Universitext

Universitext

Series Editors:

Sheldon Axler
San Francisco State University
Vincenzo Capasso
Università degli Studi di Milano
Carles Casacuberta
Universitat de Barcelona
Angus J. MacIntyre
Queen Mary, University of London
Kenneth Ribet
University of California, Berkeley
Claude Sabbah
CNRS, École Polytechnique
Endre Süli
University of Oxford
Wojbor A. Woyczynski
Case Western Reserve University

Universitext is a series of textbooks that presents material from a wide variety

of mathematical disciplines at master’s level and beyond. The books, often well
class-tested by their author, may have an informal, personal, even experimental
approach to their subject matter. Some of the most successful and established
books in the series have evolved through several editions, always following the
evolution of teaching curricula, into very polished texts.

Thus as research topics trickle down into graduate-level teaching, first textbooks
written for new, cutting-edge courses may make their way into Universitext.

For further volumes:

www.springer.com/series/223
Karl-G. Grosse-Erdmann r Alfred Peris Manguillot

Linear Chaos
Karl-G. Grosse-Erdmann Alfred Peris Manguillot
Institut de Mathématique Institut Universitari de Matemàtica
Université de Mons Pura i Aplicada
Le Pentagone, 7000 Mons Universitat Politècnica de València
Belgium Edifici 7A, 46022 València
[email protected] Spain
[email protected]

ISSN 0172-5939 e-ISSN 2191-6675

Universitext
ISBN 978-1-4471-2169-5 e-ISBN 978-1-4471-2170-1
DOI 10.1007/978-1-4471-2170-1
Springer London Dordrecht Heidelberg New York

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Control Number: 2011936377

Mathematics Subject Classification: 47-01, 37B05, 47A16, 47D06

© Springer-Verlag London Limited 2011

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as per-
mitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced,
stored or transmitted, in any form or by any means, with the prior permission in writing of the publish-
ers, or in the case of reprographic reproduction in accordance with the terms of licenses issued by the
Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to
the publishers.
The use of registered names, trademarks, etc., in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant laws and regulations and therefore free
for general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the information
contained in this book and cannot accept any legal responsibility or liability for any errors or omissions
that may be made.

Cover design: VTeX UAB, Lithuania

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

According to a widely held view, chaos is intimately linked to nonlinearity. It

is usually taken to be self-evident that a linear system behaves in a predictable
manner.
However, as early as 1929, G.D. Birkhoff obtained an example of a linear
operator that possesses an important ingredient of chaos: the existence of a
dense orbit. Later, G.R. MacLane (1952) found the same phenomenon for
the differentiation operator, which, after all, is the fundamental operator
in analysis. And S. Rolewicz (1969) showed that not only nonlinear shifts
but also linear shifts can have dense orbits. Motivated by these sporadic
examples, researchers began in the nineteen-eighties to study the dynamical
properties of general linear operators; henceforth, operators with a dense
orbit were called hypercyclic. As a first important result a useful condition
for hypercyclicity, the so-called Hypercyclicity Criterion, was obtained.
A further decisive step was taken by G. Godefroy and J.H. Shapiro (1991).
Not only did they come up with whole new classes of hypercyclic operators,
they also proposed to accept Devaney’s definition of (nonlinear) chaos as the
right definition for linear chaos: a linear operator is chaotic if it has a dense
orbit, it has a dense set of periodic points and it has sensitive dependence on
initial conditions. They then showed that many linear operators are chaotic,
including the three classical operators of Birkhoff, MacLane and Rolewicz.
The fact that chaos for linear systems has only been discovered recently
is easily explained: as Rolewicz showed, hypercyclicity, and hence also linear
chaos, requires an infinite-dimensional setting.
Over the last quarter of a century, the study of hypercyclic and chaotic
operators has turned into a fascinating and very active research area. It has
produced an astounding number of deep and beautiful results. As represen-
tative examples we mention here only Ansari’s theorem that every power of
a hypercyclic operator is hypercyclic, the Ansari–Bernal theorem that every
infinite-dimensional separable Banach space supports a hypercyclic opera-
tor, Grivaux’s theorem that every Hilbert space operator is the sum of two

v
vi Preface

chaotic operators, and the Bourdon–Feldman theorem that any orbit that is
somewhere dense is (everywhere) dense.
It seems fair to say that while research in linear dynamics is still expanding
in both depth and breadth, the foundations have reached a certain stage of
maturity. At the same time the basic ideas as well as the applications of the
field have a broad appeal also to nonspecialists.
It is therefore our aim to make the theory of hypercyclic operators and
linear chaos accessible to a wider audience. The book is aimed at advanced
undergraduate or beginning graduate students, both as a basis for a lecture
course and for self-study. We have strived at a self-contained exposition. Each
chapter contains a large number of exercises and ends with a section that gives
references and directs the reader to further literature.
We have tried to keep the necessary prerequisites for reading this book
to a minimum. Since the concept of a hypercyclic operator requires both a
topological and a linear structure, the reader is supposed to be familiar with
metric spaces (up to the Baire category theorem) and with the basic theory of
Hilbert and Banach spaces, as it is often presented in advanced undergraduate
courses on analysis. Moreover, since many examples in the theory are given
by operators on spaces of holomorphic functions the reader is also expected
to have had an introductory course on complex analysis. Additional, more
advanced tools that are only needed occasionally will be provided in the two
appendices.
The book is divided into two parts. Part I presents an introduction to the
dynamics of linear operators. Its chapters form a unity and are best studied
in that order. In contrast, Part II covers selected topics from linear dynamics.
Its chapters are largely independent so that they can be read in an arbitrary
order. An occasional cross reference should pose no problem.
More specifically, Chapter 1 introduces the reader to the fundamental con-
cepts of the theory of (not necessarily linear) dynamical systems. Its high-
lights are the Birkhoff transitivity theorem, which is of fundamental impor-
tance for all that follows, and a close study of the various concepts of maps
with complicated behaviour, including chaotic maps. In Chapter 2, the no-
tions and results from the first chapter are revisited in the context of linear-
ity. Among other things it is proved that the operators of Birkhoff, MacLane
and Rolewicz are chaotic, and that linear dynamics can be as complicated as
nonlinear dynamics. We begin the chapter with an introduction to a straight-
forward generalization of Banach spaces, the so-called Fréchet spaces; they
provide the setting for some important chaotic operators. Chapter 3 presents
several criteria for hypercyclicity and chaos, in increasing order of sophisti-
cation. It culminates in the Hypercyclicity Criterion, which is discussed in
detail. In Chapter 4, some important classes of hypercyclic and chaotic oper-
ators are described: weighted shift operators on sequence spaces, differential
and composition operators on spaces of holomorphic functions, and adjoint
multiplication operators. In addition to the shift operators, which are studied
throughout the book, the reader may want to concentrate on one or two ad-
Preface vii

ditional classes, depending on his or her own personal preference. In Chapter

5 we discuss the spectral properties of hypercyclic and chaotic operators. As
an application we derive properties that preclude hypercyclicity or chaos. Fi-
nally, Chapter 6 presents some of the deepest, most beautiful and most useful
results from linear dynamics. It contains, among other things, Ansari’s theo-
rem on the powers of hypercyclic operators, the Bourdon–Feldman theorem
on somewhere dense orbits and the León–Müller theorem on the hypercyclic-
ity of unimodular multiples of hypercyclic operators.
In the second part, Chapter 7 discusses the continuous analogue of hyper-
cyclic and chaotic operators in the form of semigroups. While the theories run
parallel in great parts, hypercyclic and chaotic semigroups have important
applications to partial differential equations. In Chapter 8 we obtain, among
other things, the Ansari–Bernal theorem on the existence of hypercyclic op-
erators. We also discuss here the richness of the set of hypercyclic operators.
The contents of Chapter 9 are motivated by recent work on the application
of ergodic theory to linear dynamics. While the technical difficulties involved
prevent us from studying these tools here, we will discuss a new concept that
has come out of these investigations, the frequently hypercyclic operators.
Chapter 10 is devoted to the question of whether there is, for a given opera-
tor, an infinite-dimensional closed subspace all of whose nonzero vectors are
hypercyclic, while Chapter 11 studies the existence of common hypercyclic
vectors for (uncountable) families of operators. The final Chapter 12 treats
hypercyclicity and linear chaos in their most natural (and most general) set-
ting, the topological vector spaces. After a brief introduction to such spaces
we revisit many of the results previously obtained in the book and show that
they hold in great generality.
At this point it seems important to add a disclaimer concerning our strat-
egy for attaching names to theorems. In keeping with the usual practice in
mathematics we have attributed results to the author(s) who first proved
them. But the form in which the result is presented in the book may well be
due to additional contributions from further authors. The reader is advised
to consult the relevant sources and comments section for complete references.
We would like to say a few words about the differences with a recent mono-
graph by F. Bayart and É. Matheron [44] on the topic of linear dynamics.
While their book is intended to be accessible to readers with a reasonable
background in functional analysis at the graduate level, we have tried to make
our text more basic and self-contained, so that students should be able to
follow it at an earlier stage of their studies without additional material. In a
certain sense, the two books are complementary. While we cover the founda-
tions and the main body of linear dynamics in detail, Bayart and Matheron
proceed to present some technically demanding topics like the counterexam-
ples to Herrero’s problem due to De la Rosa–Read and Bayart–Matheron, the
applications of ergodic theory to linear dynamics, or Read-type operators for
which every nonzero vector is hypercyclic.
viii Preface

The starting point of this book was a mini-workshop on hypercyclicity and

linear chaos at Oberwolfach in August 2006, where we agreed that a book
on this topic ought to be written. Since then, stays at the Berlin Mathemati-
cal School/Technische Universität Berlin (2007), the Centre International de
Rencontres Mathématiques in Marseille (2008, 2009, 2010), and the Universi-
tat Politècnica de València/Universitat de València (2008), made it possible,
with their support and their perfect working conditions, that the book was
written. We are, in particular, grateful to Günter M. Ziegler (Berlin), Pascal
Chossat (CIRM) and Manuel Maestre (València) for the invitations to their
institutions.
It is with great pleasure that we thank Richard Aron, Salud Bartoll, Luis
Bernal-González, Juan Bès, Manuel De la Rosa, Elisabetta Mangino, Étienne
Matheron, Quentin Menet, Raymond Mortini, Joel Shapiro, Dirk Werner, and
the referees; they have helped us, in various ways and at different stages of the
project, by providing many valuable suggestions and interesting discussions,
by detecting embarrassing mistakes and by offering constructive remarks. We
are much indebted to Klaus Cloppenburg for providing us with the index for
the book. Our thanks also go to the editors and their staff at Springer-Verlag,
London, especially Karen Borthwick, Joerg Sixt, and Lauren Stoney, for their
valuable assistance.
We acknowledge the support of MICINN and FEDER, Project MTM2007-
64222, and Generalitat Valenciana, Project PROMETEO/2008/101.
Above all we thank our partners, Klaus and Olga, for their love, encour-
agement and support. The first author wants to dedicate the book to the
memory of his father who could not live to see its completion. The second
author dedicates the book to Olga, his parents, and his family.

Mons, València Karl-G. Grosse-Erdmann

June 2011 Alfred Peris Manguillot
Contents

Part I Introduction to linear dynamics

1 Topological dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Dynamical systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Topologically transitive maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Mixing maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 Weakly mixing maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6 Universality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2 Hypercyclic and chaotic operators . . . . . . . . . . . . . . . . . . . . . . . . 31

2.1 Linear dynamical systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Hypercyclic operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 Linear chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.4 Mixing operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.5 Weakly mixing operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.6 The set of hypercyclic vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.7 Linear vs nonlinear maps, and finite vs infinite dimension . . . . 54
2.8 Hypercyclicity and complex dynamics . . . . . . . . . . . . . . . . . . . . . 58
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

3 The Hypercyclicity Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

3.1 Criteria for chaos and mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.2 Weak mixing and the Hypercyclicity Criterion . . . . . . . . . . . . . 73
3.3 Equivalent formulations of the Hypercyclicity Criterion . . . . . . 79
3.4 Hypercyclic sequences of operators . . . . . . . . . . . . . . . . . . . . . . . . 83
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

ix
x Contents

4 Classes of hypercyclic and chaotic operators . . . . . . . . . . . . . . 89

4.1 Weighted shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2 Differential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.3 Composition operators I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.4 Adjoint multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.5 Composition operators II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

5 Necessary conditions for hypercyclicity and chaos . . . . . . . . . 137

5.1 Spectral properties of hypercyclic and chaotic operators . . . . . 137
5.2 Classes of non-hypercyclic operators on Banach spaces . . . . . . 141
5.3 Classes of non-hypercyclic operators on Hilbert spaces . . . . . . 150
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

6 Connectedness arguments in linear dynamics . . . . . . . . . . . . . 161

6.1 Ansari’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.2 Somewhere dense orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.3 Multi-hypercyclic operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.4 Hypercyclic semigroup actions . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Part II Selected topics

7 Dynamics of semigroups, with applications to differential

equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.1 Semigroups of operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.2 Hypercyclic and chaotic C0 -semigroups . . . . . . . . . . . . . . . . . . . . 185
7.3 Discretizations of C0 -semigroups . . . . . . . . . . . . . . . . . . . . . . . . . 192
7.4 Criteria for the hypercyclicity and chaos of C0 -semigroups . . . 196
7.5 Applications of C0 -semigroups to differential equations . . . . . . 201
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

8 Existence of hypercyclic operators . . . . . . . . . . . . . . . . . . . . . . . . 213

8.1 Mixing perturbations of the identity . . . . . . . . . . . . . . . . . . . . . . 213
8.2 Existence of mixing operators and semigroups . . . . . . . . . . . . . . 219
8.3 Density of hypercyclic operators . . . . . . . . . . . . . . . . . . . . . . . . . . 223
8.4 Existence of hypercyclic operators with prescribed orbits . . . . 225
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Contents xi

9 Frequently hypercyclic operators . . . . . . . . . . . . . . . . . . . . . . . . . 235

9.1 Frequently recurrent orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
9.2 The Frequent Hypercyclicity Criterion . . . . . . . . . . . . . . . . . . . . 242
9.3 An eigenvalue criterion for frequent hypercyclicity . . . . . . . . . . 249
9.4 Structural properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264

10 Hypercyclic subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

10.1 Operators with hypercyclic subspaces . . . . . . . . . . . . . . . . . . . . . 267
10.2 Hypercyclic left-multiplication operators . . . . . . . . . . . . . . . . . . . 276
10.3 Operators without hypercyclic subspaces . . . . . . . . . . . . . . . . . . 282
10.4 Further operators with hypercyclic subspaces . . . . . . . . . . . . . . 286
10.5 The Fréchet space setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301

11 Common hypercyclic vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305

11.1 The Common Hypercyclicity Criterion . . . . . . . . . . . . . . . . . . . . 305
11.2 Common hypercyclic vectors for multiples of an operator . . . . 313
11.3 Further examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
11.4 Common hypercyclic subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329

12 Linear dynamics in topological vector spaces . . . . . . . . . . . . . . 331

12.1 Topological vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
12.2 Hypercyclicity, topological transitivity, and linear chaos . . . . . 334
12.3 Dynamical transference principles . . . . . . . . . . . . . . . . . . . . . . . . 338
12.4 Mixing and weakly mixing operators . . . . . . . . . . . . . . . . . . . . . . 342
12.5 Criteria for weak mixing, mixing and chaos . . . . . . . . . . . . . . . . 345
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
Sources and comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

Appendix A – Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

Appendix B – Spectral theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
 Part I
Introduction to linear dynamics
Chapter 1
Topological dynamics

This chapter provides an introduction to the theory of (not necessarily linear)

dynamical systems. Fundamental concepts such as topologically transitive,
chaotic and (weakly) mixing maps are defined and illustrated with typical
examples. The Birkhoff transitivity theorem is derived as a crucial tool for
showing that a map has a dense orbit. Moreover we obtain several character-
izations of weakly mixing maps that will be of great significance later on.
The interest in this chapter is that it derives results on dynamical systems
that do not require linearity. From Chapter 2 onwards, all our systems will
be linear.

1.1 Dynamical systems

The theory of dynamical systems studies the long-term behaviour of evolving

systems.
As a motivating example we consider the size of a population, which we
assume to be given by the value Nn at discrete times n = 0, 1, 2, . . . . In a
simple model the size at time n + 1 will only depend on the size at time n.
The population is then described by a law

Nn+1 = T (Nn ), n = 0, 1, 2, . . . ,

where T is a suitable map. It follows that

Nn = (T ◦ . . . ◦ T )(N0 ), n = 1, 2, . . .

with n applications of the map T . Thus the behaviour of the population is

completely determined by the initial population N0 and the map T .
More generally, we assume that the possible states of a (physical, biologi-
cal, economic, . . . , or abstract) system are described by the elements from a
set X and that evolution of the system is described by a map T : X → X;

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 3

DOI 10.1007/978-1-4471-2170-1_1, © Springer-Verlag London Limited 2011
4 1 Topological dynamics

that is, if xn ∈ X is the state of the system at time n ≥ 0, then

xn+1 = T (xn ), n = 0, 1, 2, . . . .

Since we want to measure changes in the values xn , we require the underlying

space to be a metric space. And since we want small changes in xn only to
result in small changes in xn+1 we require continuity of T .
Definition 1.1. A (discrete) dynamical system is a pair (X, T ) consisting of
a metric space X and a continuous map T : X → X.
Often we will simply call T (when the underlying space X is taken for
granted) or T : X → X a dynamical system. Moreover we adopt the notation
used in operator theory to write T x for T (x).
What we are interested in is the evolution of the system that starts with
a certain state x0 . For this we define the iterates T n : X → X, n ≥ 0, by the
n-fold iteration of T ,

Tn = T ◦ ... ◦ T (n times)

with
T 0 = I,
the identity on X.
Definition 1.2. Let T : X → X be a dynamical system. For x ∈ X we call

orb(x, T ) = {x, T x, T 2 x, . . .}

the orbit of x under T .

Returning to the previous discussion, suppose that the size Nn of a pop-
ulation changes proportionally to its actual size, that is, it follows the law
Nn+1 − Nn
= γ, n ≥ 0,
Nn
with some constant γ > −1. One may write this equivalently as

Nn+1 = (1 + γ)Nn ,

so that the corresponding dynamical system is given by

T : R+ → R+ , T x = (1 + γ)x.

The orbit of x ∈ R+ can be calculated explicitly as

orb(x, T ) = {(1 + γ)n x ; n ≥ 0}.

Thus, the orbit tends to 0, x and ∞ for −1 < γ < 0, γ = 0 and γ > 0,
respectively.
1.1 Dynamical systems 5

As a more realistic model for the evolution of a population the following

has been suggested. If we assume that the environment limits the size of the
population by a certain number L > 0 then we might assume the law to be
Nn+1 − Nn
= γ(L − Nn ), γ > 0.
Nn

Rescaling by Mn = (L + γ −1 )−1 Nn and setting μ = γL + 1 we obtain that

Mn+1 = μMn (1 − Mn ), n ≥ 0.

This refined model leads to the following dynamical system.

Example 1.3. (Logistic map) Let μ ∈ R. The logistic map Lμ : R → R is
given by Lμ x = μx(1 − x), x ∈ R. Figure 1.1 shows the graph of Lμ for μ = 3.

Fig. 1.1 The logistic map L3 Fig. 1.2 The tent map T

We introduce several other popular dynamical systems.

Example 1.4. (a) (Quadratic map) A quadratic map is defined by the real
dynamical system Qc : R → R, x → x2 + c, with a parameter c ∈ R, or by
the corresponding complex dynamical system Qc : C → C, z → z 2 + c, with
c ∈ C.
(b) (Doubling map on the circle) Let T : C → C denote the square
n
function T z = z 2 . Its iterates are T n z = z 2 . It follows that the orbits
for points z with |z| < 1 tend to 0, while for |z| > 1 the orbits tend to
infinity. As we will see later, the dynamics of T for points on the unit circle
T = {z ∈ C ; |z| = 1} are much more interesting. Since T (T) ⊂ T, we usually
consider the dynamical system T : T → T, z → z 2 , the so-called doubling
map. The name refers to the fact that T doubles the argument of the complex
number z.
6 1 Topological dynamics

(c) (Circle rotation) The system T : T → T, z → eiα z, α ∈ [0, 2π[,

describes the rotation of the point z on the unit circle by the angle α. We
will see that its dynamical behaviour depends to a large extent on the question
of whether the rotation is rational (α ∈ πQ) or irrational (α ∈ / πQ).
(d) (Tent map) The tent map is given by T : [0, 1] → [0, 1], T x = 2x, if
x ∈ [0, 12 ], and T x = 2 − 2x, if x ∈ ] 12 , 1]. The name derives from the shape
of its graph; see Figure 1.2.
(e) (Doubling map on the interval) We consider the interval [0, 1] in
which we identify 0 and 1; the metric on this space is given by d(x, y) =
min(|x − y|, 1 − |x − y|). Then T : [0, 1] → [0, 1], x → 2x (mod 1) describes a
dynamical system.
(f) (Shift on the interval) When we identify again 0 and 1, the map
T : [0, 1] → [0, 1], x → x + α (mod 1) with α ∈ [0, 1[ describes the shift by α,
modulo 1, of any point on the unit interval.
Every mathematical theory has its notion of isomorphism. When do we
want to consider two dynamical systems S : Y → Y and T : X → X as
equal? There should be a homeomorphism φ : Y → X such that, when
x ∈ X corresponds to y ∈ Y via φ then T x should correspond to Sy via φ. In
other words, if x = φ(y) then T x = φ(Sy). This is equivalent to saying that
T ◦ φ = φ ◦ S.
We recall that a homeomorphism is a bijective continuous map whose
inverse is also continuous. In many applications, however, it is already enough
to demand that φ is continuous with dense range.

Definition 1.5. Let S : Y → Y and T : X → X be dynamical systems.

(a) Then T is called quasiconjugate to S if there exists a continuous map
φ : Y → X with dense range such that T ◦ φ = φ ◦ S, that is, the diagram
S
Y −−−−→ Y
⏐ ⏐
⏐
φ
⏐φ

T
X −−−−→ X
commutes.
(b) If φ can be chosen to be a homeomorphism then S and T are called
conjugate.

Conjugacy is clearly an equivalence relation between dynamical systems,

and conjugate dynamical systems have the same dynamical behaviour. What
makes this notion even more interesting is the fact that it is by no means
always obvious if two systems are conjugate or not.

Example 1.6. We refer to the various dynamical systems introduced above.

(a) For any μ = 0, 2, the logistic maps Lμ and L2−μ are conjugate; one
can take an affine function x → ax + b for φ, as is easily verified. It therefore
suffices to study these maps for μ ≥ 1.
1.2 Topologically transitive maps 7

(b) Any logistic map Lμ , μ = 0, is conjugate to a suitable real quadratic

map Qc ; again, the conjugacy can be given by an affine function.
(c) The logistic map L4 , when restricted to the interval [0, 1], is conjugate
to the tent map. In fact, an easy calculation shows that one may take φ(x) =
sin2 ( π2 x).
(d) When we identify the points 0 and 1, the map φ : [0, 1] → T, x → e2πix
is a homeomorphism. It clearly defines a conjugacy between the doubling
map on [0, 1] and the doubling map on the circle; and it defines a conjugacy
between the shift by α on [0, 1] and the circle rotation by the angle 2πα.

Definition 1.7. We say that a property P for dynamical systems is preserved

under (quasi)conjugacy if the following holds: if a dynamical system S : Y →
Y has property P then every dynamical system T : X → X that is (quasi)
conjugate to S also has property P.

For example, the property of having dense range is clearly preserved under
quasiconjugacy, while the property of being surjective is preserved under
conjugacy but not under quasiconjugacy.

1.2 Topologically transitive maps

One way of defining a new dynamical system from a given dynamical system
T is by restricting it to a subset. However, one has to ensure that T maps
this subset into itself.

Definition 1.8. Let T : X → X be a dynamical system. Then a subset

Y ⊂ X is called T -invariant or invariant under T if T (Y ) ⊂ Y .

Thus, if Y ⊂ X is T -invariant, then T |Y : Y → Y is also a dynamical

system.

Example 1.9. The interval [0, 1] is invariant under the logistic map Lμ for
0 ≤ μ ≤ 4.

The study of a mathematical object is often simplified by breaking it up

into smaller parts and by studying these separately. If such a splitting is not
possible then one usually says that the object is irreducible. In the case of
dynamical systems, we might regard T : X → X as irreducible if X cannot be
divided into two T -invariant subsets with nonempty interior. In that direction
we have the following result.

Proposition 1.10. Let T : X → X be a dynamical system. Then we have

the implications (i) ⇐= (ii) ⇐⇒ (iii) ⇐⇒ (iv) ⇐⇒ (v), where
(i) X cannot be written as X = A ∪ B with disjoint T -invariant subsets
A, B such that A and B have nonempty interior;
8 1 Topological dynamics

(ii) X cannot be written as X = A ∪ B with disjoint subsets A, B such that

A is T -invariant and A and B have nonempty interior;
(iii) for any pair U, V of nonempty open subsets of X there exists some n ≥ 0
such that T n (U ) ∩ V = ∅; ∞
(iv) for any nonempty open subset U of X the set n=0 T n (U ) is dense in
X; ∞
(v) for any nonempty open subset U of X the set n=0 T −n (U ) is dense in
X.

Proof. (ii)=⇒(i) is trivial.


(ii)=⇒(iv). Let A = ∞ n=0 T (U ) and B = X \ A. Then A is T -invariant,
n

and it has nonempty interior since it contains U . By (ii), B must have empty
interior, which implies that A is dense.
(iii)=⇒(ii). Suppose that X = A ∪ B, A ∩ B = ∅ and T (A) ⊂ A. Then
int(A) and int(B) are open sets with T n (int(A)) ∩ int(B) ⊂ A ∩ B = ∅ for all
n ≥ 0. By (iii) this can only be the case if either A or B has empty interior.
We clearly have that (iii)⇐⇒(iv). For (iii)⇐⇒(v) one need only note that
T n (U ) ∩ V = ∅ is equivalent to U ∩ T −n (V ) = ∅. 


We see that condition (iii) is slightly stronger than the irreducibility of a

dynamical system; see also Exercise 1.2.1. Since this condition will turn out
to be of fundamental importance for the theory it is given its own name.
Definition 1.11. A dynamical system T : X → X is called topologically
transitive if, for any pair U, V of nonempty open subsets of X, there exists
some n ≥ 0 such that T n (U ) ∩ V = ∅.

Fig. 1.3 Topological transitivity

Example 1.12. (a) The tent map is topologically transitive. To see this, note
that T n is the piecewise linear map with T n ( 22kn ) = 0, k = 0, 1, . . . , 2n−1 ,
and T n ( 2k−1
2n ) = 1, k = 1, . . . , 2
n−1
; see Figure 1.4. Thus, let U ⊂ [0, 1] be
nonempty and open. Then U contains some interval J := [ 2mn , m+1 2n
]. But
since [0, 1] = T n (J) ⊂ T n (U ), T n (U ) in fact meets every nonempty set V .
(b) The doubling map on the circle, T : T → T, z → z 2 , is also topolog-
ically transitive. In fact, every nonempty open set U ⊂ T contains a closed
arc of angle 2π2n , for some n ≥ 1. Since the map T doubles angles, we have
1.2 Topologically transitive maps 9

that T n (U ) contains a closed arc of angle 2π, hence T n (U ) = T meets every

nonempty set V .
(c) No rational rotation T : T → T, z → eiα z, is topologically transitive.
For example, if α = 2π π
n then the iterates of an open arc γ of angle n never meet
π
the arc ei n γ. In contrast, every irrational rotation is topologically transitive;
see Exercise 1.2.9.

Fig. 1.4 The iterates T 2 and T 3 of the tent map

Proposition 1.13. Topological transitivity is preserved under quasiconjuga-

cy.

Proof. Let T : X → X be quasiconjugate to S : Y → Y via φ : Y → X,

and let U and V be nonempty open subsets of X. Since φ is continuous and
of dense range, φ−1 (U ) and φ−1 (V ) are nonempty and open. Thus there are
y ∈ φ−1 (U ) and n ≥ 0 with S n y ∈ φ−1 (V ), which implies that φ(y) ∈ U and
T n φ(y) = φ(S n y) ∈ V . 


The equivalence of conditions (iv) and (v) in Proposition 1.10 implies the
following.

Proposition 1.14. Let T : X → X be a dynamical system with continuous

inverse T −1 . Then T is topologically transitive if and only if T −1 is.

Topological transitivity can be interpreted as saying that T connects all

nontrivial parts of X. This is automatically the case whenever there is a point
x ∈ X with dense orbit under T .

Proposition 1.15. Let T be a continuous map on a metric space X without

isolated points.
(a) If x ∈ X has dense orbit under T then so does each T n x, n ≥ 1.
(b) If T has a dense orbit then it is topologically transitive.
10 1 Topological dynamics

Proof. (a) This follows easily from the fact that orb(x, T )\{x, T x, . . . , T n−1 x}
is contained in orb(T n x, T ) and that, in every metric space without isolated
points, a dense set remains dense even after removing finitely many points.
(b) Suppose that x ∈ X has dense orbit under T . Let U and V be nonempty
open sets in X. Then there is some n ≥ 0 such that T n x ∈ U . By (a), also
T n x has dense orbit, so that there is some m ≥ n such that T m x ∈ V . This
implies that T m−n (U ) ∩ V = ∅.  

What is less obvious is that, in separable complete metric spaces, the

converse of this result is also true: topologically transitive maps must have a
dense orbit. The importance of this result for the theory of dynamical systems
can hardly be overemphasized. It was first obtained in 1920 by G. D. Birkhoff
in the context of maps on compact subsets of RN .

Theorem 1.16 (Birkhoff transitivity theorem). Let T be a continuous

map on a separable complete metric space X without isolated points. Then
the following assertions are equivalent:
(i) T is topologically transitive;
(ii) there exists some x ∈ X such that orb(x, T ) is dense in X.
If one of these conditions holds then the set of points in X with dense orbit
is a dense Gδ -set.

Proof. By the previous proposition, (ii) implies (i). For the converse, let T
be topologically transitive, and let D(T ) denote the set of points in X that
have dense orbit under T . Since X has a countable dense set {yj ; j ≥ 1},
the open balls of radius m1
around the yj , m, j ≥ 1, form a countable base
(Uk )k≥1 of the topology of X. Hence, x belongs to D(T ) if and only if, for
every k ≥ 1, there is some n ≥ 0 such that T n x ∈ Uk . In other words,
∞
∞ 

D(T ) = T −n (Uk ).
k=1 n=0

By continuity of T and Proposition 1.10, each set ∞
n=0 T
−n
(Uk ), k ≥ 0, is
open and dense. The Baire category theorem then implies that D(T ) is a
dense Gδ -set, and hence nonempty. 

We note that the absence of isolated points was not needed for the proof
that (i) implies (ii).

Example 1.17. It follows from Example 1.12 that the tent map and the dou-
bling map have dense orbits. Irrational rotations (that is, with α ∈
/ πQ) have
the stronger property that each of their orbits is dense; see Exercise 1.2.9.

Let us briefly reflect on the usefulness of the transitivity theorem. Sup-

pose that we are interested in the existence of a dense orbit under a given
map. Sometimes such a point presents itself with little effort, as is the case
for irrational rotations. But what if this is not the case? Without further
1.3 Chaos 11

information we will never be likely to stumble on a point with dense orbit.

In contrast, topological transitivity seems much easier to prove, as we have
seen, for example, in the case of the tent map: we need only connect any two
nonempty open sets by suitable iterates.
We stress, however, that in more general spaces topological transitivity
and the existence of a dense orbit need not coincide.

Example 1.18. Let X be the set of all points on the unit circle that are 2n th
roots of unity, for some n ≥ 1. By Example 1.12(b) the doubling map, re-
stricted to X, is topologically transitive, but clearly has no dense orbits.

This example shows that the completeness assumption in the Birkhoff

transitivity theorem cannot be dropped.

Proposition 1.19. The property of having a dense orbit is preserved under

quasiconjugacy.

Proof. Let T : X → X be quasiconjugate to S : Y → Y via φ : Y → X,

and let y ∈ Y have dense orbit under S. If U is a nonempty open subset of
X then φ−1 (U ) is nonempty and open, so that some S n y, n ≥ 0, belongs to
φ−1 (U ). But then T n φ(y) = φ(S n y) belongs to U . 


Example 1.20. By Examples 1.6(c) and 1.17, the logistic map L4 on [0, 1] has
a dense orbit.

1.3 Chaos

What is chaos? Even when we restrict the meaning of this word to determin-
istic chaos, that is, chaotic behaviour of a dynamical system, mathematicians
have come up with different answers to this question. We will follow here the
definition that was suggested by Devaney in 1986. It has three ingredients,
which we discuss in turn.
The first ingredient tries to capture the idea of the so-called butterfly
effect: small changes in the initial state may lead, after some time, to large
discrepancies in the orbit. In order to be able to perturb points we consider
only spaces without isolated points.

Definition 1.21. Let (X, d) be a metric space without isolated points. Then
a dynamical system T : X → X is said to have sensitive dependence on
initial conditions if there exists some δ > 0 such that, for every x ∈ X and
ε > 0, there exists some y ∈ X with d(x, y) < ε such that, for some n ≥ 0,
d(T n x, T n y) > δ. The number δ is called a sensitivity constant for T .

We stress that the definition involves the metric of the space. In the fol-
lowing examples we will always work with the usual metric.
12 1 Topological dynamics

Example 1.22. (a) Using our knowledge of the iterates of the tent map (see
Example 1.12), we can easily show that it has sensitive dependence on initial
conditions with sensitivity constant 1/4, say. Indeed, if x ∈ [0, 1] and ε > 0
then there is some n ≥ 0 such that the open ball of radius ε around x contains
points y1 and y2 with T n y1 = 0 and T n y2 = 1; thus |T n x − T n yj | ≥ 1/2 for
some j ∈ {1, 2}.
(b) A similar argument, based on the fact that the doubling map doubles
angles, shows that it has sensitive dependence on initial conditions.
(c) No circle rotation has sensitive dependence on initial conditions because
we clearly have that |T n z1 − T n z2 | = |z1 − z2 | for any z1 , z2 ∈ T.
The second ingredient of chaos demands that the system is irreducible in
the sense that the map T connects any nontrivial parts of the space. We saw
in Section 1.2 that this idea is well captured by the notion of topological
transitivity of the system.
The third ingredient demands that the system has many orbits with a
regular behaviour; more precisely, there should be a dense set of points with
periodic orbit.
Definition 1.23. Let T : X → X be a dynamical system.
(a) A point x ∈ X is called a fixed point of T if T x = x.
(b) A point x ∈ X is called a periodic point of T if there is some n ≥ 1
such that T n x = x. The least such number n is called the period of x. The
set of periodic points is denoted by Per(T ).
A point is periodic if and only if it is a fixed point of some iterate T n ,
n ≥ 1. Thus, for real functions T , one easily detects them by searching for
the points where the graphs of T n and the identity function meet.
Example 1.24. (a) Considering the iterates of the tent map (see Example
1.12), we find that in every interval [ 2mn , m+1
2n ] there is a periodic point of
period n. Thus, the tent map has a dense set of periodic points.
(b) The periodic points of the doubling map on the circle are exactly the
(2n − 1)st roots of unity, n ≥ 1, so that also the doubling map has a dense
set of periodic points.
(c) For any rational rotation T there is some N ≥ 1 such that T N = I, so
that every point is periodic. In contrast, irrational rotations have no periodic
points at all.
Proposition 1.25. The property of having a dense set of periodic points is
preserved under quasiconjugacy.
Proof. Let T : X → X be quasiconjugate to S : Y → Y via φ : Y → X,
and let U ⊂ X be a nonempty open set. Then φ−1 (U ), being also open and
nonempty, contains a point y with S n y = y for some n ≥ 1. Hence φ(y) ∈ U
and T n φ(y) = φ(S n y) = φ(y). 

Summarizing, we are led to Devaney’s definition of chaos.
1.3 Chaos 13

Definition 1.26 (Devaney chaos – preliminary version). Let (X, d) be

a metric space without isolated points. Then a dynamical system T : X → X
is said to be chaotic (in the sense of Devaney) if it satisfies the following
conditions:
(i) T has sensitive dependence on initial conditions;
(ii) T is topologically transitive;
(iii) T has a dense set of periodic points.

Example 1.27. By Examples 1.12, 1.22 and 1.24, the tent map and the dou-
bling map are chaotic, but no circle rotation is chaotic.

The definition of chaos has a serious blemish: sensitive dependence on

initial conditions is not preserved under conjugacy, or, which is the same,
it depends on the metric on the underlying space. We illustrate this by an
example.

Example 1.28. Let T : ]1, ∞[ → ]1, ∞[ be given by T x = 2x. Since |T n x −

T n y| = 2n |x − y| → ∞ whenever x = y, we have that T has sensitive
dependence on initial conditions with respect to the usual metric on ]1, ∞[.
But if we define d(x, y) = | log x − log y| then d is an equivalent metric for
which d(T n x, T n y) = d(x, y) for all x, y ∈ ]1, ∞[, which shows that T does
not have sensitive dependence on initial conditions with respect to d. On the
other hand, the two versions of T are conjugate when we take the identity
map as the linking homeomorphism.

Fortunately, one can drop sensitive dependence from Devaney’s definition

because it is implied by the other two conditions.

Theorem 1.29 (Banks–Brooks–Cairns–Davis–Stacey). Let X be a

metric space without isolated points. If a dynamical system T : X → X
is topologically transitive and has a dense set of periodic points then T has
sensitive dependence on initial conditions with respect to any metric defining
the topology of X.

Proof. We fix a metric d defining the topology of X. We first show that there

exists some constant η > 0 such that, for any point x ∈ X there is a periodic
point p such that
d(x, T n p) ≥ η for all n ∈ N0 .
Indeed, since X has no isolated points it is an infinite set, so that we can find
two periodic points p1 , p2 whose orbits are disjoint. Hence,

η := inf d(T m p1 , T n p2 )/2 > 0.

m,n∈N0

It then follows from the triangle inequality that, for any x ∈ X, either for
j = 1 or for j = 2 we have that d(x, T n pj ) ≥ η for all n ∈ N0 .
14 1 Topological dynamics

We now claim that T has sensitive dependence on initial conditions with

sensitivity constant δ := η/4 > 0. To this end, let x ∈ X and ε > 0. By
assumption there is a periodic point q such that

d(x, q) < min(ε, δ). (1.1)

Let q have period N . As we have seen above there is also a periodic point p
such that
d(x, T n p) ≥ η = 4δ for n ∈ N0 . (1.2)
Since T is continuous there is some neighbourhood V of p such that

d(T n p, T n y) < δ for n = 0, 1, . . . , N and y ∈ V . (1.3)

Finally, by topological transitivity of T we can find a point z and some k ∈ N0

such that d(x, z) < ε and T k z ∈ V . Let j ∈ N0 be such that k ≤ jN < k + N .
The triangle inequality, together with (1.2), (1.3) and (1.1), then yields that
     
d T jN q, T jN z = d T jN q, T jN −k T k z = d q, T jN −k T k z
   
≥ d x, T jN −k p − d T jN −k p, T jN −k T k z − d(x, q)
> 4δ − δ − δ = 2δ.

This implies that either d(T jN x, T jN q) > δ or d(T jN x, T jN z) > δ. Since

both z and q have a distance less than ε from x, the claim follows.  

This allows us to drop sensitive dependence from Devaney’s definition of

chaos; for simplicity we also extend it to all metric spaces.
Definition 1.30 (Devaney chaos). A dynamical system T : X → X is said
to be chaotic (in the sense of Devaney) if it satisfies the following conditions:
(i) T is topologically transitive;
(ii) T has a dense set of periodic points.

Fig. 1.5 Devaney chaos

1.3 Chaos 15

By Theorem 1.29, this is consistent with Definition 1.26. Proposition 1.13

and 1.25 immediately show the following.

Proposition 1.31. Devaney chaos is preserved under quasiconjugacy.

As an application of this fact we obtain the following from Examples 1.6

and 1.27.
Example 1.32. The logistic map L4 is chaotic on [0, 1].
For the tent map and the doubling map we have shown explicitly that they
satisfy the defining conditions of chaos. But their dynamics remain myste-
rious. To get a better feeling for the origin of chaos in a dynamical system
we now want to discuss briefly a special chaotic dynamical system whose
dynamical behaviour is quite transparent. On the space

Σ2 = {(xn )n∈N0 ; xn ∈ {0, 1}}

of all 0-1-sequences we study the map

σ : Σ 2 → Σ2 , σ(x0 , x1 , x2 , . . .) = (x1 , x2 , x3 , . . .).

As usual, sequences (xn )n , (yn )n , . . . will be denoted by x, y, . . . . We define a

topology on Σ2 by the metric
∞
|xn − yn |
d(x, y) = .
n=0
2n

Under this metric, σ is clearly a continuous map.

Definition 1.33. The dynamical system (Σ2 , σ) is called the shift on two
symbols.

For topological considerations in Σ2 the following easy result will be of

constant use.

Lemma 1.34. Let x, y ∈ Σ2 .

(a) If xj = yj for j = 0, 1, . . . , n then d(x, y) ≤ 21n .
(b) If d(x, y) < 21n then xj = yj for j = 0, 1, . . . , n.

In particular, a sequence of points in Σ2 converges if and only if each

coordinate converges. It also follows easily that Σ2 is a compact metric space
without isolated points; in fact it is homeomorphic to the Cantor set.
As promised, the dynamics of this system are completely transparent.

Proposition 1.35. (a) A point x ∈ Σ2 is periodic under σ if and only if the

sequence x = (xn )n is periodic.
(b) A point x ∈ Σ2 has dense orbit under σ if and only if every finite
0-1-sequence appears as a block in x.
16 1 Topological dynamics

Proof. (a) is obvious from the definition of the map σ.

For (b), suppose that x has dense orbit. Let (y0 , . . . , ym ) be a finite 0-1-
sequence and let y = (y0 , . . . , ym , 0, 0, . . .). Since there is some n ≥ 0 such that
d(y, σ n x) < 21m , Lemma 1.34(b) implies that (xn , . . . , xn+m ) = (y0 , . . . , ym ).
The converse follows similarly with Lemma 1.34(a).  

Theorem 1.36. The shift on two symbols is chaotic.

Proof. Since the set of all finite 0-1-sequences is countable, one can construct
a 0-1-sequence that contains each finite 0-1-sequence as a block. Hence, by
Proposition 1.35(b), σ has a dense orbit, which shows that it is topologically
transitive because Σ2 has no isolated points.
Now, let x ∈ Σ2 and m ∈ N. Then the sequence y defined by yj(m+1)+k =
xk , 0 ≤ k ≤ m, j ≥ 0 is periodic, hence a periodic point for σ. Moreover, by
Lemma 1.34, we have that d(x, y) ≤ 21m . This shows that the periodic points
are dense in Σ2 .
Altogether, σ is chaotic.  

It will be a recurrent theme in this book that shifts create chaos. And in
many cases maps are chaotic precisely because there is an underlying shift.
We illustrate this by the doubling map.

Example 1.37. The doubling map is given by T z = z 2 , z ∈ T. If we write

z = exp(2πiα) with 0 ≤ α < 1 then we may represent α in binary form as

∞
xn
α= , xn ∈ {0, 1}.
n=0
2n+1

In this representation we have that T maps z into

∞
xn 
∞
xn+1 
z 2 = exp(2πi2α) = exp 2πix0 + 2πi = exp 2πi .
n=1
2n n=0
2n+1

In other words, the doubling map acts as a shift on the binary representation
of the argument of z. In this form the dynamics of the doubling map become
much clearer. To put it formally, the map

∞
xn 
φ : Σ2 → T, (xn )n → exp 2πi
n=0
2n+1

provides a quasiconjugacy from the shift on two symbols to the doubling

map. This proves once more that the doubling map is chaotic.
1.4 Mixing maps 17

1.4 Mixing maps

We return to the discussion of topologically transitive maps. As we saw in

Example 1.12, the tent map and the doubling map have a strong form of
transitivity: T n (U ) intersects V not only for some n but for all sufficiently
large n ∈ N0 . This property carries a special name.

Definition 1.38. A dynamical system T : X → X is called mixing if, for

any pair U, V of nonempty open subsets of X, there exists some N ≥ 0 such
that
T n (U ) ∩ V = ∅ for all n ≥ N .

Example 1.39. As noted above, both the tent map and the doubling map are
mixing. An example of a topologically transitive system that is not mixing
will be given in Example 1.43.

As in the case of topological transitivity one obtains the following.

Proposition 1.40. The mixing property is preserved under quasiconjugacy.

Mixing maps have a remarkable permanence property. In order to describe

this we need to define products of maps.
Let S : X → X and T : Y → Y be dynamical systems. The Cartesian
product X ×Y is endowed with the product topology, which is induced by the
metric d((x1 , y1 ), (x2 , y2 )) = dX (x1 , x2 )+dY (y1 , y2 ), where dX and dY denote
the metrics in X and Y , respectively. A base for the topology is formed by
the products U × V of nonempty open sets U ⊂ X and V ⊂ Y .

Definition 1.41. Let S : X → X and T : Y → Y be dynamical systems.

Then the map S × T is defined by

S × T : X × Y → X × Y, (S × T )(x, y) = (Sx, T y).

Then S × T is clearly continuous, and for the iterates we have that

(S × T )n = S n × T n .

Products of more than two spaces or maps are defined similarly.

Proposition 1.42. Let S : X → X and T : Y → Y be dynamical systems.

Then we have the following:
(i) if S × T has a dense orbit then so do S and T ;
(ii) if S × T is topologically transitive then so are S and T ;
(iii) if S × T is chaotic then so are S and T ;
(iv) if S and T are topologically transitive and at least one of them is mixing
then S × T is topologically transitive;
(v) S × T is mixing if and only if both S and T are.
18 1 Topological dynamics

Proof. Obviously, S and T are quasiconjugate to S × T under the maps

(x, y) → x and (x, y) → y, respectively. Thus the preservation results ob-
tained so far imply (i), (ii) and (iii).
The assertions (iv) and (v) follow directly from the identity

(S × T )n (U1 × U2 ) ∩ (V1 × V2 ) = (S n (U1 ) ∩ V1 ) × (T n (U2 ) ∩ V2 ),

the fact that topological transitivity and the mixing property can be tested
on a base of the topology, and, for (iv), the observation of Exercise 1.2.4. 


In general, however, the product of two topologically transitive maps need

not be topologically transitive, even if S = T , as the following example shows;
hence some extra condition, as in the previous proposition, is needed.

Example 1.43. Let T : T → T, z → eiα z be a circle rotation. Then (T ×

T )n (z1 , z2 ) = (einα z1 , einα z2 ), and we observe that the quotient of the two
coordinates is z1 /z2 , independent of n. This is easily seen to imply that
T × T cannot be topologically transitive. On the other hand we know that
any irrational rotation is topologically transitive; see Exercise 1.2.9. This also
shows that no irrational rotation is mixing.

1.5 Weakly mixing maps

Having looked at the question of when a product of two topologically transi-

tive systems is topologically transitive, one may wonder when the product of
a topologically transitive map with itself is again topologically transitive. We
saw in Example 1.43 that this is not always the case. On the other hand, for
any mixing map T the product T × T is topologically transitive. This leads
us to the following notion.

Definition 1.44. A dynamical system T : X → X is called weakly mixing if

T × T is topologically transitive.

Since the products U × V of nonempty open sets U, V ⊂ X form a base

of the topology of X × X, T is weakly mixing if and only if, for any 4-tuple
U1 , U2 , V1 , V2 ⊂ X of nonempty open sets, there exists some n ≥ 0 such that

T n (U1 ) ∩ V1 = ∅ and T n (U2 ) ∩ V2 = ∅.

Observation 1.45. For any dynamical system,

mixing =⇒ weak mixing =⇒ topological transitivity.

Remark 1.46. By Example 1.43, any irrational circle rotation is topologically

transitive but not weakly mixing. On the other hand, it is not an easy matter
1.5 Weakly mixing maps 19

to construct examples of weakly mixing maps that are not mixing. How-
ever, in the context of linear operators that will be discussed in the following
chapters, such examples will appear abundantly as consequences of our in-
vestigations there; see, for example, Remark 4.10.
Proposition 1.47. The weak mixing property is preserved under quasicon-
jugacy.
Proof. If φ : Y → X defines a quasiconjugacy from S : Y → Y to T : X → X,
then φ × φ defines a quasiconjugacy from S × S to T × T . Now the result
follows from Proposition 1.13. 

As a consequence, we have the following as in Proposition 1.42.
Proposition 1.48. Let S : X → X and T : Y → Y be dynamical systems. If
S × T is weakly mixing then so are S and T .
In order to formulate the arguments involving weakly mixing maps more
succinctly we introduce the following useful concept.
Definition 1.49. Let T : X → X be a dynamical system. Then, for any sets
A, B ⊂ X, the return set from A to B is defined as

NT (A, B) = N (A, B) = {n ∈ N0 ; T n (A) ∩ B = ∅}.

We usually drop the index T when this causes no ambiguity. In this nota-
tion, T is topologically transitive (or mixing) if and only if, for any pair U, V
of nonempty open subsets of X, the return set

N (U, V ) is nonempty (or cofinite, respectively);

and T is weakly mixing if and only if, for any 4-tuple U1 , U2 , V1 , V2 ⊂ X of

nonempty open sets,

N (U1 , V1 ) ∩ N (U2 , V2 ) = ∅.

Note also that if T is topologically transitive then the return sets N (U, V )
are even infinite for any nonempty open sets U, V ; see Exercise 1.2.4.
Incidentally, we observe that the larger the sets A and B are, the larger
also is the return set N (A, B).
It is our aim to give several characterizations of the weak mixing property.
To do this we will provide a useful lemma that, due to its form, we will call
the 4-set trick. We note that the return sets N (A, B) refer to T .
Lemma 1.50 (4-set trick). Let T : X → X be a dynamical system, and let
U1 , V1 , U2 , V2 ⊂ X be nonempty open sets.
(a) If there is a continuous map S : X → X commuting with T such that

S(U1 ) ∩ U2 = ∅ and S(V1 ) ∩ V2 = ∅,

20 1 Topological dynamics

then there exist nonempty open sets U1 ⊂ U1 , V1 ⊂ V1 such that

N (U1 , V1 ) ⊂ N (U2 , V2 ) and N (V1 , U1 ) ⊂ N (V2 , U2 ).

If, moreover, T is topologically transitive then N (U1 , V1 ) ∩ N (U2 , V2 ) = ∅.

(b) If T is topologically transitive then

N (U1 , U2 ) ∩ N (V1 , V2 ) = ∅ =⇒ N (U1 , V1 ) ∩ N (U2 , V2 ) = ∅.

Fig. 1.6 The 4-set trick

Proof. (a) Since S is continuous, it follows from the hypothesis that we can
find nonempty open sets U1 ⊂ U1 and V1 ⊂ V1 such that S(U1 ) ⊂ U2 and
S(V1 ) ⊂ V2 . If n ∈ N (U1 , V1 ), then there exists some x ∈ U1 with T n x ∈ V1 .
Therefore T n Sx = ST n x ∈ V2 and Sx ∈ U2 , which yields that n ∈ N (U2 , V2 ).
By symmetry we also obtain that N (V1 , U1 ) ⊂ N (V2 , U2 ). If, moreover, T is
topologically transitive then

∅ = N (U1 , V1 ) ⊂ N (U1 , V1 ) ∩ N (U2 , V2 ).

(b) If T is topologically transitive and n ∈ N (U1 , U2 ) ∩ N (V1 , V2 ), then

N (U1 , V1 ) ∩ N (U2 , V2 ) = ∅ follows if (a) is applied to S := T n . 

The 4-set trick, simple as it is, already implies an important result that is
at first sight quite surprising: as soon as the product T × T is topologically
transitive, every higher product T × · · · × T also is.
Theorem 1.51 (Furstenberg). Let T : X → X be a weakly mixing dynam-
ical system. Then the n-fold product T × · · · × T is weakly mixing for each
n ≥ 2.
Proof. Since the n-fold product being weakly mixing amounts to the 2n-fold
product being topologically transitive, it suffices to show that every n-fold
product T × · · · × T is topologically transitive for n ≥ 2.
1.5 Weakly mixing maps 21

We proceed by induction, the case of n = 2 being trivial by definition.

Thus, suppose that the n-fold product T × . . . × T is topologically transitive.
To prove topological transitivity of the corresponding (n + 1)-fold product we
need to show that, given nonempty open sets Uk , Vk ⊂ X, k = 1, . . . , n + 1,
we have that

n+1
N (Uk , Vk ) = ∅. (1.4)
k=1

Indeed, since T is weakly mixing there is some m ∈ N0 such that T m (Un ) ∩

Un+1 = ∅ and T m (Vn ) ∩ Vn+1 = ∅. The 4-set trick then yields the existence
of nonempty open sets Un ⊂ Un , Vn ⊂ Vn with N (Un , Vn ) ⊂ N (Un , Vn ) ∩
N (Un+1 , Vn+1 ). On the other hand, the induction hypothesis implies that


n−1
N (Uk , Vk ) ∩ N (Un , Vn ) = ∅,
k=1

which implies (1.4). 


The next results show that in the definition of weak mixing one may reduce
the four open sets to three, and then to two open sets.

Fig. 1.7 Weak mixing (Propositions 1.52 and 1.53)

Proposition 1.52. A dynamical system T : X → X is weakly mixing if and

only if, for any nonempty open sets U, V1 , V2 ⊂ X, we have that

N (U, V1 ) ∩ N (U, V2 ) = ∅.

Proof. We need only show sufficiency of the condition. Thus, let U1 , U2 , V1 ,

V2 ⊂ X be arbitrary nonempty open sets. By hypothesis there is some n ∈
N (U1 , U2 ) ∩ N (U1 , V2 ). In particular U := U1 ∩ T −n (U2 ) and T −n (V2 ) are
nonempty open sets. By another application of the hypothesis we find some
m ∈ N (U, V1 ) ∩ N (U, T −n (V2 )). In particular, there exists some x ∈ U with
T m x ∈ T −n (V2 ). We then conclude that T m T n x = T n T m x ∈ V2 and T n x ∈
U2 , which yields that m ∈ N (U1 , V1 ) ∩ N (U2 , V2 ). 

22 1 Topological dynamics

Proposition 1.53. A dynamical system T : X → X is weakly mixing if and

only if, for any nonempty open sets U, V ⊂ X, we have that

N (U, U ) ∩ N (U, V ) = ∅.

Proof. It suffices to show that the stated condition implies the condition of
Proposition 1.52. Thus, let U, V1 , V2 ⊂ X be arbitrary nonempty open sets.
By hypothesis there exists some n ∈ N0 such that U1 := U ∩ T −n (V1 ) is a
nonempty open set. Since topologically transitive maps have dense range, the
hypothesis also implies that T −n (V2 ) is nonempty and open, so that there
exists some m ∈ N (U1 , U1 ) ∩ N (U1 , T −n (V2 )). Therefore there are x, y ∈ U1
with T m x ∈ U1 and T n T m y ∈ V2 . We then have that T n T m x ∈ V1 , which
implies that n + m ∈ N (U, V1 ) ∩ N (U, V2 ), as desired.  

For more characterizations of weak mixing in the same spirit we refer to

Exercise 1.5.1.
We finally characterize the weak mixing property in terms of the size of the
return sets N (U, V ), or equivalently, in terms of the topological transitivity
of certain subsequences (T nk )k ; for the notion of topological transitivity for
a sequence of maps we refer to the next section.
A strictly increasing sequence (nk )k of positive integers is called syndetic
if
sup(nk+1 − nk ) < ∞.
k≥1

Likewise, a subset A of N0 is called syndetic if the increasing sequence of

positive integers forming A is syndetic, or equivalently, if its complement
does not contain arbitrarily long intervals.

Theorem 1.54. Let T : X → X be a dynamical system. Then the following

assertions are equivalent:
(i) T is weakly mixing;
(ii) for any pair U, V ⊂ X of nonempty open sets, N (U, V ) contains arbi-
trarily long intervals;
(iii) for any syndetic sequence (nk )k , the sequence (T nk )k is topologically
transitive.

Proof. (i)=⇒(ii). Let U, V ⊂ X be nonempty open sets, and let m ∈ N. As

in the proof of Proposition 1.53, each set T −k (V ), k = 1, . . . , m, is nonempty
and open. Since, by Furstenberg’s theorem, the m-fold product map T ×· · ·×T
is topologically transitive, there is some n ∈ N such that

T n (U ) ∩ T −k (V ) = ∅ for k = 1, . . . , m.

This implies that T n+k (U ) ∩ V = ∅ for k = 1, . . . , m.

(ii)=⇒(i). By Proposition 1.52 it suffices to show that, given any nonempty
open subsets U, V1 , V2 ⊂ X, N (U, V1 ) ∩ N (U, V2 ) = ∅. First, by (ii) there is
some m ∈ N (V1 , V2 ) and therefore a nonempty open set V3 ⊂ V1 such that
1.6 Universality 23

T m (V3 ) ⊂ V2 . Also by (ii) there is some k ∈ N0 such that k + j ∈ N (U, V3 )

for j = 0, 1, . . . , m. In particular we have that k + m ∈ N (U, V1 ) and

T k+m (U ) ∩ V2 ⊃ T k+m (U ) ∩ T m (V3 ) ⊃ T m (T k (U ) ∩ V3 ) = ∅.

We conclude that k + m ∈ N (U, V1 ) ∩ N (U, V2 ).

(ii)⇐⇒(iii). This follows immediately from the definitions and the fact
that a subset of N0 contains arbitrarily long intervals if and only if it meets
every syndetic sequence.  

Condition (ii) in this result shows nicely how weak mixing sits between
topological transitivity and mixing.

1.6 Universality

The basic concepts introduced so far in this chapter allow a far-reaching

generalization. The orbit of a point x under a map T is obtained by applying
the iterates T n , n = 0, 1, 2, . . ., of T to x. Instead, one could think of applying
arbitrary maps Tn , n = 0, 1, 2, . . ., to x; in this case we need not even have
that the Tn are self-maps.

Definition 1.55. Let X and Y be metric spaces, and let Tn : X → Y ,

n ∈ N0 , be continuous maps. Then the orbit of x under (Tn )n is defined as

orb(x, (Tn )) = {Tn x ; n ∈ N0 }.

An element x ∈ X is called universal for (Tn )n if it has dense orbit under

(Tn )n .

An interesting and nontrivial example is provided by universal Taylor se-

ries: it can be shown that there exists an infinitely differentiable function
f : R → R with f (0) = 0 such that, for any continuous function g : R → R
with g(0) = 0, there exists an increasing sequence (nk )k of positive integers
such that

nk
f (ν) (0)
xν → g(x) uniformly on any compact subset of R.
ν=0
ν!

In this case, Tn is the map that associates to f its Taylor polynomial of degree
n at 0.
The theory of universality will not be developed in any depth in this book.
We note that there is a difference in philosophy between universality and
topological dynamics: in the former one is interested in the universal elements
and their properties while in the latter the focus is rather on the map and
its properties.
24 1 Topological dynamics

However, occasionally the study of the dynamics of a single map requires

looking at orbits under general sequences of maps; Theorem 1.54 has already
provided such an example. For this reason we consider here briefly how the
concepts and results of this chapter can be generalized to universality.

Definition 1.56. Let Tn : X → Y , n ∈ N0 , be continuous maps between

metric spaces X and Y . Then (Tn )n is called topologically transitive if, for
any pair U ⊂ X and V ⊂ Y of nonempty open sets, there is some n ≥ 0 such
that
Tn (U ) ∩ V = ∅;
it is mixing if the same holds for all sufficiently large n, and it is weakly
mixing if (Tn × Tn )n is topologically transitive on X × X.

Now, many of the results in this chapter extend, at least under suitable
assumptions, to general sequences. We will content ourselves here with some
examples.

Theorem 1.57 (Universality Criterion). Let X be a complete metric

space, Y a separable metric space and Tn : X → Y , n ∈ N0 , continuous
maps. Then the following assertions are equivalent:
(i) (Tn )n is topologically transitive;
(ii) there exists a dense set of points x ∈ X such that orb(x, (Tn )) is dense
in Y .
If one of these conditions holds then the set of points in X with dense orbit
is a dense Gδ -set.

Proof. Suppose that (ii) holds. If U and V are nonempty open sets of X and
Y , respectively, then there exists some x ∈ U with dense orbit under (Tn )n ,
so that there exists some n ≥ 0 with Tn x ∈ V . This implies (i).
The converse implication and the fact that the set of points with dense
orbit is a dense Gδ -set can be proved exactly as in the proof of the Birkhoff
transitivity theorem.  

Typically, results on iterates of maps have a good chance of extending to

sequences (Tn )n if they consist of commuting self-maps Tn : X → X of dense
range. For example, for such sequences the Birkhoff transitivity theorem has
a perfect analogue; see Exercise 1.6.2.

Remark 1.58. If we define the return sets

N (A, B) = {n ∈ N0 ; Tn (A) ∩ B = ∅}

then part (a) of the 4-set trick remains valid for sequences (Tn )n of self-maps
if the map S commutes with all Tn , n ≥ 0, as does part (b) for commuting
self-maps. As a consequence, Furstenberg’s theorem also holds for commuting
sequences (Tn )n , as does Proposition 1.52.
Exercises 25

Exercises

Exercise 1.1.1. Show that for μ = 2 the iterates of the logistic map L2 are given by
 n 
Ln
2x =
1
2 1 − (1 − 2x)2 , n ≥ 0.

Deduce from this the long-term behaviour of the orbits orb(x, L2 ) for x ∈ R.

Exercise 1.1.2. Consider the dynamical system T : ]0, ∞[ → ]0, ∞[, T x = 12 (x + x2 ).

Show that there
√ is some q √ ∈ ]0, 1[ such that |T x − T y|√≤ q|x − y| for x, y ≥ 1. Deduce
that |T n x − 2| ≤ q n |x − 2| and hence that T n x → 2 for all x ≥ 1.

Exercise 1.1.3. Prove the statements in Example 1.6(a)–(c).

Exercise 1.1.4. Show that the logistic map L4 , when restricted to the interval [0, 1], is
quasiconjugate to the doubling map on the interval via φ(x) = sin2 (πx). Also show that
the maps are not conjugate.

Exercise 1.2.1. Show that, in general, the implication (i)=⇒(ii) does not hold in Propo-
sition 1.10.

Exercise 1.2.2. Show that the following assertions on a dynamical system T : X → X

are equivalent:
(i) T is topologically transitive;
(ii) for any open set U ⊂ X with T −1 (U ) ⊂ U , either U = ∅ or U is dense in X;
(iii) for any closed set E ⊂ X with T (E) ⊂ E, either E = X or E is nowhere dense.

Exercise 1.2.3. Suppose that X has at least one isolated point. Prove that, if there is
any topologically transitive map T : X → X, then X is finite and X = orb(x, T ) for any
x ∈ X.

Exercise 1.2.4. Show that, if T : X → X is topologically transitive, then for any pair
U, V of nonempty open subsets of X, the return set N (U, V ) is infinite; see Definition
1.49. (Hint: For the trivial case in which X has isolated points apply Exercise 1.2.3. If
X has no isolated points, then given m ∈ N (U, V ) and W := U ∩ T −m (V ), observe that
N (W, W ) ∩ N = ∅ and m + N (W, W ) ⊂ N (U, V ).)

Exercise 1.2.5. Prove that a dynamical system T : X → X on a metric space X is

topologically transitive if and only if, for any ε > 0 and any pair of points x, y ∈ X,
we can find z ∈ X and n, m ∈ N0 satisfying d(T n z, x) < ε and d(T m z, y) < ε. (Hint:
First observe that the above condition is equivalent to the fact that for any pair U, V
of nonempty open subsets of X one can find n, m ∈ N0 with T −n (U ) ∩ T −m (V ) = ∅.
This condition is obviously implied by topological transitivity. For the converse, given
nonempty open sets U, V ⊂ X, either find k ∈ N (U, V ) (in that case you are done) or,
if k ∈ N (V, U ), set W = V ∩ T −k (U ), note that N (W, W ) is infinite by Exercise 1.2.4,
and then find some j ∈ N (U, V ).)

Exercise 1.2.6. Let T be a topologically transitive dynamical system on a separable

complete metric space X without isolated points. Prove constructively, not using the
Baire category theorem, that T has a dense set of points with dense orbit. (Hint: Let
(yn )n be a dense sequence in X. Start with x0 ∈ X. Then find x1 close to x0 and a
positive integer m1 so that T m1 x1 is close to y1 . Then find x2 close to x1 and a positive
integer m2 so that T m1 x2 is close to T m1 x1 and T m1 +m2 x2 is close to y2 . Continue.)
26 1 Topological dynamics

Exercise 1.2.7. Let T be a dynamical system on a metric space X without isolated

points. A backward orbit of a vector x is a sequence (xn )n≥0 in X (if it exists!) such
that x0 = x and T xn = xn−1 , n ≥ 1. Show the following:
(i) if T is topologically transitive and X is separable and complete then there exists
a dense set of points with dense backward orbits;
(ii) if T has a dense backward orbit then T is topologically transitive.
(Hint: See the previous exercise.)

Exercise 1.2.8. Let T : X → X be a dynamical system. For x ∈ X the J-set JT (x) =

J(x) is defined as the set of all points y ∈ X for which there is a strictly increasing
sequence (nk )k of positive integers and a sequence (xk )k in X such that xk → x and
T nk xk → y as k → ∞.
(a) Show that J(x) is a closed T -invariant set.
(b) Suppose that X has no isolated points. Show that J(x) = X if and only if, for
any pair U, V of nonempty open subsets of X with x ∈ U , there exists some n ≥ 0 such
that T n (U ) ∩ V = ∅.
(c) Suppose that X has no isolated points. Show that the following assertions are
equivalent:
(i) T is topologically transitive;
(ii) for any x ∈ X, J(x) = X;
(iii) there is a dense set of points x ∈ X such that J(x) = X.

Exercise 1.2.9. Show that every orbit under an irrational rotation is dense. (Hint: Use
the pigeonhole principle to show that, for any ε > 0, some arc of angle ε must contain
two iterates of 1, T m 1 and T n 1, m > n. Then look at the iterates of T m−n .)

Exercise 1.2.10. A dynamical system T : X → X is called minimal if every orbit under

T is dense. Find a characterization of minimality in the spirit of Exercise 1.2.2.

Exercise 1.2.11. Consider the dynamical system T : [−1, 1] → [−1, 1] given by

⎧
⎨2 + 2x, if − 1 ≤ x < −1/2,
Tx = −2x, if − 1/2 ≤ x < 1/2,
⎩
−2 + 2x, if 1/2 ≤ x ≤ 1.

(a) Show that T has a dense orbit but that T 2 does not.
(b) Show that there are two points x, y ∈ [−1, 1] such that orb(x, T 2 ) ∪ orb(y, T 2 ) is
dense in [−1, 1] but neither of them has a dense orbit under T 2 .
(c) Show that there is a point x ∈ [−1, 1] such that orb(x, T 2 ) contains a nonempty
open set but x does not have a dense orbit under T 2 .
(Remark: We will prove in Chapter 6 that none of these properties can hold in a
linear setting.)

