.ca/auto/, 2016.
[Deb05] L. Debnath, Nonlinear partial differential equations for scientists and engi-
neers, second ed., Birkhäuser Boston Inc., Boston, MA, 2005.
[Den93] J. Denzler, Nonpersistence of breather families for the perturbed sine Gordon
equation, Commun. Math. Phys. 158 (1993), no. 2, 397–430.
[dES71] R. C. diPrima, W. Eckhaus, and L. A. Segel, Non-linear wave-number inter-
action in near-critical two-dimensional flows, Journal of Fluid Mechanics 49
(1971), 705–744.
[Dev89] R. L. Devaney, An introduction to chaotic dynamical systems, Addison-Wesley
Publishing Company, 1989.
Bibliography 547

[DFKM96] G. Dangelmayr, B. Fiedler, K. Kirchgässner, and A. Mielke, Dynamics of

nonlinear waves in dissipative systems: reduction, bifurcation and stability,
Pitman Research Notes in Mathematics Series, vol. 352, Longman, Harlow,
1996, With a contribution by G. Raugel.
[DG95] Ch. R. Doering and J. D. Gibbon, Applied analysis of the Navier-Stokes equa-
tions, Cambridge Univ. Press, 1995.
[DGH03] H. R. Dullin, G. A. Gottwald, and D. D. Holm, Camassa-Holm, Korteweg-
de Vries-5 and other asymptotically equivalent equations for shallow water
waves, Fluid Dynam. Res. 33 (2003), no. 1-2, 73–95.
[DH16] W.-P. Düll and M. Heß, Existence of long time solutions and validity of the
Nonlinear Schrödinger approximation for a quasilinear dispersive equation,
arXiv:1605.08704, 2016.
[DHSZ16] W.-P. Düll, A. Hermann, G. Schneider, and D. Zimmermann, Justification of
the 2D NLS equation for a fourth order nonlinear wave equation - quadratic
resonances do not matter much in case of analytic initial conditions, J. Math.
Anal. Appl. 436 (2016), no. 2, 847–867.
[DHV04] A. Doelman, G. Hek, and N. Valkhoff, Stabilization by slow diffusion in a real
Ginzburg-Landau system, J. Nonlinear Sci. 14 (2004), no. 3, 237–278.
[DIM89] Y. Demay, G. Iooss, and A. Mielke, Theory of steady Ginzburg-Landau equa-
tion in hydrodynamic stability problems, Eur. J. Mech. B 8 (1989), no. 3,
229–268.
[DJ89] P. G. Drazin and R. S. Johnson, Solitons: an introduction, Cambridge Uni-
versity Press, 1989.
[DK98] G. Dangelmayr and L. Kramer, Mathematical tools for pattern formation,
Evolution of spontaneous structures in dissipative continuous systems, Lec-
ture Notes in Phys. New Ser. m Monogr., vol. 55, Springer, Berlin, 1998,
pp. 1–85.
[DKSZ16] W.-P. Düll, Kourosh Sanei Kashani, G. Schneider, and D. Zimmermann,
Attractivity of the Ginzburg-Landau mode distribution for a pattern forming
system with marginally stable long modes, J. Differ. Equations 261 (2016),
no. 1, 319–339.
[DMCK11] A. D. Dean, P. C. Matthews, S. M. Cox, and J. R. King, Exponential asymp-
totics of homoclinic snaking., Nonlinearity 24 (2011), no. 12, 3323–3351.
[Doe07] E. J. Doedel, Lecture notes on numerical analysis of nonlinear equations,
Numerical continuation methods for dynamical systems, Springer, Dordrecht,
2007, pp. 1–49.
[DPS09] T. Dohnal, D. Pelinovsky, and G. Schneider, Coupled-mode equations and
gap solitons in a two-dimensional nonlinear elliptic problem with a separable
periodic potential, J. Nonlinear Sci. 19 (2009), no. 2, 95–131.
[DRUW14] T. Dohnal, J. Rademacher, H. Uecker, and D. Wetzel, pde2path 2.0, in ENOC
14, 2014.
[DS06] W.-P. Düll and G. Schneider, Justification of the nonlinear Schrödinger equa-
tion for a resonant Boussinesq model, Indiana Univ. Math. J. 55 (2006), no. 6,
1813–1834.
[DS09] , Validity of Whitham’s equations for the modulation of periodic trav-
eling waves in the NLS equation, J. Nonlinear Sci. 19 (2009), no. 5, 453–466.
[DSSS09] A. Doelman, B. Sandstede, A. Scheel, and G. Schneider, The dynamics of
modulated wave trains, Mem. Amer. Math. Soc. 199 (2009), no. 934.
548 Bibliography

[DSW16] W.-P. Düll, G. Schneider, and C. E. Wayne, Justification of the nonlinear

Schrödinger equation for the evolution of gravity driven 2D surface water
waves in a canal of finite depth, Arch. Ration. Mech. Anal. 220 (2016), no. 2,
543–602.
[DU09] T. Dohnal and H. Uecker, Coupled mode equations and gap solitons for the
2D Gross–Pitaevsky equation with a non-separable periodic potential, Phys.D
238 (2009), 860–879.
[DU16] , Bifurcation of Nonlinear Bloch waves from the spectrum in the non-
linear Gross-Pitaevskii equation, JNS, to appear, 2016.
[Dül12] W.-P. Düll, Validity of the Korteweg-de Vries approximation for the two-
dimensional water wave problem in the arc length formulation, Commun. Pure
Appl. Math. 65 (2012), no. 3, 381–429.
[Dül16] , Justification of the Nonlinear Schrödinger approximation for a quasi-
linear wave equation, arXiv:1602.08016, 2016.
[Duo01] J. Duoandikoetxea, Fourier analysis., Graduate Studies in Mathematics,
vol. 29, American Mathematical Society, Providence, RI, 2001.
[DV15] A. Doelman and F. Veerman, An explicit theory for pulses in two compo-
nent, singularly perturbed, reaction-diffusion equations, J. Dynam. Differen-
tial Equations 27 (2015), no. 3-4, 555–595.
[DvHX16] A. Doelman, P. van Heijster, and F. Xie, A geometric approach to stationary
defect solutions in one space dimension, SIAM J. Appl. Dyn. Syst. 15 (2016),
no. 2, 655–712.
[Eas73] M. S. P. Eastham, The spectral theory of periodic differential equations, Texts
in Mathematics (Edinburgh), Scottish Academic Press, Edinburgh; Hafner
Press, New York, 1973.
[Eck65] W. Eckhaus, Studies in non-linear stability theory, Springer Tracts in Nat.
Phil. Vol.6, 1965.
[Eck93] , The Ginzburg-Landau equation is an attractor, J. Nonlinear Science
3 (1993), 329–348.
[Ein05] A. Einstein, Über die von der molekularkinetischen Theorie der Wärme
geforderte Bewegung von in ruhenden Flüssigkeiten suspendierten Teilchen,
Ann. Phys. 322 (1905), 549–560.
[EINP03] N. Ercolani, R. Indik, A. C. Newell, and T. Passot, Global description of
patterns far from onset: a case study, Nonlinear dynamics. From lasers to
butterflies. Selected lectures from the 15th Canberra international physics
summer school, Canberra, Australia, 21 January – February 1, 2002, River
Edge, NJ: World Scientific, 2003, pp. 411–435.
[ER99] J.-P. Eckmann and J. Rougemont, Coarsening by Ginzburg-Landau dynamics,
Comm. Math. Phys. 199 (1999), 441–470.
[Erm02] B. Ermentrout, Simulating, analyzing, and animating dynamical systems. a
guide to xppaut for researchers and students, Software, Environments, and
Tools, vol. 14, SIAM, Philadelphia, PA, 2002.
[ES98] J. Escher and B. Scarpellini, On the asymptotics of solutions of parabolic
equations on unbounded domains, Nonlinear Anal., Theory Methods Appl.
33 (1998), no. 5, 483–507.
[ES00] J.-P. Eckmann and G. Schneider, Nonlinear stability of bifurcating front so-
lutions for the Taylor-Couette problem, ZAMM Z. Angew. Math. Mech. 80
Bibliography 549

(2000), no. 11-12, 745–753, Special issue on the occasion of the 125th anniver-
sary of the birth of Ludwig Prandtl.
[ES02] , Non-linear stability of modulated fronts for the Swift-Hohenberg equa-
tion, Comm. Math. Phys. 225 (2002), no. 2, 361–397.
[Eva98] L. C. Evans, Partial differential equations, Graduate Studies in Mathematics,
vol. 19, American Mathematical Society, Providence, RI, 1998.
[EvH81] W. Eckhaus and A. van Harten, The inverse scattering transformation and
the theory of solitons. An introduction, North-Holland Mathematics Studies,
vol. 50, North-Holland Publishing Co., Amsterdam, 1981.
[EW91] J.-P. Eckmann and C. E. Wayne, Propagating fronts and the center manifold
theorem, Comm. Math. Phys. 136 (1991), no. 2, 285–307.
[EW94] , The nonlinear stability of front solutions for parabolic partial differ-
ential equations, Comm. Math. Phys. 136 (1994), 285–307.
[EWW97] J.-P. Eckmann, C. E. Wayne, and P. Wittwer, Geometric stability analysis of
periodic solutions of the Swift-Hohenberg equation, Comm. Math. Phys. 190
(1997), 173–211.
[Fef06] C. L. Fefferman, Existence and smoothness of the Navier-Stokes equation,
The millennium prize problems, Clay Math. Inst., Cambridge, MA, 2006,
pp. 57–67.
[Fen79] N. Fenichel, Geometric singular perturbation theory for ordinary differential
equations, J. Differential Equations 31 (1979), no. 1, 53–98.
[FH54] K. O. Friedrichs and D. H. Hyers, The existence of solitary waves, Commun.
Pure Appl. Math. 7 (1954), 517–550.
[Fib15] G. Fibich, The nonlinear Schrödinger equation. singular solutions and optical
collapse, Applied Mathematical Sciences, vol. 192, Springer, Cham, 2015.
[Fis37] R. Fisher, The advance of advantageous genes, Ann. Eugenics 7 (1937), 355–
369.
[Fit69] R. FitzHugh, Mathematical models of excitation and propagation in nerve,
H.P. Schwan, ed., Biological Engineering, 1969, pp. 1–85.
[FL96] A. S. Fokas and Q. M. Liu, Asymptotic integrability of water waves, Phys.
Rev. Lett. 77 (1996), no. 12, 2347–2351.
[FLS64] R. P. Feynman, R. B. Leighton, and M. Sands, The Feynman lectures on
physics. Vol. 2: Mainly electromagnetism and matter, Addison-Wesley Pub-
lishing Co., Inc., Reading, Mass.-London, 1964.
[FMW92] W. E. Fitzgibbon, J. J. Morgan, and S. J. Waggoner, Weakly coupled semi-
linear parabolic evolution systems, Ann. Mat. Pura Appl. (4) 161 (1992),
213–229.
[Fow97] A. C. Fowler, Mathematical models in the applied sciences, Cambridge Uni-
versity Press, 1997.
[FPU55] E. Fermi, J. Pasta, and S. M. Ulam, Studies in nonlinear problems, Technical
Report Los Alamos Sci. Lab.., 1955.
[FRMT01] C. Foias, R. Rosa, O. Manley, and R. Temam, Navier-Stokes equations and
turbulence, Cambridge: Cambridge University Press, 2001.
[Gal94] Th. Gallay, Local stability of critical fronts in nonlinear parabolic partial dif-
ferential equations, Nonlinearity 7 (1994), no. 3, 741–764.
[Gal11] G. P. Galdi, An introduction to the mathematical theory of the Navier-Stokes
equations. Steady-state problems, second ed., Springer Monographs in Math-
ematics, Springer, New York, 2011.
550 Bibliography

[Gam01] T. W. Gamelin, Complex analysis, New York, NY: Springer, 2001.