Exercise 1.3.1. Show that none of the three conditions in Definition 1.26 alone implies
chaos.

Exercise 1.3.2. Let X be a finite set, endowed with the discrete metric. Describe all
maps on X that are chaotic. Do the same for countably infinite sets under the discrete
metric.

Exercise 1.3.3. Suppose that (X, d) is a metric space without isolated points and T :
X → X is a contracting map, that is d(T x, T y) ≤ d(x, y) for all x, y ∈ X. Show that if
T has one dense orbit then T is minimal (see Exercise 1.2.10); in particular, it cannot
be chaotic.
Exercises 27

Exercise 1.3.4. Show that T : X → X is chaotic if and only if every finite family of
nonempty open sets shares a periodic orbit, in the following sense: for each finite family
Uj ⊂ X, j = 1, . . . , n, of nonempty open sets there is a periodic point x ∈ U1 such
that T kj x ∈ Uj for some kj ≥ 0, j = 2, . . . , n; see Figure 1.8. (Hint: One implication is
trivial; for the other one use continuity of T and an induction process.)

Fig. 1.8 Exercise 1.3.4

Exercise 1.3.5. Show that the space Σ2 is a complete metric space without isolated
points. Show also that the sequences with only finitely many nonzero entries form a
dense set.

Exercise 1.3.6. Why is the quasiconjugacy of Example 1.37 between the shift on two
symbols and the doubling map not a conjugacy? Use this quasiconjugacy to find a
representation of the periodic points and the points with dense orbit for the doubling
map.

Exercise 1.4.1. Show that the shift on two symbols is mixing.

Exercise 1.4.2. Prove that a dynamical system T is mixing if and only if, for any
strictly increasing sequence (nk )k of positive integers, the sequence (T nk )k is topologi-
cally transitive; see Definition 1.56 for the notion of topological transitivity for sequences
of maps.

Exercise 1.4.3. Let X be a complete metric space. Prove that a dynamical system
T : X → X is mixing if and only if, for every sequence (xn )n in X and for every strictly
increasing sequence (nk )k of positive integers for which {xnk ; k ∈ N} is relatively com-
pact there exists a dense Gδ -set of points y ∈ X such that lim inf k→∞ d(xnk , T nk y) = 0.
(Hint: Use the previous exercise; a subset A of a metric space X is relatively compact if
and only if every sequence in A has a subsequence that converges in X.)

Exercise 1.4.4. Let T : X → X be a dynamical system. For x ∈ X, the set JTmix (x) =
J mix (x) is defined as the set of all points y ∈ X for which there is a sequence (xn )n in
X such that xn → x and T n xn → y as n → ∞; see also Exercise 1.2.8.
(a) Show that J mix (x) is a closed T -invariant set.
(b) Show that J mix (x) = X if and only if, for any pair U, V of nonempty open subsets
of X with x ∈ U , there exists some N ≥ 0 such that T n (U ) ∩ V = ∅ for all n ≥ N .
(c) Show that the following assertions are equivalent:
(i) T is mixing;
(ii) for any x ∈ X, J mix (x) = X;
(iii) there is a dense set of points x ∈ X such that J mix (x) = X.
28 1 Topological dynamics

Fig. 1.9 Weak mixing (Exercise 1.5.1(iii) and (iv))

Exercise 1.4.5. Show that minimality and mixing are independent properties of a dy-
namical system.

Exercise 1.5.1. Let T : X → X be a dynamical system. In the following, U, U1 , U2 , V

will denote arbitrary nonempty open subsets of X. Prove that any of the following
conditions is equivalent to T being weakly mixing:
(i) for any U, V ⊂ X we have N (U, V ) ∩ N (V, V ) = ∅;
(ii) for any U1 , U2 , V ⊂ X we have N (U1 , V ) ∩ N (U2 , V ) = ∅;
(iii) for any U1 , U2 , V ⊂ X we have N (U1 , U2 ) ∩ N (U2 , V ) = ∅;
(iv) for any U1 , U2 , V ⊂ X we have N (U1 , U2 ) ∩ N (V, V ) = ∅;
see Figure 1.9. (Hint: To prove sufficiency of (i) use the 4-set trick and Proposition 1.53;
this will then imply the sufficiency of the other conditions.)

Exercise 1.5.2. A dynamical system T : X → X is called totally transitive if every

power T p , p ∈ N, is topologically transitive. Show that any weakly mixing map is totally
transitive.

Exercise 1.5.3. Prove that every chaotic and totally transitive dynamical system T :
X → X is weakly mixing. (Hint: Verify the hypothesis of Proposition 1.53 by finding a
periodic point in U , of period k say, and then by using topological transitivity of T k .)

Exercise 1.5.4. A dynamical system T : X → X is called flip transitive if, for any pair
U, V ⊂ X of nonempty open sets, N (U, V ) ∩ N (V, U ) = ∅. Show that the map T of
Exercise 1.2.11 is flip transitive but not weakly mixing.

Exercise 1.5.5. Show that T is weakly mixing if and only if it is flip transitive and T 2
is topologically transitive. (Hint: To prove the weak mixing property use condition (i)
in Exercise 1.5.1. To do this, given nonempty open sets U, V ⊂ X, find k ∈ N0 with
U  := U ∩ T −2k (V ) = ∅ and then some m ∈ N (U  , T −k (V )) ∩ N (T −k (V ), U  ). Consider
m + k.)

Exercise 1.5.6. A dynamical system T : X → X is called topologically ergodic if, for

any pair U, V ⊂ X of nonempty open sets, N (U, V ) is syndetic. Prove the following:
(i) any irrational rotation is topologically ergodic but not weakly mixing;
(ii) every mixing and every chaotic dynamical system is topologically ergodic;
(iii) if T : X → X is topologically ergodic and S : Y → Y is weakly mixing, then T × S
is topologically transitive.

Exercise 1.5.7. Let T : X → X be a dynamical system. Show that any of the following
conditions is equivalent to T being weakly mixing:
(i) for any nonempty open sets U, V ⊂ X and any m ∈ N there is some k with
k, k + m ∈ N (U, V );
Exercises 29

(ii) for any m ∈ N and for any increasing sequence (nk )k with nk+1 − nk ∈ {m, 2m},
k ∈ N, we have that (T nk )k is topologically transitive.
(Hint: See the proof of Theorem 1.54.)

Exercise 1.5.8. Let T : X → X be a dynamical system. Establish the equivalence of

the following assertions:
(i) T is weakly mixing;
(ii) for any pair U, V ⊂ X of nonempty open sets, N (U, V ) contains two consecutive
integers.
(Hint: Proceeding by induction, show that N (U, V ) contains arbitrarily long intervals:
if k, k + 1 ∈ N (U, U ), set U1 := U ∩ T −k (U ), U2 := U ∩ T −k−1 (U ), apply the inductive
hypothesis to the pair (U1 , U2 ) to find an interval [j, j + m] contained in N (U1 , U2 ). By
the selection of U1 , obtain that [j, j + m + 1] ⊂ N (U, U ).)

Exercise 1.5.9. Given a metric space X, the corresponding hyperspace is defined as

K(X) = {K ⊂ X ; K is compact}. The space K(X) is endowed with the (metrizable)
Vietoris topology, for which a base of open sets is given by the family of sets

 
k 
V(U1 , . . . , Uk ) := K ∈ K(X) ; K ⊂ Uj and K ∩ Uj = ∅, j = 1, . . . , k ,
j=1

where U1 , . . . , Uk , k ∈ N, are nonempty open sets in X. If T : X → X is continuous,

then it naturally induces a continuous hyperextension T : K(X) → K(X) defined by
T (K) = T (K) = {T x ; x ∈ K}.
We say that a dynamical system T : X → X is hypertransitive if its hyperextension
T is topologically transitive. Prove that T is hypertransitive if and only if it is weakly
mixing. (Hint: For the sufficiency of weak mixing use Furstenberg’s theorem; for the
necessity use Proposition 1.52.)

Exercise 1.6.1. Let (xn )n be a dense sequence in R2 , and let yn ∈ R2 , n ≥ 1, be

vectors of length n that are orthogonal to xn . Consider the maps Tn : R2 → R2 with
Tn (α, β) = αxn + βyn . Determine all points in R2 with dense orbit under (Tn )n . Deduce
that (Tn )n has a dense orbit but is not topologically transitive.

Exercise 1.6.2. Prove the Birkhoff transitivity theorem for commuting continuous maps
Tn : X → X, n ∈ N0 , of dense range on a separable complete metric space X, that is,
that the following assertions are equivalent:
(i) (Tn )n is topologically transitive;
(ii) there exists some x ∈ X such that orb(x, (Tn )) is dense in X.
If one of these conditions holds then the set of points in X with dense orbit is a dense
Gδ -set.

Exercise 1.6.3. Let S be a mixing map on a separable complete metric space X and T
a map on a metric space Y without isolated points that admits a dense orbit orb(y, T ),
y ∈ Y . Show that there exists some x ∈ X such that (x, y) has a dense orbit under the
map S × T . (Hint: Use the previous exercise.)

Exercise 1.6.4. A sequence (Tn )n of continuous maps on a metric space X is called

hereditarily transitive with respect to an increasing sequence (nk )k of positive integers
if (Tmk )k is topologically transitive for every subsequence (mk )k of (nk )k . The sequence
(Tn )n is called hereditarily transitive if it is so with respect to some sequence (nk )k .
Prove that, if X is separable, then a commuting sequence (Tn )n is hereditarily transitive
if and only if (Tn )n is weakly mixing. (Hint: Note that X has a countable base; use
Furstenberg’s theorem for sequences of maps.)
30 1 Topological dynamics

Exercise 1.6.5. Show that Theorem 1.54 does not hold for sequences (Tn )n even if the
maps Tn : X → X commute and have dense range. (Hint: A weakly mixing sequence
(Tn )n remains weakly mixing when adding arbitrary maps.)

Sources and comments

Section 1.1. A standard reference for the theory of dynamical systems is Devaney [132].
For more recent textbooks we refer to Brin and Stuck [97] and Robinson [267], while
Gulick [188] provides an elementary introduction.

Section 1.2. Kolyada and Snoha [217] give an excellent survey on topological transi-
tivity, with many additional equivalent conditions. The original version of the Birkhoff
transitivity theorem can be found in [74, § 62].

Section 1.3. Chaos in the sense of Devaney was introduced in [132]. While there are
many other definitions of chaos (see for example Kolyada [216] or Forti [154]), Devaney’s
definition has become very popular. The theorem of Banks et al. was obtained in [31];
see also Silverman [294] and Glasner and Weiss [163].

Sections 1.4 and 1.5. For Furstenberg’s theorem see [157], which also contains the
4-set trick implicitly. Theorem 1.54 is due, independently, to Akin [4], Glasner [162],
and Peris and Saldivia [257]; in the context of linear operators on Banach spaces it was
also obtained by Grivaux [172]. The remaining characterizations of weak mixing in Sec-
tion 1.5, including Exercise 1.5.1, are due to Banks [29] and Akin [4]; more precisely,
Proposition 1.53 is called the “Furstenberg Intersection Lemma” in Akin’s book.

Section 1.6. The universal Taylor series mentioned in Section 1.6 is essentially due to
Fekete; see the discussion in [179, Section 3a]. The Universality Criterion was obtained
by Grosse-Erdmann [177].

Exercises. Exercises 1.2.8 and 1.4.4 are taken from Costakis and Manoussos [118, 119].
For Exercise 1.3.4 we refer to Touhey [299], for Exercise 1.4.3 to Moothathu [244], for
Exercise 1.5.3 to Bauer and Sigmund [32] and to Banks (attributed to Stacey) [28], for
Exercise 1.5.5 to Banks [29], and for Exercise 1.5.6 to Moothathu [245]. The two parts
of Exercise 1.5.7 are taken from Grivaux [172] and Peris and Saldivia [257], respectively;
Exercise 1.5.8 is from Grosse-Erdmann and Peris [187] (see also Grivaux [172] and Bayart
and Matheron [45]). The assertion of Exercise 1.5.9 is due to Bauer and Sigmund [32]
(one direction), and, independently, to Banks [30] and Peris [256] (the other direction).
Exercise 1.6.1 is from Godefroy and Shapiro [165], Exercise 1.6.4 from Bès and Peris [71].

Extensions. Let us add a word on the setting chosen in this chapter. Since the over-
whelming majority of linear dynamical systems studied in the literature acts on metric
spaces we have restricted our attention to such spaces. In general, however, a dynamical
system is given by a continuous map T : X → X on a topological space X. The def-
initions of topologically transitive, (weakly) mixing and chaotic maps extend verbatim
to such systems. The same applies to sequences (Tn )n of continuous maps Tn : X → Y
between arbitrary topological spaces X and Y .
Then, as the proofs show, all the results in this chapter on general dynamical systems
remain true in the setting of arbitrary topological spaces. To be more specific, this
concerns all results apart from Proposition 1.15 and the Theorems 1.16, 1.29 and 1.57.
Chapter 2
Hypercyclic and chaotic operators

In this chapter we begin our investigation of linear dynamical systems, that

is, dynamical systems that are defined by linear maps. As a simple example
one may think of the differentiation operator

D : f → f .

In the language of dynamical systems, the exponential function, for example,

is a fixed point of D, while the sine function is periodic with period 4. We
will see in this chapter that D is in fact a chaotic operator.
The linearity of a map only makes sense if the underlying space carries,
besides its topological structure, a linear structure also. Familiar examples of
such spaces are Hilbert spaces and Banach spaces. But some of our main ex-
amples demand that we go beyond Banach spaces and allow so-called Fréchet
spaces. These spaces will be introduced in the first section of this chapter.
In the subsequent sections we revisit the various topics discussed in Chap-
ter 1 under the influence of linearity, and we contrast linear with nonlinear
dynamics.

2.1 Linear dynamical systems

The dynamical systems studied in Chapter 1 were defined by continuous

maps on metric spaces. For linear dynamical systems, the underlying space
must in addition have a linear structure, as is the case for Hilbert spaces and
Banach spaces. In this book we will assume a certain familiarity with such
spaces; some of their basic properties are collected in Appendix A.
However, some interesting examples of linear dynamical systems are de-
fined on spaces of a more general type, the so-called Fréchet spaces. In this
section we introduce these spaces and describe their operators. The main
purpose will be to familiarize the reader with this new concept. Some more

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 31

DOI 10.1007/978-1-4471-2170-1_2, © Springer-Verlag London Limited 2011
32 2 Hypercyclic and chaotic operators

advanced results that are only used occasionally in this book will be covered
in Appendix A.
The following two examples will motivate the concept of a Fréchet space.

Example 2.1. The process of taking derivatives, that is, the operator D : f →
f  , provides an interesting linear dynamical system. In order to have the
powerful tools of complex analysis at our disposal we regard D as acting on
the space of entire functions,

H(C) = {f : C → C ; f holomorphic}.

The natural concept of convergence for entire functions is that of local uni-
form convergence, that is, the uniform convergence on all compact sets. In
contrast to Banach spaces, convergence is described here by a countably in-
finite collection of conditions. More precisely, we have that fk → f in H(C)
if and only if, for all n ∈ N, pn (fk − f ) → 0 as k → ∞, where

pn (f ) := sup |f (z)|.
|z|≤n

Here, (pn )n is an increasing sequence of norms.

Example 2.2. Many natural spaces of sequences are Banach spaces. But the
space of all (real or complex) sequences,

ω := KN = {(xn )n ; xn ∈ K, n ∈ N},

lies outside this framework, where K = R or C. The natural concept of con-

vergence is that of coordinatewise convergence; that is, we have that x(ν) → x
in ω if and only if, for all n ∈ N, pn (x(ν) − x) → 0 as ν → ∞, where

pn (x) := sup |xk |, x = (xk )k .

1≤k≤n

Here, (pn )n is an increasing sequence of seminorms.

We recall the notion of a seminorm.

Definition 2.3. A functional p : X → R+ on a vector space X over K = R

or C is called a seminorm if it satisfies, for all x, y ∈ X and λ ∈ K,
(i) p(x + y) ≤ p(x) + p(y),
(ii) p(λx) = |λ|p(x).

A norm is a seminorm p for which p(x) = 0 implies that x = 0. A Banach

space is a vector space X endowed with a norm, usually denoted by  · ,
whose topology is defined via the metric

d(x, y) := x − y, x, y ∈ X,
2.1 Linear dynamical systems 33

and which is complete in that metric. If, moreover, the norm derives from an
inner product  · , ·  via

x := x, x, x ∈ X,

then X is called a Hilbert space.

We recall here some classical Banach and Hilbert spaces. Further spaces
will be introduced as the need arises.

Example 2.4. (a) Let 1 ≤ p < ∞. Then the space

 ∞
 
p := x = (xn )n ∈ KN ; |xn |p < ∞
n=1

of p-summable sequences, endowed with the norm x := ( ∞ n=1 |xn | )
p 1/p
,
2
is a Banach space. Inparticular,  is a Hilbert space with inner product
∞
defined by x, y = n=1 xn yn . Occasionally we let the index start with
0. The finite sequences, that is, sequences of the form (x1 , . . . , xn , 0, 0, . . .),
n ≥ 1, constitute a dense subset. Considering only the finite sequences with
entries from Q or Q + iQ we see that any p , 1 ≤ p < ∞, is separable. The
space p (Z) of p-summable sequences, indexed over Z, is defined analogously.
(b) The space ∞ := {x = (xn )n ∈ KN ; supn∈N |xn | < ∞} of bounded
sequences, endowed with the sup-norm x := supn∈N |xn |, is a Banach space.
Since it is not separable it will be of less interest to us. Instead, its closed
subspace
c0 := {x = (xn )n ∈ KN ; lim xn = 0}
n→∞

of null sequences is a separable Banach space under the induced norm.

(c) Let a < b and 1 ≤ p < ∞. Then the space
  b 
Lp [a, b] := f : [a, b] → K ; f measurable and |f (t)|p dt < ∞
a
b
of p-integrable functions, endowed with the norm f  := ( a |f (t)|p dt)1/p , is
a Banach space; as usual, we identify functions that are equal almost every-
where. In particular, L2 [a, b] is a Hilbert space with inner product defined by
b
f, g = a f (t)g(t) dt. We will occasionally need the fact that the functions
t → √12π eint , n ∈ Z, form an orthonormal basis in L2 [0, 2π].
(d) Let a < b. Then the space
 
C[a, b] := f : [a, b] → K ; f continuous

of continuous functions, endowed with the sup-norm f  := supt∈[a,b] |f (t)|,

is a Banach space.
34 2 Hypercyclic and chaotic operators

The concept of a Fréchet space generalizes that of a Banach space by

defining the topology via a sequence (pn )n of seminorms, which we can always
assume to be increasing (by considering maxk≤n pk , if necessary). Moreover,
the sequence is supposed to be separating, that is, pn (x) = 0 for all n ≥ 1
implies that x = 0. Then it is easy to see that
∞
1
d(x, y) := min(1, pn (x − y)), x, y ∈ X (2.1)
n=1
2n

defines a metric on X; see Exercise 2.1.1. An important feature of this metric

is that it is translation-invariant, that is,

d(x, y) = d(x + z, y + z) for all x, y, z ∈ X.

Definition 2.5. A Fréchet space is a vector space X, endowed with a sep-

arating increasing sequence (pn )n of seminorms, which is complete in the
metric given by (2.1).

The following result will be of constant use. We leave its proof as a useful
exercise to the reader; see Exercise 2.1.2.

Lemma 2.6. Let X be a Fréchet space with a defining increasing sequence

(pn )n of seminorms. Let xk , x ∈ X, k ≥ 1, and U ⊂ X. Then:
(i) xk → x if and only if pn (xk − x) → 0 as k → ∞, for all n ≥ 1;
(ii) (xk )k is a Cauchy sequence if and only if pn (xk − xl ) → 0 as k, l → ∞,
for all n ≥ 1;
(iii) U is a neighbourhood of x if and only if there are n ≥ 1 and ε > 0 such
that {y ∈ X ; pn (y − x) < ε} ⊂ U .

Example 2.7. (a) Let X be a Banach space with norm  · . Setting pn =  · ,

n ≥ 1, it follows from (i) and (ii) that X is also a Fréchet space according to
Definition 2.5.
(b) With the seminorms defined in Example 2.1, the space H(C) of entire
functions is a Fréchet space; in view of Lemma 2.6(ii), completeness follows
in the usual way. Since the Taylor series expansion of an entire function
converges on every compact set, the polynomials form a dense subset of H(C).
Considering polynomials with coefficients from Q + iQ we see that H(C) is
separable.
(c) The space ω = KN of all real or complex sequences, endowed with
the seminorms given in Example 2.2, is a Fréchet space. Since the finite
sequences with entries from Q or Q + iQ form a dense subset, ω is separable.
More generally, if X is a separable Fréchet space then, in a canonical way,
also X N is a separable Fréchet space; see Exercise 2.1.3.

Further Fréchet spaces will be introduced in the course of the book.

Looking at the way the metric is defined in a Banach space it is tempting
to introduce, also in a Fréchet space X, a norm-like functional by setting
2.1 Linear dynamical systems 35

∞
1
x := n
min(1, pn (x)), x ∈ X, (2.2)
n=1
2

so that d(x, y) = x − y. We summarize its characteristic properties; see

Exercise 2.1.4.

Proposition 2.8. The functional  ·  : X → R+ given by (2.2) satisfies, for

all x, y ∈ X and λ ∈ K,
(i) x + y ≤ x + y;
(ii) λx ≤ x if |λ| ≤ 1;
(iii) limλ→0 λx = 0;
(iv) x = 0 implies that x = 0.

Definition 2.9. A functional  ·  : X → R+ on a vector space X that

satisfies conditions (i)–(iv) of Proposition 2.8 is called an F-norm.

The notion of an F-norm has the advantage that one can largely argue as
if one was working in a Banach space. One need only be aware of the fact
that the positive homogeneity of a norm is no longer available. In fact, in
many cases, this property is not needed at all or it can be replaced by the
following weaker property that follows directly from conditions (i) and (ii):
for all x ∈ X and λ ∈ K,

λx ≤ (|λ| + 1) x. (2.3)

Having discussed Fréchet spaces and their topology we now turn to the
concept of operators on them.

Definition 2.10. Let X and Y be Fréchet spaces. Then a continuous linear

map T : X → Y is called an operator. The space of all such operators is
denoted by L(X, Y ). If Y = X we say that T is an operator on X, with
L(X) = L(X, X).

The following extends a familiar result from Banach spaces to Fréchet

spaces; see also Exercise 2.1.7.

Proposition 2.11. Let X and Y be Fréchet spaces with defining increasing

sequences of seminorms (pn )n and (qn )n , respectively. Then a linear map
T : X → Y is an operator if and only if, for any m ≥ 1, there are n ≥ 1 and
M > 0 such that
qm (T x) ≤ M pn (x), x ∈ X.

Proof. The condition is obviously sufficient because, by Lemma 2.6, it implies

that when xk → x in X then T xk → T x in Y.
Conversely, let m ≥ 1. By Lemma 2.6, the set W := {y ∈ Y ; qm (y) < 1}
is a 0-neighbourhood in Y . By continuity there is a 0-neighbourhood W  in X
such that T (W  ) ⊂ W . Hence there are n ≥ 1 and ε > 0 such that pn (x) < ε
36 2 Hypercyclic and chaotic operators

implies that x ∈ W  , and therefore qm (T x) < 1. Now let x ∈ X. Then, for

any δ > 0, we have that
ε

pn x <ε
pn (x) + δ

and hence
pn (x) + δ
qm (T x) < .
ε
Since δ > 0 is arbitrary we obtain the result with M = 1/ε. 


In contrast to Banach space operators one cannot associate a norm with

a Fréchet space operator.

Example 2.12. (a) The map

D : f → f
is an operator on H(C). This follows from the Cauchy estimates by which,
for any n ≥ 1, sup|z|≤n |f  (z)| ≤ sup|z|≤n+1 |f (z)|.
(b) Let X = H(C). Then the translation map Ta is defined by

Ta f (z) = f (z + a), a ∈ C.

This is clearly an operator on X.

(c) Let X = p , 1 ≤ p < ∞, or c0 . Then the backward shift B : X → X,
defined by
B(x1 , x2 , . . . ) = (x2 , x3 , . . . )
is an operator on X, of norm B = 1, and it is also an operator on ω.

In Chapter 1 we associated to any two dynamical systems S : X → X and

T : Y → Y a new dynamical system S × T on X × Y . In a linear setting one
usually employs a different, additive notation. More specifically, let X and Y
be Fréchet spaces with defining increasing sequences of seminorms (pn )n and
(qn )n , respectively. Then the space

X ⊕ Y := {(x, y) ; x ∈ X, y ∈ Y }

will be endowed with the seminorms (x, y) → pn (x) + qn (y), n ≥ 1, which

induce the product topology on X ⊕ Y . This space then becomes a Fréchet
space, which is separable if X and Y are.

Definition 2.13. Let S : X → X and T : Y → Y be operators on Fréchet

spaces X and Y . Then the operator S ⊕ T is defined by

S ⊕ T : X ⊕ Y → X ⊕ Y, (S ⊕ T )(x, y) = (Sx, T y).

With this we end our introduction to Fréchet spaces and their operators.
We will show in Chapter 12 that several important results in linear dynamics
2.2 Hypercyclic operators 37

hold true in the wider context of operators on so-called topological vector

spaces. However, other results require the existence of a complete metric, and
most operators in linear dynamics are naturally defined on Fréchet spaces.
In addition, since we are primarily interested in operators with a dense orbit,
we will assume that the space is separable. Thus, with Chapter 12 the only
exception, we adopt the following point of view throughout this book.

Definition 2.14. A linear dynamical system is a pair (X, T ) consisting of a

separable Fréchet space X and an operator T : X → X.

Usually we simply call T or T : X → X a linear dynamical system. From

now on, all operators will be defined on separable Fréchet spaces, if nothing
else is said.

2.2 Hypercyclic operators

We begin our study of the dynamics of linear operators by considering dy-

namical systems with a dense orbit. In the presence of linearity, such systems
are given their own name.

Definition 2.15. An operator T : X → X is called hypercyclic if there is

some x ∈ X whose orbit under T is dense in X. In such a case, x is called a
hypercyclic vector for T . The set of hypercyclic vectors for T is denoted by
HC(T ).

The origin of this terminology is easily explained. For a long time, operator
theorists have been studying so-called cyclic vectors in connection with the
invariant subspace problem. Vectors with a more restrictive property were
then called supercyclic.

Definition 2.16. Let T : X → X be an operator. A vector x ∈ X is called

cyclic for T if the linear span of its orbit,

span {T n x ; n ≥ 0}

is dense in X. A vector x ∈ X is called supercyclic for T if its projective

orbit,
{λT n x ; n ≥ 0, λ ∈ K}
is dense in X.
Operators that possess a cyclic (or supercyclic) vector are called cyclic (or
supercyclic, respectively).

This suggested the name of hypercyclicity for the case when the orbit itself
is dense. Cyclic and supercyclic vectors will not be studied in detail in this
book.
38 2 Hypercyclic and chaotic operators

The invariant subspace problem, which is open to this day, asks whether
every Hilbert space operator possesses an invariant closed subspace other
than the trivial ones given by {0} and the whole space. Counterexamples do
exist for operators on non-reflexive spaces like 1 .
Obviously, the smallest closed T -invariant subspace of X that contains a
given point x coincides with the closure of the span of its orbit. Therefore, an
operator has no nontrivial invariant closed subspace precisely if every nonzero
vector is cyclic. By the same token we have a link between hypercyclicity and
the invariant subset problem: does every Hilbert space operator possess an
invariant closed subset other than the trivial ones given by {0} and the whole
space?
Observation 2.17. An operator has no nontrivial invariant closed subsets if
and only if every nonzero vector is hypercyclic.
Having explained the historical interest in hypercyclicity, our first question
has to be if hypercyclic operators exist. That is, does the additional require-
ment of linearity still allow us to find maps with dense orbits? Indeed, a very
simple operator on the Hilbert space 2 turns out to be hypercyclic.
Example 2.18. Let T : 2 → 2 be twice the backward shift, that is,

T = 2B : (x1 , x2 , x3 , . . .) → 2(x2 , x3 , x4 , . . .).

The space 2 has a countable dense set {y (k) ; k ≥ 1} consisting of finite

(k)
sequences; for each k ≥ 1, let mk be the greatest index with ymk = 0. By S
we denote half the forward shift operator,

S = 12 F : (x1 , x2 , x3 , . . .) → 12 (0, x1 , x2 , . . .).

Then, by induction, we can find a sequence (nk )k of positive integers such

that, for any k > j ≥ 1,

nk ≥ mj + nj and 2nk ≥ 2nj +k y (k) .

We claim that the vector ∞


x := S nk y (k)
k=1

is hypercyclic for T . First, since S nk y (k)  = 2−nk y (k)  ≤ 2−k for k ≥ 2,

the series converges and x ∈ 2 . Now, let k ≥ 1. Then


k−1 ∞

T nk x = 2nk −nj B nk −nj y (j) + y (k) + 2nk −nj F nj −nk y (j)
j=1 j=k+1
∞

= y (k) + 2nk −nj F nj −nk y (j) ,
j=k+1
2.2 Hypercyclic operators 39

where we have used that nk − nj ≥ mj for j < k. From

∞
 ∞
 ∞

2nk −nj F nj −nk y (j)  = 2nk −nj y (j)  ≤ 2−j = 2−k
j=k+1 j=k+1 j=k+1

we deduce that T nk x − y (k)  ≤ 2−k . Since the y (k) form a dense set, x has
a dense orbit under T .

Instead of an explicit construction of a hypercyclic vector one can also ap-

ply the Birkhoff transitivity theorem to show that an operator is hypercyclic.
This leads to more transparent proofs, in particular when the operator is
complicated. For ease of reference we restate Birkhoff’s theorem in our new
setting; note that Fréchet spaces clearly have no isolated points.

Theorem 2.19 (Birkhoff transitivity theorem). An operator T is hyper-

cyclic if and only if it is topologically transitive. In that case, the set HC(T )
of hypercyclic vectors is a dense Gδ -set.

Directly or indirectly, the transitivity theorem will be our main tool for
proving the hypercyclicity of an operator.
The first examples of hypercyclic operators were found by G.D. Birkhoff
in 1929, G.R. MacLane in 1952 and S. Rolewicz in 1969. These operators will
accompany us throughout the book as they will serve as a testing ground
for any new concept in linear dynamics; indeed, Example 2.18 was already a
special Rolewicz operator.

Example 2.20. (Birkhoff’s operators) On the space H(C) of entire func-

tions we consider the translation operators given by

Ta f (z) = f (z + a), a = 0.

Let U, V ⊂ H(C) be arbitrary nonempty open sets, and fix f ∈ U , g ∈ V . By

the definition of the topology on H(C) there is a closed disk K centred at 0
and an ε > 0 such that an entire function h belongs to U (or to V ) whenever
supz∈K |f (z)−h(z)| < ε (or supz∈K |g(z)−h(z)| < ε, respectively). Let n ∈ N
be any integer such that K and K + na are disjoint disks. Considering the
function that is defined as f on a neighbourhood of K and by z → g(z − na)
on a neighbourhood of K + na, Runge’s theorem (see Appendix A), tells us
that there exists a polynomial p such that

sup |f (z) − p(z)| < ε and sup |g(z − na) − p(z)| < ε,
z∈K z∈K+na

and hence also

sup |g(z) − (Tan p)(z)| = sup |g(z) − p(z + na)| < ε.

z∈K z∈K
40 2 Hypercyclic and chaotic operators

This shows that p ∈ U and Tan p ∈ V , so that Ta is topologically transitive.

Since H(C) is a separable Fréchet space, Ta is hypercyclic.

Example 2.21. (MacLane’s operator) We next consider the differentiation

operator
D : f → f
on H(C). Since the polynomials are dense in H(C), given arbitrary nonempty
open sets U, V ⊂ H(C), there are polynomials p ∈ U and q ∈ V , p(z) =
N N
k=0 bk z . Let n ≥ N + 1 be arbitrary. Then the
k k
k=0 ak z and q(z) =
polynomial
N
k! bk k+n
r(z) = p(z) + z
(k + n)!
k=0
n
has the property that D r = q. Moreover, for any R > 0 we have that


N
k!|bk | k+n
sup |r(z) − p(z)| ≤ R →0
|z|≤R (k + n)!
k=0

as n → ∞. Thus, if n is sufficiently large, then r ∈ U and D n r ∈ V . This

implies that D is hypercyclic.

Example 2.22. (Rolewicz’s operators) On the spaces X := p , 1 ≤ p < ∞,

or X := c0 we consider the multiple

T = λB : X → X, (x1 , x2 , x3 , . . .) → λ(x2 , x3 , x4 , . . .)

of the backward shift, where λ ∈ K. First, if |λ| ≤ 1 then T n x =

|λ|n B n x ≤ x for all x ∈ X and n ≥ 0. Thus T cannot be hypercyclic in
this case.
On the other hand, T is hypercyclic whenever |λ| > 1. Indeed, if U, V ⊂ X
are nonempty open sets, we can find x ∈ U and y ∈ V of the form

x = (x1 , x2 , . . . , xN , 0, 0, . . . ), y = (y1 , y2 , . . . , yN , 0, 0, . . . ),

for some N ∈ N. Let n ≥ N be arbitrary. Defining z ∈ X by zk = xk if

1 ≤ k ≤ N , zk = λ−n yk−n if n + 1 ≤ k ≤ n + N , and zk = 0 otherwise,
we obtain a sequence with T n z = y. Moreover, x − z = |λ|−n y → 0 as
n → ∞. Thus, if n is sufficiently large, then z ∈ U and T n z ∈ V . This shows
that T is topologically transitive; since the underlying spaces are separable
Banach spaces, T is hypercyclic.

There are various ways to derive the hypercyclicity of one operator from
that of another. The first result of this type follows from Proposition 1.14
by the Birkhoff transitivity theorem. Recall that by the inverse mapping
theorem (see Appendix A), any bijective operator has a continuous inverse
and is therefore.
2.2 Hypercyclic operators 41

Proposition 2.23. Let T be an invertible operator. Then T is hypercyclic if

and only if T −1 is.

Birkhoff’s operators provide examples of invertible hypercyclic operators.

Next, in our present context, Proposition 1.19 reads as follows.

Proposition 2.24. Hypercyclicity is preserved under quasiconjugacy.

We emphasize that the map φ defining the quasiconjugacy need not be

linear; it may, for instance, be defined between a complex space and a real
space; see Exercise 2.2.5.
We also formulate part of Proposition 1.42 in the linear setting.

Proposition 2.25. Let S : X → X and T : Y → Y be operators. If S ⊕ T is

hypercyclic then so are S and T .

We will see in Remark 4.17 that the converse fails in general. As an ap-
plication of Proposition 2.25 we obtain an interesting transference principle
that is specific to the linear setting. Let X be a real separable Fréchet space.
Then the complexification X  of X is defined formally as

 = {x + iy ; x, y ∈ X},
X

which will be identified with X ⊕ X. If multiplication by complex scalars is

defined by (a+ib)(x+iy) = (ax−by)+i(ay +bx), then X  becomes a complex
separable Fréchet space.
Moreover, let T : X → X be a (real-linear) operator on X. Then its
complexification T : X
 →X  is defined by

T(x + iy) = T x + iT y.

An easy computation shows that T is a (complex-linear) operator on X.  For


all of these statements see Exercise 2.2.7. Since T is nothing but the operator
T ⊕ T on X ⊕ X, Proposition 2.25 implies the following.

Proposition 2.26. Let T be an operator on a real separable Fréchet space.

If its complexification T is hypercyclic then so is T .

In fact, since T coincides with T ⊕ T , hypercyclicity of T is equivalent to

weak mixing of T ; see also Section 2.5.

Example 2.27. The complexification of the real Rolewicz operator T = λB,

λ ∈ R, |λ| > 1, on the spaces X := p , 1 ≤ p < ∞, or X := c0 , of real
sequences can be identified with the same operator as understood on the
corresponding spaces of complex sequences. Therefore, hypercyclicity for the
complex Rolewicz operators implies the same for the real operators.
42 2 Hypercyclic and chaotic operators

As a second application of Proposition 2.25 we consider restrictions of

hypercyclic operators to invariant subspaces. Let M1 and M2 be closed
subspaces of a (real or complex) Fréchet space X such that, algebraically,
X = M1 ⊕ M2 , that is, X = M1 + M2 and M1 ∩ M2 = {0}.
Note that, at the end of Section 2.1, we had defined M1 ⊕ M2 as the
topological product of the two spaces. But since the map φ : (x1 , x2 ) → x1 +x2
defines an algebraic isomorphism between M1 × M2 and M1 + M2 , it is also a
topological isomorphism by the inverse mapping theorem (see Appendix A),
so that the two forms of M1 ⊕ M2 can be identified.
Now suppose that T is an operator on X that leaves M1 and M2 invariant.
Then we have that

T x = T x1 + T x2 if x = x1 + x2 , x1 ∈ M1 , x2 ∈ M2 .

In the sense of the isomorphism φ we therefore have that

T = T |M1 ⊕ T |M2 ;

hence Proposition 2.25 implies the following.

Proposition 2.28. Let T : X → X be a hypercyclic operator, and let M1

and M2 be T -invariant closed subspaces of X such that X = M1 ⊕ M2 . Then
the restrictions T |M1 and T |M2 are hypercyclic.

2.3 Linear chaos

As defined in Chapter 1, chaos in the sense of Devaney consists in demanding

topological transitivity and the density of the set of periodic points. In view
of the Birkhoff transitivity theorem we can rephrase this definition in our
present setting.

Definition 2.29 (Linear chaos). An operator T is said to be chaotic if it

satisfies the following conditions:
(i) T is hypercyclic;
(ii) T has a dense set of periodic points.

We recall that sensitive dependence on initial conditions was a consequence

of chaos for metric spaces without isolated points; see Theorem 1.29. For
operators, hypercyclicity in itself already implies sensitive dependence.

Proposition 2.30. Let T be a hypercyclic operator. Then T has sensitive

dependence on initial conditions (with respect to any translation-invariant
metric defining the topology of X).
2.3 Linear chaos 43

Proof. Let d be any translation-invariant metric on X that induces its topol-

ogy. Let δ, ε > 0 and x ∈ X be arbitrary. We then consider the nonempty
open sets

U = {z ∈ X ; d(0, z) < ε}, V = {z ∈ X ; d(0, z) > δ}.

By the topological transitivity of T , there are n ∈ N0 and z ∈ U such that

T n z ∈ V . For the point y := x + z we then obtain that d(x, y) = d(0, z) < ε
and d(T n x, T n y) = d(0, T n z) > δ, which implies the result. 


Remark 2.31. Devaney’s notion of chaos has been generally accepted in linear
dynamics. There are, however, also other definitions of chaos. We mention
here that a continuous map T : X → X on a metric space (X, d) is called
chaotic in the sense of Auslander and Yorke if it is topologically transitive and
it has sensitive dependence on initial conditions. By the previous proposition,
every hypercyclic operator is Auslander–Yorke chaotic.

In some cases, the periodic points of an operator are easily determined.

As a first example we consider the multiples of backward shifts.

Example 2.32. (Rolewicz’s operators) Let T = λB, |λ| > 1, be Rolewicz’s

operator on X = p , 1 ≤ p < ∞, or X = c0 . One easily verifies that x ∈ X is
periodic if and only if there are N ∈ N and xk ∈ K, k = 1, . . . , N , such that

x = x1 , . . . , xN , λ−N x1 , . . . , λ−N xN , λ−2N x1 , . . . , λ−2N xN , . . . ).

In order to see that the set of periodic points is dense in X it suffices to ap-
proximate any finite sequence y = (y1 , . . . , yn , 0, . . .). By choosing a periodic
point whose∞ N ≥ n first coordinates coincide with those of y we see that
x − y ≤ j=1 |λ|−jN y → 0 as N → ∞. Therefore Rolewicz’s operators
are chaotic.

Let us observe that, for linear maps T on arbitrary vector spaces X, the
set of periodic points of T is a subspace of X. Indeed, let x, y ∈ X be periodic
points for T . Then we have that T n x = x and T m y = y for certain n, m ∈ N.
Thus T nm (ax + by) = a(T n )m x + b(T m )n y = ax + by, for any a, b ∈ K, so
that also ax + by is periodic.
There is, in fact, a nice and very useful description of the space of periodic
points in terms of eigenvectors to unimodular eigenvalues, that is, eigenvalues
of absolute value 1, provided that we have a complex space. The correspond-
ing result is of a purely algebraic nature.

Proposition 2.33. Let T be a linear map on a complex vector space X. Then

the set of periodic points of T is given by

Per(T ) = span{x ∈ X ; T x = eαπi x for some α ∈ Q}.

44 2 Hypercyclic and chaotic operators

Proof. If T x = eαπi x with α = nk , k ∈ Z, n ∈ N, then T 2n x = x, so that x is

periodic. This yields one inclusion.
For the other one, suppose that T n x = x, n ∈ N. We then decompose the
polynomial z n − 1 into a product of monomials,

z n − 1 = (z − λ1 )(z − λ2 ) · · · (z − λn ).

1, . . . , n, are different, the system {p1 , . . . , pn } of

Since all the roots λk , k =
polynomials with pk (z) := j=k (z − λj ), 1 ≤ k ≤ n, is a basis of the space
of polynomials of degree strictly less than n. In particular, there are αk ∈ C,
k = 1, . . . , n, such that

n
1= αk pk (z), z ∈ C.
k=1

This means that, when we substitute z by T , then


n
I= αk pk (T ).
k=1
n
We therefore have that x = k=1 αk yk with yk := pk (T )x, k = 1, . . . , n.
Since (T − λk )yk = (T n − I)x = 0 with λnk = 1, we see that x belongs to the
desired span.  

In many concrete situations, this proposition leads to a simple verification

that a given operator has a dense set of periodic points. As an example
we consider Birkhoff’s and MacLane’s operators. We first need the following
result, where eλ denotes the exponential function

eλ (z) = eλz , z ∈ C.

Lemma 2.34. Let Λ ⊂ C be a set with an accumulation point. Then the set

span{eλ ; λ ∈ Λ}

is dense in H(C).

Proof. By assumption there are λ ∈ C and λn ∈ Λ with λn → λ and λn = λ

for all n ≥ 1. Writing

(λn − λ)2 z 2
eλn z = eλz e(λn −λ)z = eλz + eλz (λn − λ)z + eλz + ... (2.4)
2!
we see that
eλn z → eλz uniformly on compact sets,
which, incidentally, also follows directly. Therefore, eλ ∈ span{eλn ; n ≥ 1}.
But now (2.4) also shows that
2.3 Linear chaos 45

eλn z − eλz (λn − λ)z 2

= eλz z + eλz + ...,
λn − λ 2!
and hence
eλn z − eλz
→ zeλz uniformly on compact sets,
λn − λ

so that also the function z → zeλz belongs to span{eλn ; n ≥ 1}.

Continuing in this way we find that all functions z → z k eλz , k ≥ 0, belong
to span{eλn ; n ≥ 1}.
Now let f ∈ H(C). Then we have that

  
∞
∞
f (z) = eλz e−λz f (z) = eλz ak z k = ak z k eλz
k=0 k=0

with suitable coefficients ak ∈ C, k ≥ 0, where convergence takes place in

H(C). Thus we also have that f ∈ span{eλn ; n ≥ 1}, which had to be shown.



The lemma allows us to show that Birkhoff’s and MacLane’s operators are
chaotic on H(C).

Example 2.35. (Birkhoff’s and MacLane’s operators) For the differenti-

ation operator D, any function eλ is an eigenvector of D to the eigenvalue λ.
Thus, since the subspace

span{eλ ; λ = eαπi for some α ∈ Q}

is dense in H(C) by Lemma 2.34, Proposition 2.33 tells us that Per(T ) is

dense. Since we already know that D is hypercyclic, it is also chaotic.
For the translation operators Ta , a ∈ C \ {0}, any function eλ is an eigen-
vector of Ta to the eigenvalue eaλ . Thus, since also the subspace

span{eλ ; eaλ = eαπi for some α ∈ Q} = span{eλ ; λ = α

a
πi, α ∈ Q}

is dense in H(C), we conclude as before that each Ta is chaotic.

The restriction that Proposition 2.33 only holds for complex spaces can
sometimes be overcome by using complexifications: an operator on a real
space is chaotic if and only if its complexification is, as we will see in Corol-
lary 2.51 below.
46 2 Hypercyclic and chaotic operators

2.4 Mixing operators

In this and the following section we study the mixing and weak mixing prop-
erties in the light of linearity.
Since we are in the setting of Fréchet spaces, the proofs could be formulated
in terms of their seminorms or their metric. However, the arguments become
particularly transparent when we use the topological language of open sets
and 0-neighbourhoods. For this we only need the following simple result.
As usual, we set A + B = {a + b ; a ∈ A, b ∈ B} for subsets A, B of a
vector space.

Lemma 2.36. Let X be a Fréchet space. If U ⊂ X is a nonempty open set

then there is a nonempty open subset U1 ⊂ U and a 0-neighbourhood W such
that U1 + W ⊂ U . If W is a 0-neighbourhood then there is a 0-neighbourhood
W1 such that W1 + W1 ⊂ W .

Proof. Let  ·  be an F-norm defining the topology of X. Then there is some

x0 ∈ U and some ε > 0 such that Uε (x0 ) = {x ∈ X ; x − x0  < ε} is
contained in U . One may then take U1 = Uε/2 (x0 ) and W = Uε/2 (0). The
second claim follows similarly upon taking x0 = 0. 

We recall that the mixing property consists in demanding the cofiniteness

of the return sets N (U, V ) for each pair U, V of nonempty open subsets of
X. For operators this requirement can be weakened.

Proposition 2.37. An operator T is mixing if and only if, for any nonempty
open set U ⊂ X and any 0-neighbourhood W , the return sets

N (U, W ) and N (W, U )

are cofinite.

Proof. It suffices to show sufficiency of the condition. Let U, V ⊂ X be

nonempty open sets. By Lemma 2.36 there are nonempty open sets U1 , V1
and a 0-neighbourhood W such that U1 + W ⊂ U and V1 + W ⊂ V . By
hypothesis, there exists some N ∈ N such that, for any n ≥ N , there are
u ∈ U1 and w ∈ W so that T n u ∈ W and T n w ∈ V1 . But then u + w ∈ U
and T n (u + w) = T n u + T n w ∈ V , which implies that N (U, V ) is cofinite.



By following the proofs of the hypercyclicity of Rolewicz’s, Birkhoff’s and

MacLane’s operators, one immediately obtains that they are even mixing.

Example 2.38. (Birkhoff’s, MacLane’s and Rolewicz’s operators) The

three classical hypercyclic operators are mixing.

We have also seen that these mixing operators are even chaotic. This is
not always the case, as the following example shows.
2.4 Mixing operators 47

Example 2.39. We consider the weighted shift T : 1 → 1 given by

 
T (x1 , x2 , . . . ) = 2x2 , 32 x3 , 43 x4 , . . . .

Let U be a nonempty open subset of 1 and W a 0-neighbourhood. By

density of the finite sequences in 1 there is a sequence of the form u =
(u1 , . . . , uN , 0, 0, . . . ) in U .
Since T n u = 0 whenever n ≥ N , the set N (U, W ) is cofinite. On the other
hand, we have that
     
T n x = (n + 1)xn+1 , n+2 2 xn+2 , n+3
3 xn+3 , . . . , n ≥ 1.

Thus, if we define w ∈ 1 by wk = k−n k uk−n for k = n + 1, . . . , n + N , and

wk = 0 otherwise, then T n w = u and
N
w ≤ u,
n+1
so that also N (W, U ) is cofinite. Hence T is a mixing operator.
We now show that T has no nontrivial periodic points and therefore cannot
be chaotic. Indeed, let us suppose that x = 0 is periodic for T , that is, there
is some n ∈ N with T n x = x, hence also T jn x = x for all j ∈ N. Using the
k xjn+k = xk , for all k, j ∈ N. Now,
above formula for T n x we obtain that jn+k
since x = 0 there is some k ∈ N with xk = 0. Hence
∞
 ∞
 k
x ≥ |xjn+k | = |xk | = ∞,
j=1 j=1
jn + k

which is a contradiction.
We reformulate part of a previous result, Proposition 1.42, for operators.
Proposition 2.40. Let S : X → X and T : Y → Y be hypercyclic operators.
If at least one of them is mixing then S ⊕ T is hypercyclic. Moreover, S ⊕ T
is mixing if and only if both S and T are.
For a later application (see Proposition 8.5), we also need to consider
direct sums of countably many operators on Banach spaces. Thus, let Tn be
operators on separable Banach spaces Xn , n ≥ 1. For 1 ≤ p < ∞, we define
the direct p -sum of these spaces as

∞
 ∞
 
Xn = (xn )n≥1 ; xn ∈ Xn , n ≥ 1, and xn p < ∞ ;
p
n=1 n=1
∞
endowed with the norm (xn )n  = ( n=1 x n )
p 1/p
this space turns into a
∞
separable Banach space. The direct c0 -sum ( n=1 Xn )c0 is defined similarly.
Now suppose that supn∈N Tn  < ∞. Then the direct sum of the operators
Tn , defined by
48 2 Hypercyclic and chaotic operators


∞

Tn (xn )n = (Tn xn )n ,
n=1
∞ ∞
is an operator on ( n=1 Xn )p and on ( n=1 Xn )c0 .

Proposition 2.41. Let Tn be operators on separable Banach

∞ spaces Xn , n ≥
1,with supn∈N Tn  < ∞. Let 1 ≤ p < ∞. Then n=1 n is mixing on
T
∞
( n=1 Xn )p if and only if each operator Tn , n ≥ 1, is mixing.
The same result holds for the direct c0 -sum.

Proof. For the necessity ∞ part one need only note that each Tn , n ≥ 1, is
quasiconjugate to k=1 Tk via the map φ : (xk )k → xn . 
Now suppose that each Tn , n ≥ 1, is mixing. Let U, V ⊂ ( ∞ n=1 Xn ) be
p

nonempty open sets. It follows from the definition of the norm on this space
that there are ε > 0, m ≥ 1, and points x := (x1 , . . . , xm , 0, 0, . . .) ∈ U and
y := (y1 , . . . , ym , 0, 0, . . .) ∈ V such that the open balls of radius ε around
these points belong to U and V , respectively. Since each Tk is mixing there is
(n)
some N ≥ 1 such that, for each 1 ≤ k ≤ m and n ≥ N , there are xk ∈ Xk
(n) (n)
such that xk − xk  < ε/m1/p and Tkn xk − yk  < ε/m1/p . Then, for all
(n) (n) ∞
n ≥ N , x(n) := ∞ (x1 , . . . , xm , 0, 0, . . .) ∈ U and ( k=1 Tk )n x(n) ∈ V , which
implies that k=1 Tk is mixing. The proof for direct c0 -sums is similar.  

2.5 Weakly mixing operators

In our present context, an operator T : X → X is weakly mixing if and only

if T ⊕ T is hypercyclic, and if and only if, for any nonempty open subsets
U1 , U2 , V1 and V2 of X, N (U1 , V1 ) ∩ N (U2 , V2 ) = ∅.

Observation 2.42. For any linear dynamical system,

mixing =⇒ weak mixing =⇒ hypercyclicity.

The study of specific hypercyclic operators in Chapter 4 will lead to many

simple examples of weakly mixing, non-mixing operators; see Remark 4.10.
In contrast, the strictness of the second implication turned out to be much
more delicate and was posed as an open problem by D. Herrero in 1992. The
problem has only recently been solved.

Theorem 2.43 (De la Rosa–Read). There are hypercyclic operators on

Banach spaces that are not weakly mixing.

Refining the techniques of De la Rosa and Read, Bayart and Matheron have
shown that such operators even exist on any of the spaces p , 1 ≤ p < ∞, and
c0 , in particular on Hilbert spaces. The proof is, however, beyond the scope
of this book.
2.5 Weakly mixing operators 49

The answer to Herrero’s question raises the problem of finding (weak)

conditions on a hypercyclic operator to be weakly mixing.
To this end we first derive a useful property of hypercyclic operators in-
volving open sets and 0-neighbourhoods.

Lemma 2.44. Let T be a hypercyclic operator. Then, for any nonempty open
sets U and V in X and any 0-neighbourhood W , there is a nonempty open
set U1 ⊂ U and a 0-neighbourhood W1 ⊂ W such that

N (U1 , W1 ) ⊂ N (V, W ) and N (W1 , U1 ) ⊂ N (W, V ).

Proof. Using topological transitivity and continuity of T one finds m ∈ N0 ,

a nonempty open set U1 ⊂ U and a 0-neighbourhood W1 ⊂ W such that
T m (U1 ) ⊂ V and T m (W1 ) ⊂ W . Now, if n ∈ N (U1 , W1 ), then there exists
some x ∈ U1 with T n x ∈ W1 . It follows that T n T m x = T m T n x ∈ W , so that
n ∈ N (V, W ). In the same way we also obtain that N (W1 , U1 ) ⊂ N (W, V ).



This proof copies our proof of the 4-set trick (see Lemma 1.50); a direct
application of that trick would not have given us that W1 contains 0.

Theorem 2.45. Let T be a hypercyclic operator. If, for any nonempty open
set U ⊂ X and any 0-neighbourhood W , there is a continuous map S : X →
X commuting with T such that

S(U ) ∩ W = ∅ and S(W ) ∩ U = ∅, (2.5)

then T is weakly mixing.

Proof. First, the 4-set trick and topological transitivity of T yield that, for
any nonempty open set U ⊂ X and for any 0-neighbourhood W ,

N (U, W ) ∩ N (W, U ) = ∅.

By Proposition 1.53 it suffices to show that, given any pair U, V of

nonempty open subsets of X, there is n ∈ N (U, U ) ∩ N (U, V ). To do
this, using Lemma 2.36, we fix nonempty open sets U1 ⊂ U , V1 ⊂ V
and a 0-neighbourhood W1 such that U1 + W1 ⊂ U and V1 + W1 ⊂ V .
Lemma 2.44 implies the existence of a 0-neighbourhood W2 ⊂ W1 and a
nonempty open set U2 ⊂ U1 such that N (W2 , U2 ) ⊂ N (W1 , V1 ). We fix n ∈
N (U2 , W2 ) ∩ N (W2 , U2 ); then there are u2 ∈ U2 with T n u2 ∈ W2 , w1 ∈ W1
with T n w1 ∈ V1 , and w2 ∈ W2 with T n w2 ∈ U2 . If we set u3 = u2 + w2 ∈ U
and u4 = u2 + w1 ∈ U , then we obtain that T n u3 ∈ W2 + U2 ⊂ U and
T n u4 ∈ W2 + V1 ⊂ V . That is, n ∈ N (U, U ) ∩ N (U, V ).  

An operator T : X → X is called flip transitive if, for any pair U, V of

nonempty open subsets of X,
50 2 Hypercyclic and chaotic operators

N (U, V ) ∩ N (V, U ) = ∅;

see also Exercises 1.5.4 and 1.5.5. By the Birkhoff transitivity theorem, such
an operator is hypercyclic. It follows from the previous result that even more
is true.

Corollary 2.46. Every flip transitive operator is weakly mixing.

As another consequence we obtain a useful characterization of the weak

mixing property; this result should also be compared with Proposition 2.37.

Fig. 2.1 Weak mixing (Theorem 2.47)

Theorem 2.47. An operator T is weakly mixing if and only if, for any
nonempty open sets U, V ⊂ X and any 0-neighbourhood W ,

N (U, W ) ∩ N (W, V ) = ∅.

Proof. It suffices to show sufficiency. As in the proof of Proposition 2.37 the

condition implies that T is topologically transitive, hence hypercyclic. Then
an application of Theorem 2.45 yields the result.  

Theorem 2.45 provides us with a rather weak condition, in terms of open

sets and 0-neighbourhoods, for making a hypercyclic operator weakly mixing.
As an application we can show that weak mixing is already implied by the
existence of a dense set of points with tame orbits. For the precise formulation
we need the notion of bounded sets in Fréchet spaces; see Appendix A.

Theorem 2.48. Let T be a hypercyclic operator. If there exists a dense subset

X0 of X such that the orbit of each x ∈ X0 is bounded, then T is weakly
mixing.

Proof. Let U be a nonempty open subset of X and W a 0-neighbourhood.

If (pn )n is an increasing sequence of seminorms defining the topology of X
then there is some k ∈ N and some ε > 0 such that pk (x) < ε implies that
2.5 Weakly mixing operators 51

x ∈ W . Now, by assumption, we can find a point x ∈ X0 ∩ U ; it follows

that M := supn∈N0 pk (T n x) < ∞. Hence 2M ε
T n x ∈ W , for all n ∈ N0 . On
the
 ε other hand,
 by topological transitivity of T , there is some n ∈ N0 with
2M T n
(W ) ∩ U ε
= T n ( 2M W ) ∩ U = ∅. Thus, condition (2.5) is satisfied for
S = 2M T . 
ε n

Of course, every periodic point has bounded orbit, as does every point
whose orbit converges. The latter holds, a fortiori, for all points from the
generalized kernel
∞

ker T n
n=0

of T . Thus we have the following.

Corollary 2.49. Any of the following operators are weakly mixing:
(i) chaotic operators;
(ii) hypercyclic operators that have a dense set of points for which the orbits
converge;
(iii) hypercyclic operators with dense generalized kernel.
A typical class of operators with dense generalized kernel is the class of
unilateral weighted shifts that will be studied in detail in Section 4.1.
For an additional characterization of weakly mixing operators in terms of
multiples of the iterates of T we refer to Theorem 12.29.
We end this section with an application to complexifications.
Proposition 2.50. Let T : X → X be an operator. Then:
(i) T ⊕ T is weakly mixing if and only if T is;
(ii) T ⊕ T is chaotic if and only if T is.
Proof. Suppose that T is chaotic. Then, by Corollary 2.49, T ⊕ T is hyper-
cyclic. Moreover, the set of all points (x, y) ∈ X ⊕ X with periodic points x
and y for T provides a dense set of periodic points for T ⊕ T . Thus T ⊕ T
is chaotic. The remaining implications are special cases of Propositions 1.42
and 1.48 and Theorem 1.51.  
More generally, an arbitrary direct sum S ⊕ T is chaotic if and only if
both S and T are; see Exercise 2.5.7. In contrast, assertion (i) cannot be
generalized in the same way; see Remark 4.17.
The discussion before Proposition 2.26, together with Proposition 2.40,
yields the following.
Corollary 2.51. An operator T on a real separable Fréchet space is mixing,
weakly mixing or chaotic, respectively, if and only if its complexification T is.
This result can be applied, for instance, to Rolewicz’s operators, just as in
Example 2.27; see also Example 3.2.
For extensions of the results of the last two sections to sequences of oper-
ators we refer to Section 3.4.
52 2 Hypercyclic and chaotic operators

2.6 The set of hypercyclic vectors

A natural question that arises in hypercyclicity is this: which kind of struc-

tures can we find in the set of hypercyclic vectors? By the Birkhoff transitivity
theorem we already know that the set HC(T ) of hypercyclic vectors of a hy-
percyclic operator T is always a dense Gδ -set. Almost trivially this leads to
a somewhat surprising representation result.

Proposition 2.52. Let T be a hypercyclic operator on X. Then

X = HC(T ) + HC(T ),

that is, every vector x ∈ X can be written as the sum of two hypercyclic
vectors.

Proof. Let x ∈ X. Since both HC(T ) and x − HC(T ) are dense Gδ -sets,
their intersection must be nonempty by the Baire category theorem, which
implies that x ∈ HC(T ) + HC(T ). 

As a consequence, the set HC(T ) of hypercyclic vectors can only then

be a linear subspace, except for the zero vector, if any nonzero vector is
hypercyclic, in which case the operator has no nontrivial invariant closed
subset; see Observation 2.17. Such an operator exists, for example, on 1 , but
the construction is highly nontrivial.
Weakening the requirement, it is natural to ask if, for a general hyper-
cyclic operator T , HC(T ) contains a large linear subspace, except for 0. In
this section we will interpret largeness as being dense. A different sense of
largeness will be studied in Chapter 10.
We first need some auxiliary results that are also important in their own
right. For the definition of the adjoint of an operator and the notation x, x∗ 
we refer to Appendix A.

Lemma 2.53. (a) Let T be a hypercyclic operator. Then its adjoint T ∗ has
no eigenvalues. Equivalently, every operator T − λI, λ ∈ K, has dense range.
(b) Let T be a hypercyclic operator on a real separable Fréchet space. Then
the adjoint T∗ of its complexification T has no eigenvalues. Equivalently,
every operator T − λI, λ ∈ C, has dense range.

Proof. (a) Let x ∈ X be a hypercyclic vector for T . Suppose, by way of

contradiction, that T ∗ has an eigenvalue λ, that is,

T ∗ x∗ = λx∗

for some x∗ ∈ X ∗ , x∗ = 0. Then we have that, for any n ≥ 0,

T n x, x∗  = x, (T ∗ )n x∗  = λn x, x∗ .
2.6 The set of hypercyclic vectors 53

Since x∗ = 0, the hypercyclicity of x implies that the left-hand side is dense in

K, while the right-hand side clearly is not, which is the desired contradiction.
Moreover, by the Hahn–Banach theorem (see Appendix A), T − λI has
dense range precisely when

x, T ∗ x∗ − λx∗  = (T − λI)x, x∗  = 0 for all x ∈ X

entails that x∗ = 0, which is equivalent to λ not being an eigenvalue of T ∗ .

(b) Now let X be a space over the real scalar field, and let T be the
complexification of T . Let x ∈ X be hypercyclic for T , and suppose that T∗
has an eigenvector x  ∗, x
∗ ∈ X ∗ = 0, to an eigenvalue λ. Then we have that

∗ | = |Tn x, x
|T n x, x ∗ | = |x, (T∗ )n x∗ | = |λ|n |x, x
∗ |, (2.6)

∗  = x1 , x
for n ≥ 0. Since x1 + ix2 , x ∗  + ix2 , x
∗  for all x1 , x2 ∈ X, and
∗
 = 0, there is some y ∈ X such that |y, x
since x ∗ | > 0. Hence | x∗ | can
take every positive value on X. By the hypercyclicity of x, the left-hand side
of (2.6) is dense in R+ , while the right-hand side clearly is not, which is a
contradiction. The remainder of the proof can be given as in (a).  

As a consequence we obtain one of the cornerstones of the theory

 of linear
dynamical systems. We recall that for any polynomial p(z) = N n
n=0 n z the
a
N n
operator p(T ) is defined as p(T ) = n=0 an T .

Theorem 2.54 (Bourdon). If T is a hypercyclic operator and p is a nonzero

polynomial, then the operator p(T ) has dense range.

Proof (complex case). We can assume that p(z) = N n=0 an z with aN = 0,
n

N ≥ 1. For spaces X over the complex field the result follows immediately
from Lemma 2.53(a) and the fact that p can be written as a product of linear
factors, so that
p(T ) = aN (T − λ1 I) · · · (T − λN I)
with certain λk ∈ C, k = 1, . . . , N .
(Real case). If X is a real space we consider the complexification T of T .
With Lemma 2.53(b), it follows as in the complex case that, for any complex
polynomial p, p(T) has dense range on X. Now if p has real coefficients, then

p(T)(x + iy) = p(T )x + ip(T )y, x, y ∈ X,

which implies that also p(T ) : X → X has dense range. 



We are now ready to deduce an important result on the algebraic structure

of the set of hypercyclic vectors.

Theorem 2.55 (Herrero–Bourdon). If x is a hypercyclic vector for T ,

then
{p(T )x ; p is a polynomial} \ {0}
54 2 Hypercyclic and chaotic operators

is a dense set of hypercyclic vectors.

In particular, any hypercyclic operator admits a dense invariant subspace
consisting, except for zero, of hypercyclic vectors.

Proof. Let x ∈ X be a hypercyclic vector for T . Then

M = {p(T )x ; p is a polynomial} = span orb(x, T )

is a dense T -invariant subspace of X. Moreover, if y = p(T )x ∈ M \ {0} then

p = 0 and
T n y = p(T )(T n x), n ∈ N0 .
Since x is hypercyclic and, by Theorem 2.54, p(T ) has dense range, also y
has dense orbit under T . 

The Herrero–Bourdon theorem allows us to deduce an additional topolog-

ical structure of the set of hypercyclic vectors: it is always a connected set.
This observation comes from the fact that, if A ⊂ B ⊂ A ⊂ X and A is
connected, then also B is connected. We apply this to A = M \ {0}, and
B = HC(T ), where M is the dense subspace of the Herrero–Bourdon theo-
rem. Note that M is of dimension greater than 1 because, otherwise, x would
be an eigenvector, which is not hypercyclic; hence A = M \ {0} is connected.

Corollary 2.56. The set HC(T ) of hypercyclic vectors for a hypercyclic op-
erator T is a connected subset of X.

2.7 Linear vs nonlinear maps, and finite vs infinite

dimension

We have seen that chaotic linear operators exist. This is contrary to the com-
mon belief that (deterministic) chaos is necessarily connected to the nonlin-
earity of a system. In this section we want to explore the connection between
linear and nonlinear chaos.
The dynamics of linear operators on a finite-dimensional space X = KN
are easy to describe, thanks to the Jordan decomposition theorem. We assume
that KN is endowed with the Euclidean norm.

Proposition 2.57. Let T be a linear operator on KN , N ≥ 1. Then, for any

x ∈ KN , either T n x → 0 or T n x → ∞ or there are m, M > 0 such that
m ≤ T n x ≤ M for all n ≥ 0.

Proof. Since every operator T : RN → RN can be regarded as an operator

on CN it suffices to consider the complex case.
By the Jordan decomposition theorem, CN has a basis with respect to
which the matrix of T is in Jordan block form. Since all norms on CN are
2.7 Linear vs nonlinear maps, and finite vs infinite dimension 55

equivalent we can assume that this basis is the canonical basis of CN , and it
suffices to show the result for each operator given by a Jordan block
⎛ ⎞
λ 1 0 ... ...
⎜0 λ 1 0 ...⎟
⎜ ⎟
⎜ ⎟
T = ⎜ . . . . . . . . . ... ⎟ : CN → CN ,
⎜ ⎟
⎝ 0 λ 1 ⎠
0 λ

with N ≥ 1. For n ≥ N − 1 we have that

⎛ n     ⎞
λ nλn−1 n2 λn−2 . . . . . . Nn−1λn−N +1
⎜ 0 λn
⎜ nλn−1 . . . . . . n
N −2
λn−N +2 ⎟
⎟
⎜ . . . .. ⎟
⎜ . . . . . . ... ⎟
Tn = ⎜   . ⎟.
⎜ n n−1 n n−2 ⎟
⎜ 0 λ nλ 2
λ ⎟
⎝ 0 λ n
nλ n−1 ⎠
n
0 λ

We apply T n to the vector x = (x1 , . . . , xN ) ∈ CN , x = 0.

Case 1 : |λ| > 1. Let xk be the last nonzero entry. Then the kth entry of
T n x is λn xk , hence T n x → ∞.
Case 2 : |λ| < 1. Then all entries of T n x tend to zero, hence T n x → 0.
Case 3 : |λ| = 1. If x = (x1 , 0, . . . , 0) then T n x = λn x = x for n ≥ 0.
Otherwise, the first entry of T n x is
 
λn x1 + nλn−1 x2 + . . . + Nn−1 λn−N +1 xN ,

which tends to infinity in absolute value; hence, again, T n x → ∞. 



As an immediate consequence we obtain the following.

Theorem 2.58. There are no hypercyclic operators on KN , N ≥ 1.

Of course, this also follows directly from Lemma 2.53 since every opera-
tor on CN has an eigenvalue. Further proofs of this result are suggested in
Exercises 2.7.1 and 2.7.2.
Since every finite-dimensional Fréchet space is isomorphic to some KN , N ≥
1 (see Appendix A), the theorem extends to such spaces.

Corollary 2.59. There are no hypercyclic operators on a finite-dimensional

Fréchet space.

The result also implies an interesting property of the orbit of a hypercyclic

vector.

Proposition 2.60. The orbit of any hypercyclic vector forms a linearly in-
dependent set.
56 2 Hypercyclic and chaotic operators

Proof. Let x be a hypercyclic vector for T and suppose that there are scalars
αk ∈ K, k = 0, . . . , N, such that


N
T N +1 x = αk T k x.
k=0

Then F := span{T k x ; k = 0, . . . , N } is a finite-dimensional T -invariant

subspace of X. Since x is hypercyclic for T , it is also hypercyclic for T |F :
F → F , which contradicts Corollary 2.59.  

Alternatively, the proof can be based on Bourdon’s theorem; see Exercise

2.7.3.
Proposition 2.57 tells us that the dynamics of linear maps on finite-
dimensional spaces are quite restrictive. On the other hand, in an infinite-
dimensional setting, linear dynamics can be arbitrarily complicated. Indeed,
as we now show, every continuous map on a compact metric space is conju-
gate to the restriction of a linear operator on some invariant set. Even more
strikingly, the same operator can be taken for all nonlinear systems, and the
operator is even chaotic. In other words: the dynamics of any (compact) non-
linear dynamical system can be described by the dynamics of a single chaotic
operator.

Theorem 2.61. There exists a chaotic operator T on a separable Hilbert

space H with the following property.
For any continuous map f on any compact metric space K there exists a
T -invariant subset L of H such that f is conjugate to the restriction T |L of
T to L.

In other words, there is a homeomorphism φ : K → L so that the diagram

f
K −−−−→ K
⏐ ⏐
⏐
φ
⏐φ

T |L
L −−−−→ L
commutes.
∞
Proof. Let H = ( n=0 2 )2 be the space of all sequences x = (xn )n≥0 of
elements xn = (xn,k )k≥0 in 2 such that

∞
1/2
x := xn 2 < ∞;
n=0

see the discussion before Proposition 2.41. Then H is a separable Hilbert

space when endowed with the canonical inner product. On H we consider
the multiple T = 2B of the backward shift B given by
2.7 Linear vs nonlinear maps, and finite vs infinite dimension 57

B(x0 , x1 , x2 , . . .) = (x1 , x2 , x3 , . . .).

The proof that T has the desired properties will be split into three steps.

Step 1. We define a suitable embedding φ : K → H.

First, we can assume that the metric d on K is bounded by 1, since oth-
erwise we can replace it by the equivalent metric d (x, y) = min(1, d(x, y)).
We then fix a dense sequence (yk )k≥0 in K; this is possible because K is a
compact metric space. Based on this sequence we define, for any x ∈ K,

1

φ(x) = n
d(yk , f (x)) ∈ H.
2k+n k n

Then we have for any x, y ∈ K and N ≥ 0 that

∞
 1
φ(x) − φ(y)2 = |d(yk , f n (x)) − d(yk , f n (y))|2
22(k+n)
k,n=0
 1  4
≤ |d(yk , f n (x)) − d(yk , f n (y))|2 + .
k,n≤N
22(k+n) k>N or
22(k+n)
n>N

This can be made arbitrarily small by first choosing N sufficiently large and
then y sufficiently close to x. Thus φ : K → H is continuous.
Moreover, φ is injective; indeed, φ(x) = φ(y) implies that
1 1
k
d(yk , x) = k d(yk , y) for all k ≥ 0,
2 2
hence x = y by density of the yk .
As a continuous injection on a compact space, φ is a homeomorphism onto
a (compact) subset, L say, of H.

Step 2. We show that φ ◦ f = T ◦ φ and that L is invariant under T .

In fact, for any x ∈ K we have that


1 n+1
φ(f (x)) = d(yk , f (x))
2k+n k n


1 n+1
=2 d(y k , f (x)) = T (φ(x)),
2k+n+1 k n

which also implies that

T (L) = T (φ(K)) = φ(f (K)) ⊂ φ(K) = L.

Step 3. The operator T is chaotic on H.

58 2 Hypercyclic and chaotic operators

The proof is the same as the proof that Rolewicz’s operators are chaotic;
see Examples 2.22 and 2.32.  

In summary we have found that

• linear chaos exists;
• linear chaos is an infinite-dimensional phenomenon;
• linear dynamics can be as complicated as nonlinear dynamics.

2.8 Hypercyclicity and complex dynamics

In this section we present a connection between the hypercyclicity of weighted

backward shifts and the Julia sets of polynomials in one complex variable.
In some sense we continue the theme of the previous section by comparing
infinite-dimensional dynamics with nonlinear but finite-dimensional dynam-
ics.
Let us first give a brief introduction to complex dynamics. Given a poly-
nomial p in one complex variable, of degree m ≥ 2, a periodic point z of p is
called repelling if |p (z)| > 1. The Julia set of p can be defined as

J (p) = {z ∈ C ; z is a repelling periodic point of p}.

While this is not the common definition of the Julia set, which is more com-
plicated and involves the behaviour of the iterates of p near J (p), a result
by Fatou and Julia tells us that the two definitions are equivalent.
For instance, for the doubling map p(z) = z 2 on C (see Example 1.37), the
periodic points are given by e2πiα with α = 2nk−1 , n, k ∈ N, and all of them
are repelling, so that J (p) = T. We have also seen that p|T is chaotic.
This feature is shared by any complex polynomial of degree m ≥ 2. More
precisely one always has that J (p) is a compact p-invariant set such that
p|J (p) is chaotic.
The dynamical behaviour of p near the Julia set consists in spreading
points. Indeed, the following property, which can be regarded as a multi-
point approximation by the iterates of p on points near J (p), characterizes
the Julia set: a point z ∈ C belongs to J (p) if and only if

∀ε > 0, ∀z1 , . . . , zk ∈ C, ∃z1 , . . . , zk ∈ C, ∃n ∈ N such that

(2.7)
|zj − z| < ε and |pn (zj ) − zj | < ε, j = 1, . . . , k.

Now, one can deduce this property, for certain polynomials and at the
point 0, from the hypercyclic behaviour of Rolewicz’s operators.

Example 2.62. Let p(z) = (z + 1)m − 1, where m ≥ 2. We consider the

following commutative diagram,
Exercises 59

mB
c0 −−−−→ c0
⏐ ⏐
⏐
φ
⏐φ

P
c0 −−−−→ c0 ,
where mB is the Rolewicz operator with λ = m on the complex space c0 ,
P (x1 , x2 , . . . ) = (p(x2 ), p(x3 ), . . . ), and φ(x1 , x2 , . . . ) = (ex1 − 1, ex2 − 1, . . . ).
It is easy to see that P and φ are continuous maps and that φ has dense
range. By Proposition 1.19, it follows from the hypercyclicity of Rolewicz’s
operators that P has a dense orbit.
In order to verify condition (2.7) at z = 0, we fix ε > 0 and arbitrary
zj ∈ C, j = 1, . . . k. Since P has a dense orbit we can find w ∈ c0 and n ≥ 0
such that

w < ε and P n w − (z1 , . . . , zk , 0, 0, . . . ) < ε.

By considering the (j + n)th coordinates wj+n of w, we deduce that

|wj+n | < ε and |pn (wj+n ) − zj | < ε, j = 1, . . . , k.

This shows that condition (2.7) holds for z = 0.

Let us mention that the map P in this example is a polynomial on c0 ; see
Exercise 2.8.1.
More generally, we have the following connection between infinite-dimen-
sional dynamics and the dynamics of arbitrary complex polynomials; its proof
is left to the reader: see Exercise 2.8.2.
Proposition 2.63. Let p be a complex polynomial of degree m ≥ 2 such that
p(0) = 0. Let P : c0 → c0 be the continuous map given by P (x1 , x2 , . . . ) :=
(p(x2 ), p(x3 ), . . . ). Then P is topologically transitive if and only if 0 belongs
to the Julia set J (p) of p.

Exercises

Exercise 2.1.1. Let (pn )n be a separating increasing sequence of seminorms on a vector

space X. Show that the map d defined in (2.1) is a metric on X. Moreover, show that
for any x, y ∈ X and n ∈ N:
(i) if pn (x − y) < 21n then d(x, y) < 22n ;
(ii) if d(x, y) < 212n then pk (x − y) < 21n for all k ≤ n.

Exercise 2.1.2. Prove Lemma 2.6. (Hint: Use the previous exercise.)

Exercise 2.1.3. Show that the spaces H(C) and ω are Fréchet spaces. Show that for
every separable Fréchet space X, X N is also a separable Fréchet space.

Exercise 2.1.4. Prove Proposition 2.8.

60 2 Hypercyclic and chaotic operators

Exercise 2.1.5. Let X = C ∞ (R) be the space of infinitely differentiable (real or com-
plex) functions f : R → K, endowed with the seminorms
 
pn (f ) = max sup f (k) (x) .
0≤k≤n |x|≤n

Show that X is a separable Fréchet space. (Hint: Use the Weierstrass approximation
theorem to approximate f (n) on [−n, n] by a polynomial.)

Exercise 2.1.6. Let 1 ≤ p < ∞. Let v : R+ → R be a strictly positive continuous

function. We define the space of weighted p-integrable functions by

X = Lpv (R+ ) = {f : R+ → K ; f is measurable and f  < ∞},

∞
where f  := ( 0 |f (x)|p v(x)dx)1/p . Then X is a separable Banach space. Show that
the translation operator T : X → X, (T f )(x) = f (x + 1), is a well-defined operator if
and only if
v(x)
sup < ∞.
x∈R+ v(x + 1)

Exercise 2.1.7. Let X be a Fréchet space with defining increasing sequence of semi-
norms (pn )n . Then a seminorm p : X → R is continuous if and only if there are n ≥ 1
and M > 0 such that
p(x) ≤ M pn (x), x ∈ X.
Show that this immediately implies the nontrivial part of Proposition 2.11.

Exercise 2.2.1. Let X = C0 (R+ ) = {f : R+ → R ; f is continuous and limx→∞ f (x) =

0}, endowed with the sup-norm f  = supx∈R+ |f (x)|. Given a > 0 and λ > 1, consider
the operator T : X → X, (T f )(x) = λf (x + a), x ∈ R+ . Show that T is hypercyclic.
(Hint: Use the fact that the continuous functions with compact support form a dense
subset of X.)

Exercise 2.2.2. Let X be a Banach space. Show that there are no operators T on X
for which λT is hypercyclic for all λ = 0. (Hint: Consider |λ| ≤ 1/T .)

Exercise 2.2.3. Given any sequence of nonzero scalars (wn )n , we define the operator
Bw : ω → ω, (x1 , x2 , x3 , . . . ) → (w2 x2 , w3 x3 , w4 x4 , . . . ). Prove that Bw is hypercyclic.

Exercise 2.2.4. Let T be the translation operator on the space Lpv (R+ ) defined in
Exercise 2.1.6. We assume that there are constants M ≥ 1 and w ∈ R such that

v(x) ≤ M ew(y−x) v(y) whenever y ≥ x ≥ 0;

in this case, v is called an admissible weight function. Show that T is hypercyclic if

and only if lim inf x→∞ v(x) = 0. (Hint: If the condition does not hold then define
g(x) = v(x)−1/p on [0, 1], and 0 otherwise, and show that, if f  is small enough, then
T n f − g ≥ 12 for n ≥ 1. For the sufficiency, use the density of the continuous functions
of compact support.)

Exercise 2.2.5. Let X = C ∞ (R) be the space of infinitely differentiable real functions
f : R → R; see Exercise 2.1.5. Show that the (real) differentiation operator D : X → X,
f → f  is hypercyclic by defining a suitable quasiconjugacy φ : H(C) → C ∞ (R).

Exercise 2.2.6. Let T : X → X be an operator. Suppose that Y ⊂ X is a T -invariant

dense subspace of X. Furthermore, suppose that Y carries a Fréchet space topology such
Exercises 61

that the embedding Y → X is continuous and such that T |Y : Y → Y is continuous.

Show that T is quasiconjugate to T |Y . In particular, if T |Y is hypercyclic, then so is
T , and T has a hypercyclic vector belonging to Y . (Remark: this result is known as the
hypercyclic comparison principle; it shows the interest of hypercyclicity on small spaces.)

Exercise 2.2.7. Show that the complexification X  of a real separable Fréchet space X
is a complex separable Fréchet space and that the complexification of a (real-linear)

operator on X is a (complex-linear) operator on X.

Exercise 2.2.8. Let T : X → X be an operator, and let M1 and M2 be T -invariant

closed subspaces of X such that X = M1 ⊕ M2 ; see Proposition 2.28. Let PM1 be the
projection X → M1 , x = x1 + x2 → x1 , where x1 ∈ M1 , x2 ∈ M2 , and similarly for M2 .
Show that, for j = 1, 2, T |Mj is quasiconjugate to T via PMj .

Exercise 2.2.9. Let T be an operator on a Banach space X and x ∈ X. Given d > 0,

let us a call the orbit of x under T d-dense if for each y ∈ X we can find n ∈ N0 such
that T n x − y < d. Show that if T admits a d-dense orbit then it is hypercyclic. (Hint:
First observe that X is separable. Then, given a vector x whose orbit is d-dense, prove
that, for each ε > 0, the vector dε x has an ε-dense orbit and conclude the result by
Exercise 1.2.5 and the Birkhoff transitivity theorem).

Exercise 2.2.10. Let ε > 0. An operator T on a Banach space X is called ε-hypercyclic

if it admits a vector x ∈ X such that, for any nonzero vector y ∈ X, we can find n ∈ N0
satisfying T n x − y ≤ εy; the vector x is then also called ε-hypercyclic. Show that if
an operator is ε-hypercyclic for all ε > 0 then it is hypercyclic. (Hint: First observe that
X is separable. Then conclude the result by Exercise 1.2.5 and the Birkhoff transitivity
theorem).

Exercise 2.2.11. An operator T on a Fréchet space X is called a J-class operator if

there is a vector x = 0 in X with J(x) = X, where J(x) is the J-set of x. By Exercise
1.2.8, every hypercyclic operator is J-class.
(a) Let B be the backward shift on 2 . Show that the operator T = 2I ⊕ 2B on
K ⊕ 2 is a J-class operator that is not hypercyclic. Deduce also that being J-class is
not preserved under quasiconjugacy.
(b) Let T be a hypercyclic operator on X. Show that, for λ ∈ K, the operator λI ⊕ T
is J-class on K ⊕ X if and only if |λ| > 1. Deduce that there is an invertible J-class
operator T whose inverse T −1 is not J-class.
(c) Show that the multiple T = 2B of the backward shift is J-class on ∞ . There-
fore there exist J-class operators on non-separable Banach spaces. (Hint: Consider
(1, 0, 0, . . .).)

Exercise 2.3.1. Let T be an operator on a Banach space X. Show that the following
assertions are equivalent:
(i) T has sensitive dependence on initial conditions with respect to the usual metric;
(ii) supn≥0 T n  = ∞;
(iii) T has an unbounded orbit.
(Hint: Use the Banach–Steinhaus theorem; see Appendix A.)

Exercise 2.3.2. Let T be the translation operator on Lpv (R+ ) with an

∞admissible weight
function v; see Exercise 2.2.4. Show that T is chaotic if and only if 0 v(x) dx < ∞.

Exercise 2.3.3. Let T : KN → KN , N ≥ 1, be an operator.

(a) Show that T has a dense set of periodic points if and only if T n = I for some
n ≥ 1. (Hint: Show that KN has a basis consisting of periodic points.)
(b) Deduce from (a) that no operator on KN can be chaotic.
62 2 Hypercyclic and chaotic operators

Exercise 2.3.4. Let H(Ω) be the Fréchet space of all holomorphic functions on a do-
main Ω in C; see Section 4.3. Then D : f → f  is an operator on H(Ω). Show that the
following assertions are equivalent:
(i) D is chaotic on H(Ω);
(ii) D is hypercyclic on H(Ω);
(iii) Ω is simply connected.
(Hint: By Runge’s theorem, the polynomials are dense in H(Ω) if Ω is simply connected.
1
On the other hand, if Ω is not simply connected, approximate a suitable function z−a
by derivatives, and integrate both over closed curves in Ω.)

Exercise 2.4.1. Modify Example 2.39 to obtain a non-chaotic mixing operator on any
space p , 1 < p < ∞.

Exercise 2.4.2. Let T be the translation operator on Lpv (R+ ) with an admissible weight
function v; see Exercise 2.2.4. Show that T is mixing if and only if limx→∞ v(x) = 0.

Exercise 2.4.3. Using Proposition 2.37, show that λD is mixing on H(C) for any λ = 0.
(This should be contrasted with Exercise 2.2.2.)

Exercise 2.4.4. Let T be an operator on a separable Banach space. It can be shown

that T ⊕ T ∗ is never hypercyclic on X ⊕ X ∗ ; see Remark 4.17. Deduce that if T is mixing
then T ∗ cannot be hypercyclic. (See Exercise 5.1.1 for a better result.)

Exercise 2.5.1. Let T be a weakly mixing (or mixing) operator and λ ∈ K, |λ| = 1.
Show that λT is also weakly mixing (or mixing, respectively). (Hint: Use Proposition
2.37 and Theorem 2.47.)

Exercise 2.5.2. Establish Theorem 2.45 without the use of Proposition 1.53, that is,
show the weak mixing property directly. (Hint: See Figure 2.2.)

Exercise 2.5.3. Let T be an operator such that, for any nonempty open sets U, V ⊂ X
and any 0-neighbourhood W , N (U, W ) is nonempty and N (W, V ) is syndetic. Then
prove that T is weakly mixing. Do likewise if the sets N (U, W ) are syndetic and the sets
N (W, V ) are nonempty. (Hint: Apply Theorem 2.47.)

Exercise 2.5.4. A subset A of N0 is called thickly syndetic if, for any n ∈ N, there is a
syndetic sequence (nk )k of positive integers such that {nk +j ; k ∈ N, j = 0, . . . , n} ⊂ A.
Show the following:
(i) the intersection of two thickly syndetic sets is thickly syndetic;
(ii) let T be an operator and U ⊂ X a nonempty open set; if, for any 0-neighbourhood
W , N (U, W ) is syndetic then these sets are all thickly syndetic; and similarly for
N (W, U ).
Deduce that, for any topologically ergodic operator T , N (U, V ) is thickly syndetic for
any nonempty open sets U, V ⊂ X; see Exercise 1.5.6 for the definition of topological
ergodicity.

Exercise 2.5.5. If S : X → X and T : Y → Y are topologically ergodic operators,

show that S ⊕ T is topologically ergodic on X ⊕ Y . In particular, topologically ergodic
operators are weakly mixing. (Hint: Use the previous exercise.)

Exercise 2.5.6. An operator T is called hereditarily ergodic if, for any pair U, V of
nonempty open subsets of X and any syndetic sequence (nk )k , there exists a subsequence
(nkj )j of (nk )k that is also syndetic such that T nkj (U ) ∩ V = ∅ for all j ≥ 1. Show
that an operator is hereditarily ergodic if and only if it is topologically ergodic. Deduce
that every power T p , p ≥ 1, of a topologically ergodic operator is topologically ergodic.
(Hint: Use Exercise 2.5.4.)
Exercises 63

Fig. 2.2 Exercise 2.5.2

Exercise 2.5.7. Prove that the direct sum S ⊕ T of two chaotic operators is chaotic.
(Hint: Use Exercises 1.5.6 and 2.5.5.)

Exercise 2.5.8. Show the following variant of Theorem 2.45: let T be an operator such
that, for any nonempty open sets U, V ⊂ X and any 0-neighbourhood W , there is a
continuous map S : X → X commuting with T such that S(U ) ∩ V = ∅, S(W ) ∩ W = ∅
and
N (U, W ) ∩ N (W, U ) = ∅.
Then T is weakly mixing. (Hint: Use the 4-set trick.)

Exercise 2.5.9. Let T be an operator such that λ1 T ⊕ λ2 T is hypercyclic (or mix-

ing), where |λ1 | ≤ |λ2 |. Show that then λT is weakly mixing (or mixing, respectively)
whenever |λ1 | ≤ |λ| ≤ |λ2 |. (Hint: Use Proposition 2.37 and Theorem 2.47.)

Exercise 2.5.10. Let P denote the set of nonzero polynomials. If T is a hypercyclic

operator such that 
ker p(T )
p∈P

is dense in X, then prove that T is weakly mixing. (Hint: Given U and W , pick u ∈ U
and p ∈ P such that p(T )u = 0. Since p(T ) has a dense range by Theorem 2.54, there is
some r > 0 with p(T )(rW ) ∩ U = ∅. Then define S = rp(T ) and apply Theorem 2.45.)

Exercise 2.5.11. An operator T on X is called upper triangular if it admits an increas-

ing sequence (En )n of invariant subspaces such that dim En = n for every n ≥ 1 and
X is the closed linear span of the finite-dimensional spaces En , n ≥ 1. If X is a Hilbert
space, T is upper triangular if and only if T has an upper triangular matrix with respect
to some orthonormal basis of X. Show that every hypercyclic upper triangular operator
64 2 Hypercyclic and chaotic operators

is weakly mixing. (Hint: Use the previous exercise and a well-known result from linear
algebra.)
Exercise 2.6.1. Show constructively, as in Example 2.18, that every vector x ∈ 2 is
the sum of two vectors that are hypercyclic for T = 2B.
Exercise 2.6.2. Let T be a hypercyclic operator on a complex space. Show that its
adjoint T ∗ has no finite-dimensional invariant subspace.
Exercise 2.6.3. Let S and T be commuting operators on X such that T has dense
range. Show that HC(S) is T -invariant.
Exercise 2.6.4. (a) Let us call an operator T (on a real or complex space) 2-hypercyclic
if there are vectors x, y ∈ X such that

{T n x + T m y ; n, m ≥ 0}

is dense in X. Show that the adjoint T ∗ of a 2-hypercyclic operator has no eigenvalues.

(Hint: Proceed as in the proof of Lemma 2.53(a).)
(b) Now let T be a hypercyclic operator on a real space. Show that its complexification
T is 2-hypercyclic, and deduce Lemma 2.53(b).
(c) Show that there are 2-hypercyclic operators that are not hypercyclic. (Hint: Con-
sider T = T1 ⊕ T2 , where T1 and T2 are hypercyclic but T is not; see Remark 4.17.)
Exercise 2.6.5. Here is an alternative proof of Bourdon’s theorem in the real case.
Supply the details.
 on R has a matrix represen-
2
(a) By the Jordan decomposition theorem,  cos
any operator
ϕ − sin ϕ
tation in one of the forms ( 0 b ), ( 0 a ) or c sin ϕ cos ϕ with a, b, c, ϕ ∈ R. Therefore
a 0 a 1

no operator on R2 can be hypercyclic.

(b) Let T be hypercyclic on X. In view of Lemma 2.53, it remains to prove that every
operator p(T ) = T 2 + bT + cI has dense range. If this is not the case there is a nonzero
x∗ ∈ X ∗ with ((T ∗ )2 + bT ∗ + cI)x∗ = 0. By Lemma 2.53, x∗ and T ∗ x∗ are linearly
independent. Then φ : X → R2 , φ(x) = (x, x∗ , x, T ∗ x∗ ) defines a quasiconjugacy
from T to an operator on R2 , which is impossible.
Exercise 2.6.6. Let T be a hypercyclic operator on X. Show that T ⊕ I is supercyclic
on X ⊕K. Show that there is no subspace of X ⊕K of dimension 2 in which every nonzero
vector is supercyclic; in particular, the Herrero–Bourdon theorem fails for supercyclicity.
Exercise 2.7.1. Let T be an operator on KN , N ≥ 1, and let y ∈ KN be a hypercyclic
vector for T .
(a) Show that the vectors y, T y, T 2 y, . . . , T N −1 y form a basis for KN .
(b) Choose (nk )k and (mk )k such that T nk y → 0 and T mk y → y. Show that
T x → 0 and T mk x → x for all x ∈ KN . Deduce that (det T )nk = det (T nk ) → 0
nk

and (det T )mk → 1.

(c) Based on (b), give a new proof of Theorem 2.58.
Exercise 2.7.2. Use the Cayley–Hamilton theorem to give a new proof of Theorem 2.58.
Exercise 2.7.3. Use Bourdon’s theorem to give a new proof of Proposition 2.60. (Hint:
N
In the notation of the stated proof, consider T N +1 − k=0 αk T k .)
Exercise 2.7.4. Let T be a hypercyclic operator. Show that its adjoint T ∗ has no
finite-dimensional invariant subspace. (This complements Exercise 2.6.2.) (Hint: Let
M = span{x∗1 , . . . , x∗N } be T ∗ -invariant; construct K = span{x1 , . . . , xN } such that
N ∗
x∗k (xj ) = δj,k ; show that φ(x) = x (x)xk provides a quasiconjugacy between T
k=1 k
and φ ◦ T |K on K.)
Sources and comments 65

Exercise 2.7.5. Show that there are supercyclic operators on R and on R2 . (Hint: For
R2 , use a rotation.)

Exercise 2.8.1. Let X be a Fréchet space. A map Q : X → X is called an m-

homogeneous polynomial, m ≥ 0, if there is a continuous multilinear map A : X m → X
such that Q(x) = A(x, . . . , x), x ∈ X (where, for m = 0, Q is understood to be a constant
N
map). A map P : X → X is called a polynomial if it can be written as P = m=0 m
Q
with m-homogeneous polynomials Qm . Show that the map P in Example 2.62 is a
polynomial on c0 .

Exercise 2.8.2. Prove Proposition 2.63.

Sources and comments

Section 2.1. For introductory texts on functional analysis that also cover Fréchet spaces
we refer to Rudin [271] and Meise and Vogt [237]. The notion of an F-norm can be found
in Kalton, Peck and Roberts [212].

Section 2.2. The term “hypercyclic” for vectors with a dense orbit was apparently first
used around 1986 (Beauzamy [46], [47], [48]) and then extended around 1988 to operators
with a dense orbit (Bourdon, Godefroy, Shapiro [94], [165]). Supercyclic vectors were
introduced by Hilden and Wallen [202] in 1973.
Beauzamy’s work was motivated by the invariant subspace problem. The negative
solution for non-Hilbert spaces is due to Enflo [142] and Read [265]. Subsequently, Read
[266] even constructed a counterexample to the invariant subset problem.

Theorem 2.64 (Read). There exists an operator on 1 all of whose nonzero vectors
are hypercyclic.

Both problems remain open for Hilbert spaces.

The three classical hypercyclic operators were found by Birkhoff [75] in 1929,
MacLane [225] in 1952 and Rolewicz [268] in 1969. Example 2.18 reproduces Rolewicz’s
original proof; Birkhoff and MacLane used very similar constructions.
The first systematic studies of hypercyclicity are due to Kitai [215] in 1982 and
Godefroy and Shapiro [165] in 1991. Though never published, Kitai’s thesis was widely
circulated. Between them, Kitai, Godefroy and Shapiro laid the foundations for what
was to become the theory of linear dynamical systems: they clarified the basic concepts,
provided a wealth of examples and introduced important techniques like criteria for
hypercyclicity that will be discussed in the next chapter. It is difficult to overemphasize
the importance of their work for the further development of linear dynamics.
Quasiconjugacies were introduced in hypercyclicity by Herrero [195], Martínez and
Peris [229]; the hypercyclic comparison principle (see Exercise 2.2.6), was formulated by
Shapiro [279]. Complexifications were first studied in hypercyclicity by Bès and Peris [71].

Section 2.3. Godefroy and Shapiro [165] suggested acceptance of Devaney’s definition of
chaos for linear operators, and they showed that the three classical operators are chaotic.
Chaos in the sense of Auslander and Yorke was introduced for continuous maps on metric
spaces in [18]. Proposition 2.30 is due to Godefroy and Shapiro [165]. Proposition 2.33
seems to be folklore, but see Herrero [195] and Bonet, Martínez and Peris [83].
We have taken the proof of Lemma 2.34 from Aron and Markose [15].

Sections 2.4. Proposition 2.37 appears in Grosse-Erdmann and Peris [187].

66 2 Hypercyclic and chaotic operators

It does not seem to be easy to find a chaotic operator that is not mixing. The first
such operator was constructed by Badea and Grivaux [19].

Sections 2.5. In 1992, Herrero [195] posed the problem of whether every hypercyclic
operator (on a Hilbert space) is weakly mixing. De la Rosa and Read [126] constructed
a counterexample in a suitable Banach space, while Bayart and Matheron [43] showed
that counterexamples exist on many classical Banach spaces like p , 1 ≤ p < ∞, c0 ,
C[0, 1] and L1 [0, 1], as well as on the Fréchet space H(C).
The results of the section appear in Grosse-Erdmann and Peris [187]; see also Bayart
and Matheron [44], [45] and Moothathu [245]. Theorem 2.47 is due to Bernal and Grosse-
Erdmann [62] and León [219]; the characterizing condition first appeared in Godefroy
and Shapiro [165]. Unlike the proofs in [62] and [219], our argument avoids the Baire
category theorem, as does another proof by Yousefi and Rezaei [303].
Theorem 2.48 is due to Grivaux [172]. The fact that every chaotic operator is weakly
mixing, Corollary 2.49, is due Bès and Peris [71]; it also follows directly from a result
that is due to Bauer and Sigmund [32] and to Stacey (see [28]) using Ansari’s theorem
(see Section 6.1).

Section 2.6. Proposition 2.52 was observed, for example, by Grosse-Erdmann [177],
Godefroy (see [94]) and Kahane (see [249]). Lemma 2.53(a) is due to Kitai [215], while
part (b) can be found in Bonet and Peris [85]. Herzog and Lemmert [199] have char-
acterized the hypercyclic operators on ω = CN ; Conejero [108, p. 123] noted that their
condition can be rephrased as σp (T ∗ ) = ∅.

Theorem 2.65 (Herzog–Lemmert). An operator T on ω = CN is hypercyclic if and

only if T ∗ has no eigenvalues.

In the complex case, Theorem 2.54 is due to Bourdon [90]. The real case was added
by Bès [66]; our proof is due to Martínez [227]. Theorem 2.55 is due to Herrero [194]
and Bourdon [90].
A considerable improvement of Corollary 2.56 is due to Fathi [146] and Godefroy (see
[44, p. 16]). They showed that for any hypercyclic operator T on a Fréchet space X, the
set HC(T ) of hypercyclic vectors for T is homeomorphic to X.

Section 2.7. Kitai [215] and Rolewicz [268] observed that there are no hypercyclic op-
erators on finite-dimensional spaces. This prompted Rolewicz to pose the problem of
whether every infinite-dimensional separable Banach space admits a hypercyclic opera-
tor, to which we will turn in Chapter 8. Theorem 2.61 is due to Feldman [147].
The title of this section was inspired by Protopopescu [262], where one finds discus-
sions on the relationship between linear and nonlinear chaos; see also Protopopescu and
Azmy [263].

Section 2.8. In this section we follow Peris [255]. For an introduction to the dynamics
of complex polynomials we refer to Devaney [132] and Carleson and Gamelin [98].

Exercises. Exercise 2.2.1 can be found in Aron, Seoane and Weber [16]. The result of
Exercise 2.2.4 is essentially due to Desch, Schappacher and Webb [131]. Exercise 2.2.9
is taken from Feldman [148], Exercise 2.2.10 from Badea, Grivaux and Müller [21],
and Exercise 2.2.11 from Costakis and Manoussos [119]. Exercise 2.3.1 can be found
in Feldman [147], Exercise 2.3.2 in deLaubenfels and Emamirad [128], Exercise 2.3.4 in
Shapiro [280], and Exercise 2.4.2 in Bermúdez, Bonilla, Conejero, and Peris [49]. Ex-
ercises 2.5.1, 2.5.3 and 2.5.8 are extracted from Grosse-Erdmann and Peris [187]. The
fact that topologically ergodic operators are weakly mixing (part of Exercise 2.5.5) is
implicit in Grosse-Erdmann and Peris [185]. The main part of Exercise 2.5.5 is due to
Sources and comments 67

Desch and Schappacher [129], who introduce the concept of operators satisfying the
Recurrent Hypercyclicity Criterion, which is equivalent to topological ergodicity. Hered-
itarily ergodic operators (see Exercise 2.5.6), were introduced as hereditarily syndetic
operators by Badea and Grivaux [19]. Exercise 2.5.9 is taken from Badea, Grivaux and
Müller [20], Exercises 2.5.10 and 2.5.11 from Grivaux [172], Exercises 2.6.5 and 2.7.1
from Bès [66], and Exercise 2.6.6 from Bourdon [90].

Extensions. Let us add again some remarks on the setting chosen for this chapter (and
for most of the book). Typically, the basic results in linear dynamics either use a Baire
category argument, in which case they are often valid in all F-spaces (see below), or
they hold in all topological vector spaces (see Chapter 12). Only in more specialized
results do structural properties such as being locally convex, being normable or having
an inner product play a role. Since our aim has not been to always provide the best
possible result but to offer a widely accessible introduction to the main ideas of linear
dynamics we have chosen to restrict ourselves to the setting of Fréchet spaces.
The larger class of F-spaces consists of all vector spaces that are endowed with an
F-norm and that are complete under the induced metric. For example, the spaces p
with 0 < p < 1 are F-spaces. One can show that a vector space is an F-space if and only
if it carries a complete translation-invariant metric; see [212].

We finish with a citation that is representative of the common, and as we now know
erroneous, belief that chaos and nonlinearity go hand in hand.

Chaotic systems not only exhibit sensitive dependence, but two other properties
as well: they are deterministic, and they are nonlinear.

(L.A. Smith, Chaos: A very short introduction, [295, p. 1]; emphasis in the original.)
Chapter 3
The Hypercyclicity Criterion

The Birkhoff transitivity theorem reduces hypercyclicity to the (formally)

simpler condition of topological transitivity. Nonetheless, in many concrete
situations it is not obvious how to verify this condition for a given operator.
The main purpose of this chapter is to derive several easily applicable cri-
teria under which an operator is chaotic, mixing or weakly mixing (and, in
particular, hypercyclic). These criteria have their origin in specific applica-
tions to operators on spaces of sequences or functions. Some of them may
appear technical, while others are easier to understand, but all of them are
very useful for particular examples. We will proceed from the easiest to the
most sophisticated criterion.
Throughout this chapter, X will denote a separable Fréchet space and T
an operator on X.

3.1 Criteria for chaos and mixing

The message of our first criterion is quite simple: a large supply of (appropri-
ate) eigenvectors is conducive to chaos. This extends the earlier observation
that linked periodic points with certain eigenvectors of modulus 1; see Propo-
sition 2.33.
The criterion, like others in this chapter, is named after its authors.

Theorem 3.1 (Godefroy–Shapiro criterion). Let T be an operator. Sup-

pose that the subspaces

X0 := span{x ∈ X ; T x = λx for some λ ∈ K with |λ| < 1},

Y0 := span{x ∈ X ; T x = λx for some λ ∈ K with |λ| > 1}

are dense in X. Then T is mixing, and in particular hypercyclic.
If, moreover, X is a complex space and also the subspace

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 69

DOI 10.1007/978-1-4471-2170-1_3, © Springer-Verlag London Limited 2011
70 3 The Hypercyclicity Criterion

Z0 := span{x ∈ X ; T x = eαπi x for some α ∈ Q}

is dense in X, then T is chaotic.

Proof. Let U, V be a pair of nonempty open subsets of X. By hypothesis we
can find x ∈ X0 ∩ U and y ∈ Y0 ∩ V . Hence these vectors can be expressed
in the form
m m
x= ak xk and y = bk yk ,
k=1 k=1

where T xk = λk xk , T yk = μk yk , for certain scalars ak , bk , λk , μk ∈ K with

|λk | < 1, |μk | > 1, k = 1, . . . , m. Since


m 
m
1
T nx = ak λnk xk → 0 and un := bk yk → 0
μnk
k=1 k=1

as n → ∞ and T un = y for all n ≥ 0, there is some N ∈ N so that, for all

n

n ≥ N,
x + un ∈ U and T n (x + un ) = T n x + y ∈ V.
This shows that T is mixing and therefore hypercyclic.
According to Proposition 2.33, in the complex case, Z0 is precisely the set
of periodic points of T . Thus, if also Z0 is dense then T is even chaotic. 

The Godefroy–Shapiro criterion provides a new proof for the chaotic be-
haviour of the three classical operators.
Example 3.2. (Rolewicz’s operators) Let T = μB, |μ| > 1, be the multiple
of the backward shift on any of the spaces X = p , 1 ≤ p < ∞, or X = c0 .
By Corollary 2.51 (see also Example 2.27), it suffices to consider the complex
case.
One easily determines the eigenvectors of B as being the nonzero multiples
of the sequences
eλ := (λ, λ2 , λ3 , . . .), |λ| < 1,
with corresponding eigenvalue λ; the condition |λ| < 1 ensures that eλ ∈ X.
Therefore, eλ is an eigenvector of T = μB to the eigenvalue μλ.
We claim that, for any subset Λ of the unit disk D = {λ ∈ C ; |λ| < 1}
that has an accumulation point inside the disk, the set

span{eλ ; λ ∈ Λ}

is dense in X. By the Hahn–Banach theorem it suffices to show that any

continuous linear functional x∗ on X that vanishes on each eλ , λ ∈ Λ, vanishes
on X. Now, since x∗ ∈ X ∗ is given, via the canonical representation, by a
sequence y = (yn )n ∈ q for a certain q with 1 ≤ q ≤ ∞, we have that
∞

x∗ (eλ ) =
eλ , x∗  = yn λn if |λ| < 1.
n=1
3.1 Criteria for chaos and mixing 71

But since (yn )n is necessarily bounded, this defines a holomorphic function in

λ, and it vanishes, by assumption, on the set Λ with an accumulation point.
The identity theorem for holomorphic functions implies that each yn is zero
and therefore x∗ = 0.
In particular, the subspace

X0 = span{x ∈ X ; T x = ηx for some η ∈ K with |η| < 1}

= span{eλ ; |λ| < 1/|μ|}

is dense in X, as are the subspaces Y0 and Z0 of the Godefroy–Shapiro crite-

rion; note that 1/|μ| < 1. This implies that Rolewicz’s operators are mixing
and chaotic.

Example 3.3. (Birkhoff’s and MacLane’s operators) Let us first consider

the translation operators Ta , a = 0, on the space H(C) of entire functions.
Here all the work has already been done in Section 2.3. Denoting by eλ the
exponential functions eλ (z) = eλz , every eλ , λ ∈ C, is an eigenvector of Ta
to the eigenvalue eaλ . Therefore the subspace X0 of the Godefroy–Shapiro
criterion contains the subspace
 
span eλ ; |eaλ | < 1 ,

which is dense in H(C) by Lemma 2.34. The density of Y0 and Z0 follows

similarly (the latter has, in fact, already been shown in Example 2.35). Hence
Birkhoff’s operators are mixing and chaotic.
In the same way one can give a new proof of the differentiation operator
D being mixing and chaotic.

The essential point in the mixing part of the proof of Theorem 3.1 was
the fact that T n x → 0, for each x ∈ X0 , and that for each y ∈ Y0 we could
find a sequence (un )n in X such that un → 0 and T n un = y for n ≥ 0. The
second condition can be achieved, for example, via a map S : Y0 → Y0 such
that S n y → 0 and T Sy = y for any y ∈ Y0 ; one need only define un = S n y.
This leads us to another important criterion in linear dynamics.

Theorem 3.4 (Kitai’s criterion). Let T be an operator. If there are dense

subsets X0 , Y0 ⊂ X and a map S : Y0 → Y0 such that, for any x ∈ X0 ,
y ∈ Y0 ,
(i) T n x → 0,
(ii) S n y → 0,
(iii) T Sy = y,
then T is mixing.

Remark 3.5. We emphasize that the map S need not have any properties
other than satisfying (ii) and (iii). In particular, it need not be a linear or
continuous map. The same remark will apply to the forthcoming criteria.
72 3 The Hypercyclicity Criterion

Our proofs of the hypercyclicity of MacLane’s and Rolewicz’s operators

in Examples 2.21 and 2.22 were based on Kitai’s criterion in disguise. Using
that criterion leads to very transparent proofs, which, in addition, differ from
the proofs based on the Godefroy–Shapiro criterion.

Example 3.6. (Rolewicz’s operators) The operator T = λB, |λ| > 1, is

mixing on any of the spaces X = p , 1 ≤ p < ∞, or c0 . In fact, taking for
X0 = Y0 the set of finite sequences, which is dense in X, and for S : Y0 →
Y0 the map S = λ1 F , where F is the forward shift F : (x1 , x2 , x3 , . . .) →
(0, x1 , x2 , . . .), the conditions of Kitai’s criterion are clearly satisfied.

Example 3.7. (MacLane’s operator) The differentiation operator D on

H(C) is mixing. In this case we take for X0 = Y0 the set of polynomi-
als, which is dense in H(C), and for S we consider the integral operator
z
Sf (z) = 0 f (ζ) dζ. While conditions (i) and (iii) are obvious, we note
for condition (ii) that it suffices to verify it for the monomials. But then
k!
S n (z k ) = (k+n)! z k+n → 0 as n → ∞, uniformly on compact sets, as re-
quired.

In contrast, the proof of hypercyclicity for Birkhoff’s operators given in

Example 2.20 does not hide a Kitai-type argument. Indeed, Kitai’s criterion
seems to be less well adapted to this operator. Still, the following provides
us with a third argument for the hypercyclicity of the translation operators.

Example 3.8. (Birkhoff’s operators) For simplicity we only show that the
translation operator T1 f (z) = f (z + 1) on H(C) is mixing. For X0 = Y0
we choose the set of all functions of the form fp,α,ν (z) = p(z)e−α(z−ν) ,
2

where p is a polynomial and α > 0, ν ∈ N0 . Since fp,α,ν → p in H(C) as

α → 0, this set is dense in H(C). Moreover, for S we consider the translation
operator Sf (z) = f (z − 1). Now, if z = x + iy with |y| ≤ 12 |x| then we have
that |e−αz | = e−α(x −y ) ≤ e− 4 αx . This implies that, for any p, α and ν,
2 2 2 3 2

fp,α,ν (z ± n) → 0 uniformly on compact sets as n → ∞, which shows that

conditions (i) and (ii) of Kitai’s criterion hold, while condition (iii) is trivial.

We end this section with an example showing that Kitai’s criterion is a

stronger result than the Godefroy–Shapiro criterion.

Example 3.9. Consider the bilateral backward shift T = B, given by

B(xn )n∈Z = (xn+1 )n∈Z ,


on the weighted space 1 (Z, v) = {(xn )n∈Z ; x := n∈Z |xn |vn < ∞},
1
where vn = |n|+1 , n ∈ Z. The equality T x = λx, x = 0, implies that
 
x = . . . , λ12 x0 , λ1 x0 , x0 , λx0 , λ2 x0 , . . . , λ = 0, x0 = 0.

But then we have that

3.2 Weak mixing and the Hypercyclicity Criterion 73

 |λ|n  1
x = |x0 | 1 + + = ∞,
n+1 |λ|n (n + 1)
n∈N n∈N

whatever the value of λ. Therefore T has no eigenvalues and, in particular,

it does not satisfy the Godefroy–Shapiro criterion.
On the other hand it satisfies Kitai’s criterion. Indeed, if we choose for
X0 = Y0 the space of bilateral finite sequences (. . . , 0, 0, x−m , . . . , xn , 0, 0, . . .),
m, n ≥ 0, and for S : Y0 → Y0 the forward shift, then one easily verifies the
conditions of Kitai’s criterion for T .

This example shows, in particular, that hypercyclic operators need not

have eigenvectors.

3.2 Weak mixing and the Hypercyclicity Criterion

The next step in the sophistication of criteria for hypercyclicity is to replace

the full sequence (n) by an increasing sequence (nk )k of positive integers for
the iterates of T and S in Kitai’s criterion. But in doing so we lose the mixing
property.

Theorem 3.10 (Gethner–Shapiro criterion). Let T be an operator. If

there are dense subsets X0 , Y0 ⊂ X, an increasing sequence (nk )k of positive
integers, and a map S : Y0 → Y0 such that, for any x ∈ X0 , y ∈ Y0 ,
(i) T nk x → 0,
(ii) S nk y → 0,
(iii) T Sy = y,
then T is weakly mixing.

Proof. Let U1 , U2 , V1 and V2 be nonempty open sets. By assumption we can

find vectors xj ∈ Uj ∩ X0 and yj ∈ Vj ∩ Y0 , j = 1, 2. Then, by (iii),

T nk (xj + S nk yj ) = T nk xj + yj , j = 1, 2.

It follows from (i) and (ii) that, for sufficiently large k, xj + S nk yj ∈ Uj and
T nk xj + yj ∈ Vj for j = 1, 2. This shows that N (U1 , V1 ) ∩ N (U2 , V2 ) = ∅. 


The requirements of this criterion are clearly weaker than those of Kitai’s
criterion. In fact we will provide an operator satisfying the Gethner–Shapiro
criterion but not Kitai’s criterion. This is also the first example that we
present of a weakly mixing operator that is not mixing; see also Remark
4.10.

Example 3.11. We consider the weighted backward shift T = Bw : c0 → c0

given by
Bw (x1 , x2 , . . . ) = (w2 x2 , w3 x3 , w4 x4 , . . . ),
74 3 The Hypercyclicity Criterion

with weight sequence

 
w = (w1 , w2 , . . . ) = 1, 2, 2−1 , 2, 2, 2−1 , 2−1 , 2, 2, 2, 2−1 , 2−1 , 2−1 , . . .

(see also Section 4.1); note that the value of w1 is irrelevant. Since w is
bounded, T is continuous. Let (mk )k be the increasing sequence of all the
integers with wmk = 2−1 and wmk +1 = 2, k ≥ 1. Since

 n+1

T nx = wν xn+1 , . . . , n ≥ 1,
ν=2

we have that T mk −1 x = (xmk , . . . ) for each k ≥ 1. In particular, if we define

U = {x ∈ c0 ; x < 1} and V = {x ∈ c0 ; |x1 | > 1}, which are nonempty
open sets, then we get that T mk −1 (U ) ∩ V = ∅, for each k ≥ 1, which shows
that T is not mixing. Therefore it cannot satisfy Kitai’s criterion.
On the other hand, let us take for X0 = Y0 the space of finite se-
quences and for S the weighted forward shift S : Y0 → Y0 , S(x1 , x2 , . . . ) =
(0, w2−1 x1 , w3−1 x2 , . . . ). It is clear that T Sy = y, y ∈ Y0 , and that T n x → 0,
x ∈ X0 . It remains to find a suitable increasing sequence (nk )k of positive
integers so that S nk y → 0 for each y ∈ Y0 in order to satisfy all the conditions
of the Gethner–Shapiro criterion. Indeed, let nk = mk + k − 1, k ∈ N, and
denote by en the vector in c0 that has 1 in the nth position, and 0 otherwise.
Then we have that
 m
k +k  
S nk
e1 = 0, . . . , 0, wν−1 , 0, . . . = 0, . . . , 0, 2−k , 0, . . . ,
ν=2

so that S e1 → 0. On the other hand, since the weights wn are bounded

nk

away from 0, S is continuous too. Therefore, for any j ≥ 1,

 
j 
j
S nk ej = S nk wν S j−1 e1 = wν S j−1 (S nk e1 ) → 0.
ν=2 ν=2

From this we conclude that S nk y → 0 for each y ∈ Y0 , which had to be

shown.
The final step in our search for a very general, but reasonable, criterion for
hypercyclicity will be to weaken the requirement of the existence of a right
inverse S for T on the dense subset Y0 . The proof of Theorem 3.10 shows
that all we need is a sequence of maps Snk with Snk y → 0 and T nk Snk y → y
for all y ∈ Y0 . These maps Snk need not even be self-maps of Y0 . We thus
have obtained the following.
Theorem 3.12 (Hypercyclicity Criterion). Let T be an operator. If there
are dense subsets X0 , Y0 ⊂ X, an increasing sequence (nk )k of positive inte-
gers, and maps Snk : Y0 → X, k ≥ 1, such that, for any x ∈ X0 , y ∈ Y0 ,
3.2 Weak mixing and the Hypercyclicity Criterion 75

(i) T nk x → 0,
(ii) Snk y → 0,
(iii) T nk Snk y → y,
then T is weakly mixing, and in particular hypercyclic.

Remark 3.13. (a) If the Gethner–Shapiro criterion or the Hypercyclicity Cri-

terion is satisfied for the full sequence (nk )k = (n)n , then the proofs show
that the operator T is even mixing; see also Exercise 3.1.1.
(b) By Furstenberg’s theorem, the Hypercyclicity Criterion even implies
that every n-fold direct sum T ⊕ . . . ⊕ T is hypercyclic. But this follows also
directly because T ⊕ . . . ⊕ T satisfies the Hypercyclicity Criterion as well.

Originally, Kitai had shown that the conditions in Theorem 3.4 imply
hypercyclicity. She did this by constructing, through a recursive procedure, a
hypercyclic vector. While the proofs we have presented here are shorter and
give better results, Kitai’s construction is useful in many other situations
where more abstract methods fail. We therefore give here a second proof of
the Hypercyclicity Criterion using Kitai’s approach.

Alternative proof of Theorem 3.12. Let  ·  denote an F-norm defining the

topology of the Fréchet space X. Since X is separable we can choose a se-
quence (yj )j from Y0 that is dense in X. We first show that there are xj in
X and positive integers kj such that

x = x1 + Snk1 y1 + x2 + Snk2 y2 + x3 + . . .

exists and is hypercyclic. To this end we construct the xj and kj recursively

such that we have, for j ≥ 1,
1 1
xj  < and T nkl xj  < (l = 1, . . . , j − 1), (3.1)
2j 2j
1 1
Snkj yj  < j and T nkl Snkj yj  < (l = 1, . . . , j − 1),
2 2j
(3.2)
 
j−1 
1   1
T nkj Snkj yj − yj  < and T nkj (xl + Snkl yl ) + xj  < j .
2j 2
l=1
(3.3)

Indeed, for j = 1 these inequalities are satisfied for x1 = 0 and for a suitable
k1 , using conditions (ii) and (iii). Now let j ≥ 2 and assume that x1 , . . . , xj−1
and k1 , . . . , kj−1 have been constructed. Then, by density of X0 , there is some
xj ∈ X such that (3.1) holds and


j−1
(xl + Snkl yl ) + xj ∈ X0 .
l=1
76 3 The Hypercyclicity Criterion

Choosing kj sufficiently large, conditions (i), (ii) and (iii) imply that one can
also achieve (3.2) and (3.3).
It then follows from the first inequalities in (3.1) and (3.2) that the series
defining x converges. The remaining inequalities imply that, for j ≥ 1,

  j−1
nkj  nkj 
T x − yj  = T (xl + Snkl yl ) + xj + T nkj Snkj yj − yj
l=1
∞
 ∞
 

+ T nkj xl + T nkj Snkl yl 
l=j+1 l=j+1
∞ 
1 1  1 1 4
≤ j + j + + = j.
2 2 2l 2l 2
l=j+1

Thus, since the yj are dense in X, x is hypercyclic for T .

To see that T is even weakly mixing we need only observe that also T ⊕ T
satisfies the conditions (i)–(iii), with the dense subsets X0 ⊕ X0 , Y0 ⊕ Y0
of X ⊕ X and the maps Snk ⊕ Snk . Thus, as we have just seen, T ⊕ T is
hypercyclic and T is therefore weakly mixing.  

What is a remarkable, and rather unexpected, feature of the Hypercyclicity

Criterion is that it actually characterizes when an operator is weakly mixing.
The proof of this result will be achieved via the intermediate notion of a
hereditarily hypercyclic operator, which, within our present framework, is
equivalent to the notion of hereditary transitivity; see Exercises 1.6.2 and
1.6.4.

Definition 3.14. Let (nk )k be an increasing sequence of positive integers.

Then an operator T is called hereditarily hypercyclic with respect to (nk )k
if, for each subsequence (mk )k of (nk )k , there is some x ∈ X such that
{T mk x ; k ≥ 1} is dense in X.
An operator T is called hereditarily hypercyclic if it is so with respect to
some sequence (nk )k .

For dynamical properties of sequences (T nk )k we use the terminology and

the results of Section 1.6; see also Section 3.4 below.

Theorem 3.15 (Bès–Peris). Let T be an operator. Then the following as-

sertions are equivalent:
(i) T satisfies the Hypercyclicity Criterion;
(ii) T is weakly mixing;
(iii) T is hereditarily hypercyclic.

Proof. (i)=⇒(ii). This is the result of Theorem 3.12.

(ii)=⇒(iii). By separability, X has a countable base (On )n of open sets.
Let (Uj , Vj )j be an enumeration of all pairs (Om , On ), m, n ≥ 1. Since T is
3.2 Weak mixing and the Hypercyclicity Criterion 77

weakly mixing, Furstenberg’s theorem implies that, for any k ≥ 1, there is a

positive integer nk such that

T nk (Uj ) ∩ Vj = ∅ for j = 1, . . . , k,

and we can choose (nk )k to be increasing. Now let (mk )k be a subsequence

of (nk )k . Let U, V be nonempty open sets. Then, by construction, there is
some l ≥ 1 such that Ul ⊂ U and Vl ⊂ V and hence

T mk (U ) ∩ V ⊃ T mk (Ul ) ∩ Vl = ∅

whenever k is sufficiently large. It follows from the Universality Criterion,

Theorem 1.57, that (T mk )k admits a dense orbit.
(iii)=⇒(i). Let T be hereditarily hypercyclic with respect to a certain
increasing sequence (mk )k of positive integers. In particular, we can find some
x ∈ X such that {T mk x ; k ≥ 1} is dense in X. Let (qk )k be a subsequence
of (mk )k such that T qk x → 0. Then we can find some y ∈ X such that
{T qk y ; k ≥ 1} is dense in X. In particular, if Uk is the open ball of radius
1/k around x then we can find a subsequence (nk )k of (qk )k such that

T nk y ∈ kUk , k ≥ 1.

Setting xk = y/k, k ≥ 1, we have that xk → 0 and T nk xk → x. Let X0 =

Y0 = orb(x, T ), which is dense in X. Since (nk )k is a subsequence of (qk )k ,
we have by continuity of T that

T nk (T n x) = T n (T nk x) → T n 0 = 0, n ≥ 0,

which gives condition (i) of the Hypercyclicity Criterion. On the other hand,
let y ∈ Y0 . Then there is some n ≥ 0 such that y = T n x, and we define
Snk y = T n xk , k ≥ 1. Then we have that

Snk y = T n xk → 0, T nk Snk y = T n (T nk xk ) → T n x = y,

yielding also conditions (ii) and (iii) of the Hypercyclicity Criterion. 



The completeness of the space X is essential in the previous theorem, as

the following example shows.

Example 3.16. Let T be a chaotic operator on a Fréchet space X (as, for

example, Birkhoff’s operator on H(C)). We consider the T -invariant subspace
Y = Per(T ) of periodic points of T . Since, by Corollary 2.49, T is weakly
mixing and Y is dense in X, we also have that T |Y : Y → Y is weakly mixing.
On the other hand, T |Y cannot satisfy the Hypercyclicity Criterion since the
only point y ∈ Y such that T nk y → 0 for some (nk )k is y = 0.

A careful look at the proof of Theorem 3.15 yields a more precise connec-
tion between hereditary hypercyclicity and the Hypercyclicity Criterion.
78 3 The Hypercyclicity Criterion

Proposition 3.17. An operator T is hereditarily hypercyclic with respect to

an increasing sequence (nk )k of positive integers if and only if every subse-
quence (mk )k of (nk )k admits a subsequence (qk )k such that T satisfies the
Hypercyclicity Criterion with respect to (qk )k .

Proof. If T is hereditarily hypercyclic with respect to (nk )k , then, given a

subsequence (mk )k of (nk )k , the argument of Theorem 3.15 yields the exis-
tence of a subsequence (qk )k of (mk )k such that T satisfies the Hypercyclicity
Criterion with respect to (qk )k .
Conversely, let (nk )k be an increasing sequence of positive integers such
that, for every subsequence (mk )k of (nk )k , T satisfies the Hypercyclicity
Criterion with respect to some subsequence (qk )k of (mk )k . But then it fol-
lows as in the proof of the Hypercyclicity Criterion (see also Theorem 3.24),
that (T qk )k admits a dense orbit, and hence so does (T mk )k . Therefore T is
hereditarily hypercyclic with respect to (nk )k . 

In view of the equivalence of the weak mixing property with the Hyper-
cyclicity Criterion, Theorem 2.48 and Corollary 2.49 provide us with a rich
source of operators that satisfy the Hypercyclicity Criterion.

Theorem 3.18. Every hypercyclic operator with a dense set of points with
bounded orbits satisfies the Hypercyclicity Criterion.
In particular, any of the following operators satisfy the Hypercyclicity Cri-
terion:
(i) chaotic operators;
(ii) hypercyclic operators that have a dense set of points for which the orbits
converge;
(iii) hypercyclic operators with dense generalized kernel.

Since the complexification T of an operator T on a real space can be

identified with T ⊕ T , Corollary 2.51 and Theorem 3.15 imply the following.

Proposition 3.19. Let T be an operator on a real separable Fréchet space

and T its complexification. Then the following assertions are equivalent:
(i) T satisfies the Hypercyclicity Criterion;
(ii) T is hypercyclic;
(iii) T satisfies the Hypercyclicity Criterion.

In particular, any hypercyclic operator on a complex space that can be

regarded as a complexification of a suitable (real) operator satisfies the Hy-
percyclicity Criterion.
In spite of all these positive results, the theorem of De la Rosa and Read,
Theorem 2.43, tells us that not every hypercyclic operator satisfies the Hy-
percyclicity Criterion. According to Bayart and Matheron, such operators
exist even on any of the spaces p , 1 ≤ p < ∞, and c0 .
3.3 Equivalent formulations of the Hypercyclicity Criterion 79

3.3 Equivalent formulations of the Hypercyclicity

Criterion

In the last two sections we have discussed some popular criteria for hyper-
cyclicity. But many other criteria have been suggested, some formally stronger
than the Hypercyclicity Criterion, some formally weaker. So the problem
arises of determining whether they are indeed stronger or weaker than the
Hypercyclicity Criterion. We address this problem here for two particular
criteria, one of which we have already encountered. Additional results along
this line will be discussed in the exercises.
Our first result says that, in the Hypercyclicity Criterion, one may, without
loss of generality, assume that the sets X0 and Y0 coincide, are linear, and
that each Snk is linear.

Proposition 3.20. An operator T satisfies the Hypercyclicity Criterion if

and only if it satisfies the following criterion.
There is a dense subspace X0 ⊂ X, an increasing sequence (nk )k of positive
integers, and linear maps Snk : X0 → X, k ≥ 1, such that, for any x ∈ X0 ,
(i) T nk x → 0,
(ii) Snk x → 0,
(iii) T nk Snk x → x.

Proof. Suppose that T satisfies the Hypercyclicity Criterion. Then, by the

proof of Theorem 3.15, there exists a hypercyclic vector x ∈ X such that T
satisfies the Hypercyclicity Criterion for the dense set orb(x, T ) (taken twice)
and certain maps Snk on orb(x, T ). But since the vectors from orb(x, T ) form
a linearly independent set (see Proposition 2.60), one can extend each Snk
linearly to span orb(x, T ), and the Hypercyclicity Criterion will hold with
X0 = Y0 := span orb(x, T ).  

A more delicate question is whether the criteria of Theorems 3.10 and

3.12 are equivalent, that is, if every operator T satisfying the Hypercyclicity
Criterion also satisfies the Gethner–Shapiro criterion. It is far from evident
whether every operator that satisfies the Hypercyclicity Criterion has a right
inverse S of the type demanded. The main purpose of this section is to show
that both criteria are indeed equivalent. For this we will need an abstract
version of the Mittag-Leffler Theorem.

Theorem 3.21 (Mittag-Leffler). Let (Xn )n be a sequence of complete met-

ric spaces and let fn : Xn+1 → Xn , n ∈ N, be continuous maps with dense
range. Then, for every nonempty open subset U ⊂ X1 , there exists a sequence
(xn )n with xn ∈ Xn , n ∈ N, such that x1 ∈ U and fn (xn+1 ) = xn , n ∈ N.

Proof. Let dn be the metric of Xn , n ∈ N. Given a nonempty open set

U ⊂ X1 , we fix ε > 0 and x ∈ U such that
80 3 The Hypercyclicity Criterion

Fig. 3.1 The Mittag-Leffler theorem

{y ∈ X1 ; d1 (x, y) < ε} ⊂ U.

Set x1,1 = x. Since f1 has dense range, there are x1,2 ∈ X1 and x2,2 ∈ X2
such that f1 (x2,2 ) = x1,2 and d1 (x1,1 , x1,2 ) < ε/2. Proceeding by induction
we will find a countable family A = {xj,k ∈ Xj ; k ∈ N, 1 ≤ j ≤ k} satisfying
ε
fj (xj+1,k+1 ) = xj,k+1 and dj (xj,k , xj,k+1 ) < , k ∈ N, 1 ≤ j ≤ k.
2k
Indeed, suppose that we already have constructed the finite family {xj,k ∈
Xj ; 1 ≤ k ≤ n, 1 ≤ j ≤ k} with the above property. Since fn has dense
range and each fj , j < n, is continuous, we can find xn,n+1 ∈ fn (Xn+1 ) close
enough to xn,n so that, for xj,n+1 := fj (xj+1,n+1 ), j < n, we have
ε
dj (xj,n , xj,n+1 ) < , 1 ≤ j ≤ n.
2n
The induction process is completed if we pick xn+1,n+1 ∈ Xn+1 with
fn (xn+1,n+1 ) = xn,n+1 .
The defining property of the family A yields that for each j ∈ N the
sequence (xj,k )k≥j is a Cauchy sequence in Xj . We define xj = limk→∞ xj,k ∈
Xj , j ∈ N, to conclude that


k−1 
k−1
ε
d1 (x, x1 ) = lim d1 (x1,1 , x1,k ) ≤ lim d1 (x1,l , x1,l+1 ) < lim = ε,
k→∞ k→∞ k→∞ 2l
l=1 l=1

so that x1 ∈ U , and

fj (xj+1 ) = fj ( lim xj+1,k+1 ) = lim xj,k+1 = xj , j ∈ N,

k→∞ k→∞

which had to be shown. 



This result implies the aforementioned equivalence of two important cri-

teria.
3.3 Equivalent formulations of the Hypercyclicity Criterion 81

Theorem 3.22. An operator satisfies the Hypercyclicity Criterion if and only

if it satisfies the Gethner–Shapiro criterion.

Proof. Let T be an operator satisfying the Hypercyclicity Criterion.

By setting Xn = X, fn = T , n ∈ N, in Mittag-Leffler’s theorem, we obtain
that

Y := {x ∈ X ; ∃(xn )n ∈ X N with x1 = x, T xn+1 = xn , n ∈ N}

is a dense subspace of X. We consider X N endowed with the product topology.

Then
X = {(xn )n ∈ X N ; T xn+1 = xn , n ∈ N}
is a closed subspace of X N and therefore a separable Fréchet space under the
induced topology; see Exercise 2.1.3.
The operator T induces an operator T : X → X ,

T (x1 , x2 , x3 , . . . ) := (T x1 , T x2 , T x3 , . . . ) = (T x1 , x1 , x2 , . . . ).

Then T is an invertible operator whose inverse is the backward shift B : X →

X , B(x1 , x2 , . . . ) = (x2 , x3 , . . . ).
By Theorem 3.15 we know that T is hereditarily hypercyclic with respect
to some increasing sequence (mk )k of positive integers. We divide the remain-
der of the proof into three steps.

Step 1. (T mk )k is a topologically transitive sequence of operators; see Section

1.6 (and Section 3.4 below).
To see this, let U, V ⊂ X be nonempty open sets. Since it suffices to
consider sets from a base of the topology of X we can assume that

U = {(xn )n ∈ X ; xn ∈ Un , n = 1, . . . , N },

V = {(xn )n ∈ X ; xn ∈ Vn , n = 1, . . . , N }
with N ∈ N and nonempty open sets Un , Vn ⊂ X, n = 1, . . . , N .
Fix x = (xn )n ∈ U and y = (yn )n ∈ V. By continuity of T we can find

open neighbourhoods UN ⊂ UN , VN ⊂ VN of xN and yN , respectively, such
that T (UN ) ⊂ UN −j and T j (VN ) ⊂ VN −j , j = 1, . . . , N − 1.
j 

Since (T mk )k is topologically transitive (see Exercise 1.6.2), there exists

  
k ≥ 1 and a nonempty open set UN ⊂ UN such that T mk (UN ) ⊂ VN . By

the density of Y there exists some uN ∈ Y ∩ UN . Hence there are un ∈ X,
n > N , such that T un+1 = un , n ≥ N ; when we now set un = T N −n uN ∈ Un ,
n = 1, . . . , N − 1, then u := (un )n ∈ U; see Figure 3.2.
On the other hand, T mk u = (T mk u1 , . . . , T mk uN , . . . ). Since T mk uN ∈
VN , we get for n = 1, . . . , N − 1 that

T mk un = T mk (T N −n uN ) = T N −n (T mk uN ) ∈ T N −n (VN ) ⊂ Vn ,
82 3 The Hypercyclicity Criterion

and therefore T mk u ∈ V.

Fig. 3.2 Proof of Theorem 3.22, Step 1

Step 2. There exists a dense subset Y0 ⊂ X, a map S : Y0 → Y0 and a

subsequence (qk )k of (mk )k such that S qk y → 0 and T Sy = y, for each
y ∈ Y0 .
Indeed, given that the backward shift B is the inverse of the operator T ,
it follows from Step 1 that, for each pair of nonempty open sets U, V ⊂ X ,
there exists k ∈ N with
T mk (U) ∩ V = ∅,
and therefore
Bmk (V) ∩ U = ∅.
By the Universality Criterion, Theorem 1.57, both sequences (T mk )k and
(Bmk )k have dense Gδ -sets of vectors with dense orbits. Hence there is a
vector (yn )n ∈ X such that

{(ymk +1 , ymk +2 , . . . ) ; k ∈ N} = X = {(T mk y1 , T mk y2 , . . . ) ; k ∈ N}.