+
[GCF 08] D. Grass, J. P. Caulkins, G. Feichtinger, G. Tragler, and D. A. Behrens, Op-
timal control of nonlinear processes: With applications in drugs, corruption,
and terror, Springer Verlag, 2008.
[Geo15] S. G. Georgiev, Theory of distributions, Springer, Cham, 2015.
[GGKM74] C. S. Gardner, J. M. Greene, M. D. Kruskal, and R. M. Miura, Korteweg-
deVries equation and generalization. VI. Methods for exact solution, Comm.
Pure Appl. Math. 27 (1974), 97–133.
[GH83] J. Guckenheimer and Ph. Holmes, Nonlinear oscillations, dynamical systems,
and bifurcations of vector fields, Applied Mathematical Sciences, 42, Springer,
New York, 1983.
[GH03] M. D. Groves and M. Haragus, A bifurcation theory for three-dimensional
oblique travelling gravity-capillary water waves, J. Nonlinear Sci. 13 (2003),
no. 4, 397–447.
[GHS02] M. D. Groves, M. Haragus, and S. M. Sun, A dimension-breaking phenomenon
in the theory of steady gravity-capillary water waves, R. Soc. Lond. Philos.
Trans. Ser. A Math. Phys. Eng. Sci. 360 (2002), no. 1799, 2189–2243, Recent
developments in the mathematical theory of water waves (Oberwolfach, 2001).
[GK14] B. Grébert and Th. Kappeler, The defocusing NLS equation and its normal
form, Zürich: European Mathematical Society (EMS), 2014.
[GL50] V. L. Ginzburg and L. D. Landau, On the theory of superconductivity, Zhurnal
Eksperimental’noi i Teoreticheskoi Fiziki 20 (1950), 1064.
[GL51] I. M. Gelf́and and B. M. Levitan, On the determination of a differential equa-
tion from its spectral function, Izv. Akad. Nauk SSSR, Ser. Mat. 15 (1951),
309–360 (Russian).
[Gle88] J. Gleick, Chaos - die Ordnung des Universums, Knaur, 1988.
[GM72] A. Gierer and H. Meinhardt, A theory of biological pattern formation, Kyber-
netik 12 (1972), 30–39.
[GM98] Th. Gallay and A. Mielke, Diffusive mixing of stable states in the Ginzburg-
Landau equation, Comm. Math. Phys. 199 (1998), 71–97.
[GMS12] P. Germain, N. Masmoudi, and J. Shatah, Global solutions for the gravity
water waves equation in dimension 3, Ann. Math. (2) 175 (2012), no. 2,
691–754.
[GP14] Yan Guo and Xueke Pu, KdV limit of the Euler-Poisson system, Arch. Ration.
Mech. Anal. 211 (2014), no. 2, 673–710.
[GPS17] S. Gilg, D. E. Pelinovsky, and G. Schneider, Validity of the NLS approxi-
mation for periodic quantum graphs, Nonl. Diff. Eq. and Appl. (2017), to
appear.
[GR12] M. Ghergu and V. D. Radulescu, Nonlinear PDEs, mathematical models in
biology, chemistry and population genetics, Springer Monographs in Mathe-
matics, Springer, Heidelberg, 2012.
[Gri96] P. Grindrod, The theory and applications of reaction-diffusion equations, pat-
terns and waves, second ed., The Clarendon Press, Oxford University Press,
New York, 1996.
[GS83] P. Gray and S. K. Scott, Autocatalytic reactions in the isothermal, continuous
stirred tank reactor, Chemical Engineering Science 38 (1983).
Bibliography 551

[GS85] M. Golubitsky and D. G. Schaeffer, Singularities and groups in bifurcation

theory. Vol. I, Applied Mathematical Sciences, vol. 51, Springer-Verlag, New
York, 1985.
[GS01a] Th. Gallay and G. Schneider, KP description of unidirectional long waves.
The model case, Proc. R. Soc. Edinb., Sect. A, Math. 131 (2001), no. 4,
885–898.
[GS01b] M. D. Groves and G. Schneider, Modulating pulse solutions for a class of
nonlinear wave equations, Commun. Math. Phys. 219 (2001), no. 3, 489–522.
[GS02] M. Golubitsky and I. Stewart, The symmetry perspective. From equilibrium
to chaos in phase space and physical space, Progress in Mathematics, vol. 200,
Birkhäuser Verlag, Basel, 2002.
[GS05] M. D. Groves and G. Schneider, Modulating pulse solutions for quasilinear
wave equations, J. Differ. Equations 219 (2005), no. 1, 221–258.
[GS08] , Modulating pulse solutions to quadratic quasilinear wave equations
over exponentially long length scales, Commun. Math. Phys. 278 (2008), no. 3,
567–625.
[GS11] Th. Gallay and A. Scheel, Diffusive stability of oscillations in reaction-
diffusion systems, Trans. Am. Math. Soc. 363 (2011), no. 5, 2571–2598.
[GSS88] M. Golubitsky, I. Stewart, and D. G. Schaeffer, Singularities and groups in
bifurcation theory. Vol. II, Applied Mathematical Sciences, vol. 69, Springer-
Verlag, New York, 1988.
[GSU04] Th. Gallay, G. Schneider, and H. Uecker, Stable transport of information near
essentially unstable localized structures, DCDS–B 4 (2004), no. 2, 349–390.
[GU17] D. Grass and H. Uecker, Optimal management and spatial patterns in a dis-
tributed shallow lake model, Electr. J. Differential Equations 2017 (2017),
no. 1, 1–21.
[Gut94] A. Guthmann, Einführung in die Himmelsmechanik und Ephemeridenrech-
nung, Spektrum–Hochschultaschenbuch, 1994.
[GW05] Th. Gallay and C. E. Wayne, Global stability of vortex solutions of the two-
dimensional Navier-Stokes equation, Comm. Math. Phys. 255 (2005), no. 1,
97–129.
[GW06a] , Long-time asymptotics of the Navier-Stokes equation in R2 and R3
[Plenary lecture presented at the 76th Annual GAMM Conference, Luxem-
bourg, 29 March–1 April 2005], ZAMM Z. Angew. Math. Mech. 86 (2006),
no. 4, 256–267.
[GW06b] , Three-dimensional stability of Burgers vortices: the low Reynolds
number case, Phys. D 213 (2006), no. 2, 164–180.
[Hal88] J. K. Hale, Asymptotic behavior of dissipative systems, Mathematical Surveys
and Monographs, vol. 25, American Mathematical Society, Providence, RI,
1988.
[Ham99] M. Hamer, Solitary killers, New Scientist 163 (1999), 18–19.
[Han13] D. Han-Kwan, From Vlasov-Poisson to Korteweg-de Vries and Zakharov-
Kuznetsov, Commun. Math. Phys. 324 (2013), no. 3, 961–993.
[HCS99] M. Haragus-Courcelle and G. Schneider, Bifurcating fronts for the Taylor
Couette problem in infinite cylinders, ZAMP 50 (1999), no. 1, 120–151.
[Hen81] D. Henry, Geometric Theory of Semilinear Parabolic Equations, Springer Lec-
ture Notes in Mathematics, Vol. 840, 1981.
552 Bibliography

[HI11] M. Haragus and G. Iooss, Local bifurcations, center manifolds, and nor-
mal forms in infinite-dimensional dynamical systems, Universitext. London:
Springer; Les Ulis: EDP Sciences. xi, 329 p. , 2011.
[HK91] J. K. Hale and H. Kocak, Dynamics and bifurcations, Springer-Verlag, 1991.
[HN01] F. Hamel and N. Nadirashvili, Travelling fronts and entire solutions of the
Fisher-KPP equation in RN , Arch. Ration. Mech. Anal. 157 (2001), no. 2,
91–163.
[HNM14] S. Hata, H. Nakao, and A. S. Mikhailov, Sufficient conditions for wave insta-
bility in three-component reaction-diffusion systems, Prog. Theor. Exp. Phys.
(2014), 013A01.
[HO98] Y. Hayase and T. Ohta, Sierpinski gasket in a reaction–diffusion system,
Phys. Rev. Let. 81 (1998), 1726.
[HO00] , Self-replicating pulses and Sierpinski gaskets in excitable media,
Phys. Rev. E 62 (2000), 5998–6003.
[Hör83] L. Hörmander, The analysis of linear partial differential operators. I. Distri-
bution theory and Fourier analysis, Grundlehren der Mathematischen Wis-
senschaften, vol. 256, Springer-Verlag, Berlin, 1983.
[Hor03] D. Horstmann, From 1970 until present: the Keller-Segel model in chemotaxis
and its consequences. I, Jahresber. Deutsch. Math.-Verein. 105 (2003), no. 3,
103–165.
[Hor04] , From 1970 until present: the Keller-Segel model in chemotaxis and
its consequences. II, Jahresber. Deutsch. Math.-Verein. 106 (2004), no. 2,
51–69.
[Hoy06] R. B. Hoyle, Pattern formation. An introduction to methods, Cambridge Uni-
versity Press, Cambridge, 2006.
[HPT00] J. K. Hale, L. A. Peletier, and W. C. Troy, Exact homoclinic and heteroclinic
solutions of the Gray-Scott model for autocatalysis, SIAM J. Appl. Math. 61
(2000), no. 1, 102–130.
[HR90] J. Hale and G. Raugel, Lower semicontinuity of the attractor for a singularly
perturbed hyperbolic equation, J. Dyn. Differ. Equations 2 (1990), no. 1, 19–67.
[HS12] M. Haragus and A. Scheel, Dislocations in an anisotropic Swift-Hohenberg
equation, Comm. Math. Phys. 315 (2012), no. 2, 311–335.
[HSD04] M. W. Hirsch, S. Smale, and R. L. Devaney, Differential equations, dynamical
systems, and an introduction to chaos, second ed., vol. 60, Elsevier/Academic
Press, Amsterdam, 2004.
[HSU12] T. Häcker, G. Schneider, and H. Uecker, Self-similar decay to the marginally
stable ground state in a model for film flow over inclined wavy bottoms, Electr.
J. Diff. Eq. 61 (2012), 1–51.
[HSZ11] T. Häcker, G. Schneider, and D. Zimmermann, Justification of the Ginzburg-
Landau approximation in case of marginally stable long waves, J. Nonlinear
Sci. 21 (2011), no. 1, 93–113.
[HT00] K.-H. Hoffmann and Qi Tang, Ginzburg-Landau Phase Transition The-
ory and Superconductivity, International Series of Numerical Mathematics,
Birkhäuser, 2000.
[HZ11] H. Hofer and E. Zehnder, Symplectic invariants and Hamiltonian dynamics,
Modern Birkhäuser Classics, Birkhäuser Verlag, Basel, 2011, Reprint of the
1994 edition.
Bibliography 553