We now set Y0 = {yn ; n ∈ N}; since {(ymk +1 , ymk +2 , . . . ) ; k ∈ N} is

dense in X , projecting onto the first coordinate yields that Y0 is a dense
subset of X. Moreover we define S : Y0 → Y0 by Syn = yn+1 , n ∈ N. This
is well defined since ym = yn if m = n. Otherwise, we would have ym = yn
for some m > n, hence T m−n yn = yn , which implies that T m−n y1 = y1 , so
that y1 would be a periodic point for T ; this would contradict the density of
{(T mk y1 , . . . ) ; k ∈ N} in X , again by projecting onto the first coordinate.
3.4 Hypercyclic sequences of operators 83

Since (yn )n ∈ X we also have that T Sy = y for all y ∈ Y0 . Finally, let (qk )k
be a subsequence of (mk )k such that Bqk y = (y1+qk , y2+qk , . . . ) → 0 in X .
Then
S qk yn = yn+qk → 0, n ∈ N.

Step 3. The operator T satisfies the Gethner–Shapiro criterion.

In fact, since T is hereditarily hypercyclic with respect to (mk )k , there
is a hypercyclic vector w ∈ X and a subsequence (nk )k of (qk )k satisfying
T nk w → 0. If we set X0 = orb(w, T ), we obtain that, for any x ∈ X0 , y ∈ Y0 ,
(i) T nk x → 0,
(ii) S nk y → 0,
(iii) T Sy = y.
This concludes the proof.  

3.4 Hypercyclic sequences of operators

We resume here the discussion of Section 1.6 in the light of linearity. As

we saw in the previous sections, the study of the dynamics of iterates of an
operator T sometimes leads to questions concerning subsequences (T nk )k ,
which forces us to consider the dynamics of general sequences of operators.
Throughout this section, X and Y will denote separable Fréchet spaces and
Tn : X → Y, n ≥ 1, operators between these spaces.
We recall that in this setting the orbit of a vector x ∈ X is defined as
orb(x, (Tn )) = {Tn x ; n ∈ N0 }; see Definition 1.55.

Definition 3.23. A sequence (Tn )n of operators is called hypercyclic if there

is some x ∈ X whose orbit under (Tn )n is dense in Y . In such a case, x is
called a hypercyclic or universal vector for (Tn )n .
The sequence (Tn )n is called hereditarily hypercyclic if there is an increas-
ing sequence (nk )k of positive integers such that (Tmk )k is hypercyclic for
every subsequence (mk )k of (nk )k .

Topological transitivity, mixing and weak mixing for (Tn )n are defined in
Definition 1.56. A notion of chaos can be defined, but seems less natural in
this context.
Now, the Birkhoff transitivity theorem extends to commuting sequences
of operators Tn : X → X with dense range; see Exercise 1.6.2. For general
sequences of operators one should instead turn to the Universality Criterion;
see Theorem 1.57.
The results of Sections 2.4 and 2.5 carry over to commuting sequences of
operators Tn : X → X (if, in Theorem 2.45, the map S : X → X commutes
with each Tn ).
The Hypercyclicity Criterion extends, with the previous proof, to arbitrary
sequences of operators.
84 3 The Hypercyclicity Criterion

Theorem 3.24 (Hypercyclicity Criterion for sequences). Let (Tn )n be

a sequence of operators. If there are dense subsets X0 ⊂ X and Y0 ⊂ Y ,
an increasing sequence (nk )k of positive integers, and maps Snk : Y0 → X,
k ≥ 1, such that, for any x ∈ X0 , y ∈ Y0 ,
(i) Tnk x → 0,
(ii) Snk y → 0,
(iii) Tnk Snk y → y,
then (Tn )n is weakly mixing, and in particular hypercyclic.

The (simple) example of Exercise 1.6.1 shows that not every hypercyclic
sequence of operators satisfies the Hypercyclicity Criterion.
The Bès–Peris theorem extends likewise to commuting sequences. It is also
of interest to add the notion of hereditary transitivity; see Exercise 1.6.4.

Theorem 3.25 (Bès–Peris). Let (Tn )n be a commuting sequence of opera-

tors Tn : X → X. Then the following assertions are equivalent:
(i) (Tn )n satisfies the Hypercyclicity Criterion;
(ii) (Tn )n is weakly mixing;
(iii) (Tn )n is hereditarily transitive;
(iv) (Tn )n is hereditarily hypercyclic.

The result breaks down for noncommuting operators; see Exercise 3.4.4.

Exercises

Exercise 3.1.1. Prove the following version of Kitai’s criterion. If there are dense sub-
sets X0 , Y0 ⊂ X such that
(i) T n x → 0 for any x ∈ X0 ,
(ii) for any y ∈ Y0 there is a sequence (un )n in X such that un → 0 and T n un → y,
then T is a mixing operator.

Exercise 3.1.2. Let T be an operator that satisfies Kitai’s criterion with X0 = Y0 .

∞sequence
Construct an increasing
nk
(nk )k of positive integers and a dense sequence (yk )k
in Y0 such that x = k=1
S y k is a hypercyclic vector for T . (Hint: See Example 2.18
and work with an F-norm on X.)

Exercise 3.1.3. Show that an operator T satisfies Kitai’s criterion if and only if there
is a syndetic increasing sequence (nk )k of positive integers such that T satisfies the
Gethner-Shapiro criterion with respect to (nk )k .

Exercise 3.1.4. Let C : Lp [0, 1] → Lp [0, 1], 1 < p < ∞, be the Cesàro operator
t
defined as Cf (t) = 1t 0 f (s) ds. Show that C is mixing and chaotic. (Hint: Use the
Hahn–Banach theorem to show the density of span{tα ; α ∈ A} in Lp [0, 1], for any set
A ⊂ H := {z ∈ C ; Re z > −1/p} with accumulation points in H. Then apply the
Godefroy–Shapiro criterion.)

Exercise 3.2.1. Give a new proof of Theorem 3.10 based on Theorem 2.47.
Exercises 85

Exercise 3.2.2. Consider the weighted backward shift Bw : c0 → c0 , Bw (x1 , x2 , . . . ) =

(w2 x2 , w3 x3 , . . . ) with weight sequence

w = (w1 , w2 , . . . ) = (1, 2, 2−1 , 2, 2, 2−2 , 2, 2, 2, 2−3 , . . . ).

Show that Bw is weakly mixing but not mixing. (Hint: Proceed as in Example 3.11;
note that this time S is not continuous, so calculate S nk ej directly.)

Exercise 3.2.3. Let T be an operator. Consider the assertions:

(i) T satisfies the Hypercyclicity Criterion with respect to the full sequence (n);
(ii) T is mixing;
(iii) T is hereditarily hypercyclic with respect to the full sequence (n).
Show that (i) =⇒ (ii) ⇐⇒ (iii). (Hint: The proof is simple.)

Exercise 3.2.4. Show that the following “blow-up/collapse” criterion is a reformulation

of the Hypercyclicity Criterion: there are dense subsets X0 , Y0 ⊂ X and an increasing
sequence (nk )k of positive integers such that
(i) T nk x → 0 for any x ∈ X0 , and
(ii) for any y ∈ Y0 there is a sequence (uk )k in X such that uk → 0 and T nk uk → y.

Exercise 3.2.5. Show that every weakly mixing operator T satisfies the Hypercyclicity
Criterion, without using the hereditary transitivity property of T . To do this, fix a
hypercyclic vector (x, y) for T ⊕ T and proceed as follows:
(i) for every k ≥ 1, the vector (x, T k y) is also hypercyclic for T ⊕ T ;
(ii) if Uk is the open ball of radius 1/k around 0, k ≥ 1, then there is an increasing
sequence (nk )k of positive integers and a sequence of vectors xk ∈ Uk ∩ orb(y, T )
such that T nk x ∈ Uk and T nk xk ∈ x + Uk , k ≥ 1;
(iii) the dense set X0 = Y0 := orb(x, T ), the sequences (nk )k and the maps Snk :
Y0 → X, Snk (T n x) := T n xk , k, n ≥ 1, satisfy the hypotheses of the Hypercyclicity
Criterion.

Exercise 3.2.6. Show that every hypercyclic operator on the space ω = CN of all com-
plex sequences satisfies the Hypercyclicity Criterion. (Hint: Use the Herzog–Lemmert
theorem, Theorem 2.65.)

Exercise 3.3.1. Prove that an operator T satisfies the Hypercyclicity Criterion if and
only if there exists an increasing sequence (nk )k of positive integers such that the fol-
lowing two conditions are satisfied:
(i) there exists a dense subset X0 ⊂ Xsuch that T nk x → 0 for any x ∈ X0 ;
∞
(ii) for any 0-neighbourhood W of X, k=1 T nk (W ) is dense in X.
∞particular case when X is a Banach space, the second condition can be simplified
In the
to k=1 T nk (BX ) being dense in X, where BX denotes the unit ball of X.

Exercise 3.3.2. A subset A of a metric space is precompact if for every ε > 0 the set A
can be covered by a finite union of open balls of radius ε in the space. If the metric space
is complete, then any sequence in a precompact set admits a convergent subsequence.
Let T be an operator satisfying the following criterion: there are dense subsets
X0 , Y0 ⊂ X, an increasing sequence (nk )k of positive integers, and maps Snk : Y0 → X,
k ∈ N, such that, for any x ∈ X0 , y ∈ Y0 ,
(i) the set {T nk x ; k ∈ N} is precompact,
(ii) Snk y → 0,
(iii) T nk Snk y → y.
Show that T satisfies the Hypercyclicity Criterion. (Hint: Prove that T is weakly mixing.)
86 3 The Hypercyclicity Criterion

Exercise 3.3.3. Consider the  bilateral backward shift T = B on the weighted space
1 (Z, v) = {(xn )n∈Z ; x := n∈Z
|xn |vn < ∞}, where vn = n1 , n ≥ 1, and vn = 1,
n ≤ 0. Show that T is not hypercyclic. Deduce that in condition (i) of Exercise 3.3.2
one cannot replace precompactness by boundedness.

Exercise 3.3.4. Repeat Exercise 3.3.2 with conditions (i) and (ii) replaced by
(i) T nk x → 0 for any x ∈ X0 ,
(ii) the set {Snk y ; k ∈ N} is precompact for any y ∈ Y0 .

Exercise 3.3.5. Let T be a hypercyclic operator with a dense subset Y0 ⊂ X such

that each y ∈ Y0 admits a precompact backward orbit; see Exercise 1.2.7. Show that T
satisfies the Hypercyclicity Criterion. (Hint: Define X0 = orb(x, T ) with a hypercyclic
vector x and apply Exercise 3.3.4.)

Exercise 3.3.6. An invertible operator T satisfies the Hypercyclicity Criterion if and

only if T −1 does. Give two proofs of 
this result, one using Theorem 3.15, another one
∞
using Theorem 3.22. (Hint: Consider n=0 T n (X0 ).)

Exercise 3.3.7. Prove the following strengthened version of the Mittag-Leffler theorem.
Let (Xn )n be a sequence of complete metric spaces, let Un ⊂ Xn , n ∈ N, be dense open
sets and let fn : Xn+1 → Xn , n ∈ N, be continuous maps with dense range. Then,
for every nonempty open subset U ⊂ X1 , there exists a sequence (xn )n with xn ∈ Un ,
n ∈ N, such that x1 ∈ U and fn (xn+1 ) = xn , n ∈ N.
Deduce from this the Baire category theorem.

Exercise 3.4.1. Two operators T1 , T2 on a separable Fréchet space X are called disjoint
hypercyclic if there is some x ∈ X such that {(T1n x, T2n x) ; n ≥ 0} is dense in X ⊕ X.
(a) Confirm the following obvious facts. No operator T is disjoint hypercyclic with
itself. The operators T1 , T2 are disjoint hypercyclic if and only if some (x, x) ∈ X ⊕ X
is hypercyclic for T1 ⊕ T2 , and if and only if the sequence (Tn )n of operators Tn : X →
X ⊕ X, Tn x = (T1n x, T2n x) is hypercyclic.
(b) Derive a corresponding transitivity theorem. The operators T1 , T2 are disjoint
hypercyclic if, for any nonempty open subsets U, V1 , V2 of X, there exists some n ≥ 0
such that T1n (U ) ∩ V1 = ∅ and T2n (U ) ∩ V2 = ∅.
(c) Derive a corresponding Disjoint Hypercyclicity Criterion. If there are dense
subsets X0 , Y1 , Y2 ⊂ X, an increasing sequence (nk )k of positive integers, and maps
Sj,nk : Yj → X, k ≥ 1, j = 1, 2, such that, for any x ∈ X0 , y1 ∈ Y1 , y2 ∈ Y2 ,
(i) Tjnk x → 0, j = 1, 2,
(ii) Sj,nk yj → 0, j = 1, 2,
(iii) Tjnk Sj,nk yj → yj , j = 1, 2,
(iv) Tlnk Sj,nk yj → 0, j = 1, l = 2 or j = 2, l = 1,
then T1 , T2 are disjoint hypercyclic.

Exercise 3.4.2. Show that the following operators are disjoint hypercyclic:
(i) λB, μB 2 , where 1 < |λ| < |μ| and B is the backward shift on any of the spaces p ,
1 ≤ p < ∞, or c0 ;
(ii) Ta , Tb , where a, b ∈ C, a = 0, b = 0, a = b, and Ta is the translation operator on
H(C) given by Ta f (z) = f (z + a).

Exercise 3.4.3. Formulate and prove an analogue of the Gethner–Shapiro criterion

for sequences of operators Tn : X → X. Show that the operators Tn : 2 → 2 ,
Tn x = 2n (xn+1 , . . . , x2n , 0, 0, . . .), n ≥ 1, satisfy the Hypercyclicity Criterion but not
the Gethner–Shapiro criterion, so that the two criteria are no longer equivalent.
Sources and comments 87

Exercise 3.4.4. Let B be the backward shift operator on 2 . Define Tn : 2 → 2

by Tn x = 2n B n+1 x + 3n (x1 , 0, 0, . . .), n ≥ 1. Characterize the hypercyclic vectors for
(Tn )n . Show that (Tn )n is hereditarily hypercyclic but not topologically transitive, and
therefore not hereditarily transitive.

Exercise 3.4.5. Let (Tn )n be a sequence of operators. Consider the assertions:

(i) (Tn )n satisfies the Hypercyclicity Criterion with respect to the full sequence (n);
(ii) (Tn )n is mixing;
(iii) (Tn )n is hereditarily transitive with respect to the full sequence (n).
Show that (i) =⇒ (ii) ⇐⇒ (iii).

Exercise 3.4.6. Let B be the backward shift operator on 2 and consider the operators
Tn : 2 → 2 , Tn x = 2n B n x − x, n ≥ 1. Show that (Tn )n is mixing and that the only
sequences x for which Tn x → 0 are multiples of the sequence ( 21n )n . Deduce that the
implication (ii) =⇒ (i) does not hold in the previous exercise.

Exercise 3.4.7. Give an example of a hypercyclic noncommuting sequence of operators

for which Lemma 2.44 does not hold.

Exercise 3.4.8. Give an example of a noncommuting sequence of operators that satis-

fies the hypothesis of Theorem 2.47, but is not weakly mixing.

Exercise 3.4.9. Let (Tj,n )n , j ≥ 1, be a countable family of hereditarily transitive

sequences of operators between separable Fréchet spaces X and Y . Show that there
exists a dense subset X0 of X and increasing sequences (nj,k )k of positive integers such
that, for any x ∈ X0 , Tj,nj,k x → 0 as k → ∞, j ≥ 1. (Hint: Use Theorem 1.57 to
produce common hypercyclic vectors for suitable sequences of operators. Note that any
hypercyclic vector has a suborbit that tends to 0.)

Exercise 3.4.10. Let (Tj,n )n , j ≥ 1, be a countable family of hereditarily transitive

sequences of operators between separable Fréchet spaces X and Y . Show that there
exists a dense subspace M of X such that every nonzero vector from M is hypercyclic
for each sequence (Tj,n )n , j ≥ 1.
Deduce that for every countable family of weakly mixing operators Tj , j ≥ 1, there
exists a dense subspace M of X such that every nonzero vector from M is hypercyclic
for each operator Tj , j ≥ 1.
(Hint: The proof is similar to that of the previous exercise; take the span of a suitably
constructed dense sequence of vectors.)

Sources and comments

Sections 3.1 and 3.2. The earliest forms of the Hypercyclicity Criterion were found
independently by Kitai [215] (who in addition demanded that X0 = Y0 ) and by Gethner
and Shapiro [161, Remark 2.3(b)]. In its general form it is due to Bès and Peris [71]. The
Godefroy–Shapiro criterion is contained implicitly in their paper [165] and was isolated
by Bernal [55].
Intended originally as simple tools for obtaining the hypercyclicity of an opera-
tor, these criteria have become tremendously important for the understanding of lin-
ear dynamics. While the Hypercyclicity Criterion, very unexpectedly, turned out to be
equivalent to the weak mixing property and therefore to be close to hypercyclicity, the
Godefroy–Shapiro criterion was further developed by Bayart and Grivaux in their study
of frequent hypercyclicity; see Chapter 9.
88 3 The Hypercyclicity Criterion

Grivaux [172] showed that there is a mixing operator T for which only x = 0 satisfies
that T n x → 0 as n → ∞. Thus, not every mixing operator satisfies the Hypercyclicity
Criterion for the full sequence (n), nor Kitai’s criterion; see Exercise 3.2.3.
We remark that the main idea of Example 3.8, that of considering the functions
z → p(z)e−αz , was taken from Birkhoff’s original proof in [75].
2

Theorem 3.15 is due Bès and Peris [71], as is Proposition 3.19. The notion of hered-
itary hypercyclicity with respect to the full sequence was introduced by Shapiro [279]
(who called it strong hypercyclicity) and by Ansari [9].

Section 3.3. Theorem 3.22 is due to Peris [253]; for the abstract Mittag-Leffler theorem
we refer to Arens [11, Theorem 2.4]; see also Esterle [145] for an interesting discussion.
It seems to be an open problem whether the strengthened conditions of the Gethner–
Shapiro criterion and of Proposition 3.20 can be combined, that is, if every operator
satisfying the Hypercyclicity Criterion admits a dense subspace X0 ⊂ X, an increasing
sequence (nk )k of positive integers, and a linear map S : X0 → X0 such that, for any
x ∈ X0 , T nk x → 0, S nk x → 0, and T Sx = x.

Section 3.4. Much closer studies of the Hypercyclicity Criterion for sequences of op-
erators were undertaken by Bernal and Grosse-Erdmann [62], Bermúdez, Bonilla, and
Peris [51] and León and Müller [223].

Exercises. For Exercise 3.1.3 we refer to Bermúdez, Bonilla, Conejero, and Peris [49].
Exercise 3.1.4 is taken from León, Piqueras and Seoane [224]; the continuity of the
Cesàro operators is due to Hardy’s inequality [193]. We will return to these operators
in Example 12.20. Exercise 3.2.5 follows the original argument in Bès and Peris [71].
Exercise 3.3.1 is taken from León and Müller [223]; see also Müller [247]. Exercises 3.3.2,
3.3.4 and 3.3.5 are due to Bermúdez, Bonilla, and Peris [51]. The notion of disjoint
hypercyclic operators is due to Bernal [58] and Bès and Peris [72], where one can also
find the results of Exercises 3.4.1 and 3.4.2. The example in Exercise 3.4.4 is a special
case of a class of operators studied by Bernal [56]. For Exercise 3.4.9 we refer to Aron,
Bès, León and Peris [14], for Exercise 3.4.10 to Bernal and Calderón [61]. Concerning
the latter exercise, it is interesting to note that by a remarkable result by Grivaux [169],
any countable family of hypercyclic operators on a Banach space has a common dense
subspace of hypercyclic vectors, except for 0.

Extensions. Repeating a remark we made in Chapter 2 we note that the results in this
chapter remain valid in separable F-spaces. For further extensions we refer to Chapter
12.
Chapter 4
Classes of hypercyclic and chaotic
operators

In this chapter we study in detail some important classes of hypercyclic and

chaotic operators. Each of them has its origin in the three classical hypercyclic
operators. Rolewicz’s multiples of backward shifts lead naturally to the study
of arbitrary weighted shifts. MacLane’s differentiation operator and Birkhoff’s
translation operators are both special cases of differential operators, while the
translation operators can also be regarded as composition operators. Finally,
Rolewicz’s operators reappear as adjoint multipliers.

4.1 Weighted shifts

The basic model of all shifts is the backward shift

B(x1 , x2 , x3 , . . .) = (x2 , x3 , x4 , . . .).

In order to distinguish this shift from the bilateral shift that we will discuss
later one also speaks of the unilateral backward shift.
Rolewicz has shown that, for any λ with |λ| > 1, the multiples of B,
λB(xn )n = (λxn+1 )n , are hypercyclic on the sequence space 2 . It is then a
small step to let the weights vary from coordinate to coordinate, which leads
to the (unilateral) weighted shift

Bw (x1 , x2 , x3 , . . .) = (w2 x2 , w3 x3 , w4 x4 , . . .),

where
w = (wn )n
is a sequence of nonzero scalars, called a weight sequence. Note that the value
of w1 is irrelevant.
We may also generalize these operators in a different direction. Rolewicz
had already replaced 2 by any of the spaces p , 1 ≤ p < ∞, and c0 . More

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 89

DOI 10.1007/978-1-4471-2170-1_4, © Springer-Verlag London Limited 2011
90 4 Classes of hypercyclic and chaotic operators

generally, one may take as the underlying space an arbitrary sequence space
X, that is, a linear space of sequences or, in other words, a subspace of
ω = KN . In addition, X should carry a topology that is compatible with
the sequence space structure of X. We interpret this as demanding that the
embedding X → ω is continuous, that is, convergence in X should imply co-
ordinatewise convergence. A Banach (Fréchet, . . . ) space of this kind is called
a Banach (Fréchet, . . . ) sequence space. The terms of a sequence x, y, z, . . .
will be denoted by xn , yn , zn , . . ., n ≥ 1.
By en , n ∈ N,

en = (δn,k )k∈N = (0, . . . , 0, 1, 0, . . .)

n

we denote the canonical unit sequences. If the en are contained in X and

span a dense subspace then an alternative way of describing weighted shifts
is by saying that

Bw en = wn en−1 , n ≥ 1, with e0 := 0.

The continuity of the embedding X → ω amounts to requiring the conti-

nuity of each coordinate functional

X → K, x → xn , n ≥ 1,

which implies that each weighted shift has closed graph. From the closed
graph theorem (see Appendix A) we thus obtain that a weighted shift defines
an operator on a Fréchet sequence space X as soon as it maps X into itself.
Proposition 4.1. Let X be a Fréchet sequence space. Then every weighted
shift Bw : X → X is continuous.
We start by studying the (unweighted) backward shift B. Our results will
then extend immediately to all weighted shifts via a simple conjugacy.
The following technical result will help us simplify the condition charac-
terizing hypercyclicity of B.
Lemma 4.2. Let X be a metric space, vn ∈ X, n ≥ 1, and v ∈ X. Suppose
that there is a strictly increasing sequence (nk )k of positive integers such that

vnk −j → v for every j ∈ N.

Then there exists a strictly increasing sequence (mk )k of positive integers

such that
vmk +j → v for every j ∈ N.
Proof. Let d denote the metric in X. It follows from the assumption that, for
any k ≥ 1, there is some Nk ≥ k + 2 such that
1
d(vNk −j , v) < , j = 1, . . . , k.
k
4.1 Weighted shifts 91

Setting mk = Nk − k − 1, k ≥ 1, we see that d(vmk +k+1−j , v) < 1

k for
j = 1, . . . , k, hence
1
d(vmk +j , v) < , j = 1, . . . , k;
k
this implies the assertion when we pass to a strictly increasing subsequence
of (mk )k , if necessary. 


We recall that the sequence (en )n is a basis in the space X if each en ,

n ∈ N, belongs to X and, for any x ∈ X,
∞

x = lim (x1 , x2 , . . . , xN , 0, 0, . . .) = xn e n .
N →∞
n=1

Clearly, (en )n is a basis in any of the sequence spaces p , 1 ≤ p < ∞, c0 and

ω.

Theorem 4.3. Let X be a Fréchet sequence space in which (en )n is a basis.

Suppose that the backward shift B is an operator on X. Then the following
assertions are equivalent:
(i) B is hypercyclic;
(ii) B is weakly mixing;
(iii) there is an increasing sequence (nk )k of positive integers such that
enk → 0 in X as k → ∞.

Proof. Let · stand for an F-norm that induces the topology of X; see
Section 2.1.
(i)=⇒(iii). Suppose that B is hypercyclic. Let N ∈ N and ε > 0. We show
that there exists some n ≥ N with en < ε.
It follows from the basis assumption that, for any x ∈ X, the sequence
(xn en )n converges to 0 in X. By the Banach–Steinhaus theorem (see Ap-
pendix A), applied to the operators x → xn en , n ≥ 1, there is some δ > 0
such that, for all x ∈ X,

x < δ =⇒ xn en < ε
2 for all n ≥ 1. (4.1)

Moreover, since convergence in X implies coordinatewise convergence,

there is some η > 0 such that, for all x ∈ X,

x < η =⇒ |x1 | ≤ 12 . (4.2)

Now, since B is hypercyclic and therefore topologically transitive, there

are x ∈ X and n ≥ N such that

x < δ and B n−1 x − e1 < η.

Hence, by (4.1) and (4.2),

92 4 Classes of hypercyclic and chaotic operators

xn en < ε
2 and |xn − 1| ≤ 12 ; (4.3)

the latter implies that xn is closer to 1 than to 0 and hence that

 −1   
xn − 1 =  1 − xn  ≤ 1.
xn
From this and (4.3) we deduce, using the properties of an F-norm, that
   −1 
en = (x−1   
n − 1)xn en + xn en ≤ (xn − 1)xn en + xn en < ε, (4.4)

which had to be shown.

(iii)=⇒(ii). We apply the Hypercyclicity Criterion. For X0 = Y0 we take
the set of finite sequences, which by the basis assumption is dense in X. For
Sn we take the nth iterate of the forward shift

F : (x1 , x2 , x3 , . . .) → (0, x1 , x2 , . . .),

that is, Sn = F n : Y0 → X, n ≥ 1. With this, conditions (i) and (iii) of the

Hypercyclicity Criterion hold even for the full sequence (n).
As for condition (ii) note that, by continuity of B,

enk −j = B j enk → 0 as k → ∞,

for all j ≥ 1. Since (nk )k must be strictly increasing, it follows from Lemma
4.2 that there is an increasing sequence (mk )k of positive integers such that

emk +j → 0 as k → ∞,

for all j ≥ 1. But since Smk ej = emk +j , we have by linearity that

Smk y → 0

for any y ∈ Y0 . This shows that conditions (i)–(iii) of the Hypercyclicity

Criterion hold for the sequence (mk )k , so that B is weakly mixing.
(ii)=⇒(i) holds for all operators on X.  

We point out that B being an operator on X is part of the hypothesis.

By Proposition 4.1 this can be restated simply as saying that (xn+1 )n ∈ X
whenever (xn )n ∈ X, which is usually easily verified for concrete spaces.

Example 4.4. (a) Let

 ∞
 
p (v) = (xn )n≥1 ; |xn |p vn < ∞ , 1 ≤ p < ∞,
n=1

be a weighted p -space, where v = (vn )n is a positive weight sequence. Then

B is an operator on p (v) if and only if there is an M > 0 such that, for all
4.1 Weighted shifts 93

x ∈ p (v),

∞ 1/p 
∞ 1/p
|xn+1 |p vn ≤M |xn |p vn ,
n=1 n=1
vn
which is equivalent to supn∈N vn+1 < ∞. Theorem 4.3 tells us that, under
this condition, the hypercyclicity of B is characterized by

inf vn = 0.
n∈N

The same conditions also characterize the continuity and hypercyclicity of

the backward shift B on the weighted c0 -space

c0 (v) = (xn )n≥1 ; lim |xn |vn = 0 .

n→∞

(b) Spaces of holomorphic functions constitute a rich and interesting source

of sequence spaces via the identification of a holomorphic function with its se-
quence of Taylor coefficients. As a first example we consider here the Bergman
space A2 of all holomorphic functions f on the unit disk D = {z ∈ C ; |z| < 1}
such that 
1
f 2 := |f (z)|2 dλ(z) < ∞,
π D
where λ denotes two-dimensional
 Lebesgue measure. Using polar coordinates
and writing f (z) = ∞ n
n=0 an z we obtain that
  
1 1 2π   2
∞
1 
|f (z)|2 dλ(z) =  an (reit )n  dt r dr
π D π 0 0 n=0
 1   2π   ∞ 2 
 1 
=2  an r n √ eint  dt r dr
0 0 n=0
2π
 1 ∞  ∞
1
=2 |an |2 r 2n r dr = |an |2 ,
0 n=0 n=0
n+1

where we have applied Parseval’s identity in L2 [0, 2π] for the orthonormal
basis ( √12π eint )n∈Z . As a consequence, A2 is isometrically isomorphic to the
1
weighted space 2 ( n+1 ) (with indices running from 0). By (a), the backward
shift is therefore hypercyclic on A2 . When acting on functions, B is the
operator
∞
 1
Bf (z) = an+1 z n = (f (z) − f (0)) with Bf (0) = f  (0).
n=0
z

Further Banach and Hilbert spaces of holomorphic functions will be studied

in Section 4.4.
94 4 Classes of hypercyclic and chaotic operators

(c) As in (b) we can consider the space H(C) of entire functions (see Ex-
ample
∞ 2.1)n as a sequence space by identifying the entire function f (z) =
n=0 an z with the sequence (an )n≥0 . By the formula for the radius of con-
vergence of Taylor series, this sequence space is given by

 ∞
 
(an )n≥0 ; lim |an |1/n = 0 = (an )n≥0 ; |an |mn < ∞, m ≥ 1 .
n→∞
n=0

Since |an+1 |1/n = (|an+1 |1/(n+1) )(n+1)/n → 0 if |an |1/n → 0, we have that
the backward shift B is an operator on H(C); see Proposition 4.1. Moreover,
the unit sequences en correspond to the monomials z → z n , n ≥ 0. It then
follows from Theorem 4.3 that B is not hypercyclic on H(C).

Using the same arguments as in the proof of Theorem 4.3, but employing
Kitai’s criterion instead of the Hypercyclicity Criterion, we obtain a charac-
terization of the mixing property for B.

Theorem 4.5. Let X be a Fréchet sequence space in which (en )n is a basis.

Suppose that the backward shift B is an operator on X. Then the following
assertions are equivalent:
(i) B is mixing;
(ii) en → 0 in X as n → ∞.

For chaos we have a curious phenomenon. Proceeding as before, but with

somewhat stronger assumptions on the space X, we easily obtain a condition
that characterizes chaos for B. But it turns out that this condition is already
implied by the existence of a single nontrivial periodic point, that is, a periodic
point other than 0. Hence, this fact alone implies chaos.
For this result we will require that (en )n is an unconditional basis, that is,
it is a basis in X such that, for any (xn )n ∈ X and any 0-1-sequence (εn )n ,
the series ∞

εn xn en
n=1

converges in X; see Appendix A.

Theorem 4.6. Let X be a Fréchet sequence space in which (en )n is an un-

conditional basis. Suppose that the backward shift B is an operator on X.
Then the following assertions are equivalent:
(i) 
B is chaotic;
∞
(ii) n=1 en converges in X;
(iii) the constant sequences belong to X;
(iv) B has a nontrivial periodic point.

Proof. (i)=⇒(iv) is trivial.

(iv)=⇒(iii). Let x = (x1 , x2 , x3 , . . .) = 0 be periodic for B, that is, a
periodic sequence. Let N be its period. Then there is some j ≤ N such that
4.1 Weighted shifts 95

xj = 0, and we have xj+νN = xj for ν ≥ 0. Setting all coordinates with

indices other than j + νN to zero and dividing the result by xj we obtain,
by unconditionality of the basis, that
∞

ej+νN ∈ X.
ν=0

Applying the backward shift N − 1 times and adding the results we obtain
(iii).
(iii)=⇒(ii) follows from our assumptions.
(ii)=⇒(i). First, by Theorem 4.3, condition (ii) implies that B is hyper-
cyclic.
Next, since (1, 1, 1, . . .) ∈ X, the unconditionality of the basis implies that
all the periodic 0-1-sequences belongs to X, and hence also all the periodic
sequences, which are exactly the periodic points for B. It remains to show
that these form a dense set in X.
To see this, let x = (xn )n ∈ X and ε > 0. Since (en )n is a basis, there is
some N ≥ 1 such that

N
x := xn e n
n=1

has distance less than ε/2 from x. The associated periodic sequence
∞ 
 N
xn en+νN
ν=0 n=1

belongs to X. The unconditionality of the basis implies that there is some

m ≥ 1 such that
 ∞  N  ε
 
 xn εn+νN en+νN  <
ν=m n=1
2

for any 0-1-sequence (εn )n ; see Theorem A.16. In particular we have that

 ∞  N  ε
 
 xn en+μmN  < .
μ=1 n=1
2

This shows that the periodic point

∞ 
 N
xn en+μmN
μ=0 n=1

, hence less than ε from x. 

has distance less than ε/2 from x 
Example 4.7. (a) We consider again the space p (v) of Example 4.4(a). Under
the assumption that B is an operator on p (v) we have that B is mixing if
96 4 Classes of hypercyclic and chaotic operators

and only if
lim vn = 0,
n→∞

and it is chaotic if and only if

∞

vn < ∞.
n=1

In this example, mixing is implied by chaos. In particular, the backward shift

on the Bergman space A2 is mixing but not chaotic.
(b) It is not difficult to give an example that shows that Theorem 4.6 does
not remain valid if one drops the unconditionality assumption on the basis
(en )n ; see Exercise 4.1.3.

It is now an easy matter to transfer our results so far to arbitrary weighted

shifts by means of a suitable conjugacy. Let Bw be a weighted shift on some
sequence space X. We define new weights vn by

n −1
vn = wν , n ≥ 1,
ν=1

and consider the sequence space

Xv = {(xn )n ; (xn vn )n ∈ X}.

The map φv : Xv → X, (xn )n → (xn vn )n is a vector space isomorphism.

We may use φv to transfer a topology from X to Xv : a set U is open in Xv
if and only if φv (U ) is open in X. If X is a Banach (Fréchet, . . . ) sequence
space then so is Xv . And if (en )n is a basis in X then it is also a basis in Xv .
Finally, a simple calculation shows that Bw ◦ φv = φv ◦ B, that is, the
following diagram commutes:
B
Xv −−−−→ Xv
⏐ ⏐
⏐
φv 
⏐φ
 v
B
X −−−w−→ X.
Thus Bw : X → X and B : Xv → Xv are conjugate operators.
Since conjugacies preserve hypercyclicity, (weak) mixing and chaos, our
previous results immediately yield the following.

Theorem 4.8. Let X be a Fréchet sequence space in which (en )n is a basis.

Suppose that the weighted shift Bw is an operator on X.
(a) The following assertions are equivalent:
(i) Bw is hypercyclic;
(ii) Bw is weakly mixing;
4.1 Weighted shifts 97

(iii) there is an increasing sequence (nk )k of positive integers such that


nk −1
wν enk → 0
ν=1

in X as k → ∞.
(b) The following assertions are equivalent:
(i) Bw is mixing;
(ii) we have that
 n −1
wν en → 0
ν=1

in X as n → ∞.
(c) Suppose that the basis (en )n is unconditional. Then the following asser-
tions are equivalent:
(i) Bw is chaotic;
(ii) the series
∞  n −1
wν en
n=1 ν=1

converges in X;
(iii) the sequence
 
n −1 
wν
n
ν=1

belongs to X;
(iv) Bw has a nontrivial periodic point.

Example 4.9. (a) A weighted shift Bw is an operator on a sequence space

p , 1 ≤ p < ∞, or c0 if and only if the weights wn , n ≥ 1, are bounded.
The respective characterizing conditions for Bw to be hypercyclic, mixing or
chaotic on p , 1 ≤ p < ∞, are


n 
n ∞
 1
sup |wν | = ∞, lim |wν | = ∞, n < ∞.
ν=1 |wν |
n→∞ p
n≥1 ν=1 ν=1 n=1

We remark that only the third condition depends on the parameter p. The
first condition also characterizes when Bw is hypercyclic on c0 , and the second
when it is mixing or, equivalently, chaotic on c0 .
nIn particular, for Rolewicz’s operator T = λB, |λ| > 1, we have that
ν=1 |wν | = λ , which implies once more that this operator is chaotic.
n

As another specificexample we consider, for α > 0, the weights wn =

n
n ) , n ≥ 1. Then
( n+1 ν=1 |wν | = (n + 1) , and the corresponding weighted
α α

shift is mixing; it is even chaotic on c0 , and it is chaotic on p exactly when

α > 1/p. We note that, for wn = ( n+1 n )
1/2
, n ≥ 1, the weighted shift Bw
98 4 Classes of hypercyclic and chaotic operators

on 2 is conjugate to the backward shift on the Bergman space; see Example

4.4(b).
(b) We consider the Fréchet space H(C) of all entire functions as a sequence
space; see Example 4.4(c). It is easy to see that a weighted shift Bw defines an
operator on H(C) if and only if supn≥1 |wn |1/n < ∞; moreover, we have that
an en → 0 in H(C) if and only if |an |1/n → 0; see Exercise 4.1.1. Theorem
4.8 then shows that a weighted shift Bw on H(C) is mixing if and only if
it is chaotic, and that the characterizing conditions for hypercyclicity and
mixing/chaos are, respectively,

n 1/n 
n 1/n
sup |wν | = ∞, lim |wν | = ∞.
n≥1 n→∞
ν=1 ν=1

In particular, for the differentiation operator we have that D( ∞ n
n=0 an z )
∞ n
= n=0 (n + 1)an+1 z , so that D is a weighted shift with weight sequence
wn = n, n ≥ 1. Since (n!)1/n → ∞ we obtain MacLane’s theorem that D is
hypercyclic; in fact, it is even a mixing and chaotic operator.
But in order to prove chaos for D it suffices, as we have seen, to come up
with a nontrivial periodic point; the easiest such example is f (z) = ez . Thus,
one might be tempted to say that the exponential
∞ function makes D chaotic.
(c) In the space ω = KN , every series n=1 an en converges. Thus, every
weighted shift Bw defines an operator on ω and is, indeed, mixing and chaotic
on ω.
Remark 4.10. By Example 4.9(a), any bounded weight sequence (wn )n with


n 
n
lim inf |wν | < lim sup |wν | = ∞
n→∞ n→∞
ν=1 ν=1

defines a weighted shift Bw on p , 1 ≤ p < ∞, or c0 that is weakly mixing but

not mixing. This provides us with a large supply of operators of this kind;
see also Example 3.11.
Remark 4.11. One might wonder why we have studied backward shifts and
not forward shifts. The simple truth is that a forward shift is never hyper-
cyclic. More precisely, a (unilateral) weighted forward shift is given by

Fw (x1 , x2 , x3 , . . .) = (0, w1 x1 , w2 x2 , . . .)

with a weight sequence w = (wn )n . The first coordinate of every point in the
orbit of x is either x1 or 0. By the assumption that convergence in the space
implies coordinatewise convergence no orbit can be dense.
Some new and interesting phenomena arise when we study shifts on se-
quence spaces indexed over Z. The bilateral backward shift is given by

B(xn )n∈Z = (xn+1 )n∈Z ,

4.1 Weighted shifts 99

and the bilateral weighted backward shifts are given by

Bw (xn )n∈Z = (wn+1 xn+1 )n∈Z ,

where w = (wn )n∈Z is a weight sequence, that is, a sequence of nonzero

scalars. The underlying space is then supposed to be a Banach (Fréchet,
. . . ) sequence space over Z, that is, a subspace of ω(Z) := KZ that carries a
Banach (Fréchet, . . . ) space topology under which the embedding X → ω(Z)
is continuous.
In this new setting, we say that the unit sequences

en = (δn,k )k∈Z , n ∈ Z,

form a basis in X if they are contained in X and if every sequence x =

(xn )n∈Z ∈ X satisfies

x= lim (. . . , 0, 0, x−M , x−M +1 , . . . , xN −1 , xN , 0, 0, . . .).

M,N →∞

The finite sequences are the sequences in span{en ; n ∈ Z}.

Theorem 4.12. Let X be a Fréchet sequence space over Z in which (en )n∈Z
is a basis. Suppose that the bilateral shift B is an operator on X.
(a) The following assertions are equivalent:
(i) B is hypercyclic;
(ii) B is weakly mixing;
(iii) there is an increasing sequence (nk )k of positive integers such that,
for any j ∈ Z, ej−nk → 0 and ej+nk → 0 in X as k → ∞.
(b) The following assertions are equivalent:
(i) B is mixing;
(ii) e−n → 0 and en → 0 in X as n → ∞.
(c) Suppose that the basis (en )n is unconditional. Then the following asser-
tions are equivalent:
(i) B
∞ is chaotic;
(ii) n=−∞ en converges in X;
(iii) the constant sequences belong to X;
(iv) B has a nontrivial periodic point.

Proof. (a), (i)=⇒(iii). Let · be an F-norm that induces the topology of

X. We will derive the following equivalent formulation of (iii): for any ε > 0
and any N ∈ N there exists some n ≥ N such that if |j| ≤ N , then

ej−n < ε and ej+n < ε.

To this end, we fix ε > 0 and N ∈ N. As in the unilateral case we can find

some δ > 0 such that, for all x ∈ X,

x < δ =⇒ xn en < ε
2 (n ∈ Z) and |xj | ≤ 1
2 (|j| ≤ N ). (4.5)
100 4 Classes of hypercyclic and chaotic operators

Now, by the topological transitivity of B, we can find some x ∈ X and

n > 2N such that

   

 n   
x − ej  < δ and B x − ej  < δ. (4.6)
|j|≤N |j|≤N

From (4.5) and (4.6) we obtain that

xn en < ε
2
(|n| > N ) and |xn+j − 1| ≤ 1
2
(|j| ≤ N ),

hence

xj+n ej+n < ε

2 (|j| ≤ N ) and |(xn+j )−1 − 1| ≤ 1 (|j| ≤ N );

here we have used that n > 2N . As in (4.4) this implies that

ej+n < ε for |j| ≤ N .

On the other hand, (4.5) and (4.6) yield that

|xj − 1| ≤ 1
2
(|j| ≤ N ) and xn+k ek < ε
2
(|k| > N ),

hence
 
(2xj )−1  ≤ 1 (|j| ≤ N ) and xj ej−n < ε
(|j| ≤ N ),
2

whence  
ej−n = (2xj )−1 2xj ej−n  < ε for |j| ≤ N .
(iii)=⇒(ii). One need only observe that for the forward shift

F (xn )n∈Z = (xn−1 )n∈Z

we have that BF x = x for any finite sequence x, and for any j ∈ Z

B nk ej = ej−nk → 0, F nk ej = ej+nk → 0,

so that the Hypercyclicity Criterion gives the required implication.

The implication (ii)=⇒(i) holds for all operators on X.
(b) The proof here is the same as that for hypercyclicity; for the sufficiency
of condition (ii) one applies Kitai’s criterion instead of the Hypercyclicity
Criterion, while the proof of the necessity of this condition simplifies as we
have only to consider the case of j = 0.
(c) This proof is much the same as that in the unilateral case.  

Using a suitable conjugacy this result can again be generalized immediately

to weighted shifts. The conjugacy here is given by
4.1 Weighted shifts 101

B
Xv −−−−→ Xv
⏐ ⏐
⏐
φv 
⏐φ
 v
B
X −−−w−→ X,
where
Xv = {(xn )n∈Z ; (xn vn )n ∈ X}
and φv : Xv → X, (xn )n∈Z → (xn vn )n∈Z with


n −1 
0
vn = wν for n ≥ 1, vn = wν for n ≤ −1, v0 = 1.
ν=1 ν=n+1

Theorem 4.13. Let X be a Fréchet sequence space over Z in which (en )n∈Z
is a basis. Suppose that the weighted shift Bw is an operator on X.
(a) The following assertions are equivalent:
(i) Bw is hypercyclic;
(ii) Bw is weakly mixing;
(iii) there is an increasing sequence (nk )k of positive integers such that,
for any j ∈ Z,

 
j   j+n
k −1
wν ej−nk → 0 and wν ej+nk → 0
ν=j−nk +1 ν=j+1

in X as k → ∞.
(b) The following assertions are equivalent:
(i) Bw is mixing;
(ii) we have that

 
0  
n −1
wν e−n → 0 and wν en → 0
ν=−n+1 ν=1

in X as n → ∞.
(c) Suppose that the basis (en )n∈Z is unconditional. Then the following as-
sertions are equivalent:
(i) Bw is chaotic;
(ii) the series


0  
0  ∞ 
 n −1
wν en + wν en
n=−∞ ν=n+1 n=1 ν=1

converges in X;
(iii) the sequence (xn )n∈Z with
102 4 Classes of hypercyclic and chaotic operators


0 
n −1
xn = wν (n ≤ 0), xn = wν (n ≥ 1)
ν=n+1 ν=1

belongs to X;
(iv) Bw has a nontrivial periodic point.

We see that the absence of an analogue of Lemma 4.2 leads to a more

complicated characterization of hypercyclic bilateral shifts. However, for in-
vertible bilateral shifts a simplified characterization is available; see Exercises
4.1.4 and 4.1.5.

Remark 4.14. In the bilateral case, forward shifts can be hypercyclic. A bi-
lateral weighted forward shift is given by an operator

Fw : X → X, (xn )n∈Z → (wn−1 xn−1 )n∈Z ,

where w = (wn )n∈Z is a weight sequence. It is easily seen to be conjugate to a

suitable backward shift. As a result one obtains, under the same assumptions
as in Theorem 4.13, that Fw is hypercyclic if and only if there is an increasing
sequence (nk )k of positive integers such that, for any j ∈ Z,

 
j−1 −1  j+n
 k −1 
wν ej−nk → 0 and wν ej+nk → 0
ν=j−nk ν=j

in X as k → ∞. The corresponding characterizations hold for the mixing

property and chaos.

Example 4.15. A weighted backward shift Bw is an operator on a sequence

space p (Z), 1 ≤ p < ∞ if and only if the weights wn , n ∈ Z, are bounded.
Such an operator is then hypercyclic, mixing or chaotic if and only if the
following conditions, respectively, are satisfied:


j k
j+n
∃(nk )k ∀j ∈ Z : lim wν = 0 and lim |wν | = ∞;
k→∞ k→∞
ν=j−nk +1 ν=j+1


0 
n
lim wν = 0 and lim |wν | = ∞;
n→∞ n→∞
ν=−n+1 ν=1
∞
 
0 ∞
 1
|wν |p < ∞ and n < ∞.
ν=1 |wν |
p
n=0 ν=−n+1 n=1

In particular, a symmetric weight (that is, one with w−n = wn for all n ≥ 0)
never defines a hypercyclic weighted shift Bw on these spaces.
As a concrete example, the weight
 
w = . . . , 12 , 12 , 12 , 2, 2, 2, . . .
4.1 Weighted shifts 103

induces a chaotic weighted backward shift on each p (Z).

One reason for studying shifts is that they provide a rich source of exam-
ples. As a first illustration we construct a hypercyclic operator whose adjoint
is also hypercyclic.
Proposition 4.16. There exists an operator T on 2 (Z) such that T and its
adjoint T ∗ are weakly mixing, and hence hypercyclic.
Proof. As usual, we identify the dual of 2 (Z) with itself; indeed, every con-
tinuous linear functional x∗ on 2 (Z) is of the form

x∗ (x) = x, x∗  = xn yn , (xn )n∈Z ∈ 2 (Z)
n∈Z

for a suitable sequence y = (yn )n∈Z ∈ 2 (Z).

Now let T = Bw be a bilateral shift. It defines an operator on 2 (Z) if and
only if the wn , n ∈ Z, are bounded. Since
 
Bw x, y = wn+1 xn+1 yn = xn wn yn−1 = x, F(wn+1 ) y,
n∈Z n∈Z

we see that the adjoint T ∗ = Bw

∗
of Bw is the forward shift F(wn+1 ) .
When we define

n −1 
0
vn = wν (n ≥ 1), vn = wν (n ≤ −1), v0 = 1,
ν=1 ν=n+1

then Theorem 4.13 and Remark 4.14 tell us that Bw and F(wn+1 ) are weakly
mixing if and only if there are increasing sequences (nk )k and (mk )k of pos-
itive integers such that, for any j ∈ Z,

vj−nk → 0, vj+nk → 0,
vj−mk → ∞, vj+mk → ∞,

and the continuity of Bw requires that vn /vn+1 , n ∈ Z, is bounded. But such

a sequence is easy to find: we choose the symmetric sequence (vn )n∈Z with
 
(vn )n≥0 = 1, 1, 2, 1, 12 , 1, 2, 4, 2, 1, 12 , 14 , 12 , 1, 2, 4, 8, 4, 2, 1, 12 , 14 , 18 , 14 , . . . ,

and the nk are the indices of the local minima, the mk the indices of the local
maxima of this sequence. Note that Bw is even invertible.  
Remark 4.17. This proposition provides us with an example of two weakly
mixing, hence hypercyclic operators S, T : X → X whose direct sum S ⊕ T
is not hypercyclic.
We show more generally that for any operator T on a Banach space X
the operator T ⊕ T ∗ cannot be hypercyclic on X ⊕ X ∗ . Indeed, suppose
104 4 Classes of hypercyclic and chaotic operators

that (x, x∗ ) is a hypercyclic vector for T ⊕ T ∗ . If we consider −x ∈ X as a

continuous linear functional on X ∗ then we have for n ≥ 0 that

(T n x, (T ∗ )n x∗ ), (x∗ , −x) = T n x, x∗  − x, (T ∗ )n x∗  = 0,

which is impossible since the left-hand side must be dense in K; note that
(x∗ , −x) cannot be the zero vector.

It was this observation that motivated Herrero’s problem if T ⊕ T is hy-

percyclic whenever T is; see Section 2.5.

4.2 Differential operators

As the last section demonstrates, Rolewicz’s result on the hypercyclicity of

multiples of the backward shift has seen far-reaching generalizations. Let us
turn, in the same spirit, to Birkhoff’s theorem and MacLane’s theorem. At
first glance, the operators

Df (z) = f  (z) and Ta f (z) = f (z + a), a ∈ C,

on the space H(C) of entire functions have little in common. But there is a
surprisingly simple connection. Since
∞ ∞
f (n) (z) n  an Dn f
f (z + a) = a = (z)
n=0
n! n=0
n!

we have, at least formally, that

Ta = eaD .

In fact, this representation can be justified rigorously. We will need the fol-
lowing notion from complex analysis: an entire function ϕ is said to be of
exponential type if there are constants M, A > 0 such that

|ϕ(z)| ≤ M eA|z|for all z ∈ C. (4.7)

∞
Lemma 4.18. An entire function ϕ(z) = n=0 an z n is of exponential type
if and only if there are M, R > 0 such that, for n ≥ 0,
Rn
|an | ≤ M . (4.8)
n!
Proof. On the one hand, if (4.7) holds, then by the Cauchy estimates we have
for any ρ > 0 that
4.2 Differential operators 105
 ϕ(n) (0) 
  1 M
|an | =   ≤ n max |ϕ(z)| ≤ n eAρ .
n! ρ |z|≤ρ ρ

Setting ρ = n/A and using Stirling’s formula we get, with some C > 0,
√ n
M An n nA (2A)n
|an | ≤ n
e ≤ CM ≤ CM .
n n! n!
Conversely, if (4.8) holds then
∞
 ∞ n
(R|z|)
|ϕ(z)| ≤ |an z n | ≤ M = M eR|z| ,
n=0 n=0
n!

so that ϕ is of exponential type. 



Proposition 4.19. Let

∞

ϕ(z) = an z n
n=0

be an entire function of exponential type. Then

∞

ϕ(D)f = an D n f
n=0

converges in H(C) for every entire function f and defines an operator on

H(C).

Proof. Let f ∈ H(C) and |z| ≤ m. By the Cauchy estimates and Lemma 4.18
there are M, R > 0 such that
   n
an f (n) (z) ≤ |an | n! max |f (ζ)| ≤ M R max |f (ζ)|. (4.9)
mn |ζ|≤2m m |ζ|≤2m
∞
Therefore, if m > R then n=0 an f
(n)
(z) converges uniformly on |z| ≤ m.
Hence
∞

ϕ(D)f = an D n f
n=0

converges in H(C). Moreover, by (4.9), writing pm (f ) = max|z|≤m |f (z)|, we

have for m > R that
1
pm (ϕ(D)f ) ≤ M p2m (f ).
1 − R/m

This shows that ϕ(D) is an operator on H(C); see Proposition 2.11. 



We will call the operators ϕ(D) simply differential operators on H(C).

They include all finite-order differential operators
106 4 Classes of hypercyclic and chaotic operators

T = a0 I + a1 D + . . . + am Dm .

Proposition 4.19, in particular, justifies our earlier calculation concerning

Birkhoff’s operators that

Ta = ϕ(D) with ϕ(z) = eaz . (4.10)

The following result gives a useful description of the differential operators

ϕ(D) among the operators on H(C).

Proposition 4.20. Let T be an operator on H(C). Then the following as-

sertions are equivalent:
(i) T = ϕ(D) for some entire function ϕ of exponential type;
(ii) T commutes with D;
(iii) T commutes with each Ta , a ∈ C.

Proof. (i)=⇒(ii). Let T = ϕ(D). By the continuity of D we have for f ∈ H(C)

∞
 ∞

T Df = an Dn (Df ) = D(an Dn f ) = DT f.
n=0 n=0

(ii)=⇒(iii). By the same token, using (4.10), we obtain for f ∈ H(C) that

∞ ∞ ∞
an n an an n
T Ta f = T D f= T Dn f = D (T f ) = Ta T f.
n=0
n! n=0
n! n=0
n!

(iii)=⇒(i). By continuity of f → (T f )(0) there is some M > 0 and some

R ∈ N such that

|(T f )(0)| ≤ M max |f (z)|, f ∈ H(C).

|z|≤R

Denoting by en , n ≥ 0, the monomials en (z) = z n we define

(T en )(0)
an =
n!
and deduce that
Rn
|an | ≤ M .
n!
∞
It follows from Lemma 4.18 and Proposition 4.19 that ϕ(z) =  n=0 an z n
∞
defines an entire function ϕ of exponential type and that ϕ(D) = n=0 an Dn
defines an operator on H(C). Then

(ϕ(D)en )(0) = an n! = (T en )(0), n ≥ 0.

Since the monomials span a dense subspace of H(C), we obtain that

4.2 Differential operators 107

(ϕ(D)f )(0) = (T f )(0) for f ∈ H(C).

By what we have shown above we also know that ϕ(D) commutes with each
Ta . Thus we get with (iii) for any z ∈ C and f ∈ H(C), using the definition
of Tz ,

(ϕ(D)f )(z) = (Tz ϕ(D)f )(0) = (ϕ(D)Tz f )(0)

= (T Tz f )(0) = (Tz T f )(0) = T f (z),

so that T = ϕ(D). 

With this in hand we can prove a remarkably general common extension
of the theorems of Birkhoff and MacLane.
Theorem 4.21 (Godefroy–Shapiro). Suppose that T : H(C) → H(C),
T = λI, is an operator that commutes with D, that is,

T D = DT.

Then T is mixing and chaotic.

Proof. By Proposition 4.20 we can write T = ϕ(D) for some entire function
∞

ϕ(z) = an z n
n=0

of exponential type. Our additional assumption implies that ϕ is nonconstant.

It is now easy to verify that T satisfies the conditions of the Godefroy–Shapiro
criterion. In fact, considering the exponential functions

eλ (z) = eλz , λ ∈ C,

we calculate that
∞

T eλ = ϕ(D)eλ = an λn eλ = ϕ(λ)eλ .
n=0

Thus each eλ is an eigenvector of T to the eigenvalue ϕ(λ). Consequently,

span{f ∈ H(C) ; T f = μf for some μ ∈ C with |μ| < 1}

contains span{eλ ; |ϕ(λ)| < 1}, which is dense in H(C) by Lemma 2.34;
indeed, since any nonconstant entire function has dense range (see Appendix
A), {λ ∈ C ; |ϕ(λ)| < 1} is a nonempty open set and therefore has an
accumulation point. For the same reason, the eigenvectors of T to eigenvalues
μ with |μ| > 1 span a dense set in H(C). For the density of

span{f ∈ H(C) ; T f = eαπi f for some α ∈ Q}

108 4 Classes of hypercyclic and chaotic operators

it suffices to observe that also the set {λ ∈ C ; ϕ(λ) = eαπi for some α ∈ Q}
has an accumulation point. Indeed, since ϕ(C) is connected and dense, it
must intersect the unit circle. And since nonconstant holomorphic functions
are open mappings, infinitely many preimages under ϕ of roots of unity lie
in some bounded subset of C and therefore have an accumulation point.  
Having established the hypercyclicity of every differential operator T =
ϕ(D) = λI we now want to focus our attention on properties of the cor-
responding hypercyclic functions. MacLane had already addressed such a
problem: he showed that there exists a D-hypercyclic entire function f of
exponential type 1, which means that for every ε > 0 there is some M > 0
such that
|f (z)| ≤ M e(1+ε)r for all z ∈ C.
Here we follow the usual convention of writing r = |z|. MacLane’s growth
condition can be improved, and one can even determine the least possible
rate of growth.
Theorem 4.22. (a) Let φ : ]0, ∞[ → [1, ∞[ be a function with φ(r) → ∞ as
r → ∞. Then there exists an entire function f that is hypercyclic for D and
that satisfies
er
|f (z)| ≤ M φ(r) √ for |z| = r > 0
r
with some M > 0.
(b) There is no entire function f that is hypercyclic for D and that satisfies

er
|f (z)| ≤ M √ for |z| = r > 0
r

with some M > 0.

Proof. (a) The assertion suggests consideration of the space
 √ 
r |f (z)|
X = f ∈ H(C) ; f := sup r
<∞ .
r=|z|>0 φ(r)e

Proving our assertion then amounts to showing that the sequence of operators

Tn : X → H(C), f → f (n) , n≥0

admits a dense orbit in the sense of Section 3.4; note that the Tn are indeed
operators because the inclusion map X → H(C) is obviously continuous.
To prove that (Tn )n is hypercyclic we apply the Hypercyclicity Criterion
for sequences of operators, Theorem 3.24.
It is an easy exercise to see that X is a Banach space. We would like
to take as X0 the set of polynomials, but we cannot guarantee that the
polynomials are dense in X. Thus we replace X by the closure X 0 of X0 in
X. Clearly Tn f → 0 for any f ∈ X0 . For Y0 we take the set of polynomials
4.2 Differential operators 109

in H(C), and we define Sn = S n : Y0 → X 0 using the antiderivative operator

z
Sf (z) = 0 f (ζ)dζ. Then Tn Sn f = f, n ∈ N0 , for any f ∈ Y0 .
It remains to show that Sn f → 0 in X 0 for any polynomial f . By linearity
we may assume that f is a monomial en (z) = z n , and because Sn ek =
k!
(n+k)! en+k = k!Sn+k e0 it suffices to consider f = e0 . For this we find that

e  r n+1/2
 n
Sn e0 =   = sup r
.
n! r>0 n!φ(r)e

A simple calculation shows that

r n+1/2 (n + 1/2)n+1/2
sup = ,
r>0 n!er n!en+1/2

and Stirling’s formula implies that this is bounded in n ≥ 0 by some constant

C. Fixing ε > 0, and letting R > 0 be such that φ(r) > 1/ε for r ≥ R we
obtain that
Rn+1/2 r n+1/2 ε Rn+1/2
Sn e0 ≤ + sup r
≤ + Cε,
n! r≥R n!e n!

which implies that Sn e0 → 0 in X and therefore in X 0 .

(b) Let f ∈ H(C). Under the assumed growth condition we have by the
Cauchy estimates that, for any n ∈ N0 and ρ > 0,
 (n) 
f (0) ≤ n! max |f (z)| ≤ M n!
√ eρ .
ρn |z|≤ρ ρn ρ

Choosing ρ = n we get
 (n)  n!
f (0) ≤ M en ,
nn+1/2
which is bounded by Stirling’s formula. Thus, f cannot be hypercyclic for D.


√
Exercise 4.2.5 explains how the critical rate of growth er / r is related to
the differentiation operator.
In contrast to the result for MacLane’s operator, entire functions that are
hypercyclic for Birkhoff’s operators can grow arbitrarily slowly. The proof
requires a different technique and will be provided in Chapter 8; see Exercise
8.1.3.
Theorem 4.23 (Duyos-Ruiz). Let a = 0. Let φ : ]0, ∞[ → [1, ∞[ be a
function so that, for any N ≥ 1, φ(r)/rN → ∞. Then there exists an entire
function f that is hypercyclic for Ta and that satisfies

|f (z)| ≤ M φ(r) for |z| = r > 0

110 4 Classes of hypercyclic and chaotic operators

with some M > 0.

By the method used in the proof of Theorem 4.21 one can derive certain
possible rates of growth for an arbitrary operator ϕ(D); see Exercise 4.2.4.

4.3 Composition operators I

As we have seen, operators may often be interpreted in various ways.

MacLane’s operator is both a differential operator and a weighted shift.
Birkhoff’s operators are differential operators as well. Here now we have an-
other interpretation of Birkhoff’s operators Ta : they are special composition
operators. Writing
τa (z) = z + a
we see that τa is an entire function such that

Ta f = f ◦ τ a .

In fact, τa is even an automorphism of C, that is, a bijective entire function.

These observations serve as the starting point of another major investigation:
the hypercyclicity of general composition operators.
Let Ω be an arbitrary domain in C, that is, a nonempty connected open
set. An automorphism of Ω is a bijective holomorphic function

ϕ : Ω → Ω;

its inverse is then also holomorphic. The set of all automorphisms of Ω is

denoted by Aut(Ω). Now, for ϕ ∈ Aut(Ω) the corresponding composition
operator is defined as
Cϕ f = f ◦ ϕ,
that is, (Cϕ f )(z) = f (ϕ(z)), z ∈ Ω.
What about the underlying space? Following Birkhoff we consider the
space H(Ω) of all holomorphic functions on Ω which we endow, as in the
case Ω = C, with the topology of local uniform convergence. To describe this
topology by seminorms we need an exhaustion of Ω by compact sets, that is,
an increasing sequence of compact sets Kn ⊂ Ω such that each compact set
K ⊂ Ω is contained in some Kn .

Lemma 4.24. Every domain Ω ⊂ C has an exhaustion of compact sets.

Proof. For each n ∈ N we consider the grid of all points x + iy in C so that

either x or y is an integer multiple of 21n ; then let Kn be the (finite) union of
all closed squares that have their sides lying on the grid and that lie entirely
in Ω ∩ {z : |z| < n}. It is obvious that (Kn )n is an exhaustion of Ω.  
4.3 Composition operators I 111

Now, if (Kn )n is an exhaustion of Ω then we endow H(Ω) with the topol-

ogy induced by the seminorms

pn (f ) = sup |f (z)|, n ∈ N.
z∈Kn

In this way H(Ω) turns into a Fréchet space; note that the topology is inde-
pendent of the chosen exhaustion. Moreover, by Runge’s theorem, H(Ω) is
separable; see Exercise 4.3.1.
Clearly, for any automorphism ϕ of Ω the composition operator Cϕ is con-
tinuous on H(Ω). Let us first note that conformal maps, that is, holomorphic
bijections between two domains, induce conjugacies between the correspond-
ing composition operators; the proof is immediate.

Proposition 4.25. Let Ω1 and Ω2 be domains in C and ψ : Ω1 → Ω2 a

conformal map. If ϕ1 and ϕ2 are automorphisms of Ω1 and Ω2 , respectively,
such that ϕ2 ◦ ψ = ψ ◦ ϕ1 then Cϕ2 and Cϕ1 are conjugate via the map
J : H(Ω2 ) → H(Ω1 ), f → f ◦ ψ, that is, the diagram
Cϕ
H(Ω2 ) −−−−
2
→ H(Ω2 )
⏐ ⏐
⏐ ⏐
J J
Cϕ
H(Ω1 ) −−−−
1
→ H(Ω1 )
commutes.

Example 4.26. Any two Birkhoff operators Ta , Tb , a, b = 0, are conjugate.

This follows immediately by taking ψ(z) = ab z, z ∈ C, since τb ◦ ψ = ψ ◦ τa .

We turn to the problem of determining which composition operators are

hypercyclic. The crucial concept will be the notion of a run-away sequence.

Definition 4.27. Let Ω be a domain in C and ϕn : Ω → Ω, n ≥ 1, holo-

morphic maps. Then the sequence (ϕn )n is called a run-away sequence if, for
any compact subset K ⊂ Ω, there is some n ∈ N such that

ϕn (K) ∩ K = ∅.

We will usually apply this definition to the sequence (ϕn )n of iterates of

an automorphism ϕ on Ω. Let us consider two examples. Another important
example will be studied below; see Proposition 4.36.

Example 4.28. (a) Let Ω = C. Then the automorphisms of C are the functions

ϕ(z) = az + b, a = 0, b ∈ C,

and (ϕn )n is run-away if and only if a = 1, b = 0.

112 4 Classes of hypercyclic and chaotic operators

Indeed, let ϕ be an automorphism of C. If ϕ is not a polynomial then,

by the Casorati–Weierstrass theorem, ϕ({z ∈ C ; |z| > 1}) is dense in C
and therefore intersects the set ϕ(D), which is open by the open mapping
theorem. Since this contradicts injectivity, ϕ must be a polynomial. Again
by injectivity, its degree must be one, so that ϕ is of the stated form. Now,
if a = 1 then ϕn (z) = z + nb, so that we have the run-away property if and
only if b = 0; while if a = 1 then (1 − a)−1 b is a fixed point of ϕ so that (ϕn )n
cannot be run-away.
(b) Let Ω = C∗ = C \ {0}, the punctured plane. An argument as in (a)
shows that the automorphisms of C∗ are the functions
a
ϕ(z) = az or ϕ(z) = , a = 0.
z
Then (ϕn )n is run-away if and only if ϕ(z) = az with |a| = 1.

We first show that the run-away property is a necessary condition for the
hypercyclicity of the composition operator.

Proposition 4.29. Let Ω be a domain in C and ϕ ∈ Aut(Ω). If Cϕ is hy-

percyclic then (ϕn )n is a run-away sequence.

Proof. If (ϕn )n is not run-away then there exists a compact set K ⊂ Ω and
elements zn ∈ K such that

ϕn (zn ) ∈ K, n ∈ N. (4.11)

Now suppose that f ∈ H(Ω) is a hypercyclic vector for Cϕ . Let M =

supz∈K |f (z)|. Then, by (4.11), we have that

inf |((Cϕ )n f )(z)| ≤ |((Cϕ )n f )(zn )| = |f (ϕn (zn ))| ≤ M,

z∈K

so that the functions (Cϕ )n f cannot approximate, for example, the constant
function M + 1 uniformly on K, a contradiction.  

Corollary 4.30. There is no automorphism of C∗ whose composition oper-

ator is hypercyclic.