[IA98] G. Iooss and M. Adelmeyer, Topics in bifurcation theory and applications. 2nd
ed., Advanced Series in Nonlinear Dynamics. 3. Singapore: World Scientific.
viii, 186 p., 1998.
[Igu01] T. Iguchi, Well-posedness of the initial value problem for capillary-gravity
waves, Funkcial. Ekvac. 44 (2001), no. 2, 219–241.
[IK90] G. Iooss and K. Kirchgaessner, Bifurcation d’ondes solitaires en présence
d’une faible tension superficielle. (Bifurcation of solitary waves, for a low
surface tension), C. R. Acad. Sci., Paris, Ser. I 311 (1990), 265–268 (French).
[IM91] G. Iooss and A. Mielke, Bifurcating time-periodic solutions of Navier-Stokes
equations in infinite cylinders, J. Nonlinear Sci. 1 (1991), no. 1, 107–146.
[Ioo95] G. Iooss, Capillary-gravity water-waves problem as a dynamical system, A.
Mielke (ed.) et al., Structure and dynamics of nonlinear waves in fluids. Pro-
ceedings of the IUTAM/ ISIMM symposium, Hannover, Germany, August
17-20, 1994. Singapore: World Scientific. Adv. Ser. Nonlinear Dyn. 7, 42-57
(1995), 1995.
[Ioo98] , Travelling water-waves, as a paradigm for bifurcations in reversible
infinite dimensional “dynamical” systems, Doc. Math., J. DMV (1998), 611–
622.
[IS16] S. Iyer and B. Sandstede, Mixing in Reaction-Diffusion Systems: Large Phase
Offsets, arXiv:1610.06527, 2016.
[Jam03] G. James, Centre manifold reduction for quasilinear discrete systems, J. Non-
linear Sci. 13 (2003), no. 1, 27–63.
[JGK93] C. K. R. T. Jones, R. Gardner, and T. Kapitula, Stability of travelling waves
for non-convex scalar viscous conservation laws, Commun. Pure Appl. Math.
46 (1993), no. 4, 505–526.
[JK84] F. John and S. Klainerman., Almost global existence to nonlinear wave equa-
tions in three space dimensions, Commun. Pure Appl. Math. 37 (1984), 443–
455.
[JNRZ14] M. A. Johnson, P. Noble, L. M. Rodrigues, and K. Zumbrun, Behavior of
periodic solutions of viscous conservation laws under localized and nonlocalized
perturbations, Invent. Math. 197 (2014), no. 1, 115–213.
[Joh91] F. John, Partial differential equations, Springer, 1991.
[Joh02] R. S. Johnson, Camassa-Holm, Korteweg-de Vries and related models for
water waves, J. Fluid Mech. 455 (2002), 63–82.
[Jon84] C. K. R. T. Jones, Stability of the travelling wave solution of the FitzHugh-
Nagumo system, Trans. Amer. Math. Soc. 286 (1984), no. 2, 431–469.
[JR94] B. Jakobsen and F. Rosendahl, The Sleipner Platform Accident, Structural
Engineering International 4 (1994), no. 3, 190–193.
[JZ10] M. A. Johnson and K. Zumbrun, Nonlinear stability of periodic traveling wave
solutions of systems of viscous conservation laws in the generic case, J. Differ.
Equations 249 (2010), no. 5, 1213–1240.
[JZ11] , Nonlinear stability of periodic traveling-wave solutions of viscous con-
servation laws in dimensions one and two, SIAM J. Appl. Dyn. Syst. 10
(2011), no. 1, 189–211.
[Kal88] L. A. Kalyakin, Asymptotic decay of a one-dimensional wavepacket in a non-
linear dispersive medium, Math. USSR Sbornik 60 (1988), no. 2, 457–483.
[Kal89] , Long wave asymptotics. Integrable equations as asymptotic limits of
non- linear systems, Russ. Math. Surv. 44 (1989), no. 1, 3–42.
554 Bibliography

[Kap94a] T. Kapitula, On the nonlinear stability of plane waves for the Ginzburg-
Landau equation, Commun. Pure Appl. Math. 47 (1994), no. 6, 831–841.
[Kap94b] , On the stability of travelling waves in weighted L∞ spaces, J. Differ.
Equations 112 (1994), no. 1, 179–215.
[Kat75] T. Kato, Quasi-linear equations of evolution, with applications to partial dif-
ferential equations, Spectral theory and differential equations (Proc. Sympos.,
Dundee, 1974; dedicated to Konrad Jörgens), Springer, Berlin, 1975, pp. 25–
70. Lecture Notes in Math., Vol. 448.
[Kat81] , The Cauchy problem for the Korteweg-de Vries equation, Nonlinear
partial differential equations and their applications, Res. Notes in Math.,
vol. 53, Pitman, 1981, pp. 293–307.
[Kat95] , Perturbation theory for linear operators. Reprint of the corr. print.
of the 2nd ed. 1980, reprint of the corr. print. of the 2nd ed. 1980 ed., Berlin:
Springer-Verlag, 1995.
[Kaw72] T. Kawahara, Oscillatory solitary waves in dispersive media, J. Phys. Soc.
Japan (1972), 260–264.
[KC13] G. Kozyreff and S. J. Chapman, Analytical results for Front Pinning between
an Hexagonal Pattern and a Uniform State in Pattern-Formation Systems,
Phys. Rev. Letters 111(5) (2013), 054501.
[KDL90] E. Knobloch and J. De Luca, Amplitude equations for travelling wave convec-
tion, Nonlinearity 3 (1990), no. 4, 975–980.
[KdV95] D. J. Korteweg and G. de Vries, On the change of form of long waves ad-
vancing in the rectangular canal and a new type of long stationary waves,
Phil.Magaz. 39 (1895), 422–443.
[Kee88] J. P. Keener, Principles of applied mathematics, Addison-Wesley, 1988.
[Kel77] H. B. Keller, Numerical solution of bifurcation and nonlinear eigenvalue prob-
lems, Application of bifurcation theory, Proc. adv. Semin., Madison/Wis.
1976, 359-384, 1977.
[Kel93] S. H. Kellert, In the wake of chaos. Unpredictable order in dynamical sys-
tems, Science and its Conceptual Foundations, University of Chicago Press,
Chicago, IL, 1993.
[KH97] A. Katok and B. Hasselblatt, Introduction to the modern theory of dynamical
systems, Cambridge University Press, 1997.
[Kir82] K. Kirchgaessner, Wave-solutions of reversible systems and applications, J.
Differ. Equations 45 (1982), 113–127.
[KN79] T. Kano and T. Nishida, Sur les ondes de surface de l’eau avec une justifi-
cation mathématique des équations des ondes en eau peu profonde, J. Math.
Kyoto Univ. 19 (1979), no. 2, 335–370.
[KN86] , A mathematical justification for Korteweg-de Vries equation and
Boussinesq equation of water surface waves, Osaka J. Math. 23 (1986), no. 2,
389–413.
[Kno92] E. Knobloch, Nonlocal amplitude equations, Pattern formation in complex
dissipative systems (Kitakyushu, 1991), World Sci. Publ., River Edge, NJ,
1992, pp. 263–274.
[Koc15] H. Koch, Nonlinear dispersive equations, H. Koch, D. Tartaru, M. Visan, eds,
Dispersive Equations and Nonlinear Waves, Birkhäuser, 2015.
[Kol27] A. Kolmogoroff, Sur la convergence des séries de fonctions orthogonales,
Math. Z. 26 (1927), no. 1, 432–441.
Bibliography 555

[KP03] T. Kappeler and J. Pöschel, KdV & KAM, Springer., 2003.

[KP13] T. Kapitula and K. Promislow, Spectral and dynamical stability of nonlinear
waves, Applied Mathematical Sciences, vol. 185, Springer, New York, 2013.
[KPP37] A. N. Kolmogorov, I. G. Petrovsky, and N. S. Piskunov, Etude de l’équation
de la diffusion avec croissance de la quantité de matière et son application
à un problème boilogique, Bulletin Université d’Etat a Moscou A 1 (1937),
1–26.
[KPV91] C. E. Kenig, G. Ponce, and L. Vega, Well-posedness of the initial value prob-
lem for the Korteweg-de Vries equation, J. Amer. Math. Soc. 4 (1991), no. 2,
323–347.
[Kre85] U. Krengel, Ergodic theorems, De Gruyter Studies in Mathematics 6. Berlin-
New York: Walter de Gruyter. VIII, 1985.
[KS05] M. Kunze and G. Schneider, Estimates for the KdV-limit of the Camassa-
Holm equation, Lett. Math. Phys. 72 (2005), no. 1, 17–26.
[KSM92] P. Kirrmann, G. Schneider, and A. Mielke, The validity of modulation equa-
tions for extended systems with cubic nonlinearities, Proc. Roy. Soc. Edin-
burgh Sect. A 122 (1992), no. 1-2, 85–91.
[Kuc93] P. Kuchment, Floquet theory for partial differential equations, Operator The-
ory: Advances and Applications. 60. Basel: Birkhäuser Verlag, 1993.
[Kue15] C. Kuehn, Multiple time scale dynamics, Applied Mathematical Sciences, vol.
191, Springer, Cham, 2015.
[Kur80] Y. Kuramoto, Instability and turbulence of wavefronts in reaction-diffusion
systems, Prog. Theor. Phys. 63 (1980), no. 6, 1885–1903.
[Kur84] , Chemical oscillations, waves, and turbulence, Springer Series in Syn-
ergetics, Vol. 19, Berlin etc.: Springer-Verlag. VIII, 156 p, 1984.
[Kuz04] Y. A. Kuznetsov, Elements of applied bifurcation theory. 3rd ed., Springer,
2004.
[KY76] Y. Kuramoto and T. Yamada, Pattern formation in oscillatory chemical re-
actions, Prog. Theoret. Phys. 56 (1976), no. 3, 724–740.
[Lan05] D. Lannes, Well-posedness of the water-waves equations, J. Amer. Math. Soc.
18 (2005), no. 3, 605–654.
[Lan13] , The water waves problem. Mathematical analysis and asymptotics,
Providence, RI: American Mathematical Society (AMS), 2013.
[Lap10] M. Lappa, Thermal convection: patterns, evolution and stability, John Wiley
& Sons, Ltd., Chichester, 2010.
[Law90] I. D. Lawrie, A unified grand tour of theoretical physics, Adam Hilger Ltd.,
Bristol, 1990.
[Lax68] P. D. Lax, Integrals of nonlinear equations of evolution and solitary waves,
Commun. Pure Appl. Math. 21 (1968), 467–490.
[LBCB+ 09] V. Lescarret, C. Blank, M. Chirilus-Bruckner, Chr. Chong, and G. Schneider,
Standing generalized modulating pulse solutions for a nonlinear wave equation
in periodic media, Nonlinearity 22 (2009), no. 8, 1869–1898.
[LE91] I. Lengyel and I. R. Epstein, Modeling of Turing structures in the chlorite-
iodide-malonic acid-starch reaction system, Science 251 (1991), 650–652.
[Lem16] P. G. Lemarié-Rieusset, The Navier-Stokes problem in the 21st century, Boca
Raton, FL: CRC Press, 2016.
556 Bibliography

[Ler34] J. Leray, Sur le mouvement d’un liquide visqueux emplissant l’espace, Acta
Math. 63 (1934), no. 1, 193–248.
[LeV92] R. J. LeVeque, Numerical methods for conservation laws, second ed.,
Birkhäuser Verlag, Basel, 1992.
[LL91] L. D. Landau and E. M. Lifschitz, Lehrbuch der theoretischen Physik
(“Landau-Lifschitz”). Band VI, fifth ed., Akademie-Verlag, Berlin, 1991, Hy-
drodynamik. [Hydrodynamics], Translated from the Russian by Wolfgang
Weller and Adolf Kühnel, Translation edited by Weller and with a foreword
by Weller and P. Ziesche.
[LLS13] D. Lannes, F. Linares, and J.-C. Saut, The Cauchy problem for the Euler-
Poisson system and derivation of the Zakharov-Kuznetsov equation, Studies
in phase space analysis with applications to PDEs. In part selected papers
based on the presentations at a meeting, Bertinoro, Italy, September 2011,
New York, NY: Birkhäuser/Springer, 2013, pp. 181–213.
[Log15a] J. D. Logan, Applied partial differential equations, Undergraduate Texts in
Mathematics, Springer, Cham, 2015, Third edition.
[Log15b] , A first course in differential equations, Undergraduate Texts in Math-
ematics, Springer, Cham, 2015, Third edition.
[Lor63] E. N. Lorenz, Deterministic non-periodic flow, J. Atmos. Sci. (1963).
[LP09] F. Linares and G. Ponce, Introduction to nonlinear dispersive equations, New
York, NY: Springer, 2009.
[LSAC08] D. J. B. Lloyd, B. Sandstede, D. Avitabile, and A. R. Champneys, Localized
hexagon patterns of the planar Swift-Hohenberg equation, SIAM J. Appl. Dyn.
Syst. 7 (2008), no. 3, 1049–1100.
[Lun95] A. Lunardi, Analytic semigroups and optimal regularity in parabolic prob-
lems, Progress in Nonlinear Differential Equations and their Applications,
16, Birkhäuser Verlag, Basel, 1995.
[Lyn04] S. Lynch, Dynamical systems with applications using MATLAB, Birkhäuser,
2004.
[Man91] B. B. Mandelbrot, Die fraktale Geometrie der Natur, Birkhaeuser, 1991.
[Man92] P. Manneville, Dissipative structures and weak turbulence, Academic Press,
1992.
[Mar55] V. A. Marchenko, On the reconstruction of the potential energy from phases of
the scattered waves, Dokl. Akad. Nauk SSSR 104 (1955), 695–698 (Russian).
[Mar07] P. A. Markowich, Applied partial differential equations: a visual approach,
Springer, Berlin, 2007.
[Maz11] Vl. Maz’ya, Sobolev spaces with applications to elliptic partial differen-
tial equations, Grundlehren der Mathematischen Wissenschaften, vol. 342,
Springer, Heidelberg, 2011.
[MB02] A. J. Majda and A. L. Bertozzi, Vorticity and incompressible flow, Cam-
bridge Texts in Applied Mathematics, vol. 27, Cambridge University Press,
Cambridge, 2002.
[MBK14] W. W. Mohammed, D. Blömker, and K. Klepel, Multi-scale analysis of SPDEs
with degenerate additive noise, J. Evol. Equ. 14 (2014), no. 1, 273–298.
[Mel98] I. Melbourne, Derivation of the time-dependent Ginzburg-Landau equation on
the line, J. Nonlinear Sci. 8 (1998), no. 1, 1–15.
Bibliography 557

[Mel00] , Ginzburg-Landau theory and symmetry, Nonlinear instability, chaos

and turbulence, Vol. II, Adv. Fluid Mech., vol. 25, WIT Press, Southampton,
2000, pp. 79–109.
[Mer15] E. Meron, Nonlinear physics of ecosystems, CRC Press, 2015.
[MH92] K. R. Meyer and G. R. Hall, Introduction to hamiltonian dynamical systems
and the n-body problem, Springer-Verlag, 1992.
[Mie88] A. Mielke, Reduction of quasilinear elliptic equations in cylindrical domains
with applications, Math. Methods Appl. Sci. 10 (1988), no. 1, 51–66.
[Mie92] , Reduction of PDEs on domains with several unbounded directions: a
first step towards modulation equations, Z. Angew. Math. Phys. 43 (1992),
no. 3, 449–470.
[Mie97a] , The complex Ginzburg-Landau equation on large and unbounded do-
mains: sharper bounds and attractors, Nonlinearity 10 (1997), no. 1, 199–222.
[Mie97b] , Instability and stability of rolls in the Swift–Hohenberg equation,
Comm. Math. Phys. 189 (1997), 829–853.
[Mie97c] , Mathematical analysis of sideband instabilities with application to
Rayleigh–Bénard convection, J. Nonlinear Science 7 (1997), 57–99.
[Mie98] , Bounds for the solutions of the complex Ginzburg-Landau equation in
terms of the dispersion parameters, Physica D 117 (1998), no. 1-4, 106–116.
[Mie02] , The Ginzburg-Landau equation in its role as a modulation equation,
Handbook of dynamical systems, Vol. 2, North-Holland, 2002, pp. 759–834.
[Mil06] P. D. Miller, Applied asymptotic analysis, Graduate Studies in Mathematics,
vol. 75, American Mathematical Society, Providence, RI, 2006.
[Miu68] R. M. Miura, Korteweg-de Vries equation and generalizations. I. A remark-
able explicit nonlinear transformation, J. Mathematical Phys. 9 (1968), 1202–
1204.
[MK15] J. Müller and Chr. Kuttler, Methods and models in mathematical biology. De-
terministic and stochastic approaches, Lecture Notes on Mathematical Mod-
elling in the Life Sciences, Springer, Heidelberg, 2015.
[MM87] R. S. MacKay and J. D. Meiss (eds.), Hamiltonian dynamical systems. A
reprint selection, IOP Publishing Ltd. Bristol, 1987.
[MN05] N. Masmoudi and K. Nakanishi, From the Klein-Gordon-Zakharov system to
the nonlinear Schrödinger equation, J. Hyperbolic Differ. Equ. 2 (2005), no. 4,
975–1008.
[MN13] , Multifrequency NLS scaling for a model equation of gravity-capillary
waves, Commun. Pure Appl. Math. 66 (2013), no. 8, 1202–1240.
[MS93] J. H. Maddocks and R. L. Sachs, On the stability of KdV multi-solitons,
Commun. Pure Appl. Math. 46 (1993), no. 6, 867–901.
[MS95] A. Mielke and G. Schneider, Attractors for modulation equations on un-
bounded domains - existence and comparison, Nonlinearity 8 (1995), no. 5,
743–768.
[MS04a] I. Melbourne and G. Schneider, Phase dynamics in the complex Ginzburg-
Landau equation, J. Differ. Equations 199 (2004), no. 1, 22–46.
[MS04b] , Phase dynamics in the real Ginzburg-Landau equation, Math. Nachr.
263-264 (2004), 171–180.
[MSU01] A. Mielke, G. Schneider, and H. Uecker, Stability and diffusive dynamics
on extended domains, Ergodic theory, analysis, and efficient simulation of
dynamical systems, Berlin: Springer, 2001, pp. 563–583.
558 Bibliography

[MU16] K. Matthies and H. Uecker, Low regularity justification results for envelope ap-
proximations of nonlinear wave packets in periodic media, Asymptotic Anal.
99 (2016), no. 1-2, 53–65.
[Mur89] J. D. Murray, Mathematical Biology, Springer, Berlin, 1989.
[MY60] L. Markus and H. Yamabe, Global stability criteria for differential systems,
Osaka Math. J. 12 (1960), 305–317.
[Nal74] V. I. Nalimov, The Cauchy-Poisson problem, Dinamika Splošn. Sredy 254
(1974), no. Vyp. 18 Dinamika Zidkost. so Svobod. Granicami, 104–210.
[New85] A. C. Newell, Solitons in mathematics and physics, vol. 48, CBMS-NSF Reg.
Conf. Ser. Appl. Math., 1985.
[New16] , Short pulse evolution equation, Laser filamentation, CRM Ser. Math.
Phys., Springer, Cham, 2016, pp. 1–17.
[Nis02] Y. Nishiura, Far-from-equilibrium dynamics, Translations of Mathematical
Monographs, vol. 209, American Mathematical Society, Providence, RI, 2002.
[Nti76] A. A. Ntinos, Lengths of instability intervals of second order periodic differ-
ential equations, Quart. J. Math. Oxford (2) 27 (1976), no. 107, 387–394.
[NTYU07] Y. Nishiura, T. Teramoto, X. Yuan, and Kei-Ichi Ueda, Dynamics of traveling
pulses in heterogeneous media, Chaos 17 (2007), no. 3, 037104, 21.
[NW69] A. C. Newell and J. A. Whitehead, Finite bandwidth, finite amplitude con-
vection, J. Fluid Mech. 38 (1969), 279–303.
[OD08] I. Oprea and G. Dangelmayr, Dynamics and bifurcations in the weak elec-
trolyte model for electroconvection of nematic liquid crystals: a Ginzburg-
Landau approach, Eur. J. Mech., B, Fluids 27 (2008), no. 6, 726–749.
[Olv14] P. J. Olver, Introduction to partial differential equations, Undergraduate Texts
in Mathematics, Springer, Cham, 2014.
[Osb10] A. R. Osborne, Nonlinear ocean waves and the inverse scattering transform,
International Geophysics Series 97. Amsterdam: Elsevier/Academic Press.
xxvi, 2010.
[Paz83] A. Pazy, Semigroups of linear operators and applications to partial differen-
tial equations, Applied Mathematical Sciences, vol. 44, Springer-Verlag, New
York, 1983.
[Pel11] D. E. Pelinovsky, Localization in periodic potentials. From Schrödinger oper-
ators to the Gross-Pitaevskii equation, London Mathematical Society Lecture
Note Series 390. Cambridge: Cambridge University Press. x, 398 p., 2011.
[Pis06] L. M. Pismen, Patterns and interfaces in dissipative dynamics, Springer Series
in Synergetics, Springer-Verlag, Berlin, 2006.
[PJS92] H.-O. Peitgen, H. Jürgens, and D. Saupe, Fractals for the classroom. Part 2:
Complex systems and the Mandelbrot set, Springer, 1992.
[PL68] I. Prigogine and R. Lefever, Symmetry breaking bifurcations in dissipative
systems II, J. Chem. Phys. 48 (1968), 1665–1700.
[Poi57] H. Poincaré, Les methodes nouvelles de la mecanique celeste, Dover Publica-
tions, 1957.
[Pom86] Y. Pomeau, Front motion, metastability and subcritical bifurcations in hydro-
dynamics, Physica D 23 (1986), 3–11.
[PS07] D. Pelinovsky and G. Schneider, Justification of the coupled-mode approxi-
mation for a nonlinear elliptic problem with a periodic potential, Appl. Anal.
86 (2007), no. 8, 1017–1036.
Bibliography 559

[PS13] D. Pelinovsky and G. Schneider, Rigorous justification of the short-pulse equa-

tion, NoDEA, Nonlinear Differ. Equ. Appl. 20 (2013), no. 3, 1277–1294.
[PW84] M. H. Protter and H. F. Weinberger, Maximum principles in differential equa-
tions, Springer-Verlag, New York, 1984.
[PW96] R. D. Pierce and C. E. Wayne, On the validity of mean field amplitude equa-
tions for counterpropagating wavetrains, Nonlinearity 8 (1996), 433–457.
[Rac92] R. Racke, Lectures on nonlinear evolution equations. Initial value problems,
Braunschweig etc.: Vieweg, 1992.
[Rad13] T. Radko, Double-diffusive convection, Cambridge University Press, Cam-
bridge, 2013.
[Rau91] J. Rauch, Partial differential equations, Graduate Texts in Mathematics. 128.
New York etc.: Springer-Verlag. x, 263 p. with 42 ill. , 1991.
[Ray76] Lord Rayleigh, On waves, Phil.Magaz. 1 (1876).
[Ren92] M. Renardy, A centre manifold theorem for hyperbolic PDEs, Proc. Roy. Soc.
Edinburgh Sect. A 122 (1992), no. 3-4, 363–377.
[Ren00] , Mathematical analysis of viscoelastic flows, CBMS-NSF Regional
Conference Series in Applied Mathematics, vol. 73, Society for Industrial
and Applied Mathematics (SIAM), Philadelphia, PA, 2000.
[Rob01] J. C. Robinson, Infinite-dimensional dynamical systems. An introduction to
dissipative parabolic PDEs and the theory of global attractors, Cambridge
Texts in Applied Mathematics, Cambridge University Press, Cambridge,
2001.
[Rob04a] , An introduction to ordinary differential equations, Cambridge Uni-
versity Press. , 2004.
[Rob04b] R. C. Robinson, An introduction to dynamical systems. Continuous and dis-
crete, Prentice Hall, 2004.
[Rös76] O. Rössler, An equation for continuous chaos, Physics Letters A 57 (1976),
397–398.
[RR04] M. Renardy and R. C. Rogers, An introduction to partial differential equa-
tions, second ed., Texts in Applied Mathematics, vol. 13, Springer-Verlag,
New York, 2004.
[RS75a] M. Reed and B. Simon, Methods of modern mathematical physics. Four vol-
umes, Academic Press, 1975.
[RS75b] , Methods of modern mathematical physics. II: Fourier analysis, self-
adjointness, Academic Press, 1975.
[RT71] D. Ruelle and F. Takens, On the nature of turbulence, Commun. Math. Phys.
20 (1971), 167–192.
[Rud73] W. Rudin, Functional analysis, McGraw-Hill, 1973.
[Rue89] D. Ruelle, Chaotic evolution and strange attractors. the statistical analysis
of time series for deterministic nonlinear systems, notes prepared and with a
foreword by S. Isola, Lezioni Lincee. [Lincei Lectures], Cambridge University
Press, Cambridge, 1989.
[Rus44] J. S. Russell, Report on waves, Rep. 14th Meet. Brit. Assoc. Sci. (1844),
311–390.
[Sal08] S. Salsa, Partial differential equations in action. From modelling to theory,
Universitext, Springer-Verlag Italia, Milan, 2008.
560 Bibliography

[San02] B. Sandstede, Stability of travelling waves, Handbook of dynamical systems.