Proof. By Proposition 4.29 and Example 4.28(b), Cϕ can only be hypercyclic

on H(C∗ ) if ϕ(z) = az with |a| = 1. Suppose that f ∈ H(C∗ ) is hypercyclic
n
for such a function ϕ. If f (z) = n∈Z cn z then
  
 c−1 
((Cϕ )n f − z1 ) dz = f (an z) dz − z dz = 2πi an − 1 ,
1
T T T

where the unit circle T is positively oriented. By hypercyclicity, there is a

sequence (nk )k for which the left-hand side converges to zero, unlike the
right-hand side, which is a contradiction. 
4.3 Composition operators I 113

Thus, when Ω is C∗ then the run-away property is not a sufficient condition

for hypercyclicity. Our goal now is to show that in essentially all other cases,
hypercyclicity is characterized by the run-away property. To this end we need
to introduce some topological properties of plane sets.
We denote by C  = C∪{∞} the one-point compactification of C. A domain

Ω is called simply connected if C\Ω is connected. A domain Ω is called finitely
connected if C \ Ω contains at most finitely many connected components,
otherwise it is infinitely connected. If M is any set in C then one also speaks
of a bounded component of C  \ M as a hole. In that sense, finitely connected
domains have only finitely many holes, a simply connected domain has no
hole.
We first deal with finitely, not simply connected domains Ω. One can show
that unless such a domain is conformally equivalent to C∗ , that is, unless
there is a holomorphic bijection between Ω and C∗ , Ω does not admit an
automorphism ϕ so that (ϕn )n is run-away; we omit the proof. By Proposition
4.29 and Corollary 4.30 we therefore have the following.
Proposition 4.31. Let Ω be a finitely connected but not simply connected
domain in C. Then Cϕ is not hypercyclic for any automorphism of Ω.
In all other cases we have the following.
Theorem 4.32. Let Ω be a domain in C that is either simply connected or
infinitely connected. Let ϕ ∈ Aut(Ω). Then Cϕ is hypercyclic if and only if
(ϕn )n is a run-away sequence.
In view of Proposition 4.29 we only have to prove sufficiency of the run-
away property. For this we need to study the geometry of domains more
closely, at least for infinitely connected domains. A compact subset K of a
domain Ω will be called Ω-convex if every hole of K contains a point of C\Ω;
see Figure 4.1. Of course, in a simply connected domain, Ω-convexity only
says that K has no holes, in other words, that its complement is connected.

Fig. 4.1 An Ω-convex set K Fig. 4.2 ϕn (K) ∪ K is Ω-convex

The following auxiliary result will be crucial for the proof of sufficiency in
Theorem 4.32. However, since its proof is rather technical we will postpone it
to the end of the section. For later use we formulate the lemma for arbitrary
sequences (ϕn )n of automorphisms.
114 4 Classes of hypercyclic and chaotic operators

Lemma 4.33. Let Ω be an infinitely connected domain in C and (ϕn )n a

run-away sequence of automorphisms of Ω. Then every compact subset of
Ω is contained in some Ω-convex compact subset K of Ω for which there is
some n ∈ N such that ϕn (K) ∩ K = ∅ and ϕn (K) ∪ K is Ω-convex.
Proof of Theorem 4.32 (sufficiency). Suppose that (ϕn )n is a run-away se-
quence. We want to show that then Cϕ is topologically transitive. Let
f, g ∈ H(Ω), let L be a compact subset of Ω and ε > 0. Then there is a
compact subset K of Ω containing L and an n ∈ N such that ϕn (K) ∩ K = ∅
and ϕn (K) ∪ K is Ω-convex (see Figure 4.2); in the simply connected case
one can take any Ω-convex compact set K containing L, in the infinitely
connected case one applies Lemma 4.33. Then the function g ◦ (ϕn )−1 is
holomorphic on some neighbourhood of ϕn (K), and f is holomorphic on
some neighbourhood of K. It follows from Runge’s theorem that there is a
function h ∈ H(Ω) such that
 
sup |f (z) − h(z)| < ε and sup g ◦ (ϕn )−1 (z) − h(z) < ε,
z∈K z∈ϕn (K)

hence

sup |f (z) − h(z)| < ε and sup |g(z) − h(ϕn (z))| < ε.
z∈L z∈L

As in Example 2.20 this implies that Cϕ is topologically transitive. 



Fig. 4.3 The set Ω (Example 4.34)

Example 4.34. We give an example of a hypercyclic composition operator

on an infinitely connected domain. We start with the unit disk D and the
automorphism
z − 12
ϕ(z) =
1 − 12 z
of D; see also Proposition 4.36. Let A = {z : |z| ≤ 10
1
}. It is easy to see that
the forward and backward iterates ϕ (A), n ∈ Z, of A are pairwise disjoint.
n

Then
4.3 Composition operators I 115

Ω := D \ ϕn (A)
n∈Z

is an infinitely connected domain (see Figure 4.3), and the restriction of ϕ to

Ω is an automorphism of Ω. Moreover, a simple calculation shows that
z − an 3n − 1
ϕn (z) = with an = , n ≥ 0,
1 − an z 3n + 1
so that limn→∞ ϕn (z) = −1, uniformly on compact subsets of Ω. Hence,
(ϕn )n is a run-away sequence on Ω, which implies that Cϕ is hypercyclic on
H(Ω).

Remark 4.35. Any hypercyclic composition operator Cϕ on a domain Ω is

even weakly mixing. To see this, let (Kn )n be an exhaustion of Ω by compact
n
m)1n is krun-away, there is some m1 such that
sets. Since (ϕ ϕm1 (K1 ) ∩ K1 = ∅.
If L = K2 ∪ k=1 ϕ (K1 ), there is some m2 such that ϕ (L) ∩ L = ∅. Then,
m2

in particular, ϕm2 (K2 ) ∩ K2 = ∅; moreover, since ϕm2 (L) contains ϕm2 (K1 )
and L contains ϕk (K1 ) for k = 1, . . . , m1 , we must have that m2 > m1 .
Proceeding inductively we obtain a strictly increasing sequence (mn )n such
that ϕmn (Kn ) ∩ Kn = ∅, for any n ∈ N; as a consequence, (ϕmn )n and any
of its subsequences is run-away. The proofs in this section then show that
every subsequence of (Cϕmn )n admits a dense orbit. This tells us that Cϕ is
hereditarily hypercyclic, and hence weakly mixing by Theorem 3.15.

We want to study the case of simply connected domains in greater detail.

If Ω = C, the automorphisms are given by

ϕ(z) = az + b, a = 0, b ∈ C,

and Cϕ is hypercyclic if and only if a = 1, b = 0; see Example 4.28(a) and

Theorem 4.32. Thus the hypercyclic composition operators on C are precisely
Birkhoff’s translation operators.
Let us now consider the simply connected domains Ω other than C. By
the Riemann mapping theorem, Ω is conformally equivalent to the unit disk,
that is, there is a conformal map ψ : D → Ω. By Proposition 4.25 it suffices
to study the case when Ω = D.

Proposition 4.36. The automorphisms of D are the linear fractional trans-

formations
a−z
ϕ(z) = b , |a| < 1, |b| = 1.
1 − az
Moreover, ϕ maps T bijectively onto itself.

Proof. We first consider the maps

a−z
ha (z) = , |a| < 1.
1 − az
116 4 Classes of hypercyclic and chaotic operators

A simple calculation shows that, for w = ha (z),

1 − |a|2
1 − |w|2 = (1 − |z|2 ). (4.12)
|1 − az|2

Hence D and T are invariant under ha . Moreover one finds that ha ◦ ha = I

on D. This implies that ha is an automorphism of D that maps T bijectively
onto itself. The same is then true for bha , |b| = 1.
Conversely, let ϕ be an automorphism of D, and let 0 = ϕ(a) with |a| < 1.
Then the map f := ϕ ◦ h−1a is also an automorphism of D with f (0) = 0. The
Schwarz lemma then implies that |f (z)| ≤ |z| for z ∈ D. The same argument
applied to the inverse of f shows that |f −1 (z)| ≤ |z|, hence |z| ≤ |f (z)| for
z ∈ D. Altogether we have that |f (z)| = |z| for z ∈ D. Again by the Schwarz
lemma, f can only be a rotation, that is, there is some b with |b| = 1 such
that f (z) = bz and therefore ϕ = bha . 

Now, linear fractional transformations are a very well understood class

of holomorphic maps; see Appendix A. Using their properties it is not diffi-
cult to determine the dynamical behaviour of the corresponding composition
operators; via conjugacy these results can then be carried over to arbitrary
simply connected domains.

Theorem 4.37. Let Ω be a simply connected domain and ϕ ∈ Aut(Ω). Then

the following assertions are equivalent:
(i) Cϕ is hypercyclic;
(ii) Cϕ is mixing;
(iii) Cϕ is chaotic;
(iv) (ϕn )n is a run-away sequence;
(v) ϕ has no fixed point in Ω;
(vi) Cϕ is quasiconjugate to a Birkhoff operator.

Proof. The implications (vi)=⇒(iii) and (vi)=⇒(ii) follow from known prop-
erties of the Birkhoff operators, (iii)=⇒(i) and (ii)=⇒(i) hold for all operators
on H(Ω), and (i)⇐⇒(iv) was proved in Theorem 4.32.
(i)=⇒(v). If ϕ has a fixed point z0 ∈ Ω then, for any f ∈ H(Ω) and n ≥ 0,
((Cϕ )n f )(z0 ) = f (ϕn (z0 )) = f (z0 ), so that f cannot have a dense orbit.
It remains to prove that (v)=⇒(vi). In the case Ω = C the result was
shown in Example 4.28(a). In the case Ω = C we can assume by the discussion
leading up to Proposition 4.36 that Ω = D. The proof then requires certain
properties of linear fractional transformations. Since we will have occasion
to study them in Section 4.5 we will postpone the proof to the end of that
section.  

In particular the final condition in the theorem is of great interest. Any

property of the Birkhoff operators that is preserved under quasiconjugacies
will transmit to all composition operators on simply connected domains.
It remains to give the proof of Lemma 4.33.
4.3 Composition operators I 117

Proof of Lemma 4.33. We consider the exhaustion of Ω by compact sets

Kn constructed in the proof of Lemma 4.24. Then each Kn is automatically
Ω-convex. But also ψ(Kn ) is Ω-convex for every automorphism ψ. Indeed,
if some hole of ψ(Kn ) contained only points of Ω then one could deform
the boundary of that hole continuously in Ω to a point in Ω; applying the
map ψ −1 , the same would then be true for the corresponding hole of Kn ,
contradicting the Ω-convexity of Kn .
By the run-away property, we can find a strictly increasing sequence (mn )n
of positive integers such that ϕmn (Kn ) ∩ Kn = ∅ for n ≥ 1; see Remark 4.35.
Now, every compact subset of Ω is contained in some KN , and since Ω is
infinitely connected we can assume that KN has at least two holes. Then, for
all n ≥ N , ϕmn (KN ) ∩ KN = ∅. Also, each ϕmn (KN ) is Ω-convex. To finish
the proof it suffices to show that there is some n ≥ N such that, in addition,
ϕmn (KN ) ∪ KN is Ω-convex.
We distinguish three cases. First, if there is some n ≥ N such that
ϕmn (KN ) lies in the unbounded component of the complement of KN and
KN lies in the unbounded component of the complement of ϕmn (KN ) then
clearly ϕmn (KN ) ∪ KN is Ω-convex.

Fig. 4.4 ϕmN +1 (KN ) ∪ KN is Ω-convex Fig. 4.5 Both ϕmN +1 (KN ) ∪ KN and
ϕmν (KN ) ∪ KN are Ω-convex

Secondly, suppose that infinitely many ϕmn (KN ), n ≥ N , lie in holes of

KN . Since KN only has a finite number of holes, infinitely many ϕmn (KN ),
n ≥ N , must lie in some fixed hole O of KN ; by passing to a subsequence we
may assume that this is true for all n > N . We then choose some ν > N such
that ϕmN +1 (KN ) ⊂ Kν . Since ϕmν (Kν ) ∩ Kν = ∅ we have that ϕmN +1 (KN )
and ϕmν (KN ) are disjoint subsets of O. Now one has three possibilities: either
ϕmN +1 (KN ) lies in a hole of ϕmν (KN ) (see Figure 4.4), or ϕmν (KN ) lies in
a hole of ϕmN +1 (KN ), or both sets lie in the unbounded component of the
complement of the other set (see Figure 4.5). Since both sets have at least
two holes one finds that in each of these cases either ϕmN +1 (KN ) ∪ KN or
ϕmν (KN ) ∪ KN is Ω-convex.
118 4 Classes of hypercyclic and chaotic operators

Finally, the remaining case is when for infinitely many n ≥ N , KN lies in

a hole of ϕmn (KN ). Again we can assume that this is true for all n > N . We
then choose some ν > N such that ϕmN +1 (KN ) ⊂ Kν . As above we find that
ϕmN +1 (KN ) and ϕmν (KN ) are disjoint sets. Since both these sets contain
KN in one of their holes, we have that either ϕmN +1 (KN ) lies in a hole of
ϕmν (KN ), or vice versa. Since both sets have two holes we find that either
ϕmν (KN ) ∪ KN or ϕmN +1 (KN ) ∪ KN is Ω-convex.  

4.4 Adjoint multipliers

In this section we consider an interesting generalization of the backward shift

operator. The underlying space will be the Hardy space H 2 . Arguably its
easiest definition is the following. If (an )n≥0 is a complex sequence such that
∞

|an |2 < ∞,
n=0

then it is, in particular, bounded, and hence

∞

f (z) = an z n , z ∈ C, |z| < 1,
n=0

defines a holomorphic function on the complex unit disk D. The Hardy space
is then defined as the space of these functions, that is,
 ∞
 ∞
 
H 2 = f : D → C ; f (z) = an z n , z ∈ D, with |an |2 < ∞ .
n=0 n=0

In other words, the Hardy space is simply the sequence space 2 (N0 ), with
its elements written as holomorphic functions. It is then clear that H 2 is a
Banach space under the norm

∞ 1/2 ∞

f = |an |2 when f (z) = an z n ,
n=0 n=0

and it is even a Hilbert space under the inner product

∞
 ∞
 ∞

f, g = an bn when f (z) = an z n , g(z) = bn z n .
n=0 n=0 n=0

The polynomials form a dense subspace of H 2 .

The following result is an immediate consequence of the definitions.
4.4 Adjoint multipliers 119

Proposition 4.38. For any λ ∈ D define kλ : D → C by

∞
 n 1
kλ (z) = λ zn = .
n=0
1 − λz

Then kλ ∈ H 2 and, for any f ∈ H 2 ,

f (λ) = f, kλ .

This implies that for any λ ∈ D the point evaluation

f → f (λ)

is a continuous linear functional on H 2 . The functions kλ , λ ∈ D, are called

reproducing kernels. They will play the same role here as the exponential
functions eλ ∈ H(C), λ ∈ C, in Section 4.2. In particular we have the following
analogue of Lemma 2.34.

Lemma 4.39. Let Λ ⊂ D be a set with an accumulation point in D. Then

the set
span{kλ ; λ ∈ Λ}
is dense in H 2 .

Proof. It suffices to show that only the zero function can be orthogonal to
span{kλ ; λ ∈ Λ}. But that is immediate by the identity theorem for holo-
morphic functions: if, for f ∈ H 2 , f, kλ  = f (λ) vanishes for all λ ∈ Λ, then
f = 0.  

The operators that we want to study are those that map f ∈ H 2 to ϕf ,

where ϕ is a bounded holomorphic function on D. In order to see that this
defines an operator on H 2 we need another representation of the space.

Proposition 4.40. A holomorphic function f : D → C belongs to H 2 if and

only if  2π
sup |f (reit )|2 dt < ∞.
0≤r<1 0

Moreover, for any f, g ∈ H , 2

  2π 1/2   2π 1/2
1 1
f = sup |f (reit )|2 dt = lim |f (reit )|2 dt
0≤r<1 2π 0 r
1 2π 0

and  2π
1
f, g = lim f (reit )g(reit ) dt.
r
1 2π 0
∞
Proof. Writing f (z) = n=0 an z n we obtain that
120 4 Classes of hypercyclic and chaotic operators
 2π  2π   ∞ 2
1 1  
|f (reit )|2 dt =  an (reit )n  dt
2π 0 2π 0 n=0
 2π  ∞ 2 ∞
 n 1 int 
=  an r √ e  dt = |an |2 r2n ,
0 n=0
2π n=0

where we have used Parseval’s identity in L2 [0, 2π] for the orthonormal basis
( √12π eint )n∈Z ; see also Example 4.4(b). Since

∞
 ∞
 ∞

sup |an |2 r2n = lim |an |2 r 2n = |an |2 = f 2 ,
0≤r<1 n=0 r
1
n=0 n=0

the first part of the assertion follows. In the same way, one obtains the second
part by using Parseval’s identity for the inner product.  
Now let ϕ be a bounded holomorphic function on D. Then, for any f ∈ H 2 ,
ϕf is holomorphic on D, and we have that
 2π  2π
1 1
sup |(ϕf )(reit )|2 dt ≤ sup |ϕ(z)|2 sup |f (reit )|2 dt,
0≤r<1 2π 0 z∈D 0≤r<1 2π 0

so that also ϕf ∈ H 2 by the previous proposition. Moreover, we see that

Mϕ f = ϕf

defines an operator on H 2 with Mϕ ≤ supz∈D |ϕ(z)|. The function ϕ is

called a multiplier of H 2 , Mϕ is called the corresponding multiplication op-
erator or briefly multiplier.
Clearly, multiplication operators are never hypercyclic. For if (Mϕ )n f =
ϕ f , n ≥ 0, formed a dense set in H 2 then, by continuity of point evaluations,
n

the same would be true of the sequence (ϕ(0)n f (0))n≥0 in C, which is never
the case. Instead, we will consider the (Hilbert space) adjoint Mϕ∗ : H 2 → H 2
of Mϕ , called an adjoint multiplication operator or adjoint multiplier.
In fact, we already know that some operators Mϕ∗ are hypercyclic, as we will
2
see now. As usual, B and F denote the backward and forward shifts ∞on  (Nn0 ),
respectively, which are operators of norm 1. Hence,
∞ if ϕ(z) = n=0 an z is
∞
holomorphic on some neighbourhood of D then n=0 an B n ≤ n=0 |an | <
∞, so that
∞
ϕ(B) = an B n
n=0
∞
defines an operator on 2 (N0 ), and the same is true for ϕ(F ) = n=0 an F n ;
see also Appendix B.
∞
Proposition 4.41. Let ϕ(z) = n=0 an z n be holomorphic on a neighbour-
∞
hood of D, and set ϕ∗ (z) = n=0 an z n . Then, via the identification of H 2
with 2 (N0 ):
4.4 Adjoint multipliers 121

(i) the multiplier Mϕ corresponds to the operator ϕ(F ) on 2 (N0 );

(ii) the adjoint multiplier Mϕ∗ corresponds to the operator ϕ∗ (B) on 2 (N0 ).

Proof. (i) On the one hand we have that for f ∈ H 2 , f (z) = ∞ k
k=0 bk z ,

∞
 ∞
 ∞ 
 n 
Mϕ f (z) = an z n bk z k = an−k bk z n .
n=0 k=0 n=0 k=0

On the other hand we have for (bk )k≥0 ∈ 2 (N0 ),

∞
 ∞

ϕ(F )(bk )k = an F n (bk )k = an (0, . . . , 0, b0 , b1 , . . .)
n=0 n=0

n 
= (a0 b0 , a1 b0 + a0 b1 , a2 b0 + a1 b1 + a0 b2 , . . .) = an−k bk .
n≥0
k=0

This implies the result.

(ii) A simple calculation shows that Bis the adjoint ofF . Hence, by (i),
∞ ∞
the adjoint Mϕ∗ corresponds to ϕ(F )∗ = ( n=0 an F n )∗ = n=0 an B n , where
we have used properties of the adjoint; see Proposition A.8.  

In particular, the adjoint multipliers Mϕ∗ with ϕ(z) = λz, |λ| > 1, corre-
spond to the Rolewicz operators λB and are therefore hypercyclic.
The Godefroy–Shapiro criterion allows us to characterize the hypercyclic
adjoint multipliers. We can exclude constant multipliers because their adjoint
multiplication operators are multiples of the identity.

Theorem 4.42. Let ϕ be a nonconstant bounded holomorphic function on D

and let Mϕ∗ be the corresponding adjoint multiplier on H 2 . Then the following
assertions are equivalent:
(i) Mϕ∗ is hypercyclic;
(ii) Mϕ∗ is mixing;
(iii) Mϕ∗ is chaotic;
(iv) ϕ(D) ∩ T = ∅.

Proof. Suppose that condition (iv) holds. Considering the reproducing ker-
nels kλ , λ ∈ D, we find that, for all f ∈ H 2 ,

f, Mϕ∗ kλ  = ϕf, kλ  = (ϕf )(λ) = f, ϕ(λ) kλ ,

which shows that

Mϕ∗ kλ = ϕ(λ) kλ .
Consequently,

span{f ∈ H 2 ; Mϕ∗ f = μf for some μ ∈ C with |μ| < 1}

122 4 Classes of hypercyclic and chaotic operators

contains span{kλ ; |ϕ(λ)| < 1}, which is dense in H 2 by Lemma 4.39; indeed,
since nonconstant holomorphic functions are open mappings, condition (iv)
implies that {λ ∈ D ; |ϕ(λ)| < 1} is nonempty and open and therefore
contains an accumulation point in D. For the same reason the eigenvectors of
Mϕ∗ to eigenvalues of modulus greater than 1 span a dense set in H 2 . Finally,
the same is true for the eigenvectors of Mϕ∗ to eigenvalues that are roots
of unity. For this it suffices to show that {λ ∈ D ; ϕ(λ) is a root of unity}
has an accumulation point. But since ϕ is an open mapping, condition (iv)
implies that infinitely many preimages of roots of unity lie in some relatively
compact subset of D and therefore have an accumulation point in D. By the
Godefroy–Shapiro criterion, therefore, (iv) implies (ii) and (iii).
To finish the proof it suffices to show that (i) implies (iv). Let us suppose
that ϕ(D) does not intersect the unit circle. Since ϕ(D) is connected, it must
lie entirely inside or entirely outside D. If ϕ(D) ⊂ D then

Mϕ∗ = Mϕ ≤ sup |ϕ(z)| ≤ 1

z∈D

(see Proposition A.8), and hence Mϕ∗ cannot be hypercyclic. On the other
hand, if ϕ(D) ⊂ C \ D then ψ := 1/ϕ is a bounded holomorphic function on
D with ψ(D) ⊂ D, which implies that Mψ∗ cannot be hypercyclic. But Mϕ
is the inverse of Mψ and therefore Mϕ∗ is the inverse of Mψ∗ ; see Proposition
A.8. By Proposition 2.23, Mϕ∗ cannot be hypercyclic.  

This result can easily be extended to more general Hilbert spaces of holo-
morphic functions, for example the Bergman space (see Exercise 4.4.3); but
see also Exercise 4.4.4. An extension to some Banach spaces X of holomor-
phic functions is also possible, in which case, of course, Mϕ∗ is the Banach
space adjoint on the dual X ∗ ; see Exercise 4.4.5.
We pass on to another, partial generalization of Theorem 4.42 that is
motivated∞by Proposition 4.41. Under the assumptions of that proposition,
ϕ(B) = n=0 an B n defines an operator on each of the spaces p , 1 ≤ p < ∞,
and c0 .

Theorem 4.43. Let X be one of the complex sequence spaces p , 1 ≤ p <

∞, or c0 . Furthermore, let ϕ be a nonconstant holomorphic function on a
neighbourhood of D. Then the following assertions are equivalent:
(i) ϕ(B) is chaotic;
(ii) ϕ(D) ∩ T = ∅;
(iii) ϕ(B) has a nontrivial periodic point.

Proof. (ii)=⇒(i). We saw in Example 3.2 that any sequence

eλ := (λ, λ2 , λ3 , . . .), |λ| < 1

is an eigenvector of B to the eigenvalue λ and that, for any set Λ ⊂ D that

has an accumulation point in D, the set
4.5 Composition operators II 123

span{eλ ; λ ∈ Λ}

is dense in X. Now, for any λ ∈ D we have that

∞
 ∞

n
ϕ(B)eλ = an B eλ = an λn eλ = ϕ(λ)eλ ,
n=0 n=0

so that each eλ is also an eigenvector of ϕ(B) to the eigenvalue ϕ(λ). From

here we proceed exactly as in the proof of Theorem 4.42 to show that ϕ(B)
is chaotic.
(i)=⇒(iii) is trivial.
(iii)=⇒(ii). For this implication we need to use results from spectral theory;
see Appendix B. By condition (iii) there is some point x = 0 from X and
some N ≥ 1 such that ϕN (B)x = ϕ(B)N x = x. Thus, 1 ∈ σp (ϕN (B)), the
point spectrum of ϕN (B). It follows from the point spectral mapping theorem
(see Theorem B.7) that 1 = ϕN (λ) for some λ ∈ σp (B) = D; see Example
3.2. Hence ϕ(λ) ∈ T, which implies (ii).  

Example 4.44. The theorem shows in particular that any operator

I + λB and eλB , λ = 0

is hypercyclic (and even chaotic) on X = p , 1 ≤ p < ∞, or c0 . In Section

8.1 we will see that much more is true: for any weight sequence w for which
the backward shift Bw is an operator on X, the operators I + Bw and eBw
are hypercyclic (and even mixing); see Theorems 8.1 and 8.2.

4.5 Composition operators II

In this section we return to the composition operators studied in Section

4.3, but we consider them now on the Hardy space H 2 . Thus, let ϕ be an
automorphism of the unit disk D and let

Cϕ f = f ◦ ϕ

be the corresponding composition operator, where we now demand that f

belongs to H 2 . The first problem is that of determining if this defines an
operator on H 2 .

Proposition 4.45. For any ϕ ∈ Aut(D), Cϕ defines an operator on H 2 .

Proof. By Proposition 4.36 there are a, b ∈ C with |a| < 1 and |b| = 1 such
that
a−z
ϕ(z) = b .
1 − az
124 4 Classes of hypercyclic and chaotic operators

First, let f be a polynomial. Then f and f ◦ ϕ are continuous on D so that,

by Proposition 4.40,

1 2π  
f ◦ ϕ = 2 f (ϕ(eit ))2 dt, (4.13)
2π 0

and similarly for f 2 . Also by Proposition 4.36, ϕ is a bijective self-map on

T so that there is some u0 ∈ R and a continuously differentiable function
u : [0, 2π] → [u0 , u0 + 2π] such that

eiu(t) = ϕ(eit ), t ∈ [0, 2π].

Differentiating with respect to t we obtain that

du
ieiu(t) = ieit ϕ (eit ), t ∈ [0, 2π],
dt
so that (4.13) and the substitution u = u(t) yield
 u0 +2π
1 1
f ◦ ϕ 2 = |f (eiu )|2 du.
2π u0 |ϕ (eit(u) )|

Since, for |z| = 1,

 aa − 1  1 − |a|2 1 − |a|
 
|ϕ (z)| = b ≥ = ,
(1 − az) 2 (1 + |a|) 2 1 + |a|

we deduce that

1 + |a| 1 u0 +2π
1 + |a|
f ◦ ϕ ≤
2
· |f (eiu )|2 du = · f 2 . (4.14)
1 − |a| 2π u0 1 − |a|

Now let f ∈ H 2 be arbitrary, and let fn , n ≥ 0, be the partial sums of its

Taylor series. By Proposition 4.40 and (4.14) we have for n ≥ 0 and 0 ≤ r < 1
that  2π
1 1 + |a|
|fn (ϕ(reit ))|2 dt ≤ fn ◦ ϕ 2 ≤ · fn 2 .
2π 0 1 − |a|
Letting n → ∞ and noting that fn → f in H 2 and locally uniformly on D
we deduce that, for 0 ≤ r < 1,

1 2π
1 + |a|
|f (ϕ(reit ))|2 dt ≤ · f 2 .
2π 0 1 − |a|

By Proposition 4.40 this shows that f ◦ ϕ ∈ H 2 and that Cϕ is continuous

on H 2 . 


The proof also gives us a norm estimate on the operator Cϕ , namely

4.5 Composition operators II 125
 1 + |a| 1/2
Cϕ ≤ .
1 − |a|

Our aim now is to characterize when Cϕ is hypercyclic on H 2 . To this end

we need to study the (nonlinear) dynamical system that is described by the
automorphism ϕ on D. It will be convenient to consider ϕ as a particular
linear fractional transformation; see Appendix A.
Indeed, let
az + b
ϕ(z) = , ad − bc = 0,
cz + d
be an arbitrary linear fractional transformation, which we consider as a map
on the extended complex plane C.  Then ϕ has either one or two fixed points
 or it is the identity.
in C,
Suppose that ϕ has a single fixed point z0 , and let σ be a linear fractional
transformation that maps z0 to ∞. Then ψ := σ ◦ ϕ ◦ σ −1 has ∞ as a unique
fixed point, which easily implies that ψ(z) = z + c for some c = 0.
Now suppose that ϕ has two distinct fixed points z0 and z1 , and let σ
be a linear fractional transformation that maps z0 to 0 and z1 to ∞. Then
ψ := σ ◦ϕ◦σ −1 has fixed points 0 and ∞, which easily implies that ψ(z) = λz
for some λ = 0. The constant λ is called the multiplier of ϕ. Replacing σ by
1/σ one sees that also 1/λ is a multiplier, which, however, causes no problem
in the following.
Definition 4.46. Let ϕ be a linear fractional transformation that is not the
identity.
(a) If ϕ has a single fixed point then it is called parabolic.
(b) Suppose that ϕ has two fixed points, and let λ be its multiplier. If
|λ| = 1 then ϕ is called elliptic; if λ > 0 then ϕ is called hyperbolic; in all
other cases, ϕ is called loxodromic.
It is now easy to deduce some important dynamical properties of auto-
morphisms ϕ of D.
Proposition 4.47. Let ϕ ∈ Aut(D), not the identity. Then we have the fol-
lowing:
(i) if ϕ is parabolic then its fixed point z0 lies in T, and ϕn (z) → z0 ,
ϕ−n (z) → z0 for all z ∈ C;
(ii) if ϕ is elliptic then it has a fixed point in D;
(iii) if ϕ is hyperbolic then it has distinct fixed points z0 and z1 in T such
that ϕn (z) → z0 for all z ∈ C, z = z1 , and ϕ−n (z) → z1 for all z ∈ C,

z = z0 ;
(iv) ϕ cannot be loxodromic.
Proof. In the various cases, let σ and ψ be the linear fractional transforma-
tions given above. Then σ provides a conjugacy between ϕ and ψ.
(i) If ϕ is parabolic then ψ n (z) = z + nc → ∞ for all z ∈ C and hence
−1  −n
ϕ (z) → σ (∞) = z0 for all z ∈ C. In the same way, ϕ (z) → z0 for all
n
126 4 Classes of hypercyclic and chaotic operators

z ∈ C.  Since ϕ maps T into T (see Proposition 4.36), we must have that

z0 ∈ T.
(ii) Let ϕ be elliptic. Since ϕ maps D onto itself, ψ maps σ(D) onto itself,
which is either a half-plane, or the interior or the exterior of a disk U . Since
ψ is a rotation and λ = 1, the first alternative is excluded and U must be
centred at 0. Thus, either 0 or ∞ lies in B = σ(D), so that either z0 or z1
belongs to D.
(iii) Let ϕ be hyperbolic. Then λ > 0, and since with λ also 1/λ is a
multiplier we can assume that λ < 1. Then ψ n (z) = λn z → 0 for all z ∈ C, 
−1
z = ∞, and therefore ϕ (z) → σ (0) = z0 for all z ∈ C,
n  z = σ (∞) = z1 .
−1

It follows as in (i) that z0 ∈ T. Moreover, we find that ψ −n (z) = λ−n z → ∞

for all z ∈ C, z = 0, and hence ϕ−n (z) → σ −1 (∞) = z1 for all z ∈ C, 
−1 −1
z = σ (0) = z0 . Since T is also invariant under ϕ we find that also
z1 ∈ T.
(iv) Let ϕ be loxodromic. As in (ii), ψ maps σ(D) onto itself, which is either
a half-plane, or the interior or the exterior of a disk. But this is incompatible
with the fact that |λ| = 1 and λ ≯ 0.  
The dynamical properties of ϕ imply the dynamical properties of Cϕ .
Theorem 4.48. Let ϕ ∈ Aut(D) and Cϕ be the corresponding composition
operator on H 2 . Then the following assertions are equivalent:
(i) Cϕ is hypercyclic;
(ii) Cϕ is mixing;
(iii) ϕ has no fixed point in D.
Proof. The implication (ii)=⇒(i) holds for all operators on H 2 , and (i)=⇒(iii)
follows as in the proof of Theorem 4.37, using the fact that point evaluations
are continuous on H 2 .
(iii)=⇒(ii). Suppose that ϕ has no fixed point in D. It suffices to show
that Cϕ satisfies Kitai’s criterion. By Proposition 4.47, ϕ is either parabolic
or hyperbolic, and in both cases there are z0 , z1 ∈ T (possibly with z0 = z1 )
such that ϕn (z) → z0 for all z ∈ T\{z1 } and ϕ−n (z) → z1 for all z ∈ T\{z0 }.
Now, for X0 we will take the subspace of H 2 of all functions that are
holomorphic on a neighbourhood ofD and that vanish at z0 . To see that X0
∞
is dense in H 2 , let f ∈ H 2 , f (z) = n=0 an z n , be orthogonal to any g ∈ X0 .
Since, for any n ≥ 0, the functions gn : z → z0 z n −z n+1 belong to X0 we have
that f, gn  = z0 an − an+1 = 0 and hence an = a0 z0 n , n ≥ 0. Since (an )n is
square summable and |z0 | = 1 we must have that a0 = 0, hence f = 0. This
implies that X0 is dense in H 2 . Moreover, let f ∈ X0 . As in (4.13) we have
that  2π
1   n  it 2
(Cϕ ) f =
n 2 f ϕ e  dt.
2π 0
Since the integrands are uniformly bounded and convergent to |f (z0 )|2 = 0,
for every t with possibly one exception, the dominated convergence theorem
implies that (Cϕ )n f → 0 for all f ∈ X0 .
Exercises 127

Next, for Y0 we take the subspace of H 2 of all functions that are holomor-
phic on a neighbourhood of D and that vanish at z1 , and for S we take the
map S = Cϕ−1 . Since z1 is a fixed point of ϕ−1 , S maps Y0 into itself, and
clearly T S = I. It follows as above that Y0 is dense in H 2 and that S n f → 0
for all f ∈ Y0 .
Therefore the conditions of Kitai’s criterion are satisfied, so that Cϕ is
mixing.  
Some concrete instances of this result are treated in Exercise 4.5.2.
We end this chapter by returning to Theorem 4.37 of Section 4.3. Propo-
sition 4.47 allows us to give the missing proof of the implication (v)=⇒(vi)
for Ω = D.
Conclusion of the proof of Theorem 4.37. Let ϕ be an automorphism of D
without fixed points in D. Again, ϕ is either parabolic or hyperbolic.
First, let ϕ be parabolic. By the discussion before Definition 4.46 there
is a linear fractional transformation σ that provides a conjugacy between ϕ
and ψ(z) = z + c, c = 0. Then ψ is an automorphism of σ(D), so that Cϕ is
conjugate to the operator Cψ on H(σ(D)) by Proposition 4.25. By Runge’s
theorem, the continuous restriction map H(C) → H(σ(D)), f → f |σ(D) has
dense range. Hence Cϕ is quasiconjugate to the Birkhoff operator Cψ on
H(C).
In the hyperbolic case, there is a linear fractional transformation σ such
that ϕ is conjugate to a dilation ψ(z) = λz, λ = 1 strictly positive, and ψ
is an automorphism of σ(D), which therefore must be a half-plane with 0 on
its boundary. After conjugation with a suitable rotation we can assume that
it is the right half-plane C+ , and ψ remains unchanged. Now, the principal
branch log of the logarithm is a conformal map from C+ to the strip S =
{z ∈ C ; |Im(z)| < π2 }, and conjugating ψ with log gives us the translation
τ (z) = z + log λ, log λ = 0, on S. We conclude as in the parabolic case that
Cϕ is quasiconjugate to the Birkhoff operator Cτ on H(C).  

Exercises

Exercise 4.1.1. Show that a weighted shift Bw defines an operator on H(C) if and only
if supn≥1 |wn |1/n < ∞, and that an en → 0 in H(C) if and only if |an |1/n → 0.
Exercise 4.1.2. Let T := Bw be a weighted shift on p , 1 ≤ p < ∞.
(a) Given an increasing sequence (nk )k of positive integers, show that the sequence
of operators (T nk )k is hypercyclic if and only if, for each j ∈ N,


j+nk

sup |wν | = ∞.
k≥1
ν=1

(b) Show that T is hereditarily hypercyclic with respect to (nk )k if and only if, for
each j ∈ N,
128 4 Classes of hypercyclic and chaotic operators


j+nk

lim |wν | = ∞.
k→∞
ν=1

Exercise 4.1.3. Let X be the Banach space of all sequences (xn )n satisfying
∞  
  xn xn+1  xn
x =  n − n + 1  < ∞ and n → 0 as n → ∞.
n=1

Show that the backward shift B is a hypercyclic operator on X and that conditions
(ii)–(iv) of Theorem 4.6 are satisfied, but that the only periodic points of B are the
constant sequences and that B is therefore not chaotic.

Exercise 4.1.4. Let X be a Fréchet sequence space over Z in which (en )n∈Z is a basis.
Suppose that the bilateral weighted shift Bw is an invertible operator on X. Show that
Bw is hypercyclic if and only if there is an increasing sequence (nk )k of positive integers
such that
 
0  
nk −1
wν e−nk → 0 and wν enk → 0
ν=−nk +1 ν=1

in X as k → ∞. (Hint: For the sufficiency, look at the proof of (iii)=⇒(ii) in Theorem

4.3.)

Exercise 4.1.5. Find a (necessarily non-invertible) bilateral weighted shift that satisfies
the condition stated in the previous exercise but that is not hypercyclic. (Hint: See the
proof of Proposition 4.16, but choose nonsymmetric vn .)

Exercise 4.1.6. Prove the results stated in Remark 4.14; instead of using a conjugacy
one may also observe that a forward shift on the basis (en )n is a backward shift on the
basis (e−n )n .

Exercise 4.1.7. Show that the characterizing conditions on a weight w to define a

hypercyclic bilateral weighted shift Bw on p (Z) can also be written as follows: for any
ε > 0 and any M, N ≥ 1 there exists some n ≥ N such that whenever |j| ≤ M then


j

j+n
1
|wν | < ε, |wν | > .
ε
ν=j−n+1 ν=j+1

Exercise 4.2.1. An entire function ϕ is of exponential type 0 if for any ε > 0 there is
some M > 0 such that
|ϕ(z)| ≤ M eε|z| for all z ∈ C.
For example, any polynomial but no exponential function z → eλz , λ
= 0, is of expo-
nential type 0.
For a domain Ω ⊂ C, let H(Ω) denote the Fréchet space of holomorphic functions
on Ω; see Section 4.3. Show the following:
∞
(i) an entire function ϕ(z) = a z n is of exponential type 0 if and only if, for
n=0 n n
any ε > 0, there is some M > 0 such that |an | ≤ M εn! ;
∞
(ii) for any domain Ω ⊂ C and any entire function ϕ(z) = a z n of exponential
∞ n
n=0 n
type 0, ϕ(D) = a D defines an operator on H(Ω);
n=0 n
(iii) for any simply connected domain Ω ⊂ C and any nonconstant entire function ϕ
of exponential type 0, ϕ(D) is chaotic on H(Ω). (Hint: Use the Godefroy–Shapiro
theorem and the fact, that, by Runge’s theorem, H(C) is dense in H(Ω).)
Exercises 129

Exercise 4.2.2. Let Ω be a domain and P a nonconstant polynomial. Show that the
following assertions are equivalent:
(i) P (D) is chaotic on H(Ω);
(ii) P (D) is hypercyclic on H(Ω);
(iii) Ω is simply connected.
 there is a smooth Jordan curve Γ in Ω sur-
(Hint: If Ω is not simply connected then
rounding some a ∈ / Ω. Show that f → Γ f (ζ) dζ is an eigenvector of P (D)∗ , and use
Lemma 2.53.)
Exercise 4.2.3. Let X = CR∞ (R) be the space of infinitely differentiable real functions
f : R → R; see Exercise 2.1.5. Show that every (real) differential operator T : X → X,
N
Tf = a f (n) , T
= a0 I, is chaotic. (Hint: See Exercise 2.2.5.)
n=0 n
Exercise 4.2.4. Let ϕ be a nonconstant entire function of exponential type and A =
min{|z| ; z ∈ C, |ϕ(z)| = 1}. Show that, for any ε > 0, there is an entire function f that
is hypercyclic for ϕ(D) such that

|f (z)| ≤ M e(A+ε)r for |z| = r > 0

with some M > 0.

For the proof consider the Hilbert spaces
 
∞ ∞ 
  
n! 2
Eτ2 = f ∈ H(C) ; f (z) = an z n , |an |2 < ∞ , τ > 0.
τn
n=0 n=0

Show that any f ∈ Eτ2 satisfies |f (z)| ≤ M eτ r ; use ideas from Example 3.2 to show
that for any Λ ⊂ Dτ with an accumulation point, span{eλ ; λ ∈ Λ} is dense in Eτ2 (see
Appendix A for the dual of Eτ2 ); show that ϕ(D) is an operator on any Eτ2 , and that
2
ϕ(D) is hypercyclic on EA+ε for any ε > 0.
Apply the result to MacLane’s and Birkhoff’s operators.

 n4.2.5. Let
Exercise
∞ −1 n
Bw be a chaotic weighted shift on H(C); see Example 4.9(b). Then
n=0
( ν=1
wν ) z is an entire function, and its maximum term is defined by

rn
μw (r) = max n , r ≥ 0.
n≥0
ν=1
|wν |

(a) Let φ : ]0, ∞[ → [1, ∞[ be a function with φ(r) → ∞ as r → ∞. Show that there
exists an entire function f that is hypercyclic for Bw and that satisfies

|f (z)| ≤ M φ(r)μw (r) for |z| = r > 0

with some M > 0.

(b) Suppose that |wn | → ∞ monotonically. Show that there is no entire function f
that is hypercyclic for Bw and that satisfies

|f (z)| ≤ M μw (r) for |z| = r > 0

with some M > 0. (Hint: Determine μw (ρ) for ρ = |wn |.)

Deduce Theorem 4.22 from this.
Exercise 4.3.1. Show in detail that H(Ω) is a separable Fréchet space and that its
topology is independent of the exhaustion chosen. (Hint: For separability, fix an ex-
haustion (Kn )n of Ω by Ω-convex compact sets. In each connected component of the
complements of the Kn fix one point outside Ω, possibly ∞; this set is denumerable.
Now use Runge’s theorem.)
130 4 Classes of hypercyclic and chaotic operators

Exercise 4.3.2. Let Ω be a domain and ϕ : Ω → Ω a holomorphic self-map that is not

necessarily an automorphism of Ω. Show that Cϕ : H(Ω) → H(Ω), Cϕ f = f ◦ ϕ, defines
an operator on H(Ω), also called a composition operator. Moreover, show the following:
(i) if Cϕ is hypercyclic then ϕ is injective;
(ii) let Ω be simply connected; then Cϕ is hypercyclic if and only if ϕ is injective and
(ϕn )n is a run-away sequence.

Exercise 4.3.3. Let Ω be a domain, ϕ : Ω → Ω an injective holomorphic self-map and

Cϕ f = f ◦ϕ the corresponding composition operator on H(Ω); see the previous exercise.
(a) Based on the proof of Theorem 4.32, find a sufficient condition under which Cϕ
is hypercyclic.
(b) Let ϕ  : D → D be given by ϕ(z) = z4 + 34 . Let K = {z ∈ C ; |z| ≤ 12 }
∞
and Ω = D \ n=0 ϕn (K). Show that the restriction of ϕ to Ω defines a hypercyclic
composition operator Cϕ on H(Ω).

Exercise 4.3.4. Let Ω = C \ Z. Then ϕ(z) = z + 1 is an automorphism of Ω. Show that

m that the 1linear span of
the composition operator Cϕ is chaotic on H(Ω). (Hint: Show
the functions eλz , eλN = 1 for some N ≥ 1, and limm→∞ ν=−m (z−k−νN )α , k ∈ Z,
α > 1, N ≥ 1, forms a dense set of periodic points.)

Exercise 4.4.1. Let ϕ be a holomorphic function on D such that ϕf ∈ H 2 for all

f ∈ H 2 . Use the closed graph theorem to show that the mapping Mϕ : f → ϕf is
continuous. Deduce that ϕ is necessarily bounded, and Mϕ  = supz∈D |ϕ(z)|. (Hint:
ϕn (z) = (Mϕ )n 1, kz .)

Exercise 4.4.2. Let Ω ⊂ C be a domain and H

= {0} a Hilbert space of holomorphic
functions on Ω. Suppose that each point evaluation f → f (λ), λ ∈ Ω, is a continuous
linear functional on H. Use the closed graph theorem to prove that the canonical em-
bedding H → H(Ω) is continuous, so that convergence in H implies locally uniform
convergence on Ω.

Exercise 4.4.3. Let Ω ⊂ C be a domain and H

= {0} a Hilbert space of holomorphic
functions on Ω such that each point evaluation f → f (λ), λ ∈ Ω, is continuous on H.
(a) By the Riesz representation theorem (see Appendix A), there is a unique function
kλ ∈ H, again called a reproducing kernel, such that

f (λ) = f, kλ , f ∈ H.

Prove an analogue of Lemma 4.39 and deduce that H is separable.

(b) Now let ϕ be a nonconstant bounded holomorphic function on Ω for which Mϕ f =
ϕf defines an operator on H. Let Mϕ∗ be the corresponding adjoint multiplier on H.
Show that Mϕ∗ is chaotic and mixing as soon as ϕ(Ω) ∩ T
= ∅. Show that, in this case,
for some λ ∈ C, λMϕ∗ is chaotic. Deduce that Mϕ∗ is supercyclic.
(c) Finally, suppose that every bounded holomorphic function ϕ on Ω defines a
multiplication operator with Mϕ  ≤ supz∈Ω |ϕ(z)|. Show that if ϕ is a nonconstant
bounded holomorphic function on Ω such that Mϕ∗ is hypercyclic then ϕ(Ω) ∩ T
= ∅.
(d) Deduce that Theorem 4.42 holds also for the Bergman space A2 ; see Example
4.4(b).

Exercise 4.4.4. The Dirichlet space D is defined as the space of all holomorphic func-
tions f on D such that

1
f 2 := |f (0)|2 + |f  (z)|2 dλ(z) < ∞,
π D

where λ denotes two-dimensional Lebesgue measure. Show the following:

Exercises 131
∞ ∞
(i) if f (z) = a z n then f 2 = |a0 |2 + n=1 n|an |2 ;
n=0 n
(ii) D is a Hilbert space with continuous point evaluations;
(iii) D ⊂ H 2 ⊂ A2 , where A2 is the Bergman space (see Example 4.4);
(iv) if ϕ is a bounded holomorphic function on D such that ϕ is also bounded then Mϕ
defines an  operator on D;  ∞
∞ ∞
(v) if ϕ(z) = b z n with
n=0 n
|b | < ∞ and
n=0 n n=0
n|bn |2 = ∞ (existence?)
then ϕ is a bounded holomorphic function on D for which Mϕ does not define an
operator on D;
(vi) if ϕ is a nonconstant bounded holomorphic function on D such that Mϕ defines an
operator on D and if ϕ(D) ∩ T
= ∅ then Mϕ∗ is mixing and chaotic;
(vii) the function ϕ(z) = z, z ∈ D, defines a hypercyclic adjoint multiplier Mϕ∗ on D,
but ϕ(D) ∩ T = ∅; is Mϕ∗ mixing or chaotic? (Hint: Identify Mϕ∗ with a weighted
shift on a weighted 2 -space.)

Exercise 4.4.5. Let X

= {0} be a Banach space of holomorphic functions on a domain
Ω ⊂ C. Suppose that X is reflexive, that is, X ∗∗ = X. Suppose further that the point-
evaluations f → f (λ), λ ∈ Ω, are continuous on X and that every bounded holomorphic
function ϕ on Ω defines a multiplication operator Mϕ with Mϕ  ≤ supz∈Ω |ϕ(z)|. Then
show the analogue of Theorem 4.42 for Mϕ∗ , the (Banach space) adjoint of Mϕ . (Hint:
Use reflexivity and the Hahn–Banach theorem to obtain the analogue of Lemma 4.39.)

Exercise 4.4.6. Let X be one of the complex spaces p , 1 ≤ p < ∞, or c0 . Let a, b ∈ C,

b
= 0. Show that the following assertions are equivalent:
(i) aI + bB is chaotic on X;
(ii) |b| > |1 − |a||.

∞ 4.4.7.p Generalize part of Theorem 4.43 as follows: let X =  (v) = {(xn )n ∈

p
Exercise
N
C ; n=1
|xn | vn < ∞}, 1 ≤ p < ∞, where v = (vn )n is a positive weight sequence
< ∞. Let R := (lim supn→∞ vn )−1 > 0, which is finite,
vn 1/n
such that M := supn∈N
∞ vn+1
and let ϕ(z) = be a nonconstant function that is holomorphic in Dr for
a zn
n=0 n ∞
some r > M . Then ϕ(B) = a B n defines an operator on X, and if
n=0 n
 
ϕ R1/p D ∩ T
= ∅ (4.15)

then ϕ(B) is a chaotic operator on X. (Hint: See Appendix A for the dual of X.)

Exercise 4.4.8. In the setting of Exercise 4.4.7, let vn = 1/n2 , n ≥ 1.

(a) Show that B is chaotic but condition (4.15) does not hold.
(b) Show that the operator 12 (I +B) has a nontrivial periodic point but is not chaotic.

Exercise 4.4.9. Let w = (wn )n be a weight sequence, and let f be a nonconstant

polynomial. Show that f (Bw ) is (well defined and) chaotic on ω = KN . (Hint: For the
density of eigenvectors for Bw show that they are contained and dense in a suitable
weighted 1 -space; see Appendix A.)

Exercise 4.5.1. The aim of this exercise is to prove Littlewood’s subordination principle:
if ϕ : D → D is a holomorphic self-map then Cϕ : f → f ◦ ϕ defines an operator on H 2
with Cϕ  ≤ ( 1−r ) , where r = |ϕ(0)|.
1+r 1/2

(a) First prove the result when ϕ(0) = 0 by proceeding as follows:

(i) show that, for any f ∈ H 2 , Cϕ f = f (0) + Mϕ Cϕ Bf , where we write B for
Mz∗ (see Proposition 4.41(ii));
(ii) using orthogonality, deduce that, for any polynomial f , Cϕ f 2 ≤ |f (0)|2 +
Cϕ Bf 2 ; n
(iii) deduce that Cϕ f 2 ≤ k=0
|(B k f )(0)|2 + Cϕ B n+1 f 2 ;
132 4 Classes of hypercyclic and chaotic operators

(iv) deduce that Cϕ f  ≤ f ;

(v) conclude that Cϕ f ∈ H 2 and Cϕ f  ≤ f  for all f ∈ H 2 .
(b) Prove the result by factorizing Cϕ = Cϕ1 Cϕ2 with ϕ1 (0) = 0 and ϕ2 ∈ Aut(D).

Exercise 4.5.2. For the following linear fractional transformations decide if they are
automorphisms of D; in the case of an automorphism, determine if the corresponding
composition operator Cϕ is hypercyclic on H 2 :
2z − 1
(i) ϕ(z) = ;
2−z
1 + (i − 1)z
(ii) ϕ(z) = ;
(i + 1) − z
4 − 5z
(iii) ϕ(z) = ;
5 − 4z
z+1
(iv) ϕ(z) = .
2

Exercise 4.5.3. Let α > −1. Then the weighted Bergman space A2α is defined as the
space of all holomorphic functions f on D such that

2 1  α
f  := |f (z)|2 1 − |z|2 dλ(z) < ∞,
π D

where λ denotes two-dimensional Lebesgue measure; see Example

4.4(b).
∞
(a) Let f be a holomorphic function on D with f (z) = a z n , z ∈ D.
∞ n=0 n
Show that f 2 = |an |2 Γ (α+1)Γ (n+1)
and deduce that f ∈ A2α if and only if
∞ n=0 Γ (α+n+2)

n=0
|an |
2 1
(n+1)α+1 < ∞. (Hint: Stirling’s formula.)
(b) Let ϕ ∈ Aut(D). Show that Cϕ is an operator on A2α with Cϕ  ≤ ( 1−r
1+r 1+α/2
) ,
 −2
where r = |ϕ(0)|. (Hint: Show that if w = ϕ(z) then dλ(z) = |ϕ (z)| dλ(w), and use
(4.12).)

Exercise 4.5.4. Let α > −1. Then the weighted Dirichlet space Dα is defined as the
space of all holomorphic functions f on D such that

  2  
f 2 := |f (0)|2 +
1 f (z) 1 − |z|2 α dλ(z) < ∞,
π D

where λ denotes two-dimensional Lebesgue measure; see Exercise 4.4.4 and the previous
exercise. ∞
(a) Let f be a holomorphic function on D with f (z) = a z n , z ∈ D. Show that
∞ n=0 n
f ∈ Dα if and only if n=0
|an | (n + 1)
2 1−α
< ∞.
(b) Let ϕ ∈ Aut(D). Show that Cϕ is an operator on Dα .
(c) Let α > 0. Show that Theorem 4.48 remains true for Dα . (Hint: Proceed as in the
proof of that theorem; use the change of variables w = ϕn (z); note that 1−|ϕ−n (w)|2 →
0.)

Exercise 4.5.5. Let ϕ ∈ Aut(D). By the previous exercise, Cϕ is an operator on the

Dirichlet space D. Show that, for any f ∈ D,

  2
Cϕ f 2 ≥
1 f (z) dλ(z).
π D

Deduce that Cϕ is not hypercyclic on D. (Hint: Change of variables w = ϕ(z).)

Sources and comments 133


Exercise 4.5.6. Let β = (βn )n≥0 be a sequence of strictly positive numbers such that
∞ −2 n
β r < ∞ whenever 0 ≤ r < 1. Then the weighted Hardy space H (β) is defined
n=0 n
2

as the space of all holomorphic functions f on D such that


∞

f 2 := |an |2 βn2 < ∞,

n=0
∞
where f (z) = a z n , z ∈ D. By the assumption on (βn )n , this condition alone
n=0 n
implies that f ∈ H(D).
Let ϕ ∈ Aut(D) and suppose that Cϕ defines an operator on H 2 (β). Show the
∞ −2
following:
(i) if β < ∞ then Cϕ is never hypercyclic on H 2 (β);
∞n=0 n
(ii) if n=0 n
β −2 = ∞ and ϕ is elliptic then Cϕ is not hypercyclic on H 2 (β).
(Hint for (i): Show that all functions in H 2 (β) have a continuous extension to D; and
use the fact that ϕ has a fixed point in D.)
Exercise 4.5.7. Let ν ∈ R. Then the space Sν is defined as the space of all holomorphic
functions f on D such that


∞

f 2 := |an |2 (n + 1)2ν < ∞,

n=0
∞
where f (z) = n=0 n
a z n , z ∈ D. In particular, S0 is the Hardy space H 2 , S−1/2 is the
Bergman space A2 , and S1/2 is the Dirichlet space D under an equivalent norm.
(a) Show that f ∈ Sν if and only if f  ∈ Sν−1 . If ν ∈ N, show that f ∈ Sν if and only
if f (ν) ∈ H 2 .
(b) Show that the multiplier Mz is an operator on each Sν , and ∞calculate the norm
of Mzn , n ≥ 0. More generally, let ϕ be holomorphic on D, ϕ(z) = b z n
, such that
∞ n=0 n

n=0 n
|b |(n + 1)ν < ∞. Show that Mϕ is an operator on Sν .
(c) Let ϕ ∈ Aut(D). Deduce from Exercise 4.5.3 that Cϕ defines an operator on Sν
for ν < 0. Use parts (a) and (b) to conclude that Cϕ defines an operator on Sν for any
ν ∈ R. (Hint: (Cϕ f ) = Mϕ Cϕ f  .)  ∞
∞ ∞
(d) Let ϕ(z) = b z n with
n=0 n n=0 n
|b | ≤ 1 and |b |2 n = ∞ (existence?).
n=0 n
Show that ϕ is a holomorphic self-map of D for which Cϕ does not define an operator
on the Dirichlet space D.
Exercise 4.5.8. Let ν ∈ R and ϕ ∈ Aut(D). By Exercise 4.5.7, Cϕ is an operator on
Sν . Deduce the following from the previous exercises:
(i) if ν ≥ 12 then Cϕ is never hypercyclic on Sν ;
(ii) if ν < 12 then Cϕ is hypercyclic on Sν if and only if ϕ is not the identity and
non-elliptic.
Spell out these results for the (weighted) Bergman and Dirichlet spaces.

Sources and comments

Section 4.1. Rolewicz’s multiples of the backward shift were the first Banach space op-
erators to be proved hypercyclic [268]. Due to its simple structure, the class of weighted
shifts is a favorite testing ground for operator-theorists (Salas [274]). Accordingly, when-
ever a new notion in linear dynamics is introduced it is usually first tested on weighted
shifts.
134 4 Classes of hypercyclic and chaotic operators

Salas [274] characterized hypercyclic and weakly mixing unilateral and bilateral
weighted shifts on 2 and 2 (Z), respectively. The characterizations for more general
sequence spaces and of chaos are due to Grosse-Erdmann [180], see also Martínez and
Peris [229] in the special case of Köthe sequence spaces, while mixing shifts on 2 and
2 (Z) were characterized by Costakis and Sambarino [124]. The approach chosen here of
first studying the unweighted shift and then using suitable conjugacies is due to Martínez
and Peris [229].
The first example of a hypercyclic operator whose adjoint is also hypercyclic (see
Proposition 4.16) was found by Salas [273]. He later showed [276] that every separa-
ble Banach space with separable dual supports such an operator. The observation that
T ⊕ T ∗ is never hypercyclic is due to Deddens; see [273].

Section 4.2. The investigation of hypercyclicity for differential operators ϕ(D) is due to
Godefroy and Shapiro [165]. Theorem 4.22 on the rate of growth of MacLane’s operator
was obtained independently by Grosse-Erdmann [178] and Shkarin [283]. The corre-
sponding result for Birkhoff’s operators was obtained by Duyos-Ruiz [137]; alternative
proofs can be found in Chan and Shapiro [106] and in Exercise 8.1.3.
Translation and differentiation operators have also been studied on spaces of har-
monic functions on RN , N ≥ 2. Hypercyclicity of these operators and corresponding
growth results have been obtained by Dzagnidze [138], Aldred and Armitage [6], [7],
[13], and Gómez, Martínez, Peris and Rodenas [166].

Section 4.3. This section draws heavily on the work of Bernal and Montes [64], who
also coined the term “run-away sequence”, and the work of Shapiro [281]. The material
up to Theorem 4.32 can be found in [64], while most of Theorem 4.37 is implicit in
[281]. We mention that Seidel and Walsh [278] were the first to study the analogue of
Birkhoff’s result in the unit disk.
For two different proofs of Proposition 4.31 we refer to Bernal and Montes [64] and to
Grosse-Erdmann and Mortini [184]. Example 4.34 is taken from Kim and Krantz [214];
see also Gorkin, León and Mortini [168].

Section 4.4. The study of the dynamical properties of adjoint multipliers was initiated
by Godefroy and Shapiro [165], who also obtained Theorem 4.42. Functions of the back-
ward shift on the spaces p and c0 were studied by deLaubenfels and Emamirad [128],
who also obtained Theorem 4.43. An interesting related investigation of functions of the
backward shift on the Bergman space is due to Bourdon and Shapiro [96].
For a more detailed introduction to Hardy spaces we refer to Duren [136] and Rudin
[270].

Section 4.5. In this section we closely follow the book of Shapiro [279]; see also Shapiro
[281]. Proposition 4.45 is a special case of the Littlewood subordination principle; see
Exercise 4.5.1. Theorem 4.48 is due to Bourdon and Shapiro [94], [95]. This result is
only the beginning of a fascinating story on the interplay between operator theory and
complex function theory. The extension of Theorem 4.48, first to non-automorphic linear
fractional transformations and then to more arbitrary holomorphic self-maps of D, can
be found in the cited work of Bourdon and Shapiro. The proofs, however, require a much
deeper understanding, for example, of the (nonlinear) dynamics of self-maps of D.
Hosokawa [204] proved that, for any automorphism ϕ of D, Cϕ is chaotic whenever
it is hypercyclic; see also Taniguchi [298]. Thus one can add chaos to the equivalent
conditions in Theorem 4.48.
We note that Gallardo and Montes [158] have obtained a complete characterization
of the cyclic, supercyclic and hypercyclic composition operators Cϕ for linear fractional
self-maps ϕ of D on any of the spaces Sν , ν ∈ R (see Exercises 4.5.7 and 4.5.8).
Sources and comments 135

For a more detailed introduction to composition operators on weighted Hardy,

Bergman and Dirichlet spaces we refer to Cowen and MacCluer [125].

Exercises. Exercise 4.1.2 is taken from Bès and Peris [71], Exercise 4.1.3 from Grosse-
Erdmann [180]. For Exercises 4.1.4 and 4.1.5 we refer to Feldman [150], Exercise 4.1.7
states the condition in the form found originally by Salas [274]. But note that the
weighted shifts considered by Feldman and Salas are forward shifts. For Exercise 4.2.1
we refer to Bernal [55] and Shapiro [280], for Exercise 4.2.2 to Shapiro [280], for Exer-
cise 4.2.4 to Chan and Shapiro [106] and to Bernal and Bonilla [60], for Exercise 4.2.5
to Grosse-Erdmann [181]. Exercises 4.3.2 and 4.3.3 follow Montes [241] and Grosse-
Erdmann and Mortini [184], while Exercise 4.3.4 is taken from Shapiro [280]. The mate-
rial for Exercises 4.4.1–4.4.4 can be found in Godefroy and Shapiro [165], with Exercise
4.4.4(vii) being taken from Chan and Seceleanu [104]; for Exercises 4.4.6–4.4.8 we refer
to deLaubenfels and Emamirad [128], for Exercise 4.4.9 to Martínez [228]. For Exercise
4.5.1 we have again followed Shapiro [279]; Exercises 4.5.4(c), 4.5.5 and 4.5.8 are taken
from Gallardo and Montes [158], Exercise 4.5.6 from Zorboska [304] and Exercise 4.5.7
from Hurst [205].
Chapter 5
Necessary conditions for hypercyclicity
and chaos

In Chapter 3 we derived various sufficient conditions for an operator to be hy-

percyclic, which we then used in Chapter 4 to obtain classes of such operators.
In this chapter we are interested in the opposite question: which conditions
on an operator rule out its hypercyclicity? In other words, we are searching
for necessary conditions for an operator to be hypercyclic. This then leads to
classes of operators that do not include any hypercyclic operators.
Many of the known necessary conditions seem to involve, in one way or
another, the spectrum of the operator. Hence we begin by studying spectral
properties of hypercyclic and chaotic operators.
Spectral considerations show their full strength only for operators on com-
plex Banach spaces. Using the complexification technique one can then often
extend the results to the real scalar case. In this chapter we therefore restrict
our attention to operators on Banach spaces; in the final section we study, in
particular, operators on Hilbert spaces.
We suppose that the reader is familiar with the basics of spectral theory
as laid out in Appendix B. The main tool needed in this chapter, the Riesz
decomposition theorem, is not usually covered in introductory courses on
functional analysis; its proof is contained in the appendix. We have also
provided some exercises on spectral theory that the reader might find useful;
see Exercises 5.0.1–5.0.9.

5.1 Spectral properties of hypercyclic and chaotic

operators

In this section we study the influence of hypercyclicity and chaos on the

spectrum of an operator.
We start with two simple observations on the adjoint T ∗ of a hypercyclic
operator, the first of which has already been observed in greater generality
in Lemma 2.53.

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 137

DOI 10.1007/978-1-4471-2170-1_5, © Springer-Verlag London Limited 2011
138 5 Necessary conditions for hypercyclicity and chaos

Proposition 5.1. Let T be a hypercyclic operator on a (real or complex)

Banach space X. Then we have:
(i) T ∗ has no eigenvalues, that is, σp (T ∗ ) = ∅;
(ii) the orbit of every x∗ = 0 in X ∗ under T ∗ is unbounded.

Proof of (ii). Suppose that there exists some x∗ ∈ X ∗ , x∗ = 0, and some

M > 0 such that (T ∗ )n x∗  ≤ M for all n ≥ 0. Let x ∈ X be a hypercyclic
vector for T . Then T n x, x∗ , n ≥ 0, forms a dense set in K. On the other
hand,
|T n x, x∗ | = |x, (T ∗ )n x∗ | ≤ M x, n ≥ 0,
which is impossible.

For mixing operators T one can even assert that (T ∗ )n x∗  → ∞ for any
nonzero element x∗ ∈ X ∗ ; see Exercise 5.1.1.
Next we study the spectrum σ(T ) of the operator T , which suggests con-
sideration of complex spaces.
Throughout the remainder of this section, T denotes an operator on a
complex Banach space X.

Lemma 5.2. Let r > 0. Then we have:

(i) if σ(T ) ⊂ {z ∈ C ; |z| < r} then there are ε > 0 and M > 0 such that
T n x ≤ M (r − ε)n x for all x ∈ X and n ∈ N0 ;
(ii) if σ(T ) ⊂ {z ∈ C ; |z| > r} then there are ε > 0 and M > 0 such that
T n x ≥ M (r + ε)n x for all x ∈ X and n ∈ N0 .

Proof. (i) Since σ(T ) is a compact set, the assumption and the spectral radius
formula imply that limn→∞ T n 1/n = r(T ) < r. Hence, if ε < r − r(T ) then
there is some M > 0 such that T n  ≤ M (r − ε)n for all n ∈ N0 , and the
result follows.
(ii) The assumption implies that 0 ∈ / σ(T ), so that T is invertible. Since
σ(T −1 ) = σ(T )−1 (see Exercise 5.0.7), we obtain that σ(T −1 ) ⊂ {z ∈
C ; |z| < 1r }. By (i) there are then some η > 0 with η < 1r and M > 0
such that (T −1 )n y ≤ M ( 1r − η)n y for all y ∈ X and n ∈ N0 . Setting
y = T n x and defining ε by 1r − η = r+ε 1
we obtain the result.

The Riesz decomposition theorem now allows us to deduce a first major

necessary condition for an operator to be hypercyclic. The result will be
superseded by Kitai’s theorem below.

Proposition 5.3. Let T be a hypercyclic operator. Then σ(T ) meets the unit
circle:
σ(T ) ∩ T = ∅.

Proof. Suppose, on the contrary, that σ(T ) does not meet the unit circle.
If σ(T ) ⊂ D then, by taking r = 1 in Lemma 5.2(i), we see that all orbits
of T tend to 0, which is impossible. Similarly, Lemma 5.2(ii) shows that
σ(T ) ⊂ C \ D is impossible. We therefore have that the sets
5.1 Spectral properties of hypercyclic and chaotic operators 139

σ1 := σ(T ) ∩ D and σ2 := σ(T ) ∩ (C \ D)

form a partition of σ(T ) into nonempty closed sets. By the Riesz decomposi-
tion theorem there then exist nontrivial T -invariant closed subspaces M1 and
M2 such that X = M1 ⊕ M2 , σ(T |M1 ) = σ1 and σ(T |M2 ) = σ2 . By Propo-
sition 2.28, T |M1 would then be a hypercyclic operator with σ(T |M1 ) ⊂ D,
which is impossible, as we saw above.

Thus, the spectrum of a hypercyclic operator must contain at least one

point from the unit circle. In Example 8.4 we will see that, indeed, the spec-
trum need not contain any additional point: there is a hypercyclic, and even
a mixing, operator T with σ(T ) = {1}.

Example 5.4. (Volterra operator) Let X be the space C[0, 1] of (complex-

valued) continuous functions on the interval [0, 1], or one of the spaces Lp [0, 1]
of (complex-valued) p-integrable functions on [0, 1], where 1 ≤ p < ∞; see
Example 2.4. For any f ∈ X we define V f by
 t
V f (t) = f (s) ds, 0 ≤ t ≤ 1.
0

This obviously defines an operator on C[0, 1] with V  ≤ 1.

On the other hand, let p > 1, and let q be the dual exponent to p defined
by p1 + 1q = 1. It follows from Hölder’s inequality that
 1  t p  1  t p/p   t p/q
|f (s)| ds dt ≤ |f (s)|p ds 1q ds dt
0 0 0 0 0
 1
1
≤ f pp t p/q
dt = p f pp .
0 q
+1

Thus V is also an operator on Lp [0, 1], and since pq + 1 = p we have that

V  ≤ p−1/p . A similar, and simpler, argument shows that the same result
is true for p = 1. The operator V is called the Volterra operator.
A simple induction shows that the nth iterate of V , n ≥ 1, is given by
 t
1
n
V f (t) = (t − s)n−1 f (s)ds, 0 ≤ t ≤ 1.
(n − 1)! 0

In the case of X = C[0, 1] we therefore have that

 t
1 1
V f  ≤n
f  max (t − s)n−1 ds = f ,
(n − 1)! 0≤t≤1 0 n!

so that
1
V n 1/n ≤ → 0.
n!1/n
140 5 Necessary conditions for hypercyclicity and chaos

The spectral radius formula implies that r(V ) = 0, so that σ(V ) = {0}. The
same result holds for any space X = Lp [0, 1], 1 ≤ p < ∞; see Exercise 5.1.2.
By Proposition 5.3, V cannot be hypercyclic on any of the spaces consid-
ered. Indeed, for the same reason, no multiple λV , λ ∈ C, can be hypercyclic.

We turn to the announced improvement of Proposition 5.3. For its proof

we need a topological lemma that we accept without proof.

Lemma 5.5. Let A and B be compact subsets of a metric space, where B is

connected. Then the following assertions are equivalent:
(i) every connected component of A meets B;
(ii) A ∪ B is connected.

Theorem 5.6 (Kitai). Let T be a hypercyclic operator. Then every con-

nected component of σ(T ) meets the unit circle.

Proof. By the preceding lemma we have to show that the compact set σ(T )∪T
is connected. If this is not the case then σ(T ) ∪ T can be partitioned into two
nonempty closed sets C1 and C2 . Since T is connected, it must lie entirely in
one of these two sets, C2 say. Then we define a partition of σ(T ) into closed
sets by setting

σ1 := C1 ∩ σ(T ) and σ2 := C2 ∩ σ(T ).

Since T is contained in C2 , we have that σ1 = C1 is nonempty. Moreover,

if C2 ∩ σ(T ) was empty then σ(T ) would be contained in C1 and therefore
disjoint from T, which is impossible by Proposition 5.3.
Now, by the Riesz decomposition theorem there exist nontrivial T -invariant
closed subspaces M1 and M2 such that X = M1 ⊕ M2 , σ(T |M1 ) = σ1 and
σ(T |M2 ) = σ2 . By Proposition 2.28, T |M1 is then a hypercyclic operator
whose spectrum does not meet T, contradicting Proposition 5.3.

In particular, the spectrum of a hypercyclic operator cannot have isolated

points outside the unit circle. For chaotic operators, the same is true unre-
strictedly.

Proposition 5.7. Let T be a chaotic operator. Then its spectrum has no

isolated points and its point spectrum contains infinitely many roots of unity.

Proof. We first show that σp (T ) contains infinitely many roots of unity. Since
T has periodic points, Proposition 2.33 shows that σp (T ) contains at least
one root of unity. Now suppose that λ1 , . . . , λN , N ≥ 1, are the only roots of
unity in the point spectrum of T . Consider the polynomial

p(z) = (z − λ1 ) · · · (z − λN ).

Then, for any eigenvector x of T to an eigenvalue λk , k = 1, . . . , N , we have

that
5.2 Classes of non-hypercyclic operators on Banach spaces 141

p(T )x = (T − λ1 I) · · · (T − λN I)x = 0.
Again by Proposition 2.33 and by linearity of p(T ) we have that p(T )x = 0
for every periodic point of T . Since these points form a dense set in X we
obtain that p(T ) = 0, which contradicts Theorem 2.54.
Now suppose that λ is an isolated point of the spectrum of T . Then by
the Riesz decomposition theorem there are nontrivial T -invariant closed sub-
spaces M1 and M2 of X such that X = M1 ⊕ M2 and σ(T |M2 ) = {λ}. It
follows as in Proposition 2.28 (see also Exercise 2.2.8), that T |M2 is chaotic,
which is impossible since its spectrum is a singleton.

5.2 Classes of non-hypercyclic operators on Banach

spaces

Based on necessary conditions for hypercyclicity we can now show that certain
classes of operators cannot contain hypercyclic operators.
Throughout this section, T denotes an operator on a (real or complex)
Banach space X.

Power-bounded operators. An operator T is called

a contraction if T  ≤ 1,
quasinilpotent if limn→∞ T n 1/n = 0,
power bounded if supn≥0 T n  < ∞.
Over complex scalars, the spectral radius formula tells us that an operator is
quasinilpotent if and only if σ(T ) = {0}.
Clearly, every contraction and every quasinilpotent operator is power
bounded, and any orbit of a power-bounded operator is bounded and there-
fore not dense. In fact, for a quasinilpotent operator, all orbits tend to 0.

Proposition 5.8. No power-bounded operator is hypercyclic. In particular,

no contraction and no quasinilpotent operator is hypercyclic.

By Theorem 4.21, every nonzero multiple λD of the differentiation opera-

tor on H(C) is hypercyclic. Such a phenomenon cannot occur in the Banach
space setting. More precisely, if T is a nonzero operator on a Banach space
and λ ∈ K with |λ| ≤ 1/T  then λT is a contraction and therefore not
hypercyclic.
We apply this proposition to a specific family of operators.

Example 5.9. (Volterra integral operator) Let k : [0, 1] × [0, 1] → K be a

continuous function. Then the operator
 t
Vk : C[0, 1] → C[0, 1], (T f )(t) = k(t, s)f (s) ds, 0 ≤ t ≤ 1,
0
142 5 Necessary conditions for hypercyclicity and chaos

is called a Volterra integral operator, and k is its kernel; note that the values
k(t, s), s > t, are irrelevant. It is obvious that Vk defines an operator on
t
C[0, 1] with Vk  ≤ maxt∈[0,1] 0 |k(t, s)| ds < ∞. In the special case when k
is identically 1 we obtain the classical Volterra operator of Example 5.4.
A simple induction shows that every power of a Volterra integral operator
is a Volterra integral operator. More precisely, if kn is the kernel of (Vk )n ,
n ≥ 1, then
 t
kn+1 (t, s) = k(t, u)kn (u, s) du, s ≤ t, n ≥ 1;
s

in addition, if M = max0≤s≤t≤1 |k(t, s)|, then

Mn
|kn (t, s)| ≤ (t − s)n−1 , s ≤ t, n ≥ 1,
(n − 1)!

and hence
Mn
(Vk )n  ≤ ,
n!
which implies that limn→∞ (Vk )n 1/n = 0. We ask the reader to verify
these statements; see Exercise 5.2.1. Thus, each Volterra integral operator is
quasinilpotent and therefore not hypercyclic.

Finite-rank operators. We already know that there are no hypercyclic

operators on finite-dimensional spaces; see Corollary 2.59. This immediately
extends to finite-rank operators, that is, operators whose range has finite
dimension.
Proposition 5.10. No finite-rank operator is hypercyclic.
Proof. Let T be an operator with finite-dimensional range ran T , and let x
be a hypercyclic vector for T . Since T x ∈ ran T is also hypercyclic for T (see
Proposition 1.15), the restriction T |ran T : ran T → ran T is hypercyclic, in
contradiction to Corollary 2.59.

Compact operators. An operator T is called compact if the image of the

closed unit ball under T is relatively compact, that is, if its closure is compact.
This is equivalent to saying that, for any sequence (xn )n in X with xn  ≤ 1,
n ≥ 1, the sequence (T xn )n has a convergent subsequence.
In KN , N ≥ 1, and therefore in any finite-dimensional space, every
bounded set is relatively compact; hence every finite-rank operator is com-
pact. The following result therefore contains Proposition 5.10.
Theorem 5.11. No compact operator is hypercyclic.
For the proof we need several lemmas. The first one is fundamental for the
study of compact operators.
5.2 Classes of non-hypercyclic operators on Banach spaces 143

Lemma 5.12 (Riesz’s lemma). Let X be a Banach space, M a proper

closed subspace of X and ε > 0. Then there is some x ∈ X \ M such that

x = 1 and x − y ≥ 1 − ε for all y ∈ M .

Fig. 5.1 The proof of Riesz’s lemma

Proof. Let 0 < ε < 1. We choose any x0 ∈ X \ M . Since M is closed we have

that
d := inf x0 − y > 0.
y∈M

Thus there is a point y0 ∈ M such that

d
x0 − y0  ≤ .
1−ε
We now consider the point
x0 − y0
x := ;
x0 − y0 

see Figure 5.1. Then x = 1, and for any y ∈ M we have that
 x −y 
 0 0 
x − y =  − y
x0 − y0 
1
= x0 − (y0 + yx0 − y0 )
x0 − y0 
d
≥ ≥ 1 − ε,
x0 − y0 

where we have used the fact that y0 + yx0 − y0  ∈ M .

Lemma 5.13. Let T be a compact operator and λ ∈ K, λ = 0. Then λI − T

has closed range.

Proof. Since λI − T = λ(I − T /λ) and T /λ is also compact we can assume

that λ = 1.
144 5 Necessary conditions for hypercyclicity and chaos

Thus let y ∈ ran(I − T ). We choose points yn ∈ ran(I − T ), n ≥ 1, so that

yn → y as n → ∞. We then set

dn = inf{x ; (I − T )x = yn },

and we choose points xn with (I − T )xn = yn and xn  ≤ dn + 1, n ≥ 1.

We first show that (dn )n must contain a bounded subsequence. Otherwise
we have that dn → ∞ as n → ∞. We then set
1
zn = xn , n ≥ 1.
dn + 1

Since xn  ≤ dn + 1, n ≥ 1, the sequence (zn )n belongs to the closed unit

ball. By compactness of T , (T zn )n contains a convergent subsequence, which
we can assume to be the full sequence. Moreover, since (yn )n converges and
dn → ∞ we conclude that
1 1
zn − T zn = (I − T )zn = (I − T )xn = yn → 0 (5.1)
dn + 1 dn + 1

as n → ∞. This implies that the sequence (zn )n also converges. We call its
limit z. It then follows from (5.1) that z − T z = 0 and hence also that

(I − T )(xn − (dn + 1)z) = (I − T )xn − (dn + 1)(I − T )z = yn

for n ≥ 1. Therefore, by the definition of dn , we have that, for all n ≥ 1,

dn ≤ xn − (dn + 1)z = (dn + 1)zn − z,

which is impossible since zn − z → 0 and dn → ∞ as n → ∞.

Consequently, (dn )n contains a bounded subsequence; we can assume that
(dn )n itself is bounded. Thus there is some δ > 0 such that any δxn , n ≥ 1,
belongs to the closed unit ball. By compactness of T , the sequence (T (δxn ))n
contains a convergent subsequence, which we can assume to be the full se-
quence; hence also (T xn )n converges. But we also have that

xn − T xn = (I − T )xn = yn → y

as n → ∞. This implies that the sequence (xn )n itself converges. Calling its
limit x we obtain that

y = lim (I − T )xn = (I − T )x,

n→∞

so that y ∈ ran(I − T ). This shows that ran(I − T ) is closed.

The next lemma tells us that the eigenvalues of a compact operator can
only have 0 as an accumulation point.
5.2 Classes of non-hypercyclic operators on Banach spaces 145

Lemma 5.14. Let T be a compact operator and ε > 0. Then T has only a
finite number of eigenvalues λ with |λ| ≥ ε.

Proof. We assume that, on the contrary, there is a sequence of pairwise dis-

tinct eigenvalues λn with |λn | ≥ ε, n ≥ 1. Let xn be an eigenvector of T to
the eigenvalue λn , n ≥ 1. It is well known from linear algebra that the xn are
linearly independent. Therefore the spaces

Mn := span{x1 , . . . , xn }, n ≥ 1

form a strictly increasing sequence of subspaces of X. Since each Mn−1 (with

M0 = {0}) is finite dimensional and hence closed in Mn (see Appendix A),
it follows from Riesz’s lemma that there are points yn ∈ Mn \ Mn−1 , n ≥ 1,
such that
yn  = 1 and yn − y ≥ 12 for all y ∈ Mn−1 .
For n > m ≥ 1 we have that

T yn − T ym  = λn yn − (λn I − T )yn − T ym .

Now, since yn is a linear combination of x1 , . . . , xn , and since (λn I −T )xn = 0,

(λn I − T )yn belongs to Mn−1 , as does T ym . Hence
1
w := ((λn I − T )yn + T ym ) ∈ Mn−1 .
λn
It follows that, for n > m ≥ 1,
ε
T yn − T ym  = λn yn − λn w = |λn |yn − w ≥ .
2
But this is impossible because (yn )n belongs to the closed unit ball so that,
by compactness of T , (T yn )n must possess a convergent subsequence.

Our final lemma shows that, in an infinite-dimensional setting, 0 belongs

to the spectrum of a compact operator.

Lemma 5.15. Let T be a compact operator on an infinite-dimensional Ba-

nach space. Then 0 ∈ σ(T ).

Proof. If 0 is not in the spectrum then T is invertible. Thus the unit ball
of X, as the homeomorphic image (under T −1 ) of a relatively compact set,
is itself relatively compact. But this can only happen in finite-dimensional
spaces; see Appendix A.

We are now in a position to show that compact operators are never hy-
percyclic. Here we only give the proof in the case of complex Banach spaces.
In the real case the proof is outlined in Exercise 5.2.5; it uses the same ideas
as in the complex case but it is less straightforward.
146 5 Necessary conditions for hypercyclicity and chaos

Proof of Theorem 5.11 for complex Banach spaces. Let T be a hypercyclic

compact operator on a complex, necessarily infinite-dimensional, Banach
space, and let λ = 0 be in the spectrum of T . By Bourdon’s theorem and
Lemma 5.13, λI − T has closed and dense range, so that it is surjective.
Therefore λ must be an eigenvalue of T .
It now follows from Lemmas 5.14 and 5.15 that the spectrum of T consists
either of a sequence of points converging to 0, together with the point 0, or of
a finite set, including 0. In both cases the spectrum has connected components
that do not meet the unit circle. But this contradicts Kitai’s theorem.

Example 5.16. In Example 5.9 we showed that the Volterra integral operators
on C[0, 1] are quasinilpotent, and therefore not hypercyclic.
We want to show here that they are also compact, implying once more
that they are not hypercyclic. To this end let (fn )n be a sequence in C[0, 1]
with fn  ≤ 1, n ≥ 1. We have to construct a subsequence (gν )ν of (fn )n
such that (T gν )ν converges.
We begin the proof by a familiar diagonal process. Let (tj )j be a dense
sequence in [0, 1]. Since T is continuous, the sequence (T fn )n is bounded in
C[0, 1], and hence the sequence ((T fn )(t1 ))n is bounded in K. We can there-
fore extract a subsequence (n1,ν )ν such that ((T fn1,ν )(t1 ))ν converges in K.
In the same way, since ((T fn1,ν )(t2 ))ν is bounded there exists a subsequence
(n2,ν )ν of (n1,ν )ν such that ((T fn2,ν )(t2 ))ν converges. Proceeding inductively,
we find subsequences (nj,ν )ν of (nj−1,ν )ν such that ((T fnj,ν )(tj ))ν converges,
for j ≥ 2. Now define mν = nν,ν . Since, for any j ≥ 1, (mν )ν≥j is a subse-
quence of (nj,ν )ν we have that ((T fmν )(tj ))ν converges.
We claim that gν = fmν , ν ≥ 1, defines the required subsequence of
(fn )n . To show that (T gν )ν converges in C[0, 1] it suffices to show that it is a
Cauchy sequence. Thus, let ε > 0. Since the kernel k of T is continuous, and
hence uniformly continuous, on [0, 1] × [0, 1] there is some δ > 0 such that, if
|τ1 − τ2 | < δ, τ1 , τ2 ∈ [0, 1], then maxs∈[0,1] |k(τ1 , s) − k(τ2 , s)| < ε. Moreover,
M := maxt,s∈[0,1] |k(t, s)| < ∞. By the density of (tj )j there is some J ∈ N
such that for any t ∈ [0, 1] there is some j ≤ J such that |t − tj | < min(δ, ε).
The definition of the sequence (gν )ν shows that there is some N ∈ N such
that, for any j ≤ J and ν, μ ≥ N ,

|(T gν )(tj ) − (T gμ )(tj )| ≤ ε.

Now let t ∈ [0, 1]. Choose j ≤ J such that |t − tj | < min(δ, ε). Then, for
any ν ≥ 1,
  tj 
 
|(T gν )(t) − (T gν )(tj )| ≤  (k(t, s) − k(tj , s))gν (s) ds
0
 t 
 
+ k(t, s)gν (s) ds
tj

≤ ε + M ε,
5.2 Classes of non-hypercyclic operators on Banach spaces 147

where we have used that gν  ≤ 1, ν ≥ 1. We therefore have that, for any
ν, μ ≥ N ,

|(T gν )(t) − (T gμ )(t)| ≤ |(T gν )(t) − (T gν )(tj )| + |(T gν )(tj ) − (T gμ )(tj )|

+ |(T gμ )(tj ) − (T gμ )(t)|
≤ (2M + 3)ε.

This shows that (T gν )ν is a Cauchy sequence in C[0, 1], as required.

Power-compact operators. An operator T is called power compact if some

power T n , n ≥ 1, of T is compact.

Proposition 5.17. No power-compact operator is hypercyclic.

The proof can be given as in the case of compact operators, taking account
of the fact that σ(T n ) = σ(T )n by the spectral mapping theorem. Alterna-
tively, the result is also an immediate consequence of the compact case and
of Ansari’s theorem that we will obtain in the next chapter; see Theorem 6.2.
By Ansari, T can only be hypercyclic if T n is, which is impossible if T n is
compact.

Finite-rank perturbations. We have noted above that finite-rank opera-

tors cannot be hypercyclic. The same is true for finite-rank perturbations of
multiples of the identity I, that is, operators of the form

λI + F, λ ∈ K,

where F is a finite-rank operator.

Proposition 5.18. No finite-rank perturbation of a multiple of the identity

is hypercyclic.

In fact, an even more general result can be established; we leave the proof
to the reader (see Exercise 5.2.8).

Compact perturbations. We next consider compact perturbations of mul-

tiples of the identity. We first show that hypercyclicity imposes a severe
restriction on such an operator.

Lemma 5.19. Let T be a hypercyclic operator of the form

T = λI + K,

where λ ∈ K and K is compact. Then |λ| = 1 and K is quasinilpotent.

We restrict our attention again to the complex case and refer to Exercise
5.2.9 for the real case.
148 5 Necessary conditions for hypercyclicity and chaos

Proof (for complex Banach spaces). Let T = λI + K be hypercyclic. For

μ ∈ C we have that
μI − T = (μ − λ)I − K.
As in the proof of Theorem 5.11 it then follows from Bourdon’s theorem and
Lemmas 5.13, 5.14 and 5.15 that λ ∈ σ(T ) and that every point μ ∈ σ(T )
with μ = λ is an isolated point of the spectrum.
Consequently, if μ ∈ σ(T ) with μ = λ then {μ} and σ(T ) \ {μ} form a
partition of σ(T ) into nonempty closed subsets. By the Riesz decomposition
theorem there are nontrivial T -invariant closed subspaces M1 and M2 of X
such that X = M1 ⊕ M2 with σ(T |M1 ) = {μ}. It follows from Proposition
2.28 that T |M1 is a hypercyclic operator and hence, by Corollary 2.59, that
M1 is infinite-dimensional.
On the other hand, M1 is also invariant under K = T − λI. Since M1 is
closed, K|M1 is a compact operator, and σ(K|M1 ) = σ(T |M1 − λI|M1 ) = {μ −
λ} does not contain 0. By Lemma 5.15 this is only possible if the underlying
space M1 is finite dimensional, a contradiction.
Therefore the spectrum of T cannot contain points other than λ, and thus
σ(T ) = {λ}. Kitai’s theorem then forces |λ| to be 1. Moreover, σ(K) = {0}.

We will show later that there are indeed hypercyclic operators of the form
T = I + K with K compact; see Example 8.4. But such an operator cannot
be chaotic.

Proposition 5.20. No compact perturbation of a multiple of the identity is

chaotic.

Proof. Since the complexification of a compact perturbation of a multiple of

the identity is again of this type it suffices, by Corollary 2.51, to study the
complex case.
Now, by Lemma 5.19, any chaotic operator T = λI + K, K compact, must
satisfy σ(T ) = {λ}, which is impossible by Proposition 5.7.

Lower triangular operators. For operators on Banach spaces with a basis

it can be very instructive to study their matrix representations. Specifically,
∗
let X be a Banach space with a basis (e n )n≥1 , and let (en )n≥1 be the cor-
∗ ∞
responding coefficient functionals en ( k=1 xk ek ) = x n , n ≥ 1; see Appendix
∞
A. Then we have for any operator T on X and x = k=1 xk ek ∈ X,
∞
∞ 
∞  ∞ 
∞ 
Tx = T x, e∗n en = T ek , e∗n xk en = ank xk en ,
n=1 n=1 k=1 n=1 k=1

where ank = T ek , e∗n , n, k ≥ 1. Therefore, in terms of the coefficient se-

quences (xn )n of x and (yn )n of y = T x, T may be represented by an infinite
matrix
5.2 Classes of non-hypercyclic operators on Banach spaces 149

A = (ank )n,k≥1 .
Now suppose that in the nth row of the matrix A, all off-diagonal
∞ elements
are 0, that is, ank = 0 for all k = n. Then we have for x = k=1 xk ek ∈ X
that ∞

∗ ∗ ∗
x, T en  = T x, en  = ank xk = ann xn = ann x, e∗n .
k=1

This shows that e∗n is an eigenvector of T ∗ . In view of Proposition 5.1, T

cannot be hypercyclic. A very particular, but quite common, class of matrices
with the stated property consists of the lower triangular matrices, that is,
matrices A = (ank )n,k≥1 with ank = 0 for k > n, n ≥ 1: here, the first row
has no nonzero off-diagonal elements. The corresponding operators are called
lower triangular operators.

Proposition 5.21. Let T be an operator on a Banach space X with a basis.

If the matrix A representing T has a row all of whose off-diagonal elements
are 0 then T is not hypercyclic. In particular, no lower triangular operator is
hypercyclic.

As an illustration, on X = p , 1 ≤ p < ∞, and c0 , the weighted forward

shifts Fw (never hypercyclic) and the weighted backward shifts Bw (hyper-
cyclic for certain weights) are represented by the matrices
⎛ ⎞ ⎛ ⎞
0 0 0 0 ... 0 w2 0 0 . . .
⎜w1 0 0 0 . . .⎟ ⎜0 0 w3 0 . . .⎟
⎜ ⎟ ⎜ ⎟
⎜ 0 w 2 0 0 . . .⎟ ⎜ ⎟
⎜ ⎟ and ⎜0 0 0 w4 . . .⎟ ,
⎜ 0 0 w3 0 . . .⎟ ⎜0 0 0 0 . . .⎟
⎝ ⎠ ⎝ ⎠
.. .. .. .. .. .. .. ..
. . . . . . . .
∞
respectively; see Section 4.1. For complex scalars, if ϕ(z) = n=0 an z n is
holomorphic on a neighbourhood of D then ϕ(B) is represented by the matrix
⎛ ⎞
a0 a1 a2 a3 ...
⎜0 . . .⎟
⎜ a0 a1 a2 ⎟
⎜ ⎟
⎜0 0 a0 a1 . . .⎟
⎜ ⎟
⎜0 0 0 a0 . . .⎟
⎝ ⎠
.. .. .. ..
. . . .

(see Section 4.4), which are special Toeplitz matrices; chaos for ϕ(B) is char-
acterized in Theorem 4.43.

n , 1 < p < ∞, or c0 , the Cesàro

p
Example 5.22. On the sequence spaces
1
operator C1 is given by C1 x = ( n k=1 xk )n . It associates to each sequence
its sequence of arithmetic means. In the canonical bases, the operator is
represented by the lower triangular matrix
150 5 Necessary conditions for hypercyclicity and chaos
⎛ ⎞
1 0 0 0 ...
⎜1 1 0
⎜2 2 0 . . .⎟
⎟
⎜1 1 1 ⎟
⎜3 3 3 0 . . .⎟ .
⎜ ⎟
⎜1 1 1 1 ⎟
⎝4 4 4 4 . . .⎠
.. .. .. ..
. . . .

The Cesàro operator is therefore not hypercyclic. Of course, it needs to be

verified that C1 is indeed an operator on the stated sequence spaces. For c0
this is easily shown, for the spaces p this is the content of the so-called Hardy
inequality.

5.3 Classes of non-hypercyclic operators on Hilbert

spaces

An important feature of Hilbert spaces is that they are self-dual; a more

precise statement is contained in the Riesz representation theorem; see Ap-
pendix A. As a consequence, the adjoint of an operator can be considered
as an operator on the space itself. This leads to some interesting classes of
operators that have no analogue for general Banach spaces.
Throughout this section, if T is an operator on a Hilbert space H then T ∗
denotes the (Hilbert space) adjoint of T ; see Appendix A.
Example 5.23. Every operator T on CN , N ≥ 1, can be represented by a
matrix A = (ank )1≤n,k≤N via the usual matrix product:
 
T x = Ax = ank xk .
n
k≥1

If CN carries its usual inner product then we have for all x, y ∈ CN ,

   
T x, y = ank xk yn = xk ank yn .
n≥1 k≥1 k≥1 n≥1

This shows that the adjoint T ∗ is represented by the conjugate transpose

t
A of the matrix A. In the same way, on the real space RN , the adjoint
corresponds to the transpose of the representing matrix.
t t
In linear algebra, a matrix A is called normal if A A = AA . In analogy,
a general Hilbert space operator T is called normal if it commutes with its
adjoint,
T ∗T = T T ∗.
An important class of normal operators are the self-adjoint operators, that
is, operators T satisfying
5.3 Classes of non-hypercyclic operators on Hilbert spaces 151

T = T ∗.
In order to explore these notions we need a preliminary result.
Lemma 5.24. Let T be an operator on a Hilbert space H.
(a) Suppose that H is a complex Hilbert space. If

T x, x ∈ R for all x ∈ H,

then T is self-adjoint. If

T x, x = 0 for all x ∈ H,

then T = 0.
(b) Suppose that H is a real Hilbert space. If T is self-adjoint and

T x, x = 0 for all x ∈ H,

then T = 0.
Proof. For any x, y ∈ H we have that

T (x + y), x + y − T x, x − T y, y = T x, y + T y, x. (5.2)

In the real case, this already implies (b). In fact, the left-hand side is zero by
assumption, and since T is self-adjoint and the inner product is symmetric,
we deduce that, for all x, y ∈ H,

0 = T x, y + T y, x = T x, y + y, T x = T x, y + T x, y = 2T x, y,

which can only be true if T = 0.

In the complex case, we replace y by iy in (5.2) to obtain that, for any
x, y ∈ H,

T (x + iy), x + iy − T x, x − T (iy), iy = i(−T x, y + T y, x). (5.3)

Under the first assumption in (a), the left-hand sides in (5.2) and (5.3) are
both real-valued. Hence we deduce that

ImT x, y = − ImT y, x and ReT x, y = ReT y, x

and therefore
T x, y = T y, x = x, T y
for all x, y ∈ H. This shows that T ∗ = T .
Under the second assumption in (a) we obtain from (5.2) and (5.3) that

T x, y + T y, x = 0 = −T x, y + T y, x

and hence
152 5 Necessary conditions for hypercyclicity and chaos

T x, y = 0
for all x, y ∈ H. This shows that T = 0.

It follows from Proposition A.8 that T ∗ T − T T ∗ is self-adjoint. Hence the

lemma tells us that, both in the real and the complex case, T ∗ T = T T ∗ if
and only if (T ∗ T − T T ∗ )x, x = 0 for all x ∈ H. But since

(T ∗ T − T T ∗ )x, x = T ∗ T x, x − T T ∗ x, x = T x, T x − T ∗ x, T ∗ x
= T x2 − T ∗ x2 , (5.4)

we have shown the following.

Proposition 5.25. An operator T on a Hilbert space H is normal if and

only if, for all x ∈ H,
T x = T ∗ x.

An operator T on a Hilbert space H is called positive if, for all x ∈ H,

T x, x ≥ 0.

One then writes

T ≥ 0.
By Lemma 5.24, every positive operator on a complex Hilbert space is self-
adjoint. However, the matrix ( 10 21 ) defines a positive operator on the real
Hilbert space R2 that is not self-adjoint; see also Example 5.33.
It seems natural to weaken the normality condition T ∗ T − T T ∗ = 0 to
a mere positivity assumption: an operator T on a Hilbert space H is called
hyponormal if
T ∗ T − T T ∗ ≥ 0.

Remark 5.26. Of course, one may just as well consider operators that satisfy

T T ∗ − T ∗ T ≥ 0.

That these are of less interest in the present context will become clear in
Exercise 5.3.5.

Clearly, every normal operator is hyponormal. Equation (5.4) implies the

following characterization.

Proposition 5.27. An operator T on a Hilbert space H is hyponormal if and

only if, for all x ∈ H,
T x ≥ T ∗ x.

Now we derive an interesting property of hyponormal operators. By the

Cauchy–Schwarz inequality, any operator T satisfies
5.3 Classes of non-hypercyclic operators on Hilbert spaces 153

T x2 = T x, T x = T ∗ T x, x ≤ T ∗ T xx

for all x ∈ H. Proposition 5.27, applied to T x, then shows that for hyponor-
mal operators we have that

T x2 ≤ T 2 xx

for all x ∈ H. Operators T satisfying this condition are called paranormal.

Thus every hyponormal operator is paranormal.
In Table 5.1 we collect the various classes of Hilbert space operators that
we have introduced, in increasing order of generality, along with their char-
acterizing conditions.

self-adjoint T = T ∗
normal T ∗T = T T ∗ ∀ x ∈ H, T x = T ∗ x
hyponormal T ∗ T − T T ∗ ≥ 0 ∀ x ∈ H, T x ≥ T ∗ x
paranormal ∀ x ∈ H, T x2 ≤ T 2 xx
Table 5.1 Classes of Hilbert space operators

 0 1
Remark 5.28. The operator on K2 that is given by the matrix A = −1 0
is normal, but not self-adjoint. For distinguishing the remaining classes of
operators we refer to Exercises 5.3.1 and 5.3.2.

The defining condition for paranormal operators is obviously not specific

to the Hilbert space setting, which leads us back to general Banach spaces.

Definition 5.29. An operator T on a Banach space X is called paranormal

if, for all x ∈ X,
T x2 ≤ T 2 xx.

The following is the main result of this section.

Theorem 5.30. No paranormal operator on a Banach space is hypercyclic.

Proof. We show that the orbits of paranormal operators are too well behaved
for the operator to be hypercyclic. Indeed, let T be a paranormal operator
on a Banach space X and x ∈ X. Suppose that, for some n ≥ 0, T n+1 x >
T n x, which implies, in particular, that T n+1 x = 0 and T n x = 0. Applying
the definition of paranormality to T n x we obtain that

T n+1 x2
T n+2 x ≥ > T n+1 x.
T n x

Therefore, the orbit of any x ∈ X is either decreasing in norm, or strictly

increasing in norm from some index on. As a consequence, no orbit can be
dense.

154 5 Necessary conditions for hypercyclicity and chaos

One may also arrive at the same conclusion from another point of view.
Considering T n x in the definition of paranormality and taking logarithms we
see that, if T is a paranormal operator, then, for any x ∈ X and n ≥ 0 with
T n+2 x = 0,
1
log T n+1 x ≤ (log T n+2 x + log T n x).
2
Thus, for any x ∈ X whose orbit does not end up in the origin, the sequence
(log T n x)n is convex. This leads to the behaviour of the sequence (T n x)n
found in the proof.
Corollary 5.31. No self-adjoint operator, no normal operator, no hyponor-
mal operator on a Hilbert space is hypercyclic. No positive operator on a
complex Hilbert space is hypercyclic.
We finish this section, and indeed the chapter, with two examples.
Example 5.32. We consider the bilateral weighted backward shift

Bw x = (wn+1 xn+1 )n∈Z

on the Hilbert space 2 (Z), where w = (wn )n∈Z is a weight sequence; see
Section 4.1. An easy computation shows that
∗
Bw = Fv ,

the weighted forward shift with weight sequence v = (wn+1 )n∈Z ; see Remark
4.14. (Note that this is the Hilbert space adjoint of Bw ; its Banach space
adjoint was determined in the proof of Proposition 4.16.)
We calculate that, for x ∈ 2 (Z) and λ ≥ 0,
 2 2    
Fv Bw − 2λFv Bw + λ2 I x, x = |wn |2 |wn−1 |2 − 2λ|wn |2 + λ2 |xn |2 .
n∈Z

Hence, by Exercise 5.3.4(b), Bw is paranormal if and only if, for all λ ≥ 0

and n ∈ Z,
|wn |2 |wn−1 |2 − 2λ|wn |2 + λ2 ≥ 0,
which by Exercise 5.3.4(a) is equivalent to

4|wn |4 ≤ 4|wn |2 |wn−1 |2 for all n ∈ Z.

Therefore, Bw is paranormal if and only if (|wn |)n∈Z is a decreasing sequence.

A comparison with Example 4.15 shows that this produces some, but not all
by far, non-hypercyclic bilateral weighted shifts.
Example 5.33. We know from Theorem 4.43 that the operator I + B, with B
the backward shift, is chaotic on 2 . This result holds for real and for complex
sequences. In the real case, however, the Cauchy–Schwarz inequality implies
that, for all x ∈ 2 ,
Exercises 155

(I + B)x, x = x2 + Bx, x ≥ x2 − Bxx ≥ 0,

so that I + B is positive. On real Hilbert spaces, therefore, positive operators

can be hypercyclic.

Exercises

Note: Exercises 5.0.1–5.0.9 concern general spectral theory; see Appendix B.

Exercise 5.0.1. Consider the operator T given by T (x1 , x2 ) = (−x2 , x1 ) on the real
Banach space R2 . Show that σ(T ) = ∅ and σ(T 2 ) = {−1}. Thus, some results of
Appendix B do not hold in the real case.

Exercise 5.0.2. Let B be the backward shift operator and F the forward shift operator
on any of the sequence spaces X = p , 1 ≤ p < ∞, or X = c0 . Show that σp (B) = D
and σ(B) = D. Determine also σp (F ) and σ(F ). (Hint: Find F ∗ .)

Exercise 5.0.3. Let Bw be a weighted shift on X = p , 1 ≤ p < ∞, or c0 , given by

Bw (xn )n = (wn+1 xn+1 )n ; see Section 4.1. Suppose that (wn )n is a positive, decreasing
sequence with limn→∞ wn = r ≥ 0. Determine σ(Bw ).

Exercise 5.0.4. Let K be a nonempty compact subset of C. Show that there is an

operator T on the Hilbert space 2 such that σ(T ) = K. (Hint: Choose T (xn )n =
(an xn )n .)

Exercise 5.0.5. Prove the spectral mapping theorem. (Hint: Look at the proof of the
point spectral mapping theorem.)

Exercise 5.0.6. Give a direct proof of the spectral mapping theorem for polynomials,
without the use of functional calculus. (Hint: For λ ∈ C there are c, μ1 , . . . , μN ∈ C such
that λ − f (z) = c(μ1 − z) · · · (μN − z); use the corresponding factorization for λI − T .)

Exercise 5.0.7. If T is an invertible operator then 0 ∈/ σ(T ) and σ(T −1 ) = σ(T )−1 :=
{ z ; z ∈ σ(T )}. Give two proofs of the identity, one direct and one using the spectral
1

mapping theorem.

Exercise 5.0.8. Let T be an operator and f a holomorphic function on a neighbourhood

of σ(T ). Show that if f (T ) = 0 then f = 0 on σ(T ). Use the Volterra operator (see
Example 5.4), to show that the converse implication is false.

Exercise 5.0.9. Give a direct proof of the point spectral mapping theorem for polyno-
mials using the ideas of Exercise 5.0.6.

Exercise 5.1.1. Let T be a mixing operator on a (real or complex) Banach space. Then,
for every nonzero vector x∗ ∈ X ∗ , (T ∗ )n x∗  → ∞ as n → ∞. (Hint: Suppose that
(T ∗ )nk x∗  ≤ M and consider U = {x ; x < 1}, V = {x ∈ X ; | x, x∗
| > M }.)

Exercise 5.1.2. (a) Show that the Volterra operator also defines an operator on the
space L∞ [0, 1] of essentially bounded measurable functions on [0, 1]. Moreover, show
that V  = 1 for V as an operator on L1 and on L∞ .
(b) Show that, on any space C[0, 1] or Lp [0, 1], 1 ≤ p ≤ ∞, the Volterra operator has
no eigenvalues.
156 5 Necessary conditions for hypercyclicity and chaos

(c) Confirm the representation of V n given in Example 5.4.

(d) For p = 1 and p = ∞, show that V n  ≤ n! 1
; deduce that σ(V ) = {0}.
1 −1/p −1/q
(e) For 1 < p < ∞, show that V  ≤ n! p
n
q (1 − 1/(pn))−1/q ; deduce that
σ(V ) = {0}.
(f) Show directly that, on C[0, 1], λI − V is injective for any λ ∈ C, and that it is
surjective precisely for λ = 0. (Hint: If λ = 0, show that λI − V maps the function
t
t → λ12 0 e(t−s)/λ g(s) ds + λ1 g(t) to g.)

Exercise 5.1.3. Deduce the following from Kitai’s theorem.

(a) Every non-invertible hypercyclic operator has an uncountable spectrum.
(b) If the spectrum of a hypercyclic operator is finite or countable then it is contained
in T.

Exercise 5.1.4. A vector x ∈ X is called an irregular vector of an operator T on X

if lim inf n→∞ T n x = 0 and lim supn→∞ T n x = ∞. For example, any hypercyclic
vector is irregular. Show the following.
(a) Let S : X → X and T : Y → Y be operators. If S ⊕ T has an irregular vector,
then so does at least one of the operators S and T .
(b) If T has irregular vectors then σ(T ) meets the unit circle.
(c) The analogue of Kitai’s theorem is not true for operators with irregular vectors.
(Hint: Consider T ⊕ (λI) with T hypercyclic.)
(d) Unilateral weighted forward shifts Fw (see Remark 4.11), can have irregular
vectors.

Exercise 5.1.5. Let T be an ε-hypercyclic operator; see Exercise 2.2.10. Show that
every connected component of σ(T ) meets the unit circle.

Exercise 5.1.6. Let T be a J-class operator; see Exercise 2.2.11. Show that σ(T ) meets
the unit circle, but that not necessarily every connected component of σ(T ) does. (Hint:
Avoid the problem presented in Exercise 2.2.11(a) by modifying the proof of Proposition
5.3.)

Exercise 5.2.1. Prove the statements on the Volterra integral operator made in Exam-
ple 5.9.

Exercise 5.2.2. Let X and Y be Banach spaces. An operator T : X → Y is called a

finite-rank operator if its range ran T is finite dimensional. Show that T : X → Y is a
∗ ∗ ∗
n if and∗ only if there are y1 , . . . , yn ∈ Y and x1 , . . . , xn ∈ X , n ≥ 1,
finite-rank operator
such that T = k=1
· , xk
yk .

Exercise 5.2.3. Let X be a Banach space and K(X) the space of compact operators
on X. Show the following:
(i) the identity operator on X is compact if and only if X is finite dimensional; (Hint:
Use a result from Appendix A.)
(ii) if K, L ∈ K(X), λ ∈ K and T ∈ L(X) then λK, K + L, T K and KT all belong to
K(X) (one says that K(X) is an ideal in L(X)).

Exercise 5.2.4. Let X be a Banach space and K(X) the space of all compact operators
on X. Show that K(X) is a closed subspace of L(X), the space of all operators on X,
endowed with the operator norm topology. Deduce that a limit, in the operator norm,
of finite-rank operators is compact. (Hint: Look at the argument contained in Example
5.16.)
Exercises 157

Exercise 5.2.5. Show that no compact operator on a real Banach space can be hyper-
cyclic by proceeding as follows. Suppose that T is a hypercyclic compact operator on an
infinite-dimensional real Banach space X, and let T : X →X  denote its complexifica-
tion. By Exercise 2.6.4, T is 2-hypercyclic.
(a) Show that T is compact.
(b) Show that, for every ε > 0, σ(T) ∩ {z ∈ C ; |z| ≥ ε} is a finite set. (Hint: Use
Exercise 2.6.4.)
(c) By Lemma 5.15, 0 ∈ σ(T). If σ(T) = {0} deduce that all orbits under T tend to
0, contradicting the 2-hypercyclicity of T.
(d) If {0}  σ(T), show that there are nontrivial T-invariant closed subspaces M1
and M2 with X  = M1 ⊕ M2 such that σ(T|M1 ) ⊂ D. Prove that T|M1 is 2-hypercyclic
and deduce a contradiction as in (c).

Exercise 5.2.6. Prove the Arzelà–Ascoli theorem: a subset A of C[0, 1] is relatively

compact if and only if it is bounded and equicontinuous. Here, A is called equicontinuous
if for any ε > 0 there is some δ > 0 such that, for any f ∈ A and any s, t ∈ [0, 1] with
|s − t| < δ, |f (s) − f (t)| < ε. (Hint: Look at the argument contained in Example 5.16.)

Exercise 5.2.7. Let ϕ : [0, 1] → [0, 1] be a continuous function. Then the Volterra
composition operator Vϕ : C[0, 1] → C[0, 1] is defined as
 ϕ(t)
Vϕ f (t) = f (s) ds, 0 ≤ t ≤ 1.
0

(a) Show that no Volterra composition operator is hypercyclic on C[0, 1]. (Hint:
Vϕ = Cϕ ◦ V .)
(b) Now let ϕ(x) = xα , α > 0, and consider Vϕ as an operator on C0 [0, 1[, the
space of continuous functions f on [0, 1[ with f (0) = 0; it is a Fréchet space under the
seminorms pn (f ) = supt∈[0,1−1/n] |f (t)|, n ≥ 1. Show the following:
(i) for α ≥ 1, Vϕ is not hypercyclic on C0 [0, 1[;
(ii) for α < 1, Vϕ is hypercyclic on C0 [0, 1[.
(Hint for (ii): Use Kitai’s criterion with X0 the set of polynomials in C0 [0, 1[ and Y0 =
span{xν ; ν > α/(1 − α)}; use the Weierstrass approximation theorem to show that Y0
is dense.)

Exercise 5.2.8. Show that no hypercyclic operator can commute with a finite-rank
operator. Deduce Proposition 5.18 from this. (Hint: Use a commutative diagram.)

Exercise 5.2.9. Prove Lemma 5.19 for real Banach spaces. (Hint: Use the ideas of
Exercise 5.2.5.)

Exercise 5.2.10. Let Bw be a weighted shift on X = p , 1 ≤ p < ∞, or c0 , given by

Bw (xn )n = (wn+1 xn+1 )n . Show that Bw is compact if and only if limn→∞ wn = 0.
Show that Bw is quasinilpotent in that case. (Hint: Use Exercise 5.2.4.)

Exercise 5.2.11. The aim of this and the next exercise is to get a better understanding
of the spectrum of compact operators. Show here that the adjoint T ∗ of a compact
operator is compact. (Hint: Let x∗n  ≤ 1. For the unit ball BX of X, K := T (BX ) is
a compact metric space. Note that x∗n |K ∈ C(K), the space of continuous functions on
K, endowed with the sup-norm. Apply the Arzelà–Ascoli theorem (see Exercise 5.2.6),
which remains true for C(K).)
158 5 Necessary conditions for hypercyclicity and chaos

Exercise 5.2.12. Let T be a compact operator on an infinite-dimensional complex Ba-

nach space. Show that its spectrum is either finite, including 0, or it consists of a se-
quence of points converging to 0, again including 0. Also, every nonzero spectral point is
an eigenvalue. (Hint: Let |λ| ≥ ε. If λ is not an eigenvalue then λI − T is non-surjective
of closed range. By the proof of Lemma 2.53, λ is an eigenvalue of the compact operator
T ∗ ; see the previous exercise. By Lemma 5.14 there are only finitely many such λ, so
that λ is isolated. Apply the Riesz decomposition theorem to find a restriction of T |M
with spectrum {λ}. By Lemma 5.15, dim M < ∞, so that λ is an eigenvalue of T .)

Exercise 5.3.1. An operator T on a Hilbert space H is called quasinormal if T (T ∗ T ) =

(T ∗ T )T . Show the following:
(i) any normal operator is quasinormal, and any quasinormal operator is hyponormal;
(ii) the unilateral forward shift on 2 , F (xn )n = (0, x1 , x2 , . . .), is quasinormal but not
normal;
(iii) the operator B + 2F on 2 , where B is the backward shift, is hyponormal but not
quasinormal, and hence not normal.
(Hint: For the nontrivial implication in (i), consider separately x ∈ ran T and x ∈
(ran T )⊥ , the orthogonal complement of ran T .)

Exercise 5.3.2. An operator T on a Hilbert space H is called quasihyponormal if

T ∗ (T ∗ T − T T ∗ )T ≥ 0. Show the following:
(i) an operator T is quasihyponormal if and only if, for all x ∈ H, T T x ≥ T ∗ T x;
(ii) any hyponormal operator is quasihyponormal, and any quasihyponormal operator
is paranormal;
(iii) let H = 2 (N, C2 ) be the Hilbert space of all sequences (xn )n≥1 with xn ∈ C2 ,
∞
n ≥ 1, such that n=1
xn 2 < ∞, where  ·  is the Euclidean norm in C2 , and
define the operator T : H → H by T (xn )n = (0, Ax1 , Bx2 , Bx3 , Bx4 , . . .), where
A = ( 10 00 ) and B = √12 ( 11 11 ); show that T is quasihyponormal and therefore also
paranormal, but not hyponormal.

Exercise 5.3.3. An operator T on a Banach space X is called power-regular if, for any
x ∈ X, (T n x1/n )n converges. It is called normaloid if, for any n ≥ 1, T n  = T n .
Show the following:
(i) every paranormal operator is power regular;
(ii) every paranormal operator is normaloid;
(iii) normaloid operators can be hypercyclic.
(Hint: In (i), study the behaviour of the sequence (T n+1 x/T n x)n≥0 ; in (ii), prove
inductively that T n x ≥ T xn , for all n ≥ 1 and all unit vectors x.)

Exercise 5.3.4. (a) Let p(λ) = aλ2 + bλ + c be a real polynomial with a = 0 and b ≤ 0.
Show that p(λ) ≥ 0 for all λ ≥ 0 if and only if a > 0 and b2 ≤ 4ac.
(b) Deduce from (a) that an operator T on a Hilbert space H is paranormal if and
only if, for all λ ≥ 0,
T ∗2 T 2 − 2λT ∗ T + λ2 I ≥ 0.

Exercise 5.3.5. An operator T on a Hilbert space H is called cohyponormal if T ∗ is

hyponormal, that is, if T T ∗ − T ∗ T ≥ 0, or, equivalently, T ∗ x ≥ T x for all x ∈ H.
Show that a cohyponormal operator can be hypercyclic.

Exercise 5.3.6. Show that for bilateral weighted backward shifts on 2 (Z), hyponor-
mality and paranormality are equivalent properties.

Exercise 5.3.7. Show that the Volterra operator on L2 [0, 1] is not paranormal.
Sources and comments 159

Sources and comments

Section 5.1. It seems fair to say that all the results in this chapter were either found,
or at least initiated, by C. Kitai in her PhD thesis [215]. Some of her results were also
found, independently, by Matache [233].
In particular, Theorem 5.6 is due to Kitai [215]. Lemma 5.5 is not as obvious as
it may at first seem. While one implication is standard (see for example [250, p. 74,
Corollary 2]), the other implication can be derived from [250, p. 83, Corollary 1].
Quite remarkably, Shkarin [287] has recently shown that Kitai’s necessary spectral
condition actually characterizes spectra of hypercyclic operators on Hilbert spaces.

Theorem 5.34. Let K ⊂ C be a nonempty compact set. There exists a hypercyclic

operator T on a complex Hilbert space such that σ(T ) = K if and only if every connected
component of K meets the unit circle.

Dilworth and Troitsky [135] have shown that Kitai’s theorem remains true for weakly
hypercyclic operators, that is, operators with a weakly dense orbit. Herrero [194] has
shown that for any supercyclic operator (on a complex Hilbert space) there is some
r ≥ 0 such that every connected component of its spectrum meets the circle |z| = r;
see also Feldman, Miller and Miller [151, Theorem 6.2]. Herrero also obtained a spec-
tral characterization of all operators on complex Hilbert spaces that are limits, in the
operator norm, of hypercyclic operators.
Concerning the other necessary conditions found in Section 5.1, Proposition 5.1, part
(i) is due to Kitai [215] while part (ii) was observed by Bourdon; see [106, p. 1446] and
[96, Lemma 3.1]. Proposition 5.7 is implicit in Bonet, Martínez and Peris [82].

Section 5.2. The main result of this section, Theorem 5.11, is once more due to Kitai
[215] (for complex Banach spaces); see also Matache [233]. The real case was provided
by Bonet and Peris [85]. Kitai’s proof is quite short but it draws heavily on the Riesz
spectral theory of compact operators by which the spectrum of any compact operator on
an infinite-dimensional complex Banach space is either finite, including 0, or it consists
of a sequence converging to 0, again including 0; see [116, Chapter VII, § 7]. In the
approach chosen here, Bourdon’s theorem allows us to avoid some parts of the Riesz
theory; but see Exercise 5.2.12.
Proposition 5.17, Proposition 5.18, Lemma 5.19 and Proposition 5.20 are due to Kitai
[215], Chan and Shapiro [106, p. 1445, p. 1446] and Martínez and Peris [229].
It is interesting to note that, by Shkarin [289], there are hypercyclic finite-rank per-
turbations of unitary operators on complex Hilbert spaces; by Grivaux [175], the per-
turbation may even be of rank 1.
Concerning specific operators, the Volterra operator (on Lp [0, 1], 1 ≤ p < ∞) and the
Cesàro operator (on p , 1 < p < ∞) are not even supercyclic; see Gallardo and Montes
[159] and León, Piqueras and Seoane [224]. As for Example 5.22, Hardy’s inequality can
be found in [193].

Section 5.3. We refer to Kubrusly [218] for a discussion of various types of normality
conditions for Hilbert space operators. It was once more Kitai [215] who showed that no
hyponormal operator can be hypercyclic. This was extended in two respects by Bourdon
[92, p. 352] who proved that no paranormal operator can even be supercyclic. Concern-
ing Example 5.33, it should be noted that some authors include self-adjointness in the
definition of positive operators on real Hilbert spaces. In that case, of course, no positive
operator is hypercyclic.

Exercises. We have taken Exercise 5.1.1 from Bonet [79] and Exercise 5.1.3 from Mat-
ache [234]. The notion of an irregular vector was introduced by Beauzamy [48, p. 41];
160 5 Necessary conditions for hypercyclicity and chaos

Exercises 5.1.4 and 5.1.5 are due to Prǎjiturǎ [261], while Exercise 5.1.6 is taken from
Costakis and Manoussos [119]. The result of Exercise 5.2.5 is due to Bonet and Peris
[85]. Exercise 5.2.7 is taken from Herzog and Weber [201], Exercise 5.2.8 from Shapiro
[281, 2.9]. The result of Exercise 5.2.11 is one half of Schauder’s theorem; see [116].
For the background to Exercises 5.3.1, 5.3.2 and 5.3.5 we refer to Kubrusly [218]. As
for Exercise 5.3.3, power-regular operators were introduced by Atzmon [17], who also
showed that operators with a countable spectrum are power regular. As a consequence,
power-regular operators can be hypercyclic; see Example 8.4. That every paranormal
operator is power regular and normaloid is due to Bourdon [92, p. 352] and Istrăţescu,
Saitô and Yoshino [206], respectively. Incidentally, it was in the latter article that the
class of paranormal operators was introduced (as operators of class (N)). The charac-
terization of paranormal operators in Exercise 5.3.4 is due to Ando [8].

Extensions. We have decided to restrict our attention in this chapter to Banach spaces.
But the main results in Section 5.2 remain true in arbitrary Fréchet spaces, and even in
all locally convex spaces; see Chapter 12. Indeed, no compact (and therefore no power-
compact) operator on a locally convex space can be hypercyclic (see Bonet and Peris
[85]), and no compact perturbation of a multiple of the identity can be chaotic; see
Martínez and Peris [229]. For an idea of the proofs we refer to Exercise 12.2.3. Proposi-
tions 5.10 and 5.18 also remain true; their proofs extend immediately.

To finish, we briefly review the hypercyclicity status of further classes of operators

that are commonly studied in operator theory. For each class we give a reference to its
definition and we state a reason why it contains hypercyclic operators or why it does
not.
We begin with classes that do not contain any hypercyclic operators:
absolutely summing operators (are power compact [134, p. 15, p. 50]);
essentially quasinilpotent operators (= Riesz operators [2, p. 302]);
Hilbert–Schmidt operators (are power compact [134, p. 50, p. 84]);
nuclear operators (are compact [134, p. 112, p. 113]);
operators of Schatten–von Neumann class (are compact [134, p. 80]);
Riesz operators (if hypercyclic then quasinilpotent [2, p. 302]);
strictly singular operators (are Riesz operators [259, 1.9.2, 26.6.5]);
strictly cosingular operators (are Riesz operators [259, 1.10.2, 26.6.10]);
trace class operators (= nuclear operators on Hilbert spaces [134, p. 123]).
We next state classes that do contain hypercyclic operators:
completely continuous operators (e.g., operators on 1 [116, p. 173]);
Dunford–Pettis operators (= completely continuous operators [2, p. 498]);
Fredholm operators (e.g., invertible operators [2, p. 156]);
Toeplitz operators (e.g., Rolewicz’s operators [191, p. 136]);
weakly compact operators (e.g., operators on Hilbert spaces [116, p. 183]).
Chapter 6
Connectedness arguments in linear
dynamics

This chapter is devoted to some of the most fundamental results in linear

dynamics. What is particularly striking is that they hold for all operators,
without further technical assumptions.
We have already obtained such a result in Chapter 2 . It says that every hy-
percyclic operator admits a dense subspace of hypercyclic vectors, except for
the zero vector. Note that this property would not make sense in a nonlinear
setting.
In this chapter we will consider the following problems, which, a priori, do
not involve linearity.
• If T has a dense orbit, does then every power T p also have a dense orbit?
• Suppose that the union of a finite collection of orbits is dense. Will then
at least one of these orbits be actually dense?
• If an orbit is somewhere dense, is it (everywhere) dense?
Each of these questions has a negative answer for arbitrary, nonlinear maps.
It is therefore even more surprising that they all have a positive answer for
(linear) operators, and that without any restrictions. The proofs depend in
a crucial way on connectedness arguments.
In the final section we will consider two more problems.
• Let T be a hypercyclic operator, and let λ ∈ K with |λ| = 1. Is then λT
also hypercyclic?
• Let (Tt )t≥0 be a hypercyclic C0 -semigroup on a Banach space. Is then
every single operator Tt , t > 0, hypercyclic?
Again we will give positive answers to these questions. The proofs can be given
within a common framework and use, once more, a connectedness argument,
this time via a suitable homotopy.

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 161

DOI 10.1007/978-1-4471-2170-1_6, © Springer-Verlag London Limited 2011
162 6 Connectedness arguments in linear dynamics

6.1 Ansari’s theorem

In this section we deal with the question of whether every power T p , p ∈ N,

of a hypercyclic operator T is again hypercyclic. Since every sequence (kp)k
is syndetic, Theorem 1.54 implies a positive answer if T is even a weakly
mixing operator on a separable Fréchet space. We will show here that the
answer is positive for all hypercyclic operators.
The following auxiliary result will be crucial.

Lemma 6.1. The T be a continuous map on a metric space X without iso-

lated points. Then the interiors of the closures of two orbits under T either
coincide, or they are disjoint.

Proof. Suppose that int(orb(x, T ))∩int(orb(y, T )) = ∅, x, y ∈ X. Then there

is some n ∈ N0 such that
T n x ∈ orb(y, T ).
Since orb(y, T ) is T -invariant, we have that T k x ∈ orb(y, T ) for k ≥ n and
therefore
{T k x ; k ≥ n} ⊂ orb(y, T ).
Since X has no isolated points, one shows easily that

int(orb(x, T )) ⊂ int({T k x ; k ≥ n});

see also Exercise 6.2.1. Hence int(orb(x, T )) ⊂ int(orb(y, T )). By symmetry,

we also have the converse inclusion, so that the two interiors coincide. 

Theorem 6.2 (Ansari). Let T be an operator on a Fréchet space. Then, for

any p ∈ N, HC(T ) = HC(T p ). In particular, if T is hypercyclic then so is
every power T p .

Proof. Let p ∈ N. We clearly have that HC(T p ) ⊂ HC(T ).

For the converse inclusion we fix x ∈ HC(T ). From Proposition 1.15 and
Corollary 2.56 we know that D := HC(T ) is a dense, T -invariant connected
subset of X; in particular, it does not have isolated points. For the remainder
of the proof we consider the map T : D → D; the topological operations of
closure and interior will be understood in D. Since D is dense in X it then
suffices to show that orb(x, T p ) = D.
To this end we define

Dj = orb(T j x, T p ), j = 0, . . . , p − 1.

We need to show that D = D0 . Observe that


p−1 
p−1
D = orb(x, T ) = orb(T j x, T p ) = Dj
j=0 j=0
6.1 Ansari’s theorem 163

and
T (Dj ) ⊂ Dj+1(mod p) .
Let F ⊂ {0, . . . , p − 1} be a set of minimal cardinality such that

D= Dj .
j∈F

Suppose that F is not a singleton. Let, in addition, int(Dj ) ∩ int(Dk ) = ∅

for some j, k ∈ F with j = k. By Lemma 6.1, int(Dj ) = int(Dk ). From
minimality we deduce that

D\ Dl
l∈F \{j}

is nonempty, and it is an open set contained in Dj and thus in int(Dj ) ⊂ Dk ,

which is not possible. Therefore, int(Dj ∩ Dk ) = ∅ for any j, k ∈ F with
j = k.
We now set Fl = F + l(mod p), l = 0, . . . , p − 1. We have D = T l (D) =
 
k∈Fl Dk , l = 0, . . . , p − 1. Since card(Fl ) = card(F ), which
l
j∈F T (Dj ) =
is minimal, we also get that int(Dj ∩ Dk ) = ∅ for any j, k ∈ Fl with j = k,
l = 0, . . . , p − 1. As a consequence, the set


p−1 
A := (Dj ∩ Dk )
l=0 j,k∈Fl
j=k

is nowhere dense as a finite union of nowhere dense sets; and it is T -invariant.

If A were nonempty, with y ∈ A, say, then

D = orb(y, T ) ⊂ A = A,

which is a contradiction. Therefore A = ∅, which implies that


D= Dj
j∈F

is a finite union of pairwise disjoint closed subsets. But this contradicts the
connectedness of D.
In conclusion, F = {j} is a singleton. Then D = Dj , and we obtain that
D = T p−j (Dj ) = D0 , which had to be shown.  

The simple example T : {−1, 1} → {−1, 1}, T x = −x, shows that Ansari’s
theorem fails for nonlinear dynamical systems; see also Exercise 1.2.11 for an
example on a metric space without isolated points. Ansari’s theorem does ex-
tend to the nonlinear setting if the set of points with dense orbit is connected;
see Exercise 6.1.7.
164 6 Connectedness arguments in linear dynamics

6.2 Somewhere dense orbits

We recall that a set is called somewhere dense if its closure contains a

nonempty open set.
It was a key point in the proof of Ansari’s theorem to write the space D
as a finite union of closures of orbits. Then one of these closures must have
an interior point, which means that the corresponding orbit is somewhere
dense. In the end we concluded that this orbit is, in fact, (everywhere) dense.
Do we have a general principle here, that is, is every somewhere dense orbit
necessarily dense? We will give a positive answer to this question.
Thus, let T be an operator on a Fréchet space X. For x ∈ X we write

D(x) = orb(x, T ) and U (x) = int D(x).

The following properties can be easily deduced from the continuity of T and
the fact that X has no isolated points (see Exercise 6.2.1):
(i) if y ∈ D(x), then D(y) ⊂ D(x);
(ii) U (x) = U (T k x) for each k ∈ N;
(iii) if R : X → X is a continuous map that commutes with T , then
R(D(x)) ⊂ D(Rx).
We first need a generalization of Theorem 2.54. An easy adaptation of the
argument used there gives the result; see Exercise 6.2.2.
Lemma 6.3. If T admits a somewhere dense orbit and p is a nonzero poly-
nomial, then the operator p(T ) has dense range.
Before proving that a vector whose orbit is somewhere dense is necessarily
hypercyclic, we will show that it is cyclic, that is, the linear span of its orbit
is dense in X.
Lemma 6.4. If orb(x, T ) is somewhere dense, then the set {p(T )x ; p =
0 a polynomial} is connected and dense in X.
Proof. The set A := {p(T )x ; p = 0 a polynomial} is path connected. Indeed,
let p, q be nonzero polynomials. If q is not a multiple of p then the straight
path t → tp(T )x + (1 − t)q(T )x, t ∈ [0, 1], is contained in A. Otherwise we
select a third nonzero polynomial r that is not a multiple of p, and therefore
not of q, and we take the union of the straight paths connecting p(T )x and
q(T )x with r(T )x.
On the other hand, A is a subspace of X that contains orb(x, T ). It follows
from the hypothesis that there is some x0 ∈ X and a 0-neighbourhood W
such that x0 + W ⊂ A. Thus, for any y ∈ X, there is a scalar λ with y ∈ λW ;
hence y ∈ λ(x0 + W ) − λx0 ⊂ A. Consequently, A is dense in X.  
Theorem 6.5 (Bourdon–Feldman). Let T be an operator on a Fréchet
space X and x ∈ X. If orb(x, T ) is somewhere dense in X, then it is dense
in X.
6.2 Somewhere dense orbits 165

Proof. We have to show that if U (x) = ∅ then D(x) = X. The proof will be
split into four steps.

Step 1. We have that T (X \ U (x)) ⊂ X \ U (x).

We show, equivalently, that T −1 (U (x)) ⊂ U (x). First, since U (x) = ∅
there is some m ∈ N0 with xm := T m x ∈ U (x).
Now let y ∈ T −1 (U (x)), and let V be an arbitrary neighbourhood of y.
Since, by property (ii), xm also has a somewhere dense orbit, Lemma 6.4
implies that we can find a polynomial p such that p(T )xm ∈ V ∩ T −1 (U (x)).
We have, using property (ii), that

p(T )xm ∈ p(T )(U (x)) = p(T )(U (T m+1 x)) ⊂ p(T )(D(T m+1 x)).

Moreover, since T p(T )xm ∈ U (x) ⊂ D(x), properties (iii) and (i) yield that

p(T )(D(T m+1 x)) ⊂ D(T p(T )xm ) ⊂ D(x).

We have therefore shown that V ∩D(x) = ∅. Since V was arbitrary and D(x)
is closed, we deduce that y ∈ D(x) and hence T −1 (U (x)) ⊂ D(x). Continuity
of T implies that T −1 (U (x)) ⊂ U (x).

Step 2. For any z ∈ X \ U (x), D(z) ⊂ X \ U (x).

By Step 1, X \ U (x) is T -invariant, and it is closed. The claim then follows
from the definition of D(z).

Step 3. For any polynomial p = 0, p(T )x ∈ X \ ∂D(x), where ∂D(x) denotes

the boundary of D(x); see Figure 6.1.
Suppose that p(T )x ∈ ∂D(x) for some polynomial p = 0. By Lemma 6.3
there is some y ∈ X such that p(T )y ∈ U (x). Since p(T )x ∈
/ U (x), property
(iii) and Step 2 imply that

p(T )(D(x)) ⊂ D(p(T )x) ⊂ X \ U (x).

We therefore have that y ∈ X \ D(x). By Lemma 6.4 there then exists a

polynomial q such that q(T )x is close enough to y to satisfy q(T )x ∈ X \
D(x) ⊂ X \ U (x) and p(T )q(T )x ∈ U (x). Since p(T )x ∈ D(x), property (iii)
and Step 2 imply that

p(T )q(T )x = q(T )p(T )x ∈ q(T )(D(x)) ⊂ D(q(T )x) ⊂ X \ U (x),

which is a contradiction. This proves the claim.

Step 4. We have that D(x) = X.

By Step 3,

A := {p(T )x ; p = 0 a polynomial} ⊂ U (x) ∪ (X \ D(x)),

166 6 Connectedness arguments in linear dynamics

Fig. 6.1 Step 3

which is a disjoint union of open sets. Since, by Lemma 6.4, A is connected

and, by density of A, A ∩ U (x) = ∅, we must have that A ∩ (X \ D(x)) = ∅.
Hence, A ⊂ D(x), which implies that D(x) = X.  

6.3 Multi-hypercyclic operators

The Bourdon–Feldman theorem provides us with a very powerful tool for

obtaining dense orbits. A particular case occurs when the union of a finite
number of orbits under T is dense in X. In this case the operator T is called
multi-hypercyclic.

Theorem 6.6 (Costakis–Peris). Let T be an operator on a Fréchet space

X and x1 , . . . , xn ∈ X. If

n
orb(xj , T )
j=1

is dense in X, then there is some j ∈ {1, . . . , n} such that orb(xj , T ) is dense

in X. In particular, every multi-hypercyclic operator is hypercyclic.

Proof. The hypothesis says that


n 
n
orb(xj , T ) = orb(xj , T ) = X.
j=1 j=1

Since a finite union of nowhere dense sets is nowhere dense, orb(xj , T ) must
be somewhere dense in X for some j ∈ {1, . . . , n}. By the Bourdon–Feldman
theorem, xj then has a dense orbit. 