Volume 2, Amsterdam: Elsevier, 2002, pp. 983–1055.
[Sat76] D. H. Sattinger, On the stability of waves of nonlinear parabolic systems,
Advances in Math. 22 (1976), no. 3, 312–355.
[Sca99] B. Scarpellini, Stability, instability, and direct integrals, Chapman & Hall,
1999.
[Sch79] J. Schnakenberg, Simple chemical reaction systems with limit cycle behaviour,
J. Theoret. Biol. 81 (1979), no. 3, 389–400.
[Sch94a] G. Schneider, Error estimates for the Ginzburg-Landau approximation, ZAMP
45 (1994), 433–457.
[Sch94b] , Global existence via Ginzburg-Landau formalism and pseudo-orbits of
Ginzburg-Landau approximations, Comm. Math. Phys. 164 (1994), 157–179.
[Sch94c] , A new estimate for the Ginzburg-Landau approximation on the real
axis, J. Nonlinear Sci. 4 (1994), no. 1, 23–34.
[Sch95a] , Analyticity of Ginzburg-Landau modes, J. Diff. Eqns. 121 (1995),
233–257.
[Sch95b] , Validity and Limitation of the Newell-Whitehead equation, Math.
Nachr. 176 (1995), 249–263.
[Sch96a] , Diffusive stability of spatial periodic solutions of the Swift-Hohenberg
equation, Comm. Math. Phys. 178 (1996), 679–702.
[Sch96b] , Limits for the Korteweg-de Vries-Approximation, CIAM/GAMM 95
Applied Analysis, ZAMM - Journal of Applied Mathematics and Mechanics
76 (1996), no. S2, 341–344.
[Sch96c] , The validity of generalized Ginzburg-Landau equations, Math. Meth-
ods Appl. Sci. 19 (1996), no. 9, 717–736.
[Sch97] , Justification of mean-field coupled modulation equations, Proc. Royal
Soc. Edinb. 127 (1997), no. A, 639–650.
[Sch98a] , Approximation of the Korteweg-de Vries equation by the Nonlinear
Schrödinger equation, JDE 147 (1998), 333–354.
[Sch98b] , Hopf bifurcation in spatially extended reaction-diffusion systems, J.
Nonlinear Sci. 8 (1998), no. 1, 17–41.
[Sch98c] , Justification of modulation equations for hyperbolic systems via nor-
mal forms, NoDEA 5 (1998), 69–82.
[Sch98d] , Nonlinear stability of Taylor-vortices in infinite cylinders, Arch. Rat.
Mech. Anal. 144 (1998), no. 2, 121–200.
[Sch99a] , Cahn-Hilliard description of secondary flows of a viscous incompress-
ible fluid in an unbounded domain, ZAMM Z. Angew. Math. Mech. 79 (1999),
no. 9, 615–626.
[Sch99b] , Global existence results in pattern forming systems - Applications to
3D Navier-Stokes problems -, J. Math. Pures Appl., IX 78 (1999), 265–312.
[Sch99c] , Some characterizations of the Taylor-Couette attractor, Differential
Integral Equations 12 (1999), no. 6, 913–926.
[Sch05] , Justification and failure of the nonlinear Schrödinger equation in case
of non-trivial quadratic resonances, J. Diff. Eq. 216 (2005), no. 2, 354–386.
[Sch10] , Bounds for the nonlinear Schrödinger approximation of the Fermi-
Pasta-Ulam system, Appl. Anal. 89 (2010), no. 9, 1523–1539.
Bibliography 561

[Sch11] , Justification of the NLS approximation for the KdV equation using
the Miura transformation, Adv. Math. Phys. (2011), Art. ID 854719, 4.
[Sch15] R. Schnaubelt, Center manifolds and attractivity for quasilinear parabolic
problems with fully nonlinear dynamical boundary conditions, Discrete Con-
tin. Dyn. Syst. 35 (2015), no. 3, 1193–1230.
[SD98] R. Schielen and A. Doelman, Modulation equations for spatially periodic sys-
tems: derivation and solutions, SIAM J. Appl. Math. 58 (1998), no. 6, 1901–
1930.
[Seg69] L. A. Segel, Distant side-walls cause slow amplitude modulation of cellular
convection, Journal of Fluid Mechanics 38(1) (1969), 203–224.
[Ser99] D. Serre, Systems of conservation laws. I: Hyperbolicity, entropies, shock
waves, Cambridge: Cambridge University Press. xxii, 263 p., 1999.
[Ser00] , Systems of conservation laws 2. Geometric structures, oscillations,
and initial-boundary value problems, Cambridge: Cambridge University
Press. xi, 269 p., 2000.
[Ser15] S. Serfaty, Coulomb gases and Ginzburg-Landau vortices, Zürich: European
Mathematical Society (EMS), 2015.
[Sey10] R. Seydel, Practical bifurcation and stability analysis. 3rd ed., Springer, 2010.
[SGKM95] A. A. Samarskii, V. Galaktionov, S. Kurdyumov, and A. Mikhailov, Blow-up
in quasilinear parabolic equations, De Gruyter, 1995.
[SH77] J. Swift and P. C. Hohenberg, Hydrodynamic fluctuations at the convective
instability, Physical Review A 15 (1977), no. 1, 319–328.
[SH94a] P. J. Schmid and D. S. Henningson, Optimal energy density growth in Hagen-
Poiseuille flow, J. Fluid Mech. 277 (1994), 197–225.
[SH94b] Y. Shi and J. E. Hearst, The Kirchhoff elastic rod, the nonlinear Schrödinger
equation, and DNA supercoiling, J. Chem. Phys. 101 (1994), no. 6, 5186–
5200.
[SH96] A. M. Stuart and A. R. Humphries, Dynamical systems and numerical anal-
ysis, Cambridge: Cambridge University Press, 1996.
[Sha85] J. Shatah, Normal forms and quadratic nonlinear Klein-Gordon equations,
Comm. Pure Appl. Math. 38 (1985), no. 5, 685–696.
[Sha10] J. Shatah, Space-time resonances, Q. Appl. Math. 68 (2010), no. 1, 161–167.
[She97] A. Shepeleva, On the validity of the degenerate Ginzburg-Landau equation,
Math. Methods Appl. Sci. 20 (1997), no. 14, 1239–1256.
[Shi76] M. Shinbrot, The initial value problem for surface waves under gravity. I. The
simplest case, Indiana Univ. Math. J. 25 (1976), no. 3, 281–300.
[Sil65] L. P. Sil’nikov, A case of the existence of a denumerable set of periodic mo-
tions, Dokl. Akad. Nauk SSSR 160 (1965), 558–561.
[Siv77] G. Sivashinsky, Nonlinear analysis of hydrodynamic instability in laminar
flames. I - Derivation of basic equations, Acta Astronautica 4 (1977), 1177–
1206.
[Smo94] J. Smoller, Shock waves and reaction-diffusion equations, Springer, New York,
1994.
[Spa82] C. Sparrow, The Lorenz equations: bifurcations, chaos, and strange attractors,
Applied Mathematical Sciences, 41. Springer-Verlag, 1982.
[SR89] X. Saint Raymond, A simple Nash-Moser implicit function theorem, Enseign.
Math. (2) 35 (1989), no. 3-4, 217–226.
562 Bibliography

[SR10] P. L. Sachdev and Ch. Srinivasa Rao, Large time asymptotics for solutions of
nonlinear partial differential equations, Springer Monographs in Mathematics,
Springer, New York, 2010.
[SS99a] B. Sandstede and A. Scheel, Essential instability of pulses and bifurcations to
modulated travelling waves, Proc. R. Soc. Edinb., Sect. A, Math. 129 (1999),
no. 6, 1263–1290.
[SS99b] C. Sulem and P.-L. Sulem, The nonlinear Schrödinger equation, Springer-
Verlag, New York, 1999.
[SS00] B. Sandstede and A. Scheel, Absolute and convective instabilities of waves on
unbounded and large bounded domains, Phys. D 145 (2000), no. 3-4, 233–277.
[SS07] E. Sandier and S. Serfaty, Vortices in the magnetic Ginzburg-Landau model,
Basel: Birkhäuser, 2007.
[SSSU12] B. Sandstede, A. Scheel, G. Schneider, and H. Uecker, Diffusive mixing of
periodic wave trains in reaction-diffusion systems, J. Differ. Equations 252
(2012), no. 5, 3541–3574.
[SSW99] B. Sandstede, A. Scheel, and C. Wulff, Bifurcations and dynamics of spiral
waves, J. Nonlinear Sci. 9 (1999), no. 4, 439–478.
[SSZ15] G. Schneider, Danish Ali Sunny, and D. Zimmermann, The NLS approxi-
mation makes wrong predictions for the water wave problem in case of small
surface tension and spatially periodic boundary conditions, J. Dyn. Differ.
Equations 27 (2015), no. 3-4, 1077–1099.
[Ste93] E. M. Stein, Harmonic analysis: Real-variable methods, orthogonality, and
oscillatory integrals. With the assistance of Timothy S. Murphy, Princeton,
NJ: Princeton University Press, 1993.
[STE+ 10] I. Staude, M. Thiel, S. Essig, C. Wolff, K. Busch, G. Von Freymann, and
M. Wegener, Fabrication and characterization of silicon woodpile photonic
crystals with a complete bandgap at telecom wavelengths, Optics letters 35
(2010), no. 7, 1094–1096.
[Str26] D. J. Struik, Détermination rigoureuse des ondes irrotationelles périodiques
dans un canal à profondeur finie, Math. Ann. 95 (1926), 595–634 (French).
[Str89] W. A. Strauss, Nonlinear wave equations. Expository lectures from the CBMS
regional conference held at George Mason University, January 16-22, 1989,
Providence, RI: American Mathematical Society, 1989.
[Str92] W. A. Strauss, Partial differential equations - an introduction, John Wiley &
Sons Inc., New York, 1992.
[Str94] S. H. Strogatz, Nonlinear dynamics and chaos, Perseus, New York, 1994.
[Str04] B. Straughan, The energy method, stability, and nonlinear convection, second
ed., Applied Mathematical Sciences, vol. 91, Springer-Verlag, New York, 2004.
[Str08] , Stability and wave motion in porous media, Applied Mathematical
Sciences, vol. 165, Springer, New York, 2008.
[SU01] G. Schneider and H. Uecker, Nonlinear coupled mode dynamics in hyperbolic
and parabolic periodically structured spatially extended systems, Asymptotic
Anal. 28 (2001), no. 2, 163–180.
[SU03a] , Almost global existence and transient self similar decay for Poiseuille
flow at criticality over exponentially long times, Physica D 185 (2003), no. 3-
4, 209–226.
[SU03b] , Existence and stability of exact pulse solutions for Maxwell’s equa-
tions describing nonlinear optics, ZAMP 54 (2003), 677–712.
Bibliography 563

[SU07] , The amplitude equations for the first instability of electro-convection

in nematic liquid crystals in the case of two unbounded space directions, Non-
linearity 20 (2007), no. 6, 1361–1386.
[SUW11] G. Schneider, H. Uecker, and M. Wand, Interaction of modulated pulses in
nonlinear oscillator chains, J. Difference Equ. Appl. 17 (2011), no. 3, 279–
298.
[SVZZ13] S. Salsa, F. M. G. Vegni, A. Zaretti, and P. Zunino, A primer on PDEs.
Models, methods, simulations. , Milano: Springer, 2013.
[SW04] T. Schäfer and C. E. Wayne, Propagation of ultra-short optical pulses in cubic
nonlinear media, Phys. D 196 (2004), no. 1-2, 90–105.
[SW00a] G. Schneider and C. E. Wayne, Counter-propagating waves on fluid surfaces
and the continuum limit of the Fermi-Pasta-Ulam model, International con-
ference on differential equations. Proceedings of the conference, Equadiff ’99,
Berlin, Germany, August 1–7, 1999. Vol. 1, Singapore: World Scientific, 2000,
pp. 390–404.
[SW00b] , The long wave limit for the water wave problem I. The case of zero
surface tension, Comm. Pure. Appl. Math. 53 (2000), no. 12, 1475–1535.
[SW01] , Kawahara dynamics in dispersive media, Phys. D 152/153 (2001),
384–394, Advances in nonlinear mathematics and science.
[SW02] , The rigorous approximation of long-wavelength capillary-gravity
waves, Arch. Rat. Mech. Anal. 162 (2002), 247–285.
[SW11] , Justification of the NLS approximation for a quasilinear water wave
model, J. Differential Equations 251 (2011), no. 2, 238–269.
[SZ13] G. Schneider and D. Zimmermann, Justification of the Ginzburg-Landau ap-
proximation for an instability as it appears for Marangoni convection, Math.
Methods Appl. Sci. 36 (2013), no. 9, 1003–1013.
[Tao06] T. Tao, Nonlinear dispersive equations, CBMS Regional Conference Series in
Mathematics, vol. 106, CBMS, Washington, DC, 2006.
[Tao09] , Why global regularity for Navier-Stokes is hard, Structure and Ran-
domness: Pages from Year One of a Mathematical Blog, AMS, 2009.
[Tay96] M. E. Taylor, Partial differential equations. 3 volumes, Applied Mathematical
Sciences, Springer-Verlag, New York, 1996.
[TBD+ 96] P. Takáč, P. Bollerman, A. Doelman, A. van Harten, and E. S. Titi, Ana-
lyticity of essentially bounded solutions to semilinear parabolic systems and
validity of the Ginzburg-Landau equation, SIAM J. Math. Anal. 27 (1996),
no. 2, 424–448.
[Tem97] R. Temam, Infinite-dimensional dynamical systems in mechanics and physics,
second ed., Applied Mathematical Sciences, vol. 68, Springer-Verlag, New
York, 1997.
[Tem01] , Navier-Stokes equations, AMS, Providence, RI, 2001.
[Tes12] G. Teschl, Ordinary differential equations and dynamical systems, Graduate
Studies in Mathematics, vol. 140, American Mathematical Society, Provi-
dence, RI, 2012.
[Thi88] W. Thirring, Lehrbuch der mathematischen Physik, Band 1: Klassische dy-
namische Systeme, Springer-Verlag, 1988.
[TM05] R. Temam and A. M. Miranville, Mathematical modelling in continuum me-
chanics, Cambridge University Press, 2005.
564 Bibliography