Ansari’s result can easily be derived from this theorem. Let x ∈ HC(T )
and p ∈ N. Since
6.4 Hypercyclic semigroup actions 167


p−1
orb(x, T ) = orb(T j x, T p )
j=0

is dense in X, Theorem 6.6 implies that there is some j ∈ {0, . . . , p − 1} such

that T j x is hypercyclic for T p . Since T p−j has dense range and

T p−j (orb(T j x, T p )) ⊂ orb(x, T p )

we obtain that x ∈ HC(T p ).

These arguments also imply that because Ansari’s theorem fails for non-
linear dynamical systems the same is true for the theorems of Costakis–Peris
and Bourdon–Feldman; see also Exercise 1.2.11.

6.4 Hypercyclic semigroup actions

In this section we will be dealing with two additional important problems in

linear dynamics.
The problem of unimodular multiples asks whether, given a hypercyclic
operator T , is every multiple λT with λ ∈ K, |λ| = 1, also hypercyclic? The
operator λT is called a rotation of T . In the real setting the answer is positive.
Indeed, one only needs to show that if T is hypercyclic then so is −T . But
this follows from Ansari’s theorem because the two operators have a common
square, T 2 = (−T )2 . Thus we will concentrate here on the complex setting.
The problem of hypercyclic discretizations of semigroups asks whether,
given a hypercyclic C0 -semigroup (Tt )t≥0 on a Banach space, is every single
operator Tt , t > 0, also hypercyclic? Although C0 -semigroups will only be
treated in the next chapter (and we ask the reader to consult the relevant
definitions there), there will be no harm in already considering the discretiza-
tion problem here. The (very basic) proof that a hypercyclic C0 -semigroup
satisfies the assumptions imposed in this section will be postponed to Chap-
ter 7.
The main aim of this section is to show that both problems have a positive
answer. In analogy with Ansari’s theorem, it will even be proved that the
corresponding sets of hypercyclic vectors coincide.

Theorem 6.7 (León–Müller). Let T be an operator on a complex Fréchet

space X. If x ∈ X is such that {λT n x ; λ ∈ C, |λ| = 1, and n ∈ N0 } is
dense in X then orb(x, λT ) is dense in X for each λ ∈ C with |λ| = 1.
In particular, for any λ ∈ C with |λ| = 1, T and λT have the same
hypercyclic vectors, that is,

HC(T ) = HC(λT ).
168 6 Connectedness arguments in linear dynamics

In Exercise 2.5.1 we saw that rotations of mixing (or weakly mixing) op-
erators are mixing (or weakly mixing, respectively). But that result did not
say anything about the sets of hypercyclic vectors.
Theorem 6.8 (Conejero–Müller–Peris). Let (Tt )t≥0 be a C0 -semigroup
on a Banach space X. If x ∈ X is hypercyclic for (Tt )t≥0 , then it is hypercyclic
for each operator Tt , t > 0.
It is particularly gratifying that the two problems can be treated within
a common framework, that of semigroup actions. We will also show that a
variant of the method leads to a new proof of Ansari’s theorem; see Exercise
6.4.5.
Throughout this section we will write

G = N0 × R+ ,

which is a semigroup under addition. If X is a Fréchet space then a map

Ψ : G → L(X)

is called a (continuous and linear) semigroup action of G on X if the following

properties hold:
(i) Ψ (0) = I;
(ii) for any g1 , g2 ∈ G, Ψ (g1 + g2 ) = Ψ (g1 )Ψ (g2 );
(iii) the map G×X → X, (g, x) → Ψ (g)x, is continuous, where G = N0 ×R+
and G × X carry the product topology.
Definition 6.9. A semigroup action Ψ on a Fréchet space X is called hyper-
cyclic if there is some x ∈ X such that {Ψ (g)x ; g ∈ G} is dense in X. The
vector x is then called hypercyclic for Ψ , and we write x ∈ HC(Ψ ).
Let us see how our two problems fit into this framework. If T is an operator
on a complex Fréchet space X, then we define

Ψ (n, t) = e2πti T n , n ∈ N0 , t ≥ 0.

In the second case, if (Tt )t≥0 is a C0 -semigroup on a Banach space X, then

we define
Ψ (n, t) = Tt , n ∈ N0 , t ≥ 0.
It is easy to see that these are semigroup actions of G on X; we refer to
Chapter 7 for the definition of a C0 -semigroup.
Moreover, in both cases, the following properties are satisfied:
(α) either Ψ (1, 0) = I or Ψ (0, 1) = I;
(β) if the semigroup action is hypercyclic then each convex combination of
Ψ (0, s) and Ψ (1, t), s, t ≥ 0, has dense range.
That property (β) is satisfied follows from a simple generalization of The-
orem 2.54 and by Theorem 7.16, respectively.
6.4 Hypercyclic semigroup actions 169

The following theorem therefore immediately implies the Theorems of

León–Müller and Conejero–Müller–Peris.

Theorem 6.10. Let Ψ be a semigroup action on an infinite-dimensional

Fréchet space X satisfying properties (α) and (β). If x ∈ X is hypercyclic
for Ψ then it is hypercyclic for each operator Ψ (1, t), t > 0.

Proof. We first note that it suffices to prove the claim for t = 1. Indeed, let
x be hypercyclic for Ψ , and let t > 0 be arbitrary. We distinguish the two
subcases of (α). If Ψ (1, 0) = I then

Ψ(n, s) := Ψ (n, st)

defines a semigroup action that satisfies (α) and (β). Since x is also hyper-
cyclic for Ψ we can conclude that x is hypercyclic for Ψ(1, 1) = Ψ (1, t). If
Ψ (0, 1) = I then we define

Ψ(n, s) = Ψ (n, nt + s),

and we can conclude as before that x is hypercyclic for Ψ(1, 1) = Ψ (1, t+1) =
Ψ (1, t).
As usual, T = {z ∈ C ; |z| = 1} is the unit circle. For ease of notation we
introduce the map ρ : R+ → T, given by ρ(t) := e2πti . We then define, for
every pair u, v ∈ X, the subset Fu,v of T by
 
Fu,v := λ ∈ T ; ∃ ((nk , tk ))k ⊂ G with Ψ (nk , tk )u → v and ρ(tk ) → λ .

The remainder of the proof will be divided into several steps.

Step 1. If u ∈ HC(Ψ ), then Fu,v = ∅ for all v ∈ X.

Since u ∈ HC(Ψ ), we can find sequences (nk )k in N0 and (tk )k in R+
such that Ψ (nk , tk )u → v. By passing to a subsequence if necessary, we may
assume that (ρ(tk ))k is convergent. Its limit is an element of Fu,v .

Step 2. If λk ∈ Fu,vk , vk → v and λk → λ, then λ ∈ Fu,v . In particular, Fu,v

is a closed set for each u, v ∈ X.
Let W be a 0-neighbourhood of X and ε > 0. There is a 0-neighbourhood
W1 such that W1 + W1 ⊂ W ; see Lemma 2.36. By assumption, there is some
k ∈ N with v−vk ∈ W1 and |λ−λk | < ε. Now, by definition, there are nk ∈ N0
and tk ∈ R+ such that vk − Ψ (nk , tk )u ∈ W1 and |λk − ρ(tk )| < ε. We then
get that v −Ψ (nk , tk )u ∈ W1 +W1 ⊂ W and |λ−ρ(tk )| < 2ε, so that λ ∈ Fu,v .

Step 3. If u, v, w ∈ X, λ ∈ Fu,v , and μ ∈ Fv,w , then λμ ∈ Fu,w .

Given a 0-neighbourhood W , take a 0-neighbourhood W1 such that W1 +
W1 ⊂ W . Let ε > 0. Then there are n1 ∈ N0 and t1 ∈ R+ such that
w −Ψ (n1 , t1 )v ∈ W1 and |μ−ρ(t1 )| < ε. One can then find a 0-neighbourhood
170 6 Connectedness arguments in linear dynamics

V , n2 ∈ N0 and t2 ∈ R+ satisfying Ψ (n1 , t1 )(V ) ⊂ W1 , v − Ψ (n2 , t2 )u ∈ V ,

and |λ − ρ(t2 )| < ε. Consequently we have for n3 := n1 + n2 and t3 := t1 + t2
that

w − Ψ (n3 , t3 )u = w − Ψ (n1 , t1 )v + Ψ (n1 , t1 )(v − Ψ (n2 , t2 )u) ∈ W1 + W1 ⊂ W,

and
|λμ − ρ(t3 )| ≤ |λ| |μ − ρ(t1 )| + |ρ(t1 )| |λ − ρ(t2 )| < 2ε.
Hence λμ ∈ Fu,w .

We now fix x ∈ HC(Ψ ). Our aim is to show that x ∈ HC(Ψ (1, 1)). By Steps
1, 2 and 3, Fx,x is a nonempty closed subsemigroup of the multiplicative
group T.

Step 4. If Fx,x = T then x is hypercyclic for Ψ (1, 1).

Suppose that Fx,x = T. Given any y ∈ X, Steps 1 and 3 imply that
Fx,y = T. In particular 1 ∈ Fx,y , which yields the existence of sequences (nk )k
in N0 and (tk )k in R+ such that Ψ (nk , tk )x → y and ρ(tk ) = e2πtk i → 1. We
can then write tk = jk − 1 + εk with jk ∈ N and εk ∈ [−1/2, 1/2], where
εk → 0.
Let W be a 0-neighbourhood, and let W1 be a 0-neighbourhood such that
W1 + W1 ⊂ W . By a standard compactness argument, the continuity of
the semigroup action implies that there is a 0-neighbourhood V such that
Ψ (0, t)(V ) ⊂ W1 if 0 ≤ t ≤ 2. Moreover, there is some k ∈ N such that
Ψ (nk , tk )x − y ∈ V and Ψ (0, 1 − εk )y − Ψ (0, 1)y ∈ W1 . Therefore

Ψ (nk , jk )x − Ψ (0, 1)y

= Ψ (0, 1 − εk )(Ψ (nk , tk )x − y) + (Ψ (0, 1 − εk ) − Ψ (0, 1))y
∈ Ψ (0, 1 − εk )(V ) + W1 ⊂ W1 + W1 ⊂ W.

Observe that, by property (α), Ψ (nk , jk )x ∈ orb(x, Ψ (1, 1)). Thus Ψ (0, 1)y ∈
orb(x, Ψ (1, 1)). Since Ψ (0, 1) has dense range by property (β), and y ∈ X is
arbitrary, x is hypercyclic for Ψ (1, 1).

For the rest of the proof we can now assume that Fx,x = T, and we will show
that this leads to a contradiction.

Step 5. There exists some m ∈ N such that, for each y ∈ HC(Ψ ), there is
λ ∈ T satisfying Fx,y = {λz ; z m = 1}.
We first note that Fx,x must be of the form Fx,x = {z ∈ T ; z m = 1} for
some m ∈ N. Indeed, if Fx,x contained points z = e2πti with t > 0 arbitrarily
small then, being a closed subsemigroup of T, Fx,x would be dense, and hence
coincide with T, which was excluded. Hence there is a minimal t0 ∈ ]0, 1] such
that z0 = e2πt0 i ∈ Fx,x . By the same argument, t0 cannot be irrational; see
6.4 Hypercyclic semigroup actions 171

Example 1.17. There is then a minimal m ∈ N with z0m = 1. The minimality

of t0 and m easily imply that Fx,x = {z ∈ T ; z m = 1}.
Now let y ∈ HC(Ψ ). By Step 1, there exist λ ∈ Fx,y and μ ∈ Fy,x . Then,
by Step 3, λFx,x ⊂ Fx,y and μFx,y ⊂ Fx,x , so that card(Fx,y ) = card(Fx,x ).
This implies that Fx,y = λFx,x .

Step 6. There is a continuous function f : HC(Ψ ) → T such that f (Ψ (0, t)x) =

e2πmti for every t ≥ 0.
Let m ∈ N be the integer given by Step 5. Then, for any y ∈ HC(Ψ ), we
define
f (y) = λm if λ ∈ Fx,y .
By Step 5, this is well defined. Moreover, f is continuous. Otherwise there
are yk ∈ HC(Ψ ) and y ∈ HC(Ψ ) such that yk → y but f (yk ) → f (y). We
choose λk ∈ Fx,yk . Passing to a subsequence if necessary, we can assume that
f (yk ) → μ = f (y) and λk → λ for some λ, μ ∈ T. It follows from Step 2 that
λ ∈ Fx,y and hence that f (yk ) = λmk → λ = f (y), which is a contradiction.
m

Now let t ≥ 0. By property (β), Ψ (0, t) has dense range and therefore
Ψ (0, t)x ∈ HC(Ψ ). Since, by definition, ρ(t) = e2πti ∈ Fx,Ψ (0,t)x we conclude
that f (Ψ (0, t)x) = e2πmti .

Step 7. There is a continuous function h : D → T, whose restriction to the

unit circle is homotopically nontrivial. A contradiction.
This is the decisive, and most difficult part of the proof. We will use here
the terminology and some results of homotopy theory; see Appendix A. In
order to define the function h we will first define a function g : T → HC(Ψ ),
where we will distinguish the two subcases of (α).
Case 1: Ψ (0, 1) = I. Here we define g : T → HC(Ψ ) by

g(e2πti ) = Ψ (0, t)x, 0 ≤ t < 1,

which is well defined by property (β), and g is continuous because Ψ (0, 1) = I.

By Step 6, the function f ◦ g : T → T satisfies f (g(e2πti )) = e2πmti , 0 ≤ t < 1,
so that the index of f ◦ g is m ≥ 1.
We extend the function g to the closed unit disk D by defining g(z) =
(1 − r)Ψ (1, 0)x + rg(e2πti ) for z = re2πti ∈ D, r ≥ 0. This extension is clearly
continuous on D. Since g(z) is a convex combination of Ψ (1, 0)x and Ψ (0, t)x
for some t ≥ 0, property (β) implies that g(z) ∈ HC(Ψ ) for every z ∈ D.
To summarize, we have found a continuous function h := f ◦ g : D → T
whose restriction to the unit circle is homotopically nontrivial. In other words,
the map H : T × [0, 1] → T, (e2πti , r) → h(re2πti ) defines a homotopy be-
tween the function h on T, which is homotopically nontrivial, and a constant
function. This is the desired contradiction.
Case 2: Ψ (1, 0) = I. Here the construction of g is slightly more delicate.
First, since f is continuous and f (x) = 1, we can find a 0-neighbourhood W
172 6 Connectedness arguments in linear dynamics

such that |f (y) − 1| < 1 if y ∈ HC(Ψ ) and y − x ∈ W . We can assume that

W is balanced, that is, μW ⊂ W whenever |μ| ≤ 1; see Lemma 2.6(iii).
Since no 0-neighbourhood in an infinite-dimensional Fréchet space can be
relatively compact (see Appendix A), the set U := W \ {x − Ψ (0, t)x ; 0 ≤
t ≤ 1} is open and nonempty. By the hypercyclicity of x there are n0 ∈ N0
and t0 ≥ 0 such that x − Ψ (n0 , t0 )x ∈ U . Since Ψ (1, 0) = I we also have that
x − Ψ (0, t0 )x ∈ U , and therefore t0 > 1 and x − Ψ (0, t0 )x ∈ W . We can now
define g : T → HC(Ψ ) by

2πti Ψ (0, 2tt0 )x if 0 ≤ t < 1/2,
g(e )=
(2t − 1)x + (2 − 2t)Ψ (0, t0 )x if 1/2 ≤ t < 1,

which is clearly continuous. The fact that g is well defined is a consequence

of property (β); note that x = Ψ (1, 0)x.

Fig. 6.2 The map h, Case 2

We consider the function f ◦ g : T → T. Then f (g(e2πti )) = e4πmtt0 i for

0 ≤ t < 1/2. Moreover, by the selection of t0 , and since W is balanced, we
obtain that |f (g(e2πti )) − 1| < 1 for 1/2 ≤ t < 1. Thus, as t moves from 0
to 1/2, f (g(e2πti )), starting from 1, moves along the unit circle in a positive
direction and covers it [mt0 ] times, finishing inside the disk of radius 1 around
1. As t then moves from 1/2 to 1, f (g(e2πti )) stays completely in that disk,
returning to 1 for t = 1. As a consequence, the path t → f (g(e2πti )) can be
deformed homotopically to the path t → e2πni with either n = [mt0 ] ≥ 1 if
Exercises 173

Im(e2πmt0 i ) ≥ 0 or with n = [mt0 ] + 1 ≥ 2 if Im(e2πmt0 i ) < 0. In any case,

the index of f ◦ g is nonzero.
We extend the function g continuously to the closed unit disk D by defining
g(z) = (1 − r)x + rg(e2πti ) for z = re2πti ∈ D, r ≥ 0. Since g(z) is a convex
combination of x = Ψ (1, 0)x and Ψ (0, s)x for some s ≥ 0, property (β) implies
that g(z) ∈ HC(Ψ ) for every z ∈ D. We define again the map h : D → T
by h = f ◦ g; see Figure 6.2. Since the restriction of h to the unit circle is
homotopically nontrivial we obtain a contradiction as in Case 1.  

When we combine Theorem 6.10 with Ansari’s theorem, by which Ψ (n, t) =

Ψ (1, t/n)n is hypercyclic whenever Ψ (1, t/n) is, we obtain the following.

Corollary 6.11. Let Ψ be a semigroup action on a Fréchet space X satisfying

properties (α) and (β). If x ∈ X is hypercyclic for Ψ then it is hypercyclic
for each operator Ψ (n, t), n, t > 0.

Exercises

Exercise 6.1.1. In a metric space, show that a finite union of nowhere dense sets is
nowhere dense.

Exercise 6.1.2. Let T : X → X be a (not necessarily linear) weakly mixing dynamical

system. Show that any T p , p ∈ N, is also weakly mixing. (Hint: Theorem 1.54.)

Exercise 6.1.3. Let T be an operator on a separable Fréchet space X that satisfies the
Hypercyclicity Criterion. Give two proofs of the fact that any T p , p ∈ N, also satisfies
the Hypercyclicity Criterion.

Exercise 6.1.4. Let T be a chaotic operator on a Fréchet space X. Without the use of
Theorem 6.2, show that any T p , p ∈ N, is also chaotic. This is not true for nonlinear
maps by the example of Exercise 1.2.11.

Exercise 6.1.5. Let S : X → X, T : Y → Y be topologically ergodic operators on

Fréchet spaces X and Y . Show that any operator S p ⊕ T q , p, q ∈ N, is topologically
ergodic on X ⊕ Y . (Hint: Exercises 2.5.5 and 2.5.6.)

Exercise 6.1.6. Let T be an operator on a Fréchet space X and x a hypercyclic vector

for T . Show that there exists an increasing sequence (nk )k of positive integers with
supk≥1 (nk+1 − nk ) = 2 such that x does not have dense orbit under (T nk )k . (Hint:
Show that there is some y ∈ X and a 0-neighbourhood W such that z ∈ y + W implies
that T z ∈
/ y + W .)

In the following two exercises, let T : X → X be a (not necessarily linear) dynamical

system, that is, a continuous map T on a metric space X. Suppose that X does not have
isolated points, and let D = {x ∈ X ; orb(x, T ) is dense in X}.

Exercise 6.1.7. Show the following generalization of Ansari’s theorem. If D contains a

connected and dense set then T and T p , p ∈ N, have the same points of dense orbits.
(Hint: Follow the proof of Ansari’s theorem and note that D itself must be connected;
see the proof of Corollary 2.56.)
174 6 Connectedness arguments in linear dynamics

Exercise 6.1.8. An alternative proof of Exercise 6.1.7 (and thus of Ansari’s theorem)
is the following. With the notation of Theorem 6.2, let

Ak := (Dj1 ∩ · · · ∩ Djk ) ,
0≤j1 <···<jk ≤p−1

where k = 1, . . . , p. Prove
p−1the following assertions:
(i) A1 = D, Ap = j=0 Dj , and Ak+1 ⊂ Ak , k = 1, . . . , p − 1;
(ii) T (Ak ) ⊂ Ak , k = 1, . . . , p;
(iii) if Ak = D, then Ak+1 = D, k = 1, . . . , p − 1.
In particular, orb(x, T p ) is dense in X for every x ∈ D. (Hint: For (iii), observe that if
Ak+1 = D, then Ak+1 = ∅, since it is closed, T -invariant, and T : D → D is minimal;
hence Ak is a finite union of pairwise disjoint closed sets.)

Exercise 6.2.1. Prove assertions (i), (ii) and (iii) before Lemma 6.3. (Hint: See the
proof of Proposition 1.15.)

Exercise 6.2.2. Prove Lemma 6.3. (Hint: Follow the argument of Theorem 2.54.)

Exercise 6.2.3. Let T be a continuous map on a metric space X without isolated

points, and let x ∈ X. With the notation of this section, prove that if U (x) = ∅ and
T (X \ U (x)) ⊂ X \ U (x), then D(x) = U (x). (Hint: Show that orb(x, T ) ⊂ U (x).)

Exercise 6.2.4. Let T be an operator on a Fréchet space X and x ∈ X. With the

notation of this section, prove directly that if X is a complex (or real) space, then
U (x) = U (λx) = λU (x) for λ = 0 (or for λ > 0, respectively). Deduce that D(x) = X if
0 ∈ U (x). (Hint: Use Lemma 6.1.)

Exercise 6.2.5. Let T : X → X be a (not necessarily linear) topologically transitive

dynamical system and x ∈ X. Show that if orb(x, T ) is somewhere dense in X, then it
is dense in X.

Exercise 6.2.6. Let T be an operator on a Fréchet space X and x a hypercyclic vector

for T . Show that there exists an increasing sequence (nk )k of positive integers with
supk≥1 (nk+1 − nk ) = 2 such that the orbit of x under (T nk )k is somewhere dense but
not dense. (Hint: See Exercise 6.1.6.)

Exercise 6.3.1. Let T be an invertible operator on a Fréchet space X and x ∈ X such

that {T n x ; n ∈ Z} is dense in X. Show that x is either hypercyclic for T or for T −1 ;
in particular, both T and T −1 are hypercyclic. For the proof,
(i) either use the Bourdon–Feldman theorem,
(ii) or proceed directly.
(Hint: For (i); see the proof of Theorem 6.6. For (ii), suppose that T n x ∈/ U for all
n ≥ 0; for any V , find k ∈ Z and U  ⊂ U such that T k (U  ) ⊂ V ; find m < −|k| such
that T m x ∈ U  ; then T m+k x ∈ V , m + k < 0.)

Exercise 6.3.2. Let T be an operator on a Fréchet space X admitting a countable set

{x1 , x2 , . . . } of vectors such that


∞

orb(xj , T ) = X.
j=1

Show that some vector xj , j ≥ 1, is hypercyclic for T . Give an example of an operator

on a normed space for which this assertion fails.
Exercises 175

Exercise 6.3.3. An operator T on a Banach space X is called countably hypercyclic if

it admits a countable bounded set {x1 , x2 , . . . } of vectors with inf j=k xj − xk > 0
such that

∞

orb(xj , T ) = X.
j=1

Show that the operator T = 2(I ⊕ B) on X = 2 ⊕ 2 is countably hypercyclic but not

hypercyclic, where B is the backward shift. (Hint: Take xj = (0, ej ) + 2−nj (I ⊕ F )nj yj ,
where F is the forward shift.)

Exercise 6.3.4. Let T be a countably hypercyclic operator on a Banach space X. Show

the following.
(a) The spectrum σ(T ) meets the unit circle.
(b) The orbit of every x∗ = 0 in X ∗ under T ∗ is unbounded.
(Hint: See Section 5.1.)

Exercise 6.3.5. Let T = Bw be a weighted backward shift on X = p , 1 ≤ p < ∞, or

c0 ; see Section 4.1. Show that if T is countably hypercyclic then it is hypercyclic. (Hint:
Apply Exercise 6.3.4(b) to x∗ = e1 .)

Exercise 6.4.1. Let T = Bw be the weighted bilateral backward shift on 2 (Z) with
weights wn = n+1n if n ≥ 1 and w−n = n+1n+2 if n ≥ 0; see Section 4.1. Show that λT ,
λ ∈ C, is hypercyclic if and only if |λ| = 1. Discuss this result in the light of Kitai’s
theorem, showing first that σ(T ) ⊂ T. (Hint: For the first part note that λBw = Bλw ;
for the second part use the spectral radius formula for T and T −1 and Exercise 5.0.7.)

Exercise 6.4.2. Let Tj be operators on complex Fréchet spaces Xj , j = 1, . . . , n, such

that T1 ⊕ · · · ⊕ Tn is hypercyclic. Show that, for any λj ∈ C with |λj | = 1, j = 1, . . . , n,
the operator λ1 T1 ⊕ · · · ⊕ λn Tn is also hypercyclic and, moreover, that it shares the set
of hypercyclic vectors with T1 ⊕ · · · ⊕ Tn . (Hint: Set Ψ (n, t) = S1n ⊕ e2πti S2n for suitable
operators S1 , S2 , deduce that HC(Ψ ) = HC(S1 ⊕ S2 ), and apply this result repeatedly.)

Exercise 6.4.3. Let Tj be operators on complex Fréchet spaces Xj , j = 1, . . . , n, and

let xj ∈ Xj , j = 1, . . . , n, be such that
 
(λ1 T1k x1 , . . . , λn Tnk xn ) ; k ∈ N0 , (λ1 , . . . , λn ) ∈ Tn

is dense in X1 ⊕ · · · ⊕ Xn . Show that x := (x1 , . . . , xn ) is hypercyclic for T1 ⊕ · · · ⊕ Tn .

(Hint: Let (Um )m be a countable base of open sets in X1 ⊕ · · · ⊕ Xn . Show that the
sets {(μ1 , . . . , μn ) ∈ Tn ; ∃ k ∈ N0 with (μk1 T1k x1 , . . . , μkn Tnk xn ) ∈ Um } are open and
dense in Tn . By a Baire argument, find (μ1 , . . . , μn ) ∈ Tn such that x is hypercyclic for
μ1 T1 ⊕ · · · ⊕ μn Tn , and conclude by using Exercise 6.4.2.)

Exercise 6.4.4. Let X = C0 (R+ ), the space of continuous functions on R+ that van-
ish at ∞, endowed with the sup-norm. Consider the semigroup action Ψ (n, t)f (x) :=
2n−t f (x + t), n ∈ N0 , t ∈ R+ . Then Ψ is hypercyclic but the operator Ψ (1, 1) is not
hypercyclic. Which hypothesis of Theorem 6.10 is not satisfied?

Exercise 6.4.5. Give a new proof of Ansari’s theorem by proceeding as follows. Let T
be a hypercyclic operator on a Fréchet space X, x a hypercyclic vector for T and p ∈ N.
For u, v ∈ X define the subset Fu,v of T by
 
Fu,v = e2πji/p ; ∃ (nk )k ⊂ N0 with T nk p+j u → v, j = 0, . . . , p − 1 .

Show the following:

176 6 Connectedness arguments in linear dynamics

(i) if u ∈ HC(T ), then Fu,v = ∅ for all v ∈ X;

(ii) if u, v, w ∈ X, λ ∈ Fu,v , and μ ∈ Fv,w , then λμ ∈ Fu,w ;
(iii) there is a divisor m ≥ 1 of p such that Fx,x = {e2πmi/p ; j = 0, . . . , p/m − 1};
(iv) for every y ∈ HC(T ) there is some j, 0 ≤ j ≤ m − 1, such that Fx,y = e2πji/p Fx,x .
Now let Dj = {y ∈ HC(T ) ; Fx,y = e2πji/p Fx,x }, j = 0, . . . , m − 1. Then finish the
proof as follows:
(v) show that the Dj form a partition of HC(T ) into closed (and open) sets;
(vi) deduce that m = 1 and hence that x ∈ HC(T p ).
In the following two exercises, let T : X → X be a (not necessarily linear) dynamical
system, where X does not have isolated points.
Exercise 6.4.6. Show the following separation theorem. If x ∈ X has dense orbit un-
der T but not under T p , p > 1, then there is a divisor m > 1 of p and a partition
D0 , . . . , Dm−1 of D = {x ∈ X ; orb(x, T ) is dense in X} into closed (and open) subsets
with the following properties:
(i) T (Dj ) ⊂ Dj+1(mod m) , j = 0, . . . , m − 1;
(ii) for j = 0, . . . , m − 1, the orbit of T j x under T p is contained and dense in Dj .
(Hint: Proceed as in the previous exercise.)

Fig. 6.3 Nonlinear dynamics if x has dense orbit under T but not under T p , m|p

Exercise 6.4.7. Show the following decomposition theorem. If x ∈ X has dense orbit
under T but not under T p , p > 1, then there is a divisor m > 1 of p and pairwise disjoint
open subsets S0 , S1 , . . . , Sm−1 of X with the following properties:
m−1
(i) S := j=0 Sj is dense in X;
(ii) T (Sj ) ⊂ Sj+1 , j = 0, . . . , m − 2, and T (Sm−1 ) ⊂ S0 ∪ (X \ S);
(iii) X \ S is invariant under T ;
(iv) for j = 0, . . . , m − 1, the orbit of T j x under T p is contained and dense in Sj ;
see Figure 6.3. m−2
(Hint: Consider the sets Dj of the previous exercise; set Sm−1 = X \ j=0 Dj , with
closure in X, and Sj = T −m+j+1 (Sm−1 ); show first that T −m (Sm−1 ) ⊂ Sm−1 , Dj ⊂ Sj ,
and that T n x ∈ Sj if and only if n = j(mod m).)
Exercise 6.4.8. Verify the results of the previous two exercises in the case of the map
of Exercise 1.2.11.
Sources and comments 177

Sources and comments

The results in this chapter have in common that their proofs use connectedness argu-
ments. But their relationship runs deeper than that. As we have seen, Ansari’s theorem
is a consequence of the Costakis–Peris theorem, which in turn follows from the Bourdon–
Feldman theorem. Moreover, the theorems of León–Müller and Conejero–Müller–Peris
are proved by a common approach. Recently, Shkarin [284] was able to unify the latter
two theorems with Ansari’s theorem by deriving them as consequences of a single, quite
general result. An alternative common framework was developed by Bayart and Math-
eron [44], which was further generalized by Matheron [235] to include Shkarin’s result.

Section 6.1. Ansari [9] showed that powers of hypercyclic operators are hypercyclic. In-
dependently, Banks [28] proved a more general result: any power of a minimal map on a
connected topological space is also minimal (see Exercise 6.1.7). We combine ideas from
Banks [28] and Peris [254] for the proof of Theorem 6.2. Lemma 6.1 is from Peris [254].

Section 6.2. Theorem 6.5 is due to Bourdon and Feldman [93], answering a question
from Peris [254]. The corresponding result for semigroups of operators (see the next
chapter for this notion) is due to Costakis and Peris [121]. It is interesting to note that
for a weighted backward shift on p , 1 ≤ p < ∞, to be hypercyclic it already suffices to
have an orbit with a nonzero limit point, as was shown by Chan and Seceleanu [105];
such an orbit though, need not be dense.

Section 6.3. The fact that multi-hypercyclic operators are hypercyclic was indepen-
dently proved by Costakis [117] and Peris [254], answering a question raised by Her-
rero [195]. The original proofs motivated the question leading to the Bourdon–Feldman
theorem.

Section 6.4. Theorem 6.7 on rotations of hypercyclic operators is due to León and
Müller [222]. Bayart and Bermúdez [37] show that the corresponding result for chaos
fails. Badea, Grivaux and Müller [20] characterize the subsets of C that can appear as
{λ ∈ C ; λT hypercyclic} for invertible operators T on a complex Hilbert space.
Theorem 6.8 on discretizations of hypercyclic C0 -semigroups is due to Conejero,
Müller and Peris [110]. Exercise 6.4.4 shows that the result fails for semigroups indexed
over N0 × R+ ; see also Shkarin [284] and Exercise 7.3.1. Bayart [36] shows that it even
fails for holomorphic groups over C. And by Bayart and Bermúdez [37] there are chaotic
C0 -semigroups on a Hilbert space for which no individual operator is chaotic.
The unified proof of Theorem 6.10 essentially follows the argument of [110]. The re-
lated approach to Ansari’s theorem in Exercise 6.4.5 is due to Grosse-Erdmann, León
and Piqueras [183]. As mentioned above, Shkarin [284], Bayart and Matheron [44] and
Matheron [235] obtain much more general results that contain the theorems of Ansari,
León–Müller and Conejero–Müller–Peris as special cases. Shkarin and Matheron point
out that the main common idea in all these proofs can already be found in a paper by
Furstenberg [156].

Exercises. Exercise 6.1.6 is taken from Montes and Salas [243], Exercise 6.1.7 from
Banks [28]. Exercise 6.1.8 outlines essentially the original proof of Ansari [9]. Exercises
6.2.2 and 6.2.3 are taken from Bourdon and Feldman [93], while the result of Exercise
6.2.6 is due to Peris and Saldivia [257]. The result of Exercise 6.3.1 is due to Herrero
and Kitai [196]. The notion of a countably hypercyclic operator (see Exercise 6.3.3), as
well as the results of Exercises 6.3.4 and 6.3.5 are due to Feldman [149]. Exercise 6.4.1
is taken from León and Müller [222], Exercise 6.4.2 from Shkarin [284]. Exercises 6.4.6
and 6.4.7 are due to Grosse-Erdmann, León and Piqueras [183] (see also Marano and
178 6 Connectedness arguments in linear dynamics

Salas [226]); the case of p = 2 was previously obtained by Bourdon [91].

Extensions. We will show in Chapter 12 that the theorems of Ansari, Bourdon–

Feldman, Costakis–Peris and León–Müller continue to hold in arbitrary topological vec-
tor spaces.
 Part II
Selected topics
Chapter 7
Dynamics of semigroups, with
applications to differential equations

In this chapter we study the dynamical properties of strongly continuous

semigroups of operators on Banach spaces, that is, of C0 -semigroups. They
can be viewed as the continuous-time analogue of the discrete-time case of
iterates of a single operator.
Specifically, we introduce and study the notions of hypercyclicity, (weak)
mixing and chaos for C0 -semigroups. We then develop corresponding criteria
such as a Hypercyclicity Criterion and an eigenvalue criterion.
For discrete linear dynamical systems the shift operators on sequence
spaces constitute one of the most important test classes. In the continu-
ous case this role is played by the translation semigroups. We will obtain, in
particular, a characterization of translation semigroups that are hypercyclic,
mixing or chaotic.
We also investigate the relationship between the dynamical properties of
a semigroup and those of its various discretizations.
C0 -semigroups describe the solutions of so-called abstract Cauchy prob-
lems. This makes the semigroup formulation of chaos applicable to linear
differential equations. Some applications of the chaotic behaviour of C0 -
semigroups to partial differential equations and to infinite linear systems of
ordinary differential equations will be discussed in the last section.
Throughout this chapter, X will denote a separable Banach space.

7.1 Semigroups of operators

A one-parameter family (Tt )t≥0 of operators on X is called a strongly con-

tinuous semigroup of operators if the following three conditions are satisfied:
(i) T0 = I;
(ii) Tt+s = Tt Ts for all s, t ≥ 0;
(iii) lims→t Ts x = Tt x for all x ∈ X and t ≥ 0.
One also refers to it as a C0 -semigroup.

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 181

DOI 10.1007/978-1-4471-2170-1_7, © Springer-Verlag London Limited 2011
182 7 Dynamics of semigroups, with applications to differential equations

Condition (iii) expresses the pointwise continuity of the semigroup. The

Banach–Steinhaus theorem (see Appendix A) yields that the family (Tt )t≥0
is then locally equicontinuous, that is, for any b > 0 we have that

sup Tt  < ∞,

t∈[0,b]

or equivalently, there exists some M > 0 such that

Tt x ≤ M x for all t ∈ [0, b], x ∈ X. (7.1)

We can express condition (iii) in some useful equivalent ways.

Lemma 7.1. Let (Tt )t≥0 be a family of operators on X. Then the following
assertions are equivalent:
(i) lims→t Ts x = Tt x for all x ∈ X and t ≥ 0;
(ii) (Tt )t≥0 is locally equicontinuous and there is a dense subset X0 of X
such that lims→t Ts x = Tt x for all x ∈ X0 and t ≥ 0;
(iii) the map
R+ × X → X, (t, x) → Tt x
is continuous.

Proof. It suffices to show that (ii) implies (iii). This implication follows im-
mediately from the identity

Tt x − Ts y = (Tt − Ts )(x − x0 ) + (Tt − Ts )x0 + Ts (x − y)

for x, y ∈ X, x0 ∈ X0 and s, t ≥ 0. 

Remark 7.2. Local equicontinuity implies, via (7.1), that Ttn xn → 0 whenever
(tn )n is bounded and xn → 0. This observation will be used repeatedly.

Moreover, one can establish an exponential bound for the operator norm
of the semigroup.

Proposition 7.3. If (Tt )t≥0 is a C0 -semigroup, then there exist M ≥ 1 and

w ∈ R such that Tt  ≤ M ewt for all t ≥ 0.

Proof. Let M = supt∈[0,1] Tt , which is finite and at least 1. Setting w =

log M and writing t ≥ 0 as t = n + s with n ∈ N0 and s ∈ [0, 1[, we obtain
that
Tt  = Ts T1n  ≤ M T1 n ≤ M enw ≤ M etw ,
where we have used the semigroup property. 

We introduce a standard example that we will revisit throughout this

chapter.
7.1 Semigroups of operators 183

Example 7.4. (Translation semigroups) Let 1 ≤ p < ∞ and let v : R+ →

R be a strictly positive locally integrable function, that is, v is measurable
b
with 0 v(x) dx < ∞ for all b > 0. We consider the space of weighted p-
integrable functions defined as

X = Lpv (R+ ) = {f : R+ → K ; f is measurable and f  < ∞},

where  ∞ 1/p
f  = |f (x)|p v(x) dx .
0
The translation semigroup is then given by

Tt f (x) = f (x + t), t, x ≥ 0.

We claim that this defines a C0 -semigroup on Lpv (R+ ) if and only if there
exist M ≥ 1 and w ∈ R such that, for all t ≥ 0,

v(x) ≤ M ewt v(x + t) for almost all x ≥ 0. (7.2)

Indeed, suppose that (7.2) is satisfied. If f ∈ X, we first observe that

 ∞  ∞
|f (x + t)|p v(x) dx ≤ M ewt |f (x + t)|p v(x + t) dx
0
0 ∞
≤ Me wt
|f (x)|p v(x) dx.
0

This shows that Tt : X → X, t ≥ 0, is well defined and continuous, and

that the family of translation operators is locally equicontinuous on X. Since
the semigroup property is obviously satisfied, it only remains to show that
lims→t Ts = Tt , for every t ≥ 0, pointwise on a dense subset of X; see Lemma
7.1. We can check it easily on the dense subspace of continuous functions of
compact support, using the fact that v is locally integrable.
Conversely, suppose that the translation semigroup (Tt )t≥0 is a C0 -semi-
group on X. Let M and w be given by Proposition 7.3. We will prove that,
for any t ≥ 0,

v(x) ≤ 2M p epwt v(x + t) for almost all x ≥ 0.

If this is not true, then there is some t0 > 0 such that

B := {x ≥ 0 ; v(x) > 2M p epwt0 v(x + t0 )}

has Lebesgue measure λ(B) > 0. Let b > 0 be such that λ(B ∩ [0, b]) > 0,
and define f (x) = v(x)11/p if x ∈ t0 + (B ∩ [0, b]), and f (x) = 0 otherwise. It
is clear that f ∈ X and f  > 0. On the other hand
184 7 Dynamics of semigroups, with applications to differential equations

p
Tt0 f  = |f (x + t0 )|p v(x) dx
B∩[0,b]

p
≥ 2M e
p pwt0
|f (x + t0 )|p v(x + t0 ) dx = 2M p epwt0 f  ,
B∩[0,b]

a contradiction with the choice of M and w. Hence (7.2) holds.

Now, in order to avoid some technical problems we will demand in the
sequel that the weight v satisfies (7.2) for any x ≥ 0. Equivalently, there are
constants M ≥ 1 and w ∈ R such that

v(x) ≤ M ew(y−x) v(y) whenever y ≥ x ≥ 0. (7.3)

In that case, v is called an admissible weight function.

We collect here some properties of C0 -semigroups. The easiest

 general con-
struction of C0 -semigroups is via an operator A on X. Since ∞ tn
n=0 n! A n
<
∞ for any t ≥ 0,
∞ n
t n
Tt = etA = A , t ≥ 0,
n=0
n!

defines operators on X, and it is easily seen that (Tt )t≥0 is a C0 -semigroup; we

even have that, for any t ≥ 0, lims→t Ts = Tt in the operator norm topology.
The semigroup is then called uniformly continuous. Moreover, for any x ∈ X,
Ax = limt→0 1t (Tt x − x). For all these statements see Exercise 7.1.6.
Now let (Tt )t≥0 be an arbitrary C0 -semigroup on X. It can be shown that

1
Ax := lim (Tt x − x)
t→0 t

exists on a dense subspace of X; the set of these x, the domain of A, is

denoted by D(A). Then A, or rather (A, D(A)), is called the (infinitesimal)
generator of the semigroup. It turns out that A : D(A) → X is a linear map
with closed graph, that is, for any sequence (xn )n in D(A), if limn→∞ xn =: x
and limn→∞ Axn =: y exist in X, then x ∈ D(A) and Ax = y. Moreover,
Tt (D(A)) ⊂ D(A) with ATt x = Tt Ax for every t ≥ 0 and x ∈ D(A). See
Exercise 7.1.7 for these statements. It can also be shown that the generator
determines the semigroup uniquely.
Another important property is provided by the point spectral mapping
theorem for semigroups (see Appendix B): if X is a complex Banach space
then, for every x ∈ X and λ ∈ C,

Ax = λx =⇒ Tt x = eλt x (7.4)

for every t ≥ 0.
As an example, the generator (A, D(A)) of the translation semigroup on
X = Lpv (R+ ), 1 ≤ p < ∞, can be shown to be
7.2 Hypercyclic and chaotic C0 -semigroups 185

D(A) = {f ∈ X ; f is absolutely continuous and f  ∈ X}, Af = f  .

The case when A is defined on all of X, that is, when A is an operator on X,

occurs precisely when (Tt )t≥0 is uniformly continuous; see Exercise 7.1.8.

Proposition 7.5. Let (Tt )t≥0 be a C0 -semigroup on X. Then the following

assertions are equivalent:
(i) the semigroup is uniformly continuous;
(ii) the generator A of the semigroup is defined everywhere;
(iii) there is an operator A on X such that Tt = etA , t ≥ 0.

We finally mention an interpretation of the generator that is the start-

ing point of the applications of semigroup theory to linear differential equa-
tions, as will be pursued in Section 7.5. Let (Tt )t≥0 be a C0 -semigroup on X,
(A, D(A)) its generator and, for any x ∈ X,

u(·, x) : R+ → X, u(t, x) := Tt x,

the function that describes the orbit of x under the semigroup. Whenever
x ∈ D(A), u(·, x) is the unique solution of the abstract Cauchy problem

dt u(t) = Au(t) for t ≥ 0,
d
(ACP)
u(0) = x.

And, for arbitrary x ∈ X, u(·, x) is the unique solution of the corresponding

integral equation  t
u(t) = A u(s) ds + x, t ≥ 0.
0

In that sense, u(·, x) is called a classical solution of (ACP) if x ∈ D(A) and

a mild solution of (ACP) if x ∈ X; (Tt )t≥0 is called the solution semigroup
of (ACP).
Conversely, suppose that A is a linear map with closed graph defined on a
dense subspace D(A) ⊂ X such that (ACP) has a unique (classical) solution
u(·, x) for every x ∈ D(A). If, moreover, for any sequence (xn )n in D(A)
with xn → 0 and any b > 0, u(t, xn ) → 0 uniformly on [0, b], then A is the
infinitesimal generator of a C0 -semigroup.

7.2 Hypercyclic and chaotic C0 -semigroups

We now begin our investigation of the dynamical properties of C0 -semigroups.

The concepts of hypercyclicity, topological transitivity, mixing, weak mixing
and chaos for operators all have a natural continuous analogue.

Definition 7.6. Let (Tt )t≥0 be a C0 -semigroup on X.

186 7 Dynamics of semigroups, with applications to differential equations

(a) For any x ∈ X we call

orb(x, (Tt )) = {Tt x ; t ≥ 0}

the orbit of x under (Tt )t≥0 .

(b) The semigroup is called hypercyclic if there is some x ∈ X whose orbit
under (Tt )t≥0 is dense in X. In such a case, x is called a hypercyclic vector
for (Tt )t≥0 .

Definition 7.7. A C0 -semigroup (Tt )t≥0 on X is called topologically transi-

tive if, for any pair U, V of nonempty open subsets of X, there exists some
t ≥ 0 such that Tt (U ) ∩ V = ∅.

Using the fact that X is separable, we easily deduce from the definitions
that hypercyclicity of a C0 -semigroup is equivalent to hypercyclicity of (Ttn )n
for some positive sequence (tn )n with tn → ∞; see Definition 3.23. Moreover,
hypercyclicity and topological transitivity for C0 -semigroups are equivalent
notions; see Exercise 7.2.1.

Definition 7.8. Let (Tt )t≥0 be a C0 -semigroup on X.

(a) The semigroup is called mixing if, for any pair U, V of nonempty open
subsets of X, there exists some t0 ≥ 0 such that Tt (U ) ∩ V = ∅ for all t ≥ t0 .
(b) The semigroup is called weakly mixing if (Tt ⊕ Tt )t≥0 is topologically
transitive on X ⊕ X.

One should note that, for any C0 -semigroups (Tt )t≥0 and (St )t≥0 on Ba-
nach spaces X and Y , respectively, (Tt ⊕ St )t≥0 is a C0 -semigroup on X ⊕ Y .
As in the discrete case, the direct sum of a mixing semigroup with a hyper-
cyclic semigroup is hypercyclic.
Finally, we also have a concept of chaos.

Definition 7.9. Let (Tt )t≥0 be a C0 -semigroup on X.

(a) A point x ∈ X is called a periodic point of (Tt )t≥0 if there is some
t > 0 such that Tt x = x. The set of periodic points for (Tt )t≥0 is denoted by
Per((Tt )).
(b) The semigroup is said to be chaotic if it is hypercyclic and its set of
periodic points is dense in X.

Example 7.10. (a) We consider the translation semigroup on the space X =

Lpv (R+ ), where 1 ≤ p < ∞ and v : R+ → R is an admissible weight function.
We claim that the following assertions are equivalent:
(i) the translation semigroup is hypercyclic;
(ii) the translation semigroup is weakly mixing;
(iii) lim inf x→∞ v(x) = 0.
We first show that (i) implies (iii). Suppose that v is bounded away from
zero. Since, by (7.3), v is bounded on [0, 1] there is then some C>0 such that
v(x) ≤ C p v(x+t) for all x ∈ [0, 1], t ≥ 0. Let g be defined by g(x) = v(x)−1/p
7.2 Hypercyclic and chaotic C0 -semigroups 187

on [0, 1], and 0 otherwise. Then we have for any f ∈ X with f  < (2C)−1
and any t ≥ 0 that
 1 1/p
Tt f − g ≥ |f (x + t) − g(x)|p v(x) dx
0
 1 1/p   1 1/p
≥ |g(x)|p v(x) dx − |f (x + t)|p v(x) dx
0 0
 1 1/p 1
≥1−M |f (x + t)|p v(x + t) dx ≥ 1 − Cf  > .
0 2

Consequently, (Tt )t≥0 cannot be hypercyclic.

Next we will show that the semigroup is topologically transitive under con-
dition (iii). Let X0 be the dense subspace of continuous functions of compact
support. It suffices to show that, for any f1 , f2 ∈ X0 and ε > 0, there is some
t ≥ 0 and some g ∈ X such that

f1 − g < ε and Tt g = f2 .

To this end, let

t0 = max supp(f1 ) ∪ supp(f2 ) .

Then, for any t > t0 , we define gt (x) = f1 (x) for x < t and gt (x) = f2 (x − t)
for x ≥ t. We clearly have that Tt gt = f2 . By (iii) there is some t ≥ t0 such
that
p
M 2 ewt0 f2  v(t0 + t) < v(0)ε,
where M and w are the parameters associated with the admissibility of v;
see (7.3). Therefore,
 ∞  ∞
p
f1 − gt  = |f2 (x − t)| v(x) dx =
p
|f2 (x)|p v(x + t) dx
t
 ∞ 0

≤ M v(t0 + t) |f2 (x)|p ew(t0 −x) dx

0
 ∞
2 wt0 v(t0 + t)
≤M e |f2 (x)|p ewx v(x)e−wx dx
v(0) 0

2 wt0 v(t0 + t)
=M e f2 p < ε,
v(0)

which had to be shown.

Indeed, by the same argument, the direct sum (Tt ⊕ Tt )t≥0 is topologically
transitive, which shows (ii).
(b) Continuing the study of the translation semigroup of (a), we claim
that the following assertions are equivalent:
(i) the translation semigroup is mixing;
(ii) limx→∞ v(x) = 0.
188 7 Dynamics of semigroups, with applications to differential equations

We need only modify the arguments in (a) slightly. Thus, suppose that (ii)
fails, that is, there is some sequence (tn )n with tn → ∞ such that (v(tn ))n is
bounded away from zero. The ∞admissibility condition (7.3) implies that v is
bounded away from zero on n=1 [tn , tn + 1]. But then there is some C > 0
such that v(x) ≤ C p v(x + tn ) for all x ∈ [0, 1], n ≥ 1, and the argument in
(a) shows that the sequence of operators (Ttn )n cannot be hypercyclic, which
contradicts (i). The proof that (ii) implies (i) is exactly as in (a).
(c) Let X = C0 (R+ ) be the space of continuous functions f : R+ → K
with limx→∞ f (x) = 0, endowed with the sup-norm f  = supx∈R+ |f (x)|.
Given a constant w > 0, we consider the family of operators defined by
(Tt f )(x) = ewt f (x + t), x ∈ R+ , which is easily seen to define a C0 -semigroup
on X.
We claim that this semigroup is mixing and chaotic. To this end, let X0
be the dense subspace of continuous functions with compact support, and let
f1 ∈ X0 , f2 ∈ X and ε > 0. For any t > t0 := max supp(f1 ) we define gt ∈ X
by ⎧
⎪f1 (x)
⎨ if x ≤ t0 ,
gt (x) = e−wt f2 (0) x−t 0
if t0 < x < t,
⎪
⎩ −wt
t−t0
e f2 (x − t) if x ≥ t.
Then we have that Tt gt = f2 and

f1 − gt  = e−wt f2  < ε

for any t > t0 sufficiently large. As in (a), this yields the mixing property for
the semigroup.
In order to show that the semigroup is even chaotic, it suffices to prove
that, for any f ∈ X0 and ε > 0, there is a periodic point h ∈ X with
f − h < ε. To this end, for any t > t0 := max supp(f ), we define the
function ht ∈ X by

e−wkt f (x − kt) if kt ≤ x ≤ kt + t0 , k ∈ N0 ,
ht (x) = −w(k+1)t
e x−kt−t0
f (0) t−t0 if kt + t0 < x < (k + 1)t, k ∈ N0 .

Then Tt ht = ht , so that ht is a periodic point. On the other hand,

f − ht  = sup |ht (x)| ≤ max e−wkt f  < ε

x≥t0 k≥1

for any t > t0 sufficiently large, which had to be shown.

In the previous chapters we realized how important quasiconjugacies are

for discrete dynamical systems. The same is true in the continuous case.

Definition 7.11. Let (Tt )t≥0 and (St )t≥0 be C0 -semigroups on separable
Banach spaces X and Y , respectively.
7.2 Hypercyclic and chaotic C0 -semigroups 189

(a) Then (Tt )t≥0 is called quasiconjugate to (St )t≥0 if there exists a con-
tinuous map φ : Y → X with dense range such that Tt ◦ φ = φ ◦ St for all
t ≥ 0. If φ can be chosen to be a homeomorphism then (Tt )t≥0 and (St )t≥0
are called conjugate.
(b) A property P of C0 -semigroups is said to be preserved under (quasi)
conjugacy if any C0 -semigroup that is (quasi)conjugate to a C0 -semigroup
with property P also possesses property P.

Proposition 7.12. Hypercyclicity, mixing, weak mixing and chaos for a C0 -

semigroup are preserved under quasiconjugacy.

As a special case we consider again the notion of complexification; see

the discussion before Proposition 2.26. If X is a real Banach space and X 

is its complexification then the family (Tt )t≥0 of complexifications of a C0 -
semigroup (Tt )t≥0 on X is a C0 -semigroup on X. Since (Tt )t≥0 is quasicon-

jugate to (Tt )t≥0 via the canonical projection of X onto X, we obtain the
following.

Corollary 7.13. If the complexification (Tt )t≥0 is hypercyclic (or weakly

mixing, mixing or chaotic, respectively) then so is (Tt )t≥0 .

In the discrete case we saw that adjoints of hypercyclic operators have an

empty point spectrum. The same happens for the infinitesimal generator A
of a hypercyclic C0 -semigroup. In this context we call λ ∈ K an eigenvalue
of A∗ if there is some x∗ ∈ X ∗ , x∗ = 0, such that Ax, x∗  = λx, x∗  for all
x ∈ D(A). Note that we do not need to define A∗ .

Lemma 7.14. Let (Tt )t≥0 be a hypercyclic C0 -semigroup on X with infinites-

imal generator (A, D(A)).
(a) Then, for every t > 0, the adjoint Tt∗ has no eigenvalues. Equiva-
lently, every operator Tt − λI, λ ∈ K, has dense range. Moreover, A∗ has no
eigenvalues.
(b) If X is a real Banach space, then, for every t > 0, the adjoint Tt∗ has
no eigenvalues. Equivalently, every operator Tt − λI, λ ∈ C, has dense range.

Proof. (a) Proceeding by contradiction, we suppose that orb(x, (Tt )) is dense

in X for some x ∈ X and that there are t > 0, λ ∈ K, and x∗ ∈ X ∗ \ {0}
such that Tt∗ x∗ = λx∗ . We will distinguish two cases:
Case 1: |λ| < 1. In this case we fix y ∈ X such that y, x∗  = 1. Then
there is a sequence (tn )n that tends to infinity such that limn→∞ Ttn x = y.
We write tn = mn t + sn with mn ∈ N0 and sn ∈ [0, t[, n ∈ N, and we obtain
that

1 = lim Ttn x, x∗  = lim Tsn x, (Tt∗ )mn x∗  = lim λmn Tsn x, x∗  = 0,

n→∞ n→∞ n→∞

since the sequence (Tsn x)n is bounded in X. This is a contradiction.

190 7 Dynamics of semigroups, with applications to differential equations

Case 2: |λ| ≥ 1. On the one hand, let r > 0 be such that |Tr x, x∗ | > 1.
On the other hand, by equicontinuity of (Ts )s∈[0,t] , there is some ε > 0 such
that |Ts y, x∗ | < 1 if s ∈ [0, t] and y < ε. We can also find t > r such that
Tt x < ε. If we write t = mt − s + r for some m ∈ N and s ∈ [0, t], then,
putting all these facts together, we obtain that

1 > |Ts (Tt x), x∗ | = |Tr x, (Tt∗ )m x∗ | = |λ|m |Tr x, x∗ | > 1,

which is a contradiction.
Next, let (A, D(A)) be the generator of (Tt )t≥0 , and suppose that A∗ has
an eigenvalue λ ∈ K. Let x∗ ∈ X ∗ \ {0} be such that Ax, x∗  = λx, x∗ 
for all x ∈ D(A). Given x ∈ D(A), we define h(t) = Tt x, x∗ , t ≥ 0. Then
Tt x ∈ D(A) and h (t) = ATt x, x∗  = λTt x, x∗  = λh(t) for all t ≥ 0; see
Exercise 7.1.7. Therefore,

x, Tt∗ x∗  = h(t) = h(0)eλt = eλt x, x∗ , t ≥ 0.

Since x ∈ D(A) is arbitrary and D(A) is dense in X, we conclude that x∗ is

an eigenvector of every Tt∗ , and we know that this is impossible.
The remaining claim follows from the Hahn–Banach theorem; see the proof
of Lemma 2.53.
(b) It is easy to adapt the above argument. One need only note that, for
any x∗ ∈ X ∗ \ {0} there is some y ∈ X such that |y, x
∗ | > 0; see the proof
of Lemma 2.53(b). 

Since all operators on CN have eigenvalues, we deduce the following as in

the discrete case; see Corollary 2.59.

Theorem 7.15. There are no hypercyclic C0 -semigroups on a finite-dimen-

sional Banach space.

Furthermore, we obtain the analogue of Bourdon’s theorem for C0 -semi-

groups of operators. The proof is identical.

Theorem 7.16. If (Tt )t≥0 is a hypercyclic C0 -semigroup and p a nonzero

polynomial, then every operator p(Tt ), t > 0, has dense range.

From this we can deduce the continuous analogue of the Herrero–Bourdon

theorem. The proof is the same as in the discrete case. However, one needs to
add that, by Theorem 6.8, any vector that is hypercyclic for the semigroup
(Tt )t≥0 is also hypercyclic for every operator Tt , t > 0. It is important to
realize that we have no vicious circle here: the proof of Theorem 6.8 used
Theorem 7.16, but not Theorem 7.17.

Theorem 7.17. If (Tt )t≥0 is a C0 -semigroup and x is a hypercyclic vector,

then, for any t > 0,

{p(Tt )x ; p is a polynomial} \ {0}

7.2 Hypercyclic and chaotic C0 -semigroups 191

is a dense set of hypercyclic vectors.

In particular, any hypercyclic C0 -semigroup admits a dense subspace of
vectors consisting, except for zero, of hypercyclic vectors.

We can also apply Theorem 7.16 to obtain a necessary condition for chaos
of a C0 -semigroup on complex Banach spaces; compare with Proposition 5.7.

Proposition 7.18. If (A, D(A)) is the generator of a chaotic C0 -semigroup

on a complex Banach space X, then

σp (A) ∩ iR

is infinite and, moreover,


X = span ker(λI − A).
λ∈iR

Proof. By hypothesis, the set Per((Tt )) of periodic points of the semigroup

is dense in X. By the point spectral mapping theorem for semigroups (see
Appendix B), we know that
    
2πni
Per((Tt )) = ker(I − Tt ) = span ker I −A
t>0 t>0
t
n∈Z
   
2πni
⊂ span ker I −A .
t>0
t
n∈Z

The density of periodic points implies that X = span λ∈iR ker(λI − A).
On the other hand, if σp (A) ∩ iR was finite, then there would exist finite
collections of nk ∈ Z, tk > 0, k = 1, . . . , N , such that


N  
2πnk i
Per((Tt )) ⊂ span ker I −A . (7.5)
tk
k=1

Moreover, by the point spectral mapping theorem for semigroups, we have

that, for k = 1, . . . , N ,
 
2πnk i
ker I − A ⊂ ker(I − Ttk ). (7.6)
tk

It follows from (7.5) and (7.6) that the operator T := (I − Tt1 ) · · · (I − TtN )
vanishes on Per((Tt )). Since this set is dense we have that T = 0. But this
contradicts Theorem 7.16 by which T has dense range. 

Example 7.19. We consider again the translation semigroup on the space X =

Lpv (R+ ), 1 ≤ p < ∞, for an admissible weight v. Then the following assertions
are equivalent:
192 7 Dynamics of semigroups, with applications to differential equations

(i) the
 ∞ translation semigroup is chaotic;
(ii) 0 v(x) dx < ∞.
We will see below that a semigroup on a real Banach space is chaotic if and
only if its complexification is; see Corollary 7.24. Therefore we may assume
that we are dealing with the complex space X. Thus, if the semigroup is
chaotic, Proposition 7.18 implies that its generator has an eigenvalue on the
imaginary axis. Since the generator is the derivative, we can find some s ∈ R
such that the function x → eisx , x ∈ R+ , belongs to X. This yields (ii).
Conversely, suppose that v is integrable on R+ . By Example 7.10(a), the
semigroup is then hypercyclic. Now let f be a continuous function of compact
support on R+ and ∞ ε > 0. Setting K = 1 + maxx∈R+ |f (x)|, we can find some
t > 0 such that t v(x) dx < (ε/2K)p . We define g(x + nt) = f (x), x ∈ [0, t[,
n ∈ N0 . The function g is bounded, thus g ∈ X by the integrability of the
weight. Moreover, Tt g = g and
 ∞ 1/p   ∞ 1/p
f − g ≤ |f (x)|p v(x) dx + |g(x)|p v(x) dx
t
 ∞
t
1/p
≤ 2K v(x) dx < ε.
t

Since the continuous functions of compact support are dense in X, the semi-
group has a dense set of periodic points and is therefore chaotic.

7.3 Discretizations of C0 -semigroups

In this section we study the relationship between hypercyclicity, mixing and

weak mixing of a semigroup and the corresponding properties of its discretiza-
tions. In this way, we will establish a direct link between the continuous and
the discrete case.
A discretization of a semigroup (Tt )t≥0 is a sequence of operators (Ttn )n
with tn → ∞. If there is some t0 > 0 such that tn = nt0 for n ∈ N, then
(Ttn )n = (Ttn0 )n is called an autonomous discretization of (Tt )t≥0 .
As we have already observed, a C0 -semigroup is hypercyclic if and only
if it admits a hypercyclic discretization (Ttn )n . The following results pro-
vide characterizations of mixing and weakly mixing semigroups in terms of
discretizations.

Proposition 7.20. Let (Tt )t≥0 be a C0 -semigroup on X. Then the following

assertions are equivalent:
(i) (Tt )t≥0 is weakly mixing;
(ii) some discretization of (Tt )t≥0 is mixing;
(iii) some discretization of (Tt )t≥0 is weakly mixing.
7.3 Discretizations of C0 -semigroups 193

Proof. If the semigroup is weakly mixing, then there exists a hypercyclic

discretization (Ttn ⊕ Ttn )n of (Tt ⊕ Tt )t≥0 . Since the operators Ttn , n ∈ N,
commute, Theorem 3.25 (see also Exercise 1.4.2) implies that there exists a
mixing subsequence (Ttnk )k of (Ttn )n . The other implications are trivial. 

Proposition 7.21. Let (Tt )t≥0 be a C0 -semigroup on X. Then the following

assertions are equivalent:
(i) (Tt )t≥0 is mixing;
(ii) every discretization of (Tt )t≥0 is mixing;
(iii) every discretization of (Tt )t≥0 is weakly mixing;
(iv) every discretization of (Tt )t≥0 is hypercyclic;
(v) every autonomous discretization of (Tt )t≥0 is mixing;
(vi) some autonomous discretization of (Tt )t≥0 is mixing.

Proof. The implications (i) =⇒ (ii) =⇒ (iii) =⇒ (iv) and (ii) =⇒ (v) =⇒
(vi) are trivial.
(iv) =⇒ (i). Suppose that (Tt )t≥0 is not mixing. Then there exists a pair
of nonempty open subsets U, V of X and a sequence (tn )n tending to ∞ such
that Ttn (U ) ∩ V = ∅ for all n ≥ 1. Then (Ttn )n cannot be topologically
transitive.
(vi) =⇒ (i). Let (Tnt0 )n be a mixing autonomous discretization of (Tt )t≥0 .
We fix arbitrary nonempty open sets U, V ⊂ X. By Lemma 2.36 there are
nonempty open sets U1 ⊂ U , V1 ⊂ V and a 0-neighbourhood W such that
U1 + W ⊂ U and V1 + W ⊂ V . By local equicontinuity of the semigroup
there is a 0-neighbourhood W1 such that Ts (W1 ) ⊂ W for all s ∈ [0, t0 ]. The
hypothesis then implies the existence of some N ∈ N such that

Tnt0 (U1 ) ∩ W1 = ∅ and Tnt0 (W1 ) ∩ V1 = ∅

for every n ≥ N . Given any t ≥ N t0 , we write t = nt0 + s with n ∈ N,

n ≥ N , and s ∈ [0, t0 ]. We can then find u1 ∈ U1 and w1 ∈ W1 such that
Tnt0 u1 ∈ W1 and T(n+1)t0 w1 ∈ V1 . Thus

Tt u1 = Ts Tnt0 u1 ∈ Ts (W1 ) ⊂ W,

and for w := Tt0 −s w1 ∈ W we have that

Tt w = T(n+1)t0 w1 ∈ V1 ,

that is, Tt (u1 + w) ∈ V with u1 + w ∈ U . This shows (i). 

In this book we are mainly interested in autonomous discretizations. As we

have noted, hypercyclicity is inherited by the semigroup if some autonomous
discretization has this property, and the same is true for weak mixing, mixing
and chaos. The natural question is whether, conversely, we can deduce hyper-
cyclicity of autonomous discretizations from hypercyclicity of the semigroup.
The Conejero–Müller–Peris theorem (see Chapter 6) gives a positive answer.
194 7 Dynamics of semigroups, with applications to differential equations

But since its proof is highly nontrivial we will provide here a simpler proof of
a classical result that is somewhat weaker but still strong enough for many
applications.

Theorem 7.22 (Oxtoby–Ulam). If (Tt )t≥0 is a C0 -semigroup on X and

x ∈ X is a hypercyclic vector, then there is a dense Gδ -set J ⊂ ]0, ∞[ such
that, for every t ∈ J, x is hypercyclic for Tt .

Proof. Let x ∈ X be a hypercyclic vector for the C0 -semigroup. We fix a

countable base of nonempty open sets (Uk )k in X and set

Jk = {t ∈ ]0, ∞[ ; Tnt x ∈ Uk for some n ∈ N}.

Each Jk is an open subset of ]0, ∞[. We will see that it is also dense. Indeed,
if 0 < a < b < ∞, there is some n0 ∈ N such that n0 b > (n0 + 1)a. Then

]na, nb[ = ]n0 a, ∞[.
n≥n0

Let s > n0 a be such that Ts x ∈ Uk . Then there exists some n ≥ n0 with

s ∈ ]na, nb[. If we define t = s/n ∈ ]a, b[, then we get that Tnt x ∈ Uk , so that
t ∈ Jk ∩ ]a, b[. Thus, each set Jk is dense in ]0, ∞[.
To finish the proof we consider
∞

J= Jk ,
k=1

which is a dense Gδ -subset of ]0, ∞[; note that the Baire category theorem is
applicable since ]0, ∞[ is isomorphic to R. Let t ∈ J. Then, for every k ∈ N,
there exists some n ∈ N such that Tnt x ∈ Uk . That is, x is a hypercyclic
vector for Tt . 

The theorem yields that, as in the discrete case, chaos implies weak mixing.

Theorem 7.23. Let (Tt )t≥0 be a hypercyclic C0 -semigroup on X. If there

exists a dense subset X0 of X such that the orbit of each x ∈ X0 is bounded,
then (Tt )t≥0 is weakly mixing.
In particular, every chaotic C0 -semigroup is weakly mixing.

Proof. By the Oxtoby–Ulam theorem, some operator Ts , s > 0, is hypercyclic.

On the other hand, the orbit of every x ∈ X0 under Ts is contained in the
orbit of x under (Tt )t≥0 , which is a bounded set. Thus, by Theorem 2.48, Ts
is weakly mixing, which implies the same property for the semigroup. 