[Tot15] N. Totz, A justification of the modulation approximation to the 3D full water

wave problem, Commun. Math. Phys. 335 (2015), no. 1, 369–443.
[Tuc02] W. Tucker, A rigorous ODE solver and Smale’s 14th problem, Found. Com-
put. Math. 2 (2002), no. 1, 53–117.
[Tur52] A. Turing, The chemical basis of morphogenesis, Phil. Trans. Roy. Soc. B 237
(1952), 37–72.
[TW12] N. Totz and Sijue Wu, A rigorous justification of the modulation approxima-
tion to the 2D full water wave problem, Commun. Math. Phys. 310 (2012),
no. 3, 817–883.
[TW15] Y. Tao and M. Winkler, Boundedness vs. blow-up in a two-species chemotaxis
system with two chemicals, Discrete Contin. Dyn. Syst. Ser. B 20 (2015), no. 9,
3165–3183.
[TZ10] D. Turaev and S. Zelik, Analytical proof of space-time chaos in Ginzburg-
Landau equations, Discrete Contin. Dyn. Syst. 28 (2010), no. 4, 1713–1751.
[Uec99] H. Uecker, Diffusive stability of rolls in the two-dimensional real and complex
Swift-Hohenberg equation, Comm. Partial Differential Equations 24 (1999),
no. 11-12, 2109–2146.
[Uec03] , Approximation of the Integral Boundary Layer equation by the
Kuramoto-Sivashinsky equation, SIAM J. Appl. Math. 63 (2003), no. 4, 1359–
1377.
[Uec07] , Self-similar decay of spatially localized perturbations of the Nusselt
solution for the inclined film problem, Arch. Ration. Mech. Anal. 184 (2007),
no. 3, 401–447.
[Uec09] , A short ad hoc introduction to spectral methods for parabolic PDE and
the Navier-Stokes equations, Modern Computational Science 09 (R. Leidl and
A. Hartmann, eds.), Universitätsverlag Oldenburg, Oldenburg, 2009, pp. 169–
212.
[Uec16] , Optimal harvesting and spatial patterns in a semi arid vegetation
system, Natural Resource Modelling 29 (2016), no. 2, 229–258.
[UGS+ 15] H. Uecker, D. Grieser, Z. Sobirov, D. Babajanov, and D. Matrasulov, Soliton
transport in tubular networks: transmission at vertices in the shrinking limit,
Phys. Rev. E 91 (2015), 023209.
[UW14] H. Uecker and D. Wetzel, Numerical results for snaking of patterns over pat-
terns in some 2D Selkov-Schnakenberg Reaction-Diffusion systems, SIADS
13-1 (2014), 94–128.
[UWR14] H. Uecker, D. Wetzel, and J. Rademacher, pde2path – a Matlab package for
continuation and bifurcation in 2D elliptic systems, NMTMA 7 (2014), 58–
106, see also www.staff.uni-oldenburg.de/hannes.uecker/pde2path.
[Van89] A. Vanderbauwhede, Centre manifolds, normal forms and elementary bifur-
cations, Dynamics reported, Vol. 2, Dynam. Report. Ser. Dynam. Systems
Appl., vol. 2, Wiley, Chichester, 1989, pp. 89–169.
[Vas15] A. Vasy, Partial differential equations. An accessible route through theory and
applications., Providence, RI: American Mathematical Society (AMS), 2015.
[VdP26] B. Van der Pol, On relaxation-oscillations, The London, Edinburgh and
Dublin Phil. Mag. & J. of Sci. 2 (1926), no. 7, 978–992.
[Ver96] F. Verhulst, Nonlinear differential equations and dynamical systems, second
ed., Springer-Verlag, Berlin, 1996.
Bibliography 565

[vH91] A. van Harten, On the validity of Ginzburg-Landau’s equation, J. Nonlinear

Science 1 (1991), 397–422.
[VI92] A. Vanderbauwhede and G. Iooss, Center manifolds in infinite dimensions,
Springer, Dynamics reported 1 (1992), 125–163.
[vS95] W. van Saarloos, The complex Ginzburg-Landau equation for beginners,
Spatio-temporal patterns in nonequilibrium complex systems (Santa Fe, NM,
1993), Santa Fe Inst. Stud. Sci. Complexity Proc., XXI, Addison-Wesley,
Reading, MA, 1995, pp. 19–31.
[vS03] W. van Saarloos, Front propagation into unstable states, Physics Reports 386
(2003), 29–222.
[vSH92] W. van Saarloos and P. C. Hohenberg, Fronts, pulses, sources and sinks in
generalized complex Ginzburg–Landau equations, Physica D 56 (1992), 303–
367.
[VVV94] A. I. Volpert, V. A. Volpert, and V. A. Volpert, Traveling wave solutions
of parabolic systems, Translations of Mathematical Monographs, vol. 140,
American Mathematical Society, Providence, RI, 1994.
[vW85] W. von Wahl, The equations of Navier-Stokes and abstract parabolic equa-
tions, Vieweg, 1985.
[Wal82] P. Walters, An introduction to ergodic theory, Graduate Texts in Mathemat-
ics, Vol. 79. New York - Heidelberg -Berlin: Springer-Verlag. IX, 1982.
[Way96] C. E. Wayne, An introduction to KAM theory, Dynamical systems and prob-
abilistic methods in partial differential equations. 1994 summer seminar on
dynamical systems and probabilistic methods for nonlinear waves (Percy (ed.)
et al. Deift, ed.), American Mathematical Society, 1996, pp. 3–29.
[Way97] , Invariant manifolds for parabolic partial differential equations on un-
bounded domains, Arch. Rat. Mech. Anal. 138 (1997), 279–306.
[Wei81] F. B. Weissler, Existence and nonexistence of global solutions for a semilinear
heat equation, Israel Journal of Mathematics 38 (1981), 29–40.
[Wer00] D. Werner, Funktionalanalysis, extended ed., Springer-Verlag, Berlin, 2000.
[Whi99] G. B. Whitham, Linear and nonlinear waves, Pure and Applied Mathematics
(New York), John Wiley & Sons, Inc., New York, 1999, Reprint of the 1974
original, A Wiley-Interscience Publication.
[Wie99] M. Wiegner, The Navier-Stokes equations-a neverending challenge?, Jahres-
ber. Deutsch. Math.-Verein. 101 (1999), no. 1, 1–25.
[Wig88] S. Wiggins, Global bifurcations and chaos. analytical methods, Applied Math-
ematical Sciences, 73. New York etc.: Springer-Verlag., 1988.
[Wig03] , Introduction to applied nonlinear dynamical systems and chaos, sec-
ond ed., Texts in Applied Mathematics, vol. 2, Springer-Verlag, New York,
2003.
[Wlo87] J. Wloka, Partial differential equations, Cambridge University Press, Cam-
bridge, 1987.
[Wri06] J. D. Wright, Corrections to the KdV approximation for water waves, SIAM
J. Math. Anal. 37 (2006), no. 4, 1161–1206.
[Wu97] Sijue Wu, Well-posedness in Sobolev spaces of the full water wave problem in
2-D, Invent. Math. 130 (1997), no. 1, 39–72.
[Wu99] , Well-posedness in Sobolev spaces of the full water wave problem in
3-D, J. Amer. Math. Soc. 12 (1999), no. 2, 445–495.
566 Bibliography

[Wu09] , Almost global wellposedness of the 2-D full water wave problem, In-
vent. Math. 177 (2009), no. 1, 45–135.
[Wu11] , Global wellposedness of the 3-D full water wave problem, Invent.
Math. 184 (2011), no. 1, 125–220.
[WW02] C. E. Wayne and J. D. Wright, Higher-order modulation equations for a
Boussinesq equation, SIAM J. Appl. Dyn. Syst. 1 (2002), no. 2, 271–302.
[WW15] C. E. Wayne and M. I. Weinstein, Dynamics of partial differential equa-
tions, Frontiers in Applied Dynamical Systems: Reviews and Tutorials, vol. 3,
Springer, Cham, 2015.
[Yan10] J. Yang, Nonlinear waves in integrable and nonintegrable systems, SIAM,
2010.
[YLWM13] S. Yanchuk, L. Lücken, M. Wolfrum, and A. Mielke, Spectrum and ampli-
tude equations for scalar delay-differential equations with large delay, Discrete
Contin. Dyn. Syst. 35 (2015), no. 1, 537–553.
[Yos71] K. Yosida, Functional analysis, Springer, New York, 1971.
[Yos82] H. Yosihara, Gravity waves on the free surface of an incompressible perfect
fluid of finite despth, Publ. Res. Inst. Math. Sci. 18 (1982), 49–96.
[Yos83] , Capillary-gravity waves for an incompressible ideal fluid, J. Math.
Kyoto Univ. 23 (1983), 649–694.
[Zak68] V. E. Zakharov, Stability of periodic waves of finite amplitude on the surface
of a deep fluid, Sov. Phys. J. Appl. Mech. Tech. Phys 4 (1968), 190–194.
[Zei71] E. Zeidler, Existenzbeweis für cnoidal waves unter Berücksichtigung der
Oberflächenspannung. (Existence proof for cnoidal waves under considera-
tion of the surface tension.), Arch. Ration. Mech. Anal. 41 (1971), 81–107
(German).
[ZF71] V. E. Zakharov and L. D. Fadeev, The Korteweg-de Vries equation is a fully
integrable Hamiltonian system, Functional Anal. Appl. 5 (1971), 280–287.
[Zim16] D. Zimmermann, Justification of the Ginzburg-Landau approximation for
quasilinear problems - including an application to Bénard-Marangoni con-
vection, Preprint Universität Stuttgart (2016), 89 pages.
[ZK65] N. J. Zabusky and M. D. Kruskal, Interactions of solitons in a collisionless
plasma and the recurrence of initial states, Phys. Rev. Lett. 15 (1965), 240–
243.
List of symbols

Acronyms C0∞ (Ω, R), infinitely differentiable

BBM Benjamin-Bona-Mahony, 452 functions with compact support,
GL Ginzburg-Landau, 9 153
KdV Korteweg-de Vries, 9 C k,α (Ω, R), uC k,α , Hölder spaces
KPP Kolmogorov-Petrovsky- and norm, 153
Piskounov, C m (Ω, R), continuously differentiable
12 functions, 152
KS Kuramoto-Shivashinsky, 332 Cbm (Ω, R), uCbm , bounded
NLS Nonlinear Schrödinger, 8 continuously differentiable
SH Swift-Hohenberg, 10 functions, and their norm, 152
Dynamicals systems m
Hper , Sobolev space on torus, 158
(Z, Zρ )-attractor, attractor Hm r
, weighted Sobolev space of
w.r.t. different norms, 286 (differentiation) order r, 236
Ws , Wc , Wu , stable, center and L ((−1/2, 1/2], H θ ([0, 2π), C)), basic
2
unstable manifolds, 46 Bloch space, 434
A = ω(B), attractor as ω-limit set of L2m , weighted L2 space, 235
set B, 51
L2loc , local L2 spaces, 277
St : X → X, general evolution
Lp (Ω, R), uLp , Lebesgue spaces
semigroup, 49
and norms, 153
γ+ (u0 ), γ− (u0 ), forward, backward
orbit through u0 , 38 W m,p (Ω, R), uW m,p , H m (Ω, R),
ω(u0 ), ω-limit set of (the orbit uH m , Sobolev spaces and
through) u0 , 50 norms, 154
u(t, t0 , u0 ) [u(t, u0 )], solution of initial W m,p (Rd , R), H m (Rd , R), Sobolev
value problem [autonomous spaces over Rd , 235
case], 15 D, D  , space of test functions, and its
Function spaces and norms dual, i.e., space of distributions,
C 0 (Ω, R), uC 0 , continuous 155
S, S  , Schwartz space, and dual, i.e.,
b
functions, and their norm, 152
Cb0 (Ω, R), bounded continuous the space of tempered
functions, 152 distributions, 229
0
Cb,unif (Ω, R), bounded uniformly  ∂x ), Bloch operators, 435
L( ,
continous functions, 152 RN , RZ , R∞ , spaces of sequences, 135

567
568 List of symbols

p,θ (RZ ), p,θ , up,θ (RZ ) , weighted p

sequence spaces, and norm, 135
Hulθ
, ·H θ , uniformly local Sobolev
ul
spaces and norm, 278
L2ul , uL2 , uniformly local L2 spaces
ul
and norm, 277
·ρ , localized norm, 287
u∞,θ (RZ ) , weighted ∞ spaces, and
norm, 135
distX (a, B), distX (A, B), distance in
the Banach space X, 287
Miscellanous
|n| = n1 + . . . + nd , for n a multi
index, 151
BX (r, u0 ) ball with radius r around
u0 in the Banach space X, 286
Res(εψGL ), Res(εΨGL ) residual of the
modulation ansatz, 319
, convolution in Bloch space, 434
k , Fourier coefficients of u (Fourier
u
series), 146
u ∗ v, convolution of functions u, v,
231
supp(u), support of a function, 153
Erf, erf (error functions), 245
Operators
n
∂xn = ∂xn11 . . . ∂xdd , for n a multi index,
151
Ec , Es , Ech , Esh , mode filters, defined
via cutoff in Fourier space, 338
M = F −1 M F , multiplier operator in
physical space, 280
Sε , scaling operators, 338
Tu regular distribution defined by
function u, 155
Ty , translation operator, 277
F , F −1 , Fourier transform and
inverse, 227
T , T −1 Bloch transform and inverse,
434
Δ = ∂x21 + . . . + ∂x2d , Laplace
operator, 1
etA exponential matrix, 24
etΛ , semigroup generated by operator
Λ, 338
M = F M F −1 , multiplier operator in
Fourier space, 280
Index

absorbing set, 51 Newton-polygon, 81

abstract pattern forming system, 339 numerical analysis, 96
action and angle variables, 114 pitchfork, 76
Allen-Cahn equation, 222 reduced equation, 84
bounded domain, 148, 167 saddle-node, 77
front,kink,interface, 223 scaling arguments, 80
local bifurcations, 478 subcritical, 76
attractor, 51 supercritical, 76
(Zu , Zρ ), 287 transcritical, 77
(Z, Zρ ), 286 Turing, 309
abstract existence theorem, 51 Turing-Hopf, 309
Chafee-Infante, 169 unfolding, 83
for heat equation over R, 285, 289 Biot-Savart law, 184
for pattern forming systems, 378 Bloch
for PDEs on unbounded domains, 285 convolution, 434
for the SH equation over R, 379 modes, 429
global, 51 operators, 435
lower semicontinuity, 54 space, 434
Navier-Stokes, 196 transform, 434, 518
shadowing, 52, 382 Bose-Einstein condensates, 250
upper semicontinuity, 54 boundary conditions, 2
autonomous, 15 Dirichlet, 2
Neumann, 2
Bénard’s problem, 361 periodic, 158, 179, 292, 397
Bénard-Marangoni system, 396 Boussinesq approximation, 356, 361
basin of attraction, 51 Boussinesq equation, 423, 451, 472
Benjamin-Feir instability, 290, 294 breathers, 491
Bessel’s equality, 159 generalized, 493
bifurcation, 76 sine-Gordon, 492
Bogdanov-Takens, 95, 307 Brillouin zone, 430
genericity, 82 Brownian motion, 207
global, 102 Brusselator, 295
Hopf, 78, 91 Burgers equation, 10, 12, 237, 465, 526

569
570 Index

bounded domain, 148, 167 diffusion, 207

integrated, 244 fundamental solution, 209
limit profile, 526 nonlinear, 498
perturbed, 527 on the real line, 208
polynomial decay, 497
Cantor set, 61 selfsimilar decay, 498
carrier wave, 8, 401 semigroup, 211
celestial mechanics, 116 smoothing properties, 210
center manifold diffusive mixing, 245
cut off, 475 diffusive stability
for SH equation, 480 in the GL equation, 507, 517
center manifold, 47, 85, 86 of Swift-Hohenberg rolls, 517
examples, 86–89 of Taylor vortices, 523
attractivity, 89 dispersion, 251, 401
for maps, 89 dispersion relation, 310, 356, 402, 404,
in infinite dimensions, 473 424, 432, 460
non-smoothness, 88 dispersive system
non-uniqueness, 87, 88 general, 412
Chafee-Infante problem, 167 distribution
characteristics, 241 δ, 234
chemostat, 105 Fourier transform of, 233
Cole-Hopf transformation, 244 regular, 233
compactness, 149
scaling of, 248
competition, 68
tempered, 234
conducting state, 361
dynamical system, 6, 38
conservation law, 262
bistable, 305
conservation laws (Navier Stokes), 180
chaotic, 59
contracting set, 300
completely integrable, 121
convection, 354
conservative, 109
Bénard, 361
continuous, 38
cell, 355
discrete, 38
porous media, 355
dissipative, 51
rolls, 361
equivariant, 371
convective stability, 222
excitable, 303
convolution
Hamiltonian, 109
in Bloch space, 434
measure preserving, 112
in Fourier space, 231
oscillatory, 305
of Fourier series, 146
relaxational, 304
Couette-Taylor problem, 370
shift, 59
counter-example
for Ginzburg-Landau approximation,
365 Eckhaus instability, 290
for NLS approximation, 427 eigenmode, 4
coupled mode equations, 440, 448 electro convection, 355
critical nonlinearity, 524, 526 electromagnetic pulse, 8
embedding
Darcy’s law, 355, 361 compact, 150
degree of irrelevance, 507 embedding, 281
derivative Rellich, 150
Gateaux and Fréchet, 140 Sobolev, 147, 150, 154, 156, 165, 200,
distributional, 155 235
weak, 154 energy
Index 571

kinetic, potential, total, for Swift-Hohenberg, see also

pendulum, 49 Swift-Hohenberg equation
enstrophy, 203 transport, 211
equation Euler-Poisson system, 464
Airy, 258, 259 exponential matrix, 23
Allen-Cahn, see also Allen-Cahn exponentially long transients, 524
equation
Feigenbaum constant, 101
Benjamin-Bona-Mahony, 466
finite wave length instability, 9
Boussinesq, see also Boussinesq
FitzHugh-Nagumo model, 296, 302
equation
fixed point, 27, 38
Burgers, see also Burgers equation,
asymptotically stable, 39
533
hyperbolic, 43
Camassa-Holm, 470 stability for iterations, 45
delay-differential, 397 stable, 26, 39
diffusion, heat, 3 unstable, 39
elliptic PDE, 2 Floquet
Euler, 182 exponents, 30, 431
Fisher, see also KPP equation multipliers, 30, 46, 431
Ginzburg-Landau, see also flow
Ginzburg-Landau equation conjugated, 59
Gross-Pitaevsky, 250, 439 Couette, 185, 371
hyperbolic PDE, 4 inclined film, 396
Kadomtsev-Petviashvili, 472 irrotational, 185
Kawahara, 462, 472 Poiseuille, 185
Klein-Gordon, 402, 535 shear, 185
Kolmogorov-Petrovsky-Piskounov, strain, 185
see also KPP equation flux function, 237
Korteweg-de Vries, see also KdV Fourier series, 146, 158
equation C m - 1,m estimate, 161
KS-KdV, 399 H m - 2,m isomorphy, 161
Kuramoto-Sivashinsky, see also coefficient, 146
Kuramoto-Sivashinsky equation Gibbs phenomenon, 164
Laplace, 1 main theorem, 160
lattice differential, 396 point-wise convergence, 164
real, 163
Maxwell’s, see also Maxwell’s
Fourier transform, 225
equations
convolution, 231
mKdV, 447
in S and Lp , 228
Navier-Stokes, see also Navier-Stokes
in L1 , 226
equations
inverse, 228
Newell-Whitehead, 364
isomorphism property, 236
Nonlinear Schrödinger, see also NLS of distributions, 233
equation smoothness and decay, 228
parabolic PDE, 3 Fourier’s law, 208
peakon, 471 FPU model, 264, 447, 451, 462
phase diffusion, 529 derivation of NLS and KdV, 464
plasma physics, 464 poly-atomic, 426
reaction-diffusion, 212 front, 216
regularized long wave, 467, 471 modulating, 322
short pulse, 448 stability and instability, 216
sine-Gordon, 492 fundamental matrix, 22
572 Index

Galilei invariance, 293 hexagons (planar pattern), 363

gap solitons, 439, 447 high speed ferry, 9
Gelfand-Levitan equation, 271 Hodgkin-Huxley model, 302
Gierer-Meinhard model, 296
Ginzburg-Landau approximation, 315 inclined film problem, 529
abstract approximation theorem, 341 inequality
attractivity, 383 Gagliardo-Nirenberg, 201, 247
counter-example, 365 Gronwall, see also Gronwall
for porous media flow, 361 Hausdorff-Young, 163, 231
for Bénard convection, 363 Poincaré, 164
for Cahn-Hilliard problems, 396 Young’s for convolutions, 146, 232
for degenerated problems, 395 invariance (translational, S 1 , Galilei,
for delay-differential equations, 397 scaling), 293
for reaction diffusion systems, 347, invariant manifolds
351 for maps, 49
for spatially periodic problems, 395 for ODEs, 47
for systems with neutral modes, 396 invariant region, 300
for the Couette-Taylor problem, 373, irrelevant nonlinearity, 500
375 dispersive case, 535
for the Hopf bifurcation case, 354 in Bloch space, 520
for the KS equation, 344 in Fourier space, 511
for the Schnakenberg model, 350
for the SH equation, 330 Jordan normal form, 24
for time periodic systems, 366
Fourier mode distribution, 390 KdV approximation
Ginzburg-Landau equation, 9, 12, 275 for bi-directional models, 466
bounded domain, 148, 167 for the Boussinesq equation, 452
derivation in general systems, 335 for the Euler-Poisson system, 464
derivation via solvability condition, for the FPU model, 464
349 for water waves, 461
front solutions, 322 KdV equation, 9, 12, 259
global existence, bounded domain, 2-soliton, 272
283 as a Hamiltonian system, 261
mean field coupling, 377 bounded domain, 148, 167
roll solutions, 289 hierarchy, 266
stochastic version, 397 inverse scattering, 271
universality of, 332 scattering data, 269
glass fibers, 440 systems of, 468
global existence via GL approximation, universality, 456
384 Kepler’s laws, 116
gradient system, 56, 168 kink, see also front, 495
Gray Scott model, 306 Kirkwood gaps, 127
Greeks and Trojans, 121 Kolmogorov equations, 68, 296
Gronwall’s inequality Kolmogorov-Petrovsky-Piskounov
differential version, 37 equation, see also KPP equation
integral version, 21 Korteweg-de Vries equation, see also
group velocity, 253, 401, 404 KdV equation
KPP equation, 12, 206, 299
Hölder spaces, 152 bounded domain, 148, 167
Hamiltonian system, 48, 109 traveling wave, 217
Heisenberg uncertainty principle, 247 Kuramoto-Shivashinsky equation, 11,
Hermite polynomial, 499 332
Index 573