Corollary 7.24. A C0 -semigroup on a real Banach space is chaotic if and

only if its complexification is.
7.3 Discretizations of C0 -semigroups 195

Proof. Suppose that the semigroup (Tt )t≥0 on a real Banach space X is
chaotic. Since its complexification (Tt )t≥0 can be identified with the direct
sum (Tt ⊕ Tt )t≥0 it is hypercyclic by the preceding result. Moreover, if U
and V are nonempty open subsets of X then, by topological transitivity and
continuity, there is a nonempty open subset U1 ⊂ U and some t ≥ 0 such
that Tt (U1 ) ⊂ V . We can then find some periodic point x ∈ U1 . Let Ts x = x,
s > 0. Then y := Tt x ∈ V and Ts y = Tt Ts x = y. Therefore, (x, y) ∈ U × V
is a periodic point for (Tt ⊕ Tt )t≥0 , which concludes the proof that (Tt )t≥0 is
chaotic. The converse implication was obtained in Corollary 7.13. 

The Oxtoby–Ulam theorem also implies that weakly mixing C0 -semigroups

contain weakly mixing autonomous discretizations. However, a stronger result
can be derived.

Proposition 7.25. Let (Tt )t≥0 be a C0 -semigroup on X. Then the following

assertions are equivalent:
(i) (Tt )t≥0 is weakly mixing;
(ii) some autonomous discretization of (Tt )t≥0 is weakly mixing;
(iii) every autonomous discretization of (Tt )t≥0 is weakly mixing.

Proof. It suffices to show that (ii) implies (iii). In view of the Bès–Peris
theorem we have to show that if some Tt0 , t0 > 0, satisfies the Hypercyclicity
Criterion, then so does every Tt1 , t1 > 0. Since Tt1 = Ttp1 /p , p ∈ N, we may
assume that t1 < t0 ; see Exercise 6.1.3.
Now, Theorem 3.22 tells us that Tt0 even satisfies the Gethner–Shapiro
criterion; that is, there are dense subsets X0 , Y0 of X, an increasing sequence
(nk )k of positive integers and a mapping St0 : Y0 → Y0 such that, for any
x ∈ X0 and y ∈ Y0 , Ttn0k x → 0, Stn0k y → 0, and Tt0 St0 y = y. As the proof of
Theorem 3.22 shows, Y0 can be taken in the form Y0 = {yn ; n ∈ N}, where
the yn are pairwise distinct and St0 yn = yn+1 for all n ∈ N.
Since t1 < t0 there are mk ∈ N0 and sk ∈ [0, t0 [, k ∈ N, such that

mk t1 = nk t0 + sk , k ∈ N.

We claim that also Tt1 satisfies the Hypercyclicity Criterion, with respect to
X0 , Y0 , and the sequence (mk )k .
Indeed, for every x ∈ X0 we have that

Ttm
1
k
x = Tsk Ttn0k x → 0

as k → ∞, where we use the fact that the family (Ts )s∈[0,t0 ] is equicontinuous.
Next we define a family (St )t≥0 of maps on Y0 by setting St yn = Ts yn+k
if t = kt0 − s for some k ∈ N and 0 < s ≤ t0 ; note that this is consistent with
St0 . Then we have for every yn ∈ Y0 , n ∈ N, that

Smk t1 yn = S(nk +1)t0 −(t0 −sk ) yn = Tt0 −sk yn+nk +1 = Tt0 −sk Stn0k yn+1 → 0
196 7 Dynamics of semigroups, with applications to differential equations

as k → ∞, again by equicontinuity.
Finally, for every yn ∈ Y0 , n ∈ N, we obtain that

Ttm
1
k
Smk t1 yn = Tmk t1 Tt0 −sk Stn0k yn+1 = Ttn0k +1 Stn0k +1 yn = yn .

Altogether we find that Tt1 satisfies the Hypercyclicity Criterion.



To finish this section we reformulate the Conejero–Müller–Peris theorem in

the language of discretizations. The theorem contains both the Oxtoby–Ulam
theorem and Proposition 7.25.

Theorem 7.26. Let (Tt )t≥0 be a C0 -semigroup on X and x ∈ X. Then the

following assertions are equivalent:
(i) x is hypercyclic for (Tt )t≥0 ;
(ii) x is hypercyclic for some discretization of (Tt )t≥0 ;
(iii) x is hypercyclic for some autonomous discretization of (Tt )t≥0 ;
(iv) x is hypercyclic for every autonomous discretization of (Tt )t≥0 .

7.4 Criteria for the hypercyclicity and chaos of

C0 -semigroups

In this section we derive, in analogy to the discrete case, the Hypercyclicity

Criterion and various other criteria for weak mixing, mixing and chaos of
semigroups of operators.

Theorem 7.27 (Hypercyclicity Criterion for semigroups). Let (Tt )t≥0

be a C0 -semigroup on X. If there are dense subsets X0 , Y0 ⊂ X, a sequence
(tn )n in R+ with tn → ∞, and maps Stn : Y0 → X, n ∈ N, such that, for
any x ∈ X0 , y ∈ Y0 ,
(i) Ttn x → 0,
(ii) Stn y → 0,
(iii) Ttn Stn y → y,
then (Tt )t≥0 is weakly mixing, and in particular hypercyclic.

Proof. This is an immediate consequence of the Hypercyclicity Criterion for

sequences of operators, Theorem 3.24, applied to the sequence (Ttn )n . 

Calling a semigroup (Tt )t≥0 hereditarily hypercyclic if there is a sequence

(tn )n in R+ with tn → ∞ such that, for every subsequence (tnk )k , the se-
quence of operators (Ttnk )k admits a dense orbit, we have the following ana-
logue of the Bès–Peris theorem.

Theorem 7.28. Let (Tt )t≥0 be a C0 -semigroup on X. Then the following

assertions are equivalent:
(i) (Tt )t≥0 satisfies the Hypercyclicity Criterion;
7.4 Criteria for the hypercyclicity and chaos of C0 -semigroups 197

(ii) (Tt )t≥0 is weakly mixing;

(iii) (Tt )t≥0 is hereditarily hypercyclic.

Proof. Suppose that (Tt )t≥0 is weakly mixing. Then, by Proposition 7.25, the
operator T1 is weakly mixing and therefore hereditarily hypercyclic, so that
the same holds for (Tt )t≥0 .
On the other hand, if (Tt )t≥0 is hereditarily hypercyclic then there is some
sequence (tn )n in R+ with tn → ∞ such that (Ttn )n is a hereditarily transitive
sequence of operators. By Theorem 3.25, (Tn )n satisfies the Hypercyclicity
Criterion for sequences of operators (Theorem 3.24), which implies that the
semigroup (Tt )t≥0 satisfies the Hypercyclicity Criterion. 

If, in the Hypercyclicity Criterion, one has convergence along the whole
real line then we obtain a criterion for mixing.

Theorem 7.29. Let (Tt )t≥0 be a C0 -semigroup on X. If there are dense sub-
sets X0 , Y0 ⊂ X, and maps St : Y0 → X, t ≥ 0, such that, for any x ∈ X0 ,
y ∈ Y0 ,
(i) Tt x → 0,
(ii) St y → 0,
(iii) Tt St y → y,
then (Tt )t≥0 is mixing.

Proof. By the Hypercyclicity Criterion for sequences of operators (Theorem

3.24), every discretization (Ttn )n of (Tt )t≥0 is weakly mixing. It then suffices
to apply Proposition 7.21. Of course, a simple direct proof can also be given;
see the proof of Kitai’s theorem. 

We will now consider complex Banach spaces. In the discrete case, the
Godefroy–Shapiro criterion tells us that a large supply of eigenvectors of an
operator T to eigenvalues in an open set intersecting the unit circle leads
to mixing and chaos of T . An analogous result is true when one studies
eigenvectors of the generator A to eigenvalues in an open set that intersects
the imaginary axis.
By a weakly holomorphic function f : U → X on an open set U ⊂ C we
understand an X-valued function such that, for every x∗ ∈ X ∗ , the complex-
valued function z → f (z), x∗  is holomorphic on U . In the sequel, J is a
nonempty index set.

Theorem 7.30. Let X be a complex separable Banach space, and (Tt )t≥0 a
C0 -semigroup on X with generator (A, D(A)). Assume that there exists an
open connected subset U and weakly holomorphic functions fj : U → X,
j ∈ J, such that
(i) U ∩ iR = ∅,
(ii) fj (λ) ∈ ker(λI − A) for every λ ∈ U , j ∈ J,
(iii) for any x∗ ∈ X ∗ , if fj (λ), x∗  = 0 for all λ ∈ U and j ∈ J then x∗ = 0.
Then the semigroup (Tt )t≥0 is mixing and chaotic.
198 7 Dynamics of semigroups, with applications to differential equations

Proof. We will use the Hahn–Banach theorem to prove the density of certain
subspaces generated by eigenvectors, and then apply the Godefroy–Shapiro
criterion.
For a fixed t > 0, we consider the subspaces

X0 = span{fj (λ) ; j ∈ J, λ ∈ U with Re λ > 0},

Y0 = span{fj (λ) ; j ∈ J, λ ∈ U with Re λ < 0},

and

Z0 = span{fj (λ) ; j ∈ J, λ ∈ U with λt = απi for some α ∈ Q}.

We first show that these spaces are dense in X. To this end, suppose that
x∗ ∈ X ∗ is identically 0 on X0 . For any j ∈ J, the function λ → fj (λ), x∗ 
is holomorphic on U , and, in view of (i), it vanishes on some nonempty open
subset of U . Therefore it vanishes on U , for any j ∈ J. By (iii) we have that
x∗ = 0. The Hahn–Banach theorem then implies that X0 is dense in X. In
the same way it follows that also Y0 and Z0 are dense.
Now, by (7.4), we have that

Tt x = eλt x

for all x ∈ ker(λI −A). Therefore, by (ii), X0 and Y0 are contained in the span
of the eigenvectors of Tt to eigenvalues of modulus greater than 1 and smaller
than 1, respectively, and Z0 is contained in the span of the eigenvectors of
Tt to eigenvalues that are roots of unity. The Godefroy–Shapiro criterion
yields that Tt is mixing and chaotic, properties that are inherited by the
C0 -semigroup; see Proposition 7.21. 

One can strengthen the above criterion in the sense that the selection map
f need not be defined on an open subset of C, and we can replace weak
holomorphy by continuity, as we will see next.
First of all, for any continuous vector-valued function f : [a, b] → X one
can define the Riemann integral
 b
f (t) dt
a

in the usual way by its Riemann sums; see Appendix A for more details and
basic properties.

Lemma 7.31 (Riemann–Lebesgue lemma). Let X be a complex Banach

b
space and f : [a, b] → X a continuous function. Then a eirt f (t) dt → 0 as
r → ±∞.

Proof. Let ε > 0. We choose a partition a = t0 < t1 < . . . < tN = b of [a, b]

such that f (t) − f (tj ) < ε whenever tj−1 ≤ t ≤ tj , j = 1, . . . , N . Let g be
7.4 Criteria for the hypercyclicity and chaos of C0 -semigroups 199

the step function that takes the value f (tj ) on ]tj−1 , tj ], j = 1, . . . , N . Then
we find that
 b   b   b 
     
 eirt
f (t) dt ≤
  e irt
(f (t) − g(t)) dt + eirt g(t) dt
a a a

N  tj 
 
≤ (b − a)ε + f (tj ) eirt dt.
j=1 tj−1

 tj
Now, since tj−1
eirt dt → 0 as r → ±∞, the claim follows. 

Theorem 7.32. Let X be a complex separable Banach space, and (Tt )t≥0 a
C0 -semigroup on X with generator (A, D(A)). Assume that there are a < b
and continuous functions fj : [a, b] → X, j ∈ J, such that
(i) fj (s) ∈ ker(isI − A) for every s ∈ [a, b], j ∈ J,
(ii) span{fj (s) ; s ∈ [a, b], j ∈ J} is dense in X.
Then the semigroup (Tt )t≥0 is mixing and chaotic.
Proof. We apply Theorem 7.29 to show that the semigroup is mixing. For
b
any r ∈ R, j ∈ J, we define xr,j = a eirs fj (s) ds, and we set

X0 = Y0 = span{xr,j ; r ∈ R, j ∈ J}.

To prove that this subspace is dense in X we will again use the Hahn–Banach
theorem. Thus let x∗ be a continuous linear functional on X such that, for
all r ∈ R, j ∈ J,
 b
xr,j , x∗  = eirs fj (s), x∗  ds = 0.
a

The functions s → fj (s), x∗  are continuous on [a, b] and therefore belong to
1 2π
L2 [a, b]. Since ( √b−a exp ( b−a ikt))k∈Z is an orthonormal basis in this Hilbert
space we deduce that, in view of continuity,

fj (s), x∗  = 0 for all s ∈ [a, b], j ∈ J.

Hence x∗ vanishes on span{fj (s) ; s ∈ [a, b], j ∈ J}, which is dense in X by

(ii). Thus x∗ = 0, which implies that X0 = Y0 is dense.
Now, by (i) and (7.4) we have that, for any t ≥ 0,

Tt fj (s) = eist fj (s) for s ∈ [a, b], j ∈ J, (7.7)

so that
 b  b
irs
Tt xr,j = e Tt fj (s) ds = ei(t+r)s fj (s) ds = xt+r,j . (7.8)
a a

The Riemann–Lebesgue lemma then implies that Tt x → 0 for every x ∈ X0 .

200 7 Dynamics of semigroups, with applications to differential equations

Next we would like to define maps St on Y0 = X0 by demanding that

St xr,j = xr−t,j , r ∈ R, j ∈ J, and then extending linearly. However this
will run into difficulties if the xr,j are not linearly independent. We therefore
observe, as we have already done in a similar situation (see Exercise 3.1.1),
that instead of verifying conditions (ii) and (iii) in Theorem 7.29 it suffices
to produce, for any y ∈ Y0 , a family (ut )t≥0 in X such that ut → 0 and
→ y as t → ∞. To this end, for any y ∈ Y0 = X0 we fix a representation
Tt ut 
y= m k=1 ak xrk ,jk and define


m
ut = ak xrk −t,jk .
k=1

By (7.8) we then have that Tt ut = y for all t ≥ 0 and, again by the Riemann–
Lebesgue lemma, that ut → 0 as t → ∞.
Thus, (Tt )t≥0 is a mixing semigroup.
Finally, by continuity of the functions fj , also

Z0 := span{fj (s) ; s ∈ [a, b] ∩ Q, s = 0, j ∈ J}

is dense in X. By (7.7), every vector fj (s), s = 0, is a fixed point of T2π/s ,

and thus any point in Z0 is a periodic point for (Tt )t ; note that any finitely
many rational numbers are integer multiples of a common rational number.
Hence the semigroup is also chaotic. 

Example 7.33. Let (Tt )t≥0 be the translation semigroup on the complex Ba-
nach space X = Lpv (R+ ), 1 ≤ p < ∞, where v(x) = e−x . The dual of X
is given by the space Lqw (R+ ), where 1 < q ≤ ∞ satisfies p1 + 1q = 1, with
w = v −q/p for p > 1 and w = v−1 for p = 1 (see Example A.3(c)); note
that L∞w (R+ ) is the space of measurable functions g on R+ such that gw is
essentially bounded.
We define f : D → X by f (λ)(x) = eλx , λ ∈ D, x ∈ R+ . Then f is well
defined and weakly holomorphic on D. Clearly Af (λ) = dx d
f (λ) = λf (λ). If
∗
a continuous linear functional on X , represented by g ∈ Lqw (R+ ), satisfies
 ∞
f (λ), g = g(x)eλx dx = 0
0

for all λ ∈ D, then we obtain, by taking the nth derivative with respect to λ,
 ∞
xn g(x)eλx dx = 0, n ∈ N0 , λ ∈ D.
0

Therefore, letting λ = 0 and taking linear combinations, we obtain that

 ∞
g(x)p(x) dx = 0
0
7.5 Applications of C0 -semigroups to differential equations 201

for every polynomial p. Since the polynomials form a dense subspace of X,

we conclude that g = 0.
Thus, Theorem 7.30 shows that the translation semigroup is mixing and
chaotic, confirming our earlier findings in Examples 7.10 and 7.19.
We note that for C0 -semigroups on real Banach spaces one can still apply
the above eigenvalue criteria. Indeed, if they are satisfied for the complexifica-
tion of the semigroup then the original semigroup is also chaotic and mixing
by Corollary 7.13.

7.5 Applications of C0 -semigroups to differential

equations

The purpose of this section is to present some applications of C0 -semigroups

to the study of the asymptotic behaviour of solutions of linear partial dif-
ferential equations (PDEs) or of infinite linear systems of ordinary differ-
ential equations (ODEs). We do not pretend to obtain a full description of
all possible situations, but we want to show the main techniques in certain
representative cases.

A first-order PDE. We start with the following first-order abstract Cauchy

problem on the space X = L1 (R+ ):
⎧
⎪
⎨ ∂u = ∂u + 2x u,
∂t ∂x 1 + x2 (7.9)
⎪
⎩u(0, x) = ϕ(x), x ∈ R .
+

The solution C0 -semigroup is given by

1 + (x + t)2
Tt ϕ(x) = ϕ(x + t), x, t ∈ R+ .
1 + x2
That is, the action of the semigroup is a translation, together with multipli-
2
cation by the function ht (x) = 1+(x+t)
1+x2
. The fact that ht (x) → ∞ as t → ∞
opens up the possibility of approximating arbitrary functions by small func-
tions under the action of Tt , and thus of obtaining the mixing property for
the semigroup. On the other hand, the fact that ht (x) → 1 as x → ∞ pre-
vents eigenvectors of Tt to eigenvalues of modulus greater than 1. Therefore
the eigenvalue criterion of Theorem 7.30 cannot be applied.
Proposition 7.34. The solution semigroup (Tt )t≥0 of (7.9) is mixing and
chaotic on L1 (R+ ).

Proof. We first apply the mixing criterion, Theorem 7.29. We take for X0
the dense subspace of X = L1 (R+ ) of functions with compact support. If the
202 7 Dynamics of semigroups, with applications to differential equations

support of ϕ ∈ X0 is contained in [0, b], then Tt ϕ = 0 for every t ≥ b, and

condition (i) of the criterion is satisfied. For conditions (ii) and (iii) we take
Y0 = X and define the maps St : X → X by

1 + (x − t)2
St ϕ(x) = ϕ(x − t), for x ≥ t ≥ 0,
1 + x2
and St (x) = 0 for 0 ≤ x < t. Then Tt St ϕ = ϕ for every ϕ ∈ X. Moreover,
 ∞  ∞
1 + (x − t)2 1 + x2
St ϕ = 2
|ϕ(x − t)| dx = |ϕ(x)| dx → 0
t 1+x 0 1 + (x + t)2

as t → ∞ by the dominated convergence theorem. Thus the semigroup is

mixing.
As for chaos, let ϕ ∈ X0 and ε > 0. Let the support of ϕ be contained in
∞ 2 ∞
[0, b], and fix t ≥ b such that n=1 1+bn2 t2 ϕ < ε. We define ψ = n=0 Snt ϕ.
Since
∞ ∞  b ∞
1 + x2 1 + b2
Snt ϕ = |ϕ(x)| dx ≤ ϕ < ε,
n=1 n=1 0
1 + (x + nt)2 n=1
n 2 t2

ψ is well defined and ϕ − ψ < ε. Moreover, ψ is periodic for Tt , so that the

semigroup has a dense set of periodic points, implying that it is chaotic. 

A second-order PDE. In our first example we obtained mixing and chaos

for a first-order PDE by analysing the corresponding solution C0 -semigroup.
A similar treatment is possible for higher-order PDEs.
As a specific example we study the chaotic behaviour of the solutions of
an abstract Cauchy problem that is given by the hyperbolic heat transfer
equation in the absence of internal heat sources:
⎧ 2
⎪ ∂ u ∂u ∂2u
⎪τ
⎪ + = α ,
⎪
⎨ ∂t2 ∂t ∂x2
u(0, x) = ϕ1 (x), x ∈ R, (HHTE)
⎪
⎪
⎪ ∂u
⎪
⎩ (0, x) = ϕ (x), x ∈ R,
2
∂t
where ϕ1 and ϕ2 represent the initial temperature and the initial variation
of temperature, respectively, α > 0 is the thermal diffusivity, and τ > 0 is
the thermal relaxation time.
We will express this system as a first-order equation by representing it as
a C0 -semigroup on the product of a certain function space with itself. To do
this we set u1 = u and u2 = ∂u
∂t . Then the associated first-order equation is
7.5 Applications of C0 -semigroups to differential equations 203
⎧   ⎛ ⎞ 
⎪
⎪ 0 I
⎪
⎪ ∂ u1 u1
⎪
⎪ = ⎝ α ∂2 1 ⎠ ,
⎨ ∂t u2 − I u2
τ ∂x 2 τ
    (7.10)
⎪
⎪
⎪
⎪ u1 (0, x) ϕ1 (x)
⎪
⎩ u (0, x) = ϕ (x) , x ∈ R.
⎪
2 2

We fix ρ > 0 and consider the space

 ∞
an ρ n n 
Xρ = f : R → C ; f (x) = x , (an )n≥0 ∈ c0 , (7.11)
n=0
n!

endowed with the norm f  = supn≥0 |an |, where c0 is the Banach space
of complex sequences tending to 0. Then Xρ is a Banach space of analytic
functions with a certain growth control. By its definition it is isometrically
isomorphic to c0 .
Since ⎛ ⎞
0 I
A := ⎝ α ∂ 2 1 ⎠
− I
τ ∂x2 τ
is easily seen to be an operator on X := Xρ ⊕ Xρ , we have that (etA )t≥0 is a
C0 -semigroup on X, which is the solution semigroup of (7.10) on X.

Proposition 7.35. Let ρ > 0 be such that ατ ρ2 > 2. Then the solution
semigroup (etA )t≥0 of (7.10) is mixing and chaotic on Xρ ⊕ Xρ .

Proof. Given λ ∈ C, z0 , z1 ∈ R, we set Rλ = (τ λ2 + λ)/α and define

∞ ∞
Rλn x2n Rλn x2n+1
ϕλ,z0 ,z1 (x) = z0 + z1 , x ∈ R.
n=0
(2n)! n=0
(2n + 1)!

Let U be the open disk of radius r = αρ2 /2τ > 0 centred at zero. If λ ∈ U
then
2
αρ2 !
|Rλ | τ αρ
2τ + 2τ 1 1 2
< = + < 1, (7.12)
ρ2 αρ2 2 2 ατ ρ2
so that ϕλ,z0 ,z1 ∈ Xρ . We now consider the functions fz0 ,z1 : U → X, z0 , z1 ∈
R, given by  
ϕλ,z0 ,z1
fz0 ,z1 (λ) = .
λϕλ,z0 ,z1
Then one finds that

fz0 ,z1 (λ) ∈ ker(λI − A), λ ∈ U, z0 , z1 ∈ R.

204 7 Dynamics of semigroups, with applications to differential equations

According to Theorem 7.30 it remains to prove that, for any x∗ ∈ Xρ∗ ⊕ Xρ∗ ,
the functions λ → fz0 ,z1 (λ), x∗ , z0 , z1 ∈ R, are holomorphic on U , and that
if they all vanish on U then x∗ = 0.
Thus, let x∗ ∈ Xρ∗ ⊕ Xρ∗ . By the isomorphism of Xρ with c0 , x∗ can be
represented in a canonical way by a pair ((bn )n≥0 , (cn )n≥0 ) ∈ 1 ⊕ 1 . We
then have that, for λ ∈ U ,

fz0 ,z1 (λ), x∗  =

∞ ∞ ∞ ∞
b2n b2n+1 c2n c2n+1
z0 Rλn 2n + z1 Rλn 2n+1 + λz0 Rλn 2n + λz1 Rλn 2n+1 .
n=0
ρ n=0
ρ n=0
ρ n=0
ρ

By (7.12), these series converge uniformly on U , which implies that each

function fz0 ,z1 is weakly holomorphic on U .
Finally, suppose that all the functions λ → fz0 ,z1 (λ), x∗ , z0 , z1 ∈ R,
vanish on U . Considering, in particular, the values λ = 0 and λ = −1/τ in
U we obtain that z0 b0 + z1 b1 /ρ = 0 and z0 c0 + z1 c1 /ρ = 0 for all z0 , z1 ∈ R,
hence b0 = b1 = c0 = c1 = 0. As a consequence, we may divide out Rλ and
still retain holomorphic functions that vanish on U . Considering again λ = 0
and λ = −1/τ we obtain in the same way that b2 = b3 = c2 = c3 = 0.
Proceeding inductively we deduce that bn = cn = 0 for all n ≥ 0, and hence
that x∗ = 0. This had to be shown. 

An infinite system of ODEs. We conclude this section by providing an ap-

plication of C0 -semigroups to an infinite linear system of ordinary differential
equations.
Let us consider the following infinite system of ODEs associated with a
linear kinetic model:
⎧
⎪
⎨ dfn = −αn fn + βn fn+1 , n ≥ 1,
dt (7.13)
⎪
⎩f (0) = a , n ≥ 1,
n n

where (αn )n , (βn )n are bounded positive sequences and (an )n ∈ 1 is a real
sequence.
In this case we consider the real Banach space X = 1 and the map A
given by
Af = (−αn fn + βn fn+1 )n for f = (fn )n .
Since A is an operator on 1 , it generates a C0 -semigroup (Tt )t≥0 , which is
then the solution semigroup of the abstract Cuachy problem (7.13).

Proposition 7.36. Let αn > 0 and βn ∈ R, n ∈ N, be such that

sup αn < lim inf βn . (7.14)

n≥1 n→∞
Exercises 205

Then the solution semigroup (Tt )t≥0 of (7.13) is mixing and chaotic on 1 .

Proof. Let α = supn≥1 αn , β = lim inf n→∞ βn and α/2 < r < β/2. We fix
U ⊂ C as the open disk of radius r centred at −α/2, which intersects the
imaginary axis iR. We can calculate the eigenvectors of A to the eigenvalues
λ ∈ U . Indeed, suppose that f = (fn )n satisfies Af = λf , λ ∈ U . Then

λfn = −αn fn + βn fn+1 , n ≥ 1,

and therefore
"
n−1
λ + αk
fn = γn f1 with γn = , n ≥ 1,
βk
k=1

where γ1 = 1. Conversely, f (λ) := (γn )n satisfies Af (λ) = λf (λ); we still

need to verify that f (λ) ∈ 1 .
To this end let δ ∈ ]2r, β[. Then there is some n0 ∈ N such that βn > δ for
all n ≥ n0 . Thus, since our assumptions imply that −αn ∈ U for all n ∈ N,
we have that
|λ + αn | 2r
≤ <1 (7.15)
βn δ
for all n ≥ n0 , which implies that f (λ) ∈ X = 1 . We can therefore consider
the function
f : U → X, λ → f (λ).
Let x∗ ∈ X ∗ = ∞ be given by some bounded sequence (bn )n . Then

∞  n−1
" λ + αk 
f (λ), x∗  = bn
n=1
βk
k=1

converges uniformly for λ ∈ U by (7.15), which implies that f is weakly

holomorphic on U .
Finally, let x∗ ∈ X ∗ , given by (bn )n ∈ ∞ , be such that f (λ), x∗  = 0 for
all λ ∈ U . Considering, in particular, λ = −α1 ∈ U we obtain that b1 = 0.
As a consequence we may divide out λ + α1 and still retain a holomorphic
function on U . Then taking λ = −α2 ∈ U we obtain that b2 = 0. Proceeding
inductively we deduce that bn = 0 for all n ≥ 1, and therefore x∗ = 0.
We have shown that all the hypotheses of Theorem 7.30 are satisfied, which
implies the result. 

Exercises

Exercise 7.1.1. Let Δ = Δ(α) := {reiθ ; r ≥ 0, |θ| ≤ α}, 0 < α ≤ π/2, be a sector
in the complex plane, and v : Δ → R a strictly positive measurable function for which
206 7 Dynamics of semigroups, with applications to differential equations

there exist constants M ≥ 1 and w ∈ R such that v(z1 ) ≤ M ew|z2 | v(z1 + z2 ) for all
z1 , z2 ∈ Δ, called an admissible weight function on Δ. For 1 ≤ p < ∞ we define the
space
Lpv (Δ) = {f : Δ → K ; f is measurable and f  < ∞}

with f  = ( Δ |f (z)|p v(z) dλ(z))1/p , where λ denotes two-dimensional Lebesgue mea-
sure. Show that, for every z ∈ Δ, the translation semigroup (Ttz )t≥0 given by Ttz f (ζ) =
f (ζ + tz) is a C0 -semigroup on Lpv (Δ).

Exercise 7.1.2. Let v : R+ → R be a strictly positive continuous function, and define

the space X = C0,v (R+ ) by

C0,v (R+ ) = {f : R+ → K ; f is continuous and limx→∞ f (x)v(x) = 0},

endowed with the norm f  = supx≥0 |f (x)|v(x). Consider the translation semigroup
given by Tt f (x) = f (x + t), t, x ≥ 0. Show that this defines a C0 -semigroup on X if and
only if v is admissible in the sense of (7.3).

Exercise 7.1.3. For the operator A on R2 given by the matrix 01 −1 0 determine the
semigroup (etA )t≥0 , and verify that A is its infinitesimal

generator. Describe the orbits
{etA x ; t ≥ 0}, x ∈ R2 . Do the same for −1 1 1
−1 .

Exercise 7.1.4. Let ϕ be a bounded holomorphic function on the unit disk D. Then
Tt f = etϕ f , t ≥ 0, defines multiplication operators on the Hardy space H 2 ; see Section
4.4. Show that (Tt )t≥0 is a uniformly continuous semigroup on H 2 , and identify its
generator. Do the same for the family (Tt∗ )t≥0 of Hilbert space adjoints.

Exercise 7.1.5. Let T be an operator on a Banach space X such that T  < 1. Then
∞ (−1)n+1 n
the series R := n=1 n T converges in the operator norm, so that R ∈ L(X).
Prove that eR = I + T .

Exercise 7.1.6. Let A be an operator on a Banach space X. Show that the operators
Tt = etA , t ≥ 0, define a C0 -semigroup on X. Moreover, show that lims→t Ts − Tt  = 0
and Ax = limt→0 1t (Tt x − x) for all x ∈ X.

Exercise 7.1.7. Let (Tt )t≥0 be a C0 -semigroup on a Banach space X and (A, D(A))
its infinitesimal generator. Show the following for any t ≥ 0:
(i) for any x ∈ D(A), Tt x ∈ D(A) and ATt x = Tt Ax;
(ii) for any x ∈ D(A), lims→t Ts x−T tx
= ATt x;
t s−t
t
(iii) for any x ∈ X, Ts x ds ∈ D(A) and A Ts x ds = Tt x − x;
0 t 0
(iv) for any x ∈ D(A), 0 Ts Ax ds = Tt x − x,
where the integral is the vector-valued Riemann integral; see Appendix B. Deduce that
 t+h t
D(A) is dense and that A has closed graph. (Hint for (ii): h Ts x ds − 0 Ts x ds =
 t+h h
t
Ts x ds − 0
Ts x ds.)

Exercise 7.1.8. Prove Proposition 7.5. (Hint: For (i) =⇒ (ii), use Lemma 8.20 to show
t
that 1t 0 Ts ds is an invertible operator if t > 0 is sufficiently small, and use Exercise
7.1.7(ii). For (ii) =⇒ (iii), use the closed graph theorem and the fact that the infinitesimal
generator determines the C0 -semigroup.)

Exercise 7.2.1. Let (Tt )t≥0 be a C0 -semigroup on a separable Banach space X.

(a) If x ∈ X is hypercyclic for (Tt )t≥0 and U a nonempty open subset of X then, for
any t0 ≥ 0, there is some t ≥ t0 such that Tt x ∈ U . (Hint: In an infinite-dimensional
space, the unit ball is not relatively compact.)
Exercises 207

(b) Prove the Birkhoff transitivity theorem for C0 -semigroups, that is: (Tt )t≥0 is
hypercyclic if and only if it is topologically transitive. In that case, the set of hypercyclic
vectors for (Tt )t≥0 is a dense Gδ -set.

Exercise 7.2.2. A one-parameter family (Tt )t∈R of operators on a Banach space X is

called a C0 -group of operators if T0 = I, Tt+s = Tt Ts for all s, t ∈ R, and lims→t Ts x =
Tt x for all x ∈ X and t ∈ R. The orbit of a vector x ∈ X is the set {Tt x ; t ∈ R} and
the C0 -group is called hypercyclic if it admits a dense orbit. Prove that if a C0 -group
(Tt )t∈R of operators is hypercyclic, then so are the C0 -semigroups (Tt )t≥0 and (T−t )t≥0 .
(Hint: Proceed as in Exercise 6.3.1(ii).)

Exercise 7.2.3. Let (Tt )t≥0 be the translation semigroup on X = Lpv (R+ ) for an ad-
missible weight v. Establish the equivalence of the following assertions:
(i) 
(Tt )t≥0 is chaotic;
∞
(ii) v(n) < ∞;
n=1 ∞
(iii) for all ε > 0 and t > 0, there is some s > 0 such that n=1
v(t + ns) < ε.

Exercise 7.2.4. Let (Tt )t≥0 be the translation semigroup on X = C0,v (R+ ) for an
admissible continuous weight function v; see Exercise 7.1.2. Show that (Tt )t≥0 is hyper-
cyclic if and only if lim inf x→∞ v(x)=0. And that it is mixing, if and only if it is chaotic,
and if and only if limx→∞ v(x)=0. (Hint: Use the fact that the generator A is given by
Af = f  on the subspace of continuously differentiable functions f ∈ X with f  ∈ X.)

Exercise 7.2.5. Let (Tt )t≥0 be the translation semigroup on the space X = Lpv (Δ); see
Exercise 7.1.1. Let Δm = {z ∈ Δ ; Im(z) ∈ [−m, m]}, m ∈ N. Demonstrate that the
following are equivalent:
 translation semigroup is chaotic on X;
(i) the
(ii) Δ v(z) dλ(z) < ∞ for every m ∈ N.
m

Exercise 7.3.1. Let Δ = Δ(π/4), v : Δ → R the √ admissible weight function defined by

v(x+iy) = (x+y+1)2 (x−y+1)−2 if x+y+1 ≥ x − y + 1, and v(x+iy) := (x+y+1)−2
otherwise; see Figure 7.1. Let Tz , z ∈ Δ, be the translation operators on the space
X = Lpv (Δ); see Exercise 7.1.1. The dynamical notions of C0 -semigroups introduced so
far naturally extend to the semigroup (Tz )z∈Δ .

Fig. 7.1 The weight function on Δ(π/4)

Show that there is a mixing nonautonomous discretization (Tzn )n of (Tz )z∈Δ , but
that no C0 -subsemigroup (Ttz )t≥0 , z ∈ Δ, is hypercyclic. In particular, no autonomous
208 7 Dynamics of semigroups, with applications to differential equations

√ of (Tz )z∈Δ is hypercyclic. (Hint: Take a discretization within the curve

discretization
x + y + 1 = x − y + 1. On the other hand, observe that the weight v is asymptotically
bounded below by 1 on any ray in Δ passing through 0.)

Exercise 7.3.2. Given a C0 -semigroup (Tt )t≥0 and z1 , z2 ∈ K with Re z1 ≤ Re z2 , we

consider the rescaled semigroups (ez1 t Tt )t≥0 and (ez2 t Tt )t≥0 . Show that, if the direct
sum semigroup (ez1 t Tt ⊕ez2 t Tt )t≥0 is hypercyclic, then the rescaled semigroup (ezt Tt )t≥0
is weakly mixing for every z ∈ K such that Re z1 ≤ Re z ≤ Re z2 . (Hint: Use the Oxtoby–
Ulam theorem and Exercise 2.5.9.)

Exercise 7.3.3. Let (Tt )t≥0 be a C0 -semigroup on X. Given a pair U, V ⊂ X of open

sets, we define
R(U, V ) = {t ≥ 0 ; Tt (U ) ∩ V = ∅}.
Prove that a C0 -semigroup is weakly mixing if and only if, for any 0-neighbourhood W
and for any pair U, V ⊂ X of nonempty open sets, we have that R(U, W )∩R(W, V ) = ∅.
(Hint: One implication is trivial; for the other one use the Oxtoby–Ulam theorem and
Theorem 2.45.)

Exercise 7.3.4. Prove that a C0 -semigroup (Tt )t≥0 is weakly mixing if and only if, for
every pair U, V ⊂ X of nonempty open sets, R(U, V ) contains arbitrarily long intervals.
(Hint: For one implication take any autonomous discretization and apply Theorem 1.54;
the other implication is based on the local equicontinuity of the semigroup and, again,
on Theorem 1.54.)

Exercise 7.3.5. Prove that a C0 -semigroup (Tt )t≥0 is weakly mixing if and only if there
exists some ε > 0 such that, for any pair U, V ⊂ X of nonempty open sets, R(U, V )
contains some interval of length ε. (Hint: Exercise 7.3.4, and Exercise 1.5.8 for Tt0 with
t0 ∈ ]0, ε/2[.)

Exercise 7.3.6. A C0 -semigroup (Tt )t≥0 of operators is topologically ergodic if, for ev-
ery pair U, V ⊂ X of nonempty open sets, R(U, V ) is syndetic, that is, R+ \ R(U, V )
does not contain arbitrarily long intervals. Show that every autonomous discretization
of a topologically ergodic C0 -semigroup is topologically ergodic. (Hint: Given t > 0,
nonempty open sets U, V ⊂ X, and an arbitrary 0-neighbourhood W , use local equicon-
tinuity to prove that N (U, W ) = {n ∈ N0 ; nt ∈ R(U, W )} and N (W, V ) = {n ∈
N0 ; nt ∈ R(W, V )} are syndetic. Then show that N (U, W ) ∩ N (W, V ) is (nonempty
and) syndetic.)

Exercise 7.3.7. Prove that, if (Tt )t≥0 and (St )t≥0 are topologically ergodic C0 -semi-
groups, then (Tt ⊕ St )t≥0 is also topologically ergodic. (Hint: Exercises 7.3.6 and 2.5.5.)

Exercise 7.3.8. Show the following:

(i) every chaotic C0 -semigroup is topologically ergodic;
(ii) if (Tt )t≥0 is a topologically ergodic C0 -semigroup and (St )t≥0 is weakly mixing,
then (Tt ⊕ St )t≥0 is weakly mixing.
(Hint to (ii): Proposition 7.25 and Exercises 7.3.6, 2.5.5 and 1.5.6.)

Exercise 7.4.1. Show directly that under the hypotheses of Theorem 7.27 the semi-
group is topologically transitive. Prove Theorem 7.29 in the same way.

Exercise 7.4.2. By using a direct argument, prove that if a C0 -semigroup (Tt )t≥0 sat-
isfies the Hypercyclicity Criterion then the operator T1 satisfies the Hypercyclicity Cri-
terion. (Hint: See the proof of Proposition 7.25.)

Exercise 7.4.3. Let (Tt )t≥0 be a C0 -semigroup. Show that the following assertions are
equivalent:
Exercises 209

(i) (Tt )t≥0 satisfies the criterion for mixing of Theorem 7.29;
(ii) every discretization of (Tt )t≥0 satisfies the Hypercyclicity Criterion for sequences
of operators for the full sequence (n);
(iii) some autonomous discretization of (Tt )t≥0 satisfies the Hypercyclicity Criterion
for sequences of operators for the full sequence (n).

Exercise 7.4.4. Suppose that a C0 -semigroup (Tt )t≥0 admits a discretization (Ttn )n∈N
with supn≥1 (tn+1 − tn ) < ∞ satisfying the Hypercyclicity Criterion for sequences of
operators for the full sequence (n). Prove that (Tt )t≥0 satisfies the criterion for mixing
of Theorem 7.29.

Exercise 7.4.5. Show the following generalization of Theorem 7.32. Let X be a complex
separable Banach space, and (Tt )t≥0 a C0 -semigroup on X with generator (A, D(A)).
Assume that there are compact intervals Ij = [aj , bj ] with aj < bj and continuous
functions fj : Ij → X, j ∈ J, such that
(i) fj (s) ∈ ker(isI − A) for every s ∈ Ij , j ∈ J,
(ii) span{fj (s) ; s ∈ Ij , j ∈ J} is dense in X.
Then the semigroup (Tt )t≥0 is mixing and chaotic.

Exercise 7.5.1. Provide an alternative proof of Proposition 7.34 for complex functions
by applying Exercise 7.4.5 to fj : [−j, j] → X, fj (s)(x) = exp (isx)
1+x2 , j ∈ N. (Hint: For
condition (ii), use the Hahn–Banach theorem and the fact that if the Fourier transform
of an integrable function vanishes identically then so does the function itself.)

Exercise 7.5.2. Consider the following abstract Cauchy problem on X = Lp (R+ ), 1 ≤

p < ∞, or on X = C0 (R+ ), that generalizes (7.9):
⎧
⎨ ∂u = ∂u + h(x)u,
∂t ∂x
⎩u(0, x) = ϕ(x), x ∈ R ,
+

where h : R+ → R is a bounded continuous function. The solution semigroup is given

by  x+t
h(s) ds
Tt ϕ(x) = e x ϕ(x + t), x, t ∈ R+ .
∞ x

Show that, on Lp (R+ ), the semigroup is chaotic if and only if 0 exp −p 0 h(s) ds dx
x
<∞, mixing if and only if limx→∞ 0 h(s) ds = ∞, and hypercyclic if and only if
x
supx≥0 0 h(s) ds = ∞. Find the corresponding characterizations on C0 (R+ ). (Hint:
Determine an admissible weight function v so that the semigroup is conjugate to the
translation semigroup on Lpv (R+ ) or on C0,v (R+ ); see Exercise 7.2.4.)

Exercise 7.5.3. Consider the following wave equation:

⎧ 2
⎪
⎪ ∂ u ∂2u
⎪
⎨ ∂t2
= α 2,
∂x
⎪ u(0, x) = ϕ1 (x), x ∈ R,
⎪
⎪
⎩ ∂u (0, x) = ϕ2 (x), x ∈ R,
∂t
where α > 0 is the square of the speed of wave propagation. As for the hyperbolic heat
transfer equation (HHTE), by setting u1 = u and u2 = ∂u ∂t , one can reduce it to a
first-order equation. Show that the solution semigroup to the corresponding first-order
equation is chaotic on Xρ ⊕Xρ , for any ρ > 0, where Xρ is the space of analytic functions
defined in (7.11).
210 7 Dynamics of semigroups, with applications to differential equations

Exercise 7.5.4. The classical heat equation

⎧
⎨ ∂u = ∂ u ,
2

∂t ∂x2
⎩
u(0, x) = ϕ(x), x ∈ R,

on the real line admits a solution semigroup given by

 ∞  
1 (x − s)2
Tt ϕ(x) = √ exp − ϕ(s) ds, t ∈ R+ , x ∈ R.
2 πt −∞
4t

∂2
If we consider the space Xρ , ρ > 0, defined in (7.11), then D2 = ∂x2 is an operator on
2
Xρ and Tt = etD . Show that Tt is chaotic on Xρ for any t > 0.
Exercise 7.5.5. For the abstract Cauchy problem associated with the linear kinetic
model (7.13), let α1 = 3, αn = 1, n ≥ 2, and βn = 2, n ∈ N. Prove that the solution
semigroup is not hypercyclic on 1 . This shows that condition (7.14) cannot be replaced
by lim supn→∞ αn < lim inf n→∞ βn .

Sources and comments

Section 7.1. For the theory of C0 -semigroups we refer to the books by Engel and Nagel
[143], [144]. All the results mentioned in this section can be found there.

Section 7.2. Rolewicz [268] was the first to observe the existence of a dense orbit under
a C0 -semigroup. A systematic study of the dynamical properties of semigroups, however,
was only started by Desch, Schappacher and Webb [131]. In particular, they introduced
the notions of hypercyclicity and chaos for semigroups. Several other results in this
chapter are due to them, for example the characterization of hypercyclic translation
semigroups on Lpv (R+ ). The characterization of the corresponding mixing property was
added by Bermúdez, Bonilla, Conejero, and Peris [49].
Lemma 7.14 on the necessity of empty point spectrum for the adjoint operators and
for the adjoint of the generator of a hypercyclic semigroup was obtained by Costakis
and Peris [121] and by Desch, Schappacher and Webb [131], respectively. Theorem 7.16
is due to Costakis and Peris [121].
Chaos for the translation semigroup was characterized by deLaubenfels and Emami-
rad [128]. Kalmes [209], [210] has undertaken a systematic study of hypercyclicity, weak
mixing, mixing and chaos of C0 -semigroups that are induced by solution semiflows of
differential equations on open subsets of Rd .

Section 7.3. Theorem 7.22 is due to Oxtoby and Ulam [251]. As for the discretization re-
sults, Proposition 7.20 is due to Conejero and Peris [113], Proposition 7.21 to Bermúdez,
Bonilla, Conejero, and Peris [49] and Conejero and Peris [113], and Proposition 7.25 to
Desch and Schappacher [129]; see also Conejero and Peris [113]. Theorem 7.23 can be
found in Bernal and Grosse-Erdmann [63]. As we already mentioned in Chapter 6, the
discretization results fail for chaotic semigroups; in fact, Bayart and Bermúdez [37] have
constructed a chaotic semigroup on a Hilbert space for which no autonomous discretiza-
tion is chaotic.

Section 7.4. The first criteria for hypercyclicity of C0 -semigroups were found by Desch,
Schappacher and Webb [131]. In the form given here, the Hypercyclicity Criterion, The-
Sources and comments 211

orem 7.27, is due to Conejero and Peris [111] and El Mourchid [139], while the criterion
for mixing, Theorem 7.29, is due to Bermúdez, Bonilla, Conejero, and Peris [49]. See
also Conejero and Peris [113].
As for the eigenvalue criteria, Theorem 7.30 is due to Desch, Schappacher and
Webb [131], while Theorem 7.32 is due to El Mourchid [140]; see also Conejero and
Mangino [109]. We mention that the latter result was motivated by a similar criterion
for frequent hypercyclicity; see Section 9.3.
Theorem 7.28, together with further equivalent forms of the Hypercyclicity Criterion,
can be found in Conejero and Peris [113]. Incidentally, Herrero’s problem if hypercyclic-
ity implies weak mixing (see Section 2.5), also has a negative answer for semigroups, as
was proved by Shkarin [292].

Section 7.5. The three applications treated in this section are, in order, from El
Mourchid [140], Conejero, Peris and Trujillo [114], and Banasiak and Lachowicz [24].
The first results on the hypercyclic behaviour of the heat equation were given by
Herzog [198]. Generalizations of Herzog’s results were accomplished by Betancor and
Bonilla [73]. Some recent developments on the heat dynamics are due to Ji and Weber
[207], [208].
The method presented here for the hyperbolic heat transfer equation is based on the
conversion of the second-order Cauchy problem to first order so that the theory of C0 -
semigroups can be applied. Alternatively, it is possible to keep the second-order problem
and solve it directly, by applying the theory of hypercyclic cosine operator functions; this
theory was initiated by Bonilla and Miana [89], and thoroughly developed by Kalmes
[211].
The chaotic dynamics associated with infinite systems of differential equations mod-
elling birth-and-death processes (evolution of cell population, etc.) have been systemat-
ically studied by Banasiak et al. in [24], [25], [22], [23], [26], [27].
We also mention an interesting recent application of the methods in this chapter to
the Black–Scholes formula. Emamirad, Goldstein and Goldstein [141] have shown that
the Black–Scholes semigroup is chaotic for certain choices of the parameters.

Exercises. The weight condition (iii) in Exercise 7.2.3 is due to Matsui, Yamada and
Takeo [236]. For Exercise 7.2.5 we refer to Conejero and Peris [112]. The results of
Exercise 7.2.4 are due to Desch, Schappacher and Webb [131] (hypercyclicity), Bermúdez,
Bonilla, Conejero and Peris [49] (mixing) and Matsui, Yamada and Takeo [236] (chaos).
Exercise 7.3.1 can be found in Conejero and Peris [113]. Exercises 7.3.3 and 7.3.4 are
taken from Bernal and Grosse-Erdmann [63]. Topologically ergodic semigroups were
considered by Desch and Schappacher in [129] under the name of semigroups satisfying
the Recurrent Hypercyclicity Criterion. Exercises 7.3.6, 7.3.7 and 7.3.8 are extracted
from their paper. Exercise 7.4.3 is taken from Conejero and Peris [113], Exercise 7.4.4
from Bermúdez, Bonilla, Conejero and Peris [49], and Exercise 7.4.5 from Conejero and
Mangino [109]. Exercise 7.5.1 is extracted from El Mourchid [140], Exercise 7.5.2 is
taken from, and improves on, Takeo [297], while Exercises 7.5.3 and 7.5.4 are taken from
Conejero, Peris and Trujillo [114] and Herzog [198], respectively.
Chapter 8
Existence of hypercyclic operators

In the previous chapters we saw that there exist hypercyclic operators on a

large variety of spaces. This suggests the question of whether every Fréchet
space supports a hypercyclic operator. We will show here that this is indeed
the case when one takes account of two natural restrictions: the space has
to be separable and, by a result in Chapter 2, also infinite dimensional. In
contrast, we will see that not every such Banach space supports a chaotic
operator. We also study the corresponding questions for semigroups of oper-
ators.
We then ask how large the set of hypercyclic operators is. The answer
depends on the underlying notion of size. We will show that if there exist
hypercyclic operators on a given space then, under a suitable topology, they
form a dense set.
In the final section, the flexibility of hypercyclic operators will be demon-
strated in yet another way. We know that the orbit of a hypercyclic vector
is a (dense) linearly independent set. Conversely, it will be shown that ev-
ery dense linearly independent sequence in a Banach space is the orbit of a
(necessarily) hypercyclic vector under a certain operator.

8.1 Mixing perturbations of the identity

In this section we prepare the ground for the proofs of the existence of hy-
percyclic operators and C0 -semigroups on arbitrary infinite-dimensional sep-
arable Fréchet and Banach spaces, respectively. What we will need are hy-
percyclic operators on 1 that are “small” perturbations of the identity or,
equivalently, perturbations of the identity on weighted 1 -spaces with rapidly
growing weights. Such operators are also interesting in their own right.
Let v = (vn )n be a positive weight sequence, that is, a sequence of strictly
positive numbers, and let X be one of the weighted spaces

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 213

DOI 10.1007/978-1-4471-2170-1_8, © Springer-Verlag London Limited 2011
214 8 Existence of hypercyclic operators

 ∞
 
p (v) = (xn )n≥1 ; |xn |p vn < ∞ , 1 ≤ p < ∞,
n=1
 
c0 (v) = (xn )n≥1 ; lim |xn |vn = 0 .
n→∞

By Example 4.4(a), the backward shift B defines an operator on X if and

vn
only if supn∈N vn+1 < ∞. Under this assumption,

∞
 1 k
eB = B
k!
k=0

is a well-defined operator on X; see the discussion before Proposition 4.41 or

Appendix B. Moreover, its powers are given by
∞
 nk
(eB )n = enB = Bk , n ≥ 1.
k!
k=0

Theorem 8.1. Let X be one of the spaces p (v), 1 ≤ p < ∞, or c0 (v). If

vn
supn∈N vn+1 < ∞, then the operator T := eB is mixing on X.

Proof. Since 1 (v) is densely and continuously contained in c0 (v) and in

p ((vnp )n ), the hypercyclic comparison principle (see Exercise 2.2.6) tells us
that it suffices to show the result for 1 (v).
Let X0 be the subspace of finite sequences, which is dense in 1 (v). We fix
arbitrary vectors x = (xk )k , y = (yk )k ∈ X0 and ε > 0, and we choose m ∈ N
such that xk = yk = 0 for k > m. Then the matrix
⎛ ⎞
1
m!
1
(m+1)! ··· 1
(2m−1)!
⎜ 1 1
··· 1 ⎟
⎜ (m−1)! m! (2m−2)! ⎟
W =⎜ .. .. .. .. ⎟
⎝ . . . . ⎠
1
1!
1
2! ··· 1
m!

is easily seen to be invertible; see Exercise 8.1.1. We let


m 
m 
C = W −1  |yk | + m |xk | , (8.1)
k=1 k=1

where W −1  denotes the norm of W −1 as an operator from (Km ,  · 1 ) to

(Km ,  · ∞ ). Finally, let N ∈ N be such that


2m
ε
N m−k vk < . (8.2)
eC
k=m+1

We will show that, for any n ≥ N , there is some z ∈ 1 (v) such that
8.1 Mixing perturbations of the identity 215

x − z < ε and y − T n z < ε,

which will imply that T is mixing.

Indeed, fix n ≥ N , and let z be a finite sequence such that zk = xk for
k = 1, . . . , m and zk = 0 for k ≥ 2m + 1. We claim that the missing m entries

of z can be chosen in such a way that the first m entries of T n z = ∞ nj j
j=0 j! B z
coincide with those of y, that is
∞
 
2m
nj nj−k
yk = zk+j = zj , k = 1, . . . , m.
j! (j − k)!
j=0 j=k

This is equivalent to solving the system

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
x1 zm+1 y1
⎜ ⎟ ⎜ ⎟ ⎜ .. ⎟ ,
A ⎝ ... ⎠ + V ⎝ ... ⎠ = ⎝ . ⎠ (8.3)
xm z2m ym

where
⎛ ⎞ ⎛ nm nm+1 n2m−1
⎞
nm−1 ···
1 n
1!
··· (m−1)! m! (m+1)! (2m−1)!
⎜ ⎟ ⎜ nm−1 nm n2m−2 ⎟
⎜0 1 ··· nm−2 ⎟ ⎜ ··· ⎟
A=⎜ (m−2)! ⎟, V =⎜
(m−1)! m! (2m−2)! ⎟.
⎜. .. . . .. ⎟ ⎜ .. .. .. .. ⎟
⎝ .. . . . ⎠ ⎝ . . . . ⎠
2
0 0 ··· nm
1 n
1!
n
2!
··· m!

Observe that V = D1 W D2 , where

⎛ m−1 ⎞ ⎛ ⎞
n 0 ··· 0 n 0 ··· 0
⎜ 0 nm−2 · · · 0 ⎟ ⎜0 n2 · · · 0 ⎟
⎜ ⎟ ⎜ ⎟
D1 = ⎜ . .. . . .. ⎟ , D2 = ⎜ . .. . . .. ⎟.
⎝ .. . . .⎠ ⎝ .. . . . ⎠
0 0 ··· 1 0 0 ··· nm

Thus the solution of system (8.3) is given by

⎛ ⎞ ⎡⎛ ⎞ ⎛ ⎞⎤
zm+1 y1 x1
⎜ .. ⎟ −1 −1 −1 ⎢⎜ . ⎟ ⎜ . ⎟⎥
⎝ . ⎠ = D2 W D1 ⎣⎝ .. ⎠ − A ⎝ .. ⎠⎦ , (8.4)
z2m ym xm

where
⎛ ⎞ ⎛ ⎞
n−1 0 · · · 0 n−m+1 0 ··· 0
⎜ 0 n−2 · · · 0 ⎟ ⎜ 0 n−m+2
··· 0⎟
⎜ ⎟ ⎜ ⎟
D2−1 =⎜ . .. . . .. ⎟ , D1−1 =⎜ .. .. .. .. ⎟ .
⎝ .. . . . ⎠ ⎝ . . . .⎠
0 0 · · · n−m 0 0 ··· 1
216 8 Existence of hypercyclic operators

This shows that the sequence z with the stated properties exists. Moreover,
since the entries of D1−1 and of D1−1 A are bounded by 1, we have that


m 
m
D1−1 [(y1 , . . . , ym )t − A(x1 , . . . , xm )t ]1 ≤ |yk | + m |xk |.
k=1 k=1

With (8.1) and (8.4) we deduce that

|zk | ≤ Cnm−k , k = m + 1, . . . , 2m. (8.5)

Thus (8.2) and (8.5) imply that


2m 
2m
x − z = |zk |vk ≤ C nm−k vk < ε.
k=m+1 k=m+1

On the other hand, the entries of T n z of index k ≥ 2m + 1 vanish, so that,

again by (8.2) and (8.5),
2m    2m 
  nm−k 
2m 2m
 nj−k 
y − T z =
n
 zj  v k ≤ C vk
(j − k)! (j − k)!
k=m+1 j=k k=m+1 j=k

 
2m

≤ eC nm−k vk < ε.
k=m+1

This had to be shown. 



As a first application we also obtain that the operator I + B is always

mixing.

Theorem 8.2. Let X be one of the spaces p (v), 1 ≤ p < ∞, or c0 (v). If

vn
supn∈N vn+1 < ∞, then the operator T := I + B is mixing on X.

Proof. Again, it suffices to prove the result for 1 (v). We claim that there is a
positive weight sequence w = (wn )n and an operator φ : 1 (w) → 1 (v) with
dense range such that the following diagram commutes:

eB
1 (w) −−−−→ 1 (w)
⏐ ⏐
φ
⏐ ⏐φ

I+B
1 (v) −−−−→ 1 (v).

To this end, let A = (ank )n≥1;k≤n be an infinite lower triangular matrix

with ann
= 0 for n ≥ 1, and define φ by φ(en ) = nk=1 ank ek , n ≥ 1, where
en denotes the canonical unit sequence, n ≥ 1. We want to choose A in such
a way that (I + B)φ(en ) = φ(eB en ) for n ≥ 1. On the one hand we have that
8.1 Mixing perturbations of the identity 217


n 
n−1
(I + B)φ(en ) = ank (I + B)ek = ann en + (ank + an,k+1 )ek .
k=1 k=1

On the other hand,

n−1
 1  n
1 
φ(eB en ) = φ en−k = φ ej
k! j=1
(n − j)!
k=0


n
1 
j    a n  n
jk
= ajk ek = ek .
j=1
(n − j)! (n − j)!
k=1 k=1 j=k

Setting these expressions equal and comparing the coefficients of en−1 yields
that an,n−1 + ann = an−1,n−1 + an,n−1 for n ≥ 2, so that the elements on the
diagonal must coincide; comparing the remaining coefficients implies that,
for k = 1, . . . , n − 2,

n
ajk
ank + an,k+1 = ,
(n − j)!
j=k

hence

n−1
ajk
an,k+1 = ,
(n − j)!
j=k

a system that we can solve inductively for k = 1, . . . , n − 2, where n ≥ 3.

Observe that the elements an1 , n ≥ 2, are free, since no restriction on them
is given. We can set them, for example, equal to 0. Similarly, the common
value on the diagonal is free, which we may set equal to 1.
Once the matrix A = (ank )n≥1;k≤n is fixed, we can find a sufficiently
rapidly increasing sequence w = (wn )n with supn∈N wwn+1 n
< ∞ such that
φ :  (w) →  (v) is a well-defined operator. By construction, (I + B) ◦ φ =
1 1

φ ◦ eB ; moreover, the range of φ contains any finite sequence and is therefore

dense. The result is now a consequence of Theorem 8.1 and Proposition 1.40.



We note that, conversely, Theorem 8.2 can also be derived from Theorem
8.1 (see Exercise 8.1.5), so that the two results are in fact equivalent. Another
equivalent statement will be discussed in Exercise 8.1.2.
Using a suitable conjugacy one can reformulate the previous two results in
terms of weighted shifts on unweighted spaces. We recall that, given a weight
sequence w = (wn )n , the weighted shift Bw is a well-defined operator on p ,
1 ≤ p < ∞, or on c0 if and only if supn∈N |wn | < ∞; see Example 4.9(a).

Corollary 8.3. Let X be one of the spaces p , 1 ≤ p < ∞, or c0 , and let

w = (wn )n be a weight sequence such that supn∈N |wn | < ∞. Then T = I +Bw
and T = eBw are mixing operators on X.
218 8 Existence of hypercyclic operators
n
Proof. Let X = p , 1 ≤ p < ∞, and set un = ( ν=1 wν )−1 , vn = |un |p . Then
vn
supn∈N vn+1 < ∞, and φ : p (v) → p , (xn )n → (xn un )n defines a conjugacy
between B : p (v) → p (v) and Bw : p → p . Linearity and continuity of φ
imply that the operators I + Bw and eBw on p are also conjugate to I + B
and to eB respectively, on p (v). We then conclude with Theorems 8.2 and
8.1. The proof for X = c0 is similar. 

The corollary furnishes an interesting example in connection with Kitai’s

theorem. Kitai tells us that, for any hypercyclic operator T on a complex
Banach space, every connected component of the spectrum of T meets the
unit circle. We will now see that, on the other hand, the spectrum may consist
of a single point on the unit circle.

Example 8.4. Let Bw be the weighted shift with w = (1/n)n on a complex

space p or c0 . Since
 
(Bw )n x = 2·3·...·(n+1)
1 1
xn+1 , 3·4·...·(n+2) xn+2 , . . . ,

we have that (Bw )n 1/n ≤ (n!)−1/n → 0 as n → ∞. Hence Bw is quasinilpo-

tent, so that, by the spectral radius formula, σ(Bw ) = {0}. Thus, T = I + Bw
is a mixing operator with σ(T ) = {1}. Note also that Bw is compact; see Ex-
ercise 5.2.10.

We also consider a variant of the operator I + Bw that will be of interest

later on; see Theorem 8.13.

Proposition 8.5. Let X be one of the spaces p , 1 ≤ p < ∞, or c0 . Let

w = (wn )n be a weight sequence such that supn∈N |wn | < ∞ and Dw the
operator on X given by

Dw (xn )n = (w2 x2 , w4 x4 , w6 x6 , . . .).

Then the operator T := I + Dw is mixing on X.

Proof. Once more it suffices to perform the proof for X = 1 . The operator
Dw is best understood by its action on the unit sequences en , n ≥ 1. If we
write n ≥ 1 as

n = m2k−1 with m ≥ 1 odd and k ≥ 1,

then
T en = 0 if k = 1, and T en = wm2k−1 em2k−2 if k ≥ 2.
In this way it becomes obvious that Dw is conjugate to a countable direct
sum of weighted shifts; see the explanations before Proposition 2.41. More
precisely, if we define weights w (m) = (wm2k−1 )k≥1 for m ≥ 1 odd, and the
map
8.2 Existence of mixing operators and semigroups 219
   
φ: 1 → 1 , (xm,k )k≥1 m odd
→ (yn )n ,
1
m odd

where yn = xm,k if n = m2k−1 , then φ is an isometric isomorphism, and the

following diagram commutes:

  Bw(m)  
1 −−−m
−− −−−−−→
odd
1
1 1
m odd
⏐ m odd
⏐
⏐
φ
⏐φ

D
1 −−−w−→ 1 .
 
Then also I + Dw and I + m odd Bw(m) = m odd (I + Bw(m) ) are conjugate
via φ. Since, by Corollary 8.3,each operator I + Bw(m) is mixing on 1 ,
Proposition 2.41 implies that m odd (I + Bw(m) ) is mixing, and therefore
also I + Dw . 
To conclude this section we mention that all the results obtained so far
are special cases of a single very general theorem. Its proof uses, in essence,
the main argument of the proof of Theorem 8.1. To formulate the result we
need the notion of an extended backward shift, that is, an operator T on a
Banach space for which

∞ 
span (ker T n ∩ ran T n )
n=0

is dense. If T is surjective then the condition simply says that T has dense
generalized kernel. Moreover, any weighted backward shift on a Banach se-
quence space in which the finite sequences are dense is an extended backward
shift.
Theorem 8.6. Let A be an extended backward shift on a Banach space X.
Then the operators eA and I + A are mixing.
Since this result will not be needed in the sequel we omit the proof; but
see Exercise 8.1.9.

8.2 Existence of mixing operators and semigroups

In this section we will prove the existence of mixing, and therefore hypercyclic,
operators on arbitrary infinite-dimensional separable Fréchet spaces. For the
particular case of Banach spaces we even show the existence of mixing C0 -
semigroups.
We first need two auxiliary results for Fréchet spaces. The first one is
beyond the scope of this book and is therefore stated without proof.
220 8 Existence of hypercyclic operators

Lemma 8.7. Every separable Fréchet space that is not isomorphic to ω has
a dense subspace that admits a continuous norm.
Here, a continuous norm is exactly what it says: a norm functional that is
continuous on the underlying space.
The second lemma is the crucial tool for transferring the mixing operators
of the previous section to more general Fréchet spaces.
Lemma 8.8. Let X be an infinite-dimensional separable Fréchet space that
is not isomorphic to ω. Then there are sequences (xn )n in X and (x∗n )n in
X ∗ such that
(i) (xn )n converges to 0, and span{xn ; n ∈ N} is dense in X,
(ii) (x∗n )n is equicontinuous,
(iii) xk , x∗n  = 0 if k
= n, and 0 < xn , x∗n  ≤ 1, n ≥ 1.
Proof. By Lemma 8.7 there exists a dense subspace M of X that admits a
continuous norm | · |. We select a linearly independent sequence (zn )n in M
whose linear span is dense in M , and therefore in X. By the Hahn–Banach
theorem (see Appendix A) there are functionals yn∗ ∈ (M, | · |)∗ such that,
for all n ≥ 1, zn , yn∗  = 1 and zk , yn∗  = 0 for k < n. Following a Gram–
n−1
Schmidt procedure in setting y1 = z1 and yn = zn − k=1 zn , yk∗ yk for
n > 1, we obtain a sequence (yn )n in M such that span{yn ; n ∈ N} =
span{zn ; n ∈ N} is dense in X and yk , yn∗  = δk,n , k, n ∈ N.
By continuity there are Kn ≥ 1, n ≥ 1, such that | x, yn∗ | ≤ Kn |x| for
all x ∈ M . Since M is dense in X, there is a (unique) continuous seminorm p
on X whose restriction to M coincides with | · |, and each yn∗ has a unique
continuous linear extension to X, which we denote again by yn∗ . Clearly,
| x, yn∗ | ≤ Kn p(x) for x ∈ X, n ≥ 1. This implies that {Kn−1 yn∗ ; n ∈ N}
is equicontinuous. Since X is metrizable, there are numbers αn ∈ ]0, 1] such
that (αn yn )n converges to 0 in X. Setting xn = αn yn and x∗n = Kn−1 yn∗ ,
n ≥ 1, the claim follows.  
We are now ready for the main result of this section.
Theorem 8.9 (Ansari–Bernal). Every infinite-dimensional separable Fré-
chet space supports a mixing, and therefore hypercyclic, operator.
Proof. On X = ω, the backward shift is a mixing operator; see Example
4.9(c). We may therefore suppose that X is an infinite-dimensional separable
Fréchet space that is not isomorphic to ω. Applying Lemma 8.8 we find
sequences (xn )n in X and (x∗n )n in X ∗ with the specified properties. Let us
consider T : X → X,
∞

Tx = x + 2−n x, x∗n+1 xn , x ∈ X.
n=1

The equicontinuity of (x∗n )n and the fact that (xn )n tends to 0 easily imply
that T is a well-defined operator on X.
8.2 Existence of mixing operators and semigroups 221

On the other hand, the operator

x2 , x∗2  x3 , x∗3  
S : 1 →  1 , S((αn )n ) = α1 + α2 , α2 + 2
α3 , . . .
2 2
is a perturbation of the identity by a weighted backward shift and there-
∞ mixing by Corollary 8.3. The