Lagrange multiplier, 274 for cubic nonlinearities, 407

Laplace’s principle, 6 for multiplexing, 445
lattice differential equation, 447, 462 for nonlinear wave equations, 410
law of mass action, 67, 297 for photonic crystals, 439
Lax pair, 264 for quadratic systems, 416
Lengyel-Epstein model, 295 for quasilinear systems, 428
Lennart-Jones potential, 130 for spatially periodic problems, 429
lift (from GL to patterns), 383 for the Boussinesq model, 424
linear planar systems, 26 for the FPU model, 464
center, 28 for the mKdV equation, 447
degenerate node, 29 NLS equation, 12, 250, 402
saddle, 28 as a Hamiltonian system, 255
source, unstable node, 28 bounded domain, 148, 167
vortex, 28 derivation from general system, 412
Lipschitz continuous, 34 discrete, 293
liquid crystal, 365 solitons, 251
localized pattern, 398 systems of, 445
logarithmic time scale, 500 universality, 411
logistic map, 99 non-resonance conditions, 94, 406, 415
Lorenz attractor, 6, 101 weakening, 423
Lotka-Volterra system, 66 Nonlinear Schrödinger equation, see
Lyapunov exponent, 60 also NLS equation
Lyapunov function, 57, 502 norm, 18
Lyapunov-Schmidt reduction, 83, 395 compatible, 19
equivalent, 18, 134
Mandelbrot set, 106 operator, 19
manifold normal form transforms, 92, 417
center, see also center manifold nullclines, 27
stable, 46 numerical path following, 96
unstable, 46 Nusselt solution, 529
Marangoni convection, 355
Marchenko equation, 271 ODE
mathematical pendulum, 47 nth order linear, 23
maximal regularity, 157 periodic solution, 43
maximum principle, 215, 391 autonomous, 17
Maxwell’s equations, 5, 440 blow up example, 36
Maxwell-Lorentz system, 444 characteristic equation, 24
mean flow, 188 direction field, 26
Millennium problem, 5, 179, 196 example for non-unique solutions, 36
Miura transformation, 263, 447 first order linear system, 23
mode filters, 338 flow, 27
modulation equation, 7 homogeneous, 17
momentum, 181 initial value problem, 15, 17
monodromy matrix, 30, 431 integrated, 20
multi index, 151 singularly perturbed, 219
multiplexing, 445 vector field, 26, 34
multiplier theory, 279 omega-limit set, 50
operator
Navier-Stokes equations, 5, 179, 370 elliptic, 2, 14
Newton-polygon, 81 hyperbolic, 14
NLS approximation parabolic, 14
counter-example, 427 optical fiber, 7
574 Index

optimal control, 313 reversibility, 65

orbit, 38 Reynolds number, 5, 183, 371
heteroclinic, homoclinic, 46 roll solution, 289
orbital stability, 273 rolls (as planar patterns), 364
orthonormal system, 158
oscillator scattering transform
harmonic, 114, 130 AKNS scheme, 292
model for polarization, 442 ZS scheme, 292
nonlinear, 47, 250 Schnakenberg model, 295, 309
van der Pol, 70 Schrödinger equation (linear), 251
weakly damped, 32 Schwartz space, 229
sectorial operator, 197
Parseval’s equality semigroup, 38
for Fourier series, 159, 160 diffusion, 211
for Fourier transform, 230 infinitesimal generator, 136
pattern forming systems, 9 linear wave equation, 409
period doubling, 98 of bounded operators, 136
phase locking, 398 operator-continuous, 137
phase portrait, 26 smoothing, 144
phase space, 38 strongly continuous, 137
phase velocity, 253, 401 translation, 211
photonic crystals, 429, 440, 447 sensitive dependence, 60
eigenvalue problem, 430 shadowing, 53
planar pattern (rolls, hexagons and shock, 241
squares), 363 SI and SIS deseases, 67
Poincaré map, 45 Silnikov chaos, 62
Poiseuille flow, 524 Sleipner crash, 6
polarization, 441 Smale’s horseshoe, 60, 128
predator-prey system, 66, 68 small divisor problem, 125
pressure, 181, 187 snaking, 398
pseudo orbits Sobolev spaces, 153, 235
for the SH and the KS equation, 386 uniformly local, 276
for ODEs, 52 weighted, 236
solitary wave, 462
quadratic autocatalysis, 298 KdV, 260
stability, 273
Rössler system, 99 water wave problem, 484
random walk, 208 soliton, 9, 264, see also solitary wave
Rankine-Hugoniot, 248 solution
rarefaction wave, 242 mild, 143, 166, 213
Rayleigh number, 355 strong, 143, 166
Reaction-Diffusion systems, 295 spatial dynamics, 473, 480
reduced flow, 86 (generalized) breathers, 491
relevant nonlinearity, 507 for elliptic problems in a strip, 482
renormalization modulating front, 489
continuous, 499 water wave problem, 484
discrete, 504 spatially extended domain, 7
resonance, see also non-resonance spectral gaps, 433
conditions squares (planar pattern), 363
non–trivial, 426 stability
trivial at k = 0, 423 asymptotic, 40
restricted 3-body problem, 119 diffusive, 290, 497
Index 575

exponential, 40 van der Corput Lemma, 257

linearized, 40 van der Pol oscillator, 70
stationary phase, 257, 259 variation of constant, 22
stream function, 356 viscosity solution, 242
superconductivity, 275 vorticity, 184
surface tension , 458
Swift-Hohenberg equation, 10, 319, 480 water wave problem, 448, 458
linear, 460
attractor upper semicontinuity, 394
solitary wave, 484
cubic-quintic, 397
wave equation
global semiflow, absorbing ball, 380
linear, 3, 166, 206, 451
modulating front, 489
nonlinear, 402
spatially periodic equilibria, 480
nonlinear, 2D, 447
symbiosis, 68
nonlinear, in Fourier space, 411
symplectic
with periodic coefficients, 430, 494
structure, 122
Whitham system, 465
transformation, 122

theorem
Beppo-Levi, 226
center manifold, 86, 476
contraction mapping, 20
dominated convergence, 226
Floquet, 30
Fubini, 226
Hartman-Grobman, 43
Hausdorff-Young, 231
Heine Borel, 150
KAM, 126
Liouville, 111, 123
Lyapunov’s subcenter, 115
Nekhoroshov, 127
Paley-Wiener, 233
Picard-Lindelöf, 34
Poincaré-Bendixson, 55
Sarkovskii, 101
Smale-Birkhoff, 128
three wave interaction, 427, 448
traffic flow, 237
trapping region, 299
traveling wave
for reaction diffusion, 299
for the Burgers equation, 243
for the GL equation, 322
for the KdV equation, 261, 293
tsunami, 9
turbulence, 98
Turing instability, 307, 316, 347
Turing space, 310
Turing-Hopf, 375

ultra-short pulses, 446

Selected Published Titles in This Series
182 Guido Schneider and Hannes Uecker, Nonlinear PDEs, 2017
180 Joseph J. Rotman, Advanced Modern Algebra: Third Edition, Part 2, 2017
179 Henri Cohen and Fredrik Strömberg, Modular Forms, 2017
178 Jeanne N. Clelland, From Frenet to Cartan: The Method of Moving Frames, 2017
177 Jacques Sauloy, Differential Galois Theory through Riemann-Hilbert Correspondence,
2016
176 Adam Clay and Dale Rolfsen, Ordered Groups and Topology, 2016
175 Thomas A. Ivey and Joseph M. Landsberg, Cartan for Beginners: Differential
Geometry via Moving Frames and Exterior Differential Systems, Second Edition, 2016
174 Alexander Kirillov Jr., Quiver Representations and Quiver Varieties, 2016
173 Lan Wen, Differentiable Dynamical Systems, 2016
172 Jinho Baik, Percy Deift, and Toufic Suidan, Combinatorics and Random Matrix
Theory, 2016
171 Qing Han, Nonlinear Elliptic Equations of the Second Order, 2016
170 Donald Yau, Colored Operads, 2016
169 András Vasy, Partial Differential Equations, 2015
168 Michael Aizenman and Simone Warzel, Random Operators, 2015
167 John C. Neu, Singular Perturbation in the Physical Sciences, 2015
166 Alberto Torchinsky, Problems in Real and Functional Analysis, 2015
165 Joseph J. Rotman, Advanced Modern Algebra: Third Edition, Part 1, 2015
164 Terence Tao, Expansion in Finite Simple Groups of Lie Type, 2015
163 Gérald Tenenbaum, Introduction to Analytic and Probabilistic Number Theory, Third
Edition, 2015
162 Firas Rassoul-Agha and Timo Seppäläinen, A Course on Large Deviations with an
Introduction to Gibbs Measures, 2015
161 Diane Maclagan and Bernd Sturmfels, Introduction to Tropical Geometry, 2015
160 Marius Overholt, A Course in Analytic Number Theory, 2014
159 John R. Faulkner, The Role of Nonassociative Algebra in Projective Geometry, 2014
158 Fritz Colonius and Wolfgang Kliemann, Dynamical Systems and Linear Algebra,
2014
157 Gerald Teschl, Mathematical Methods in Quantum Mechanics: With Applications to
Schrödinger Operators, Second Edition, 2014
156 Markus Haase, Functional Analysis, 2014
155 Emmanuel Kowalski, An Introduction to the Representation Theory of Groups, 2014
154 Wilhelm Schlag, A Course in Complex Analysis and Riemann Surfaces, 2014
153 Terence Tao, Hilbert’s Fifth Problem and Related Topics, 2014
152 Gábor Székelyhidi, An Introduction to Extremal Kähler Metrics, 2014
151 Jennifer Schultens, Introduction to 3-Manifolds, 2014
150 Joe Diestel and Angela Spalsbury, The Joys of Haar Measure, 2013
149 Daniel W. Stroock, Mathematics of Probability, 2013
148 Luis Barreira and Yakov Pesin, Introduction to Smooth Ergodic Theory, 2013
147 Xingzhi Zhan, Matrix Theory, 2013
146 Aaron N. Siegel, Combinatorial Game Theory, 2013
145 Charles A. Weibel, The K-book, 2013
144 Shun-Jen Cheng and Weiqiang Wang, Dualities and Representations of Lie
Superalgebras, 2012

For a complete list of titles in this series, visit the

AMS Bookstore at www.ams.org/bookstore/gsmseries/.
This is an introductory textbook about nonlinear dynamics of PDEs, with a focus
on problems over unbounded domains and modulation equations. The presenta-
tion is example-oriented, and new mathematical tools are developed step by step,
giving insight into some important classes of nonlinear PDEs and nonlinear dynamics
phenomena which may occur in PDEs.
8LI FSSO GSRWMWXW SJ JSYV TEVXW 4EVXW - ERH -- EVI MRXVSHYGXMSRW XS ½RMXI  ERH
MR½RMXIHMQIRWMSREPH]REQMGWHI½RIHF]3()WERHF]4()WSZIVFSYRHIHHSQEMRW
respectively, including the basics of bifurcation and attractor theory. Part III intro-
duces PDEs on the real line, including the Korteweg-de Vries equation, the Nonlinear
Schrödinger equation and the Ginzburg-Landau equation. These examples often
occur as simplest possible models, namely as amplitude or modulation equations, for
some real world phenomena such as nonlinear waves and pattern formation. Part IV
explores in more detail the connections between such complicated physical systems
ERHXLIVIHYGIHQSHIPW*SVQER]QSHIPWEQEXLIQEXMGEPP]VMKSVSYWNYWXM½GEXMSRF]
approximation results is given.
The parts of the book are kept as self-contained as possible. The book is suitable for
self-study, and there are various possibilities to build one- or two-semester courses
from the book.

For additional information

and updates on this book, visit
www.ams.org/bookpages/gsm-182

www.ams.org
GSM/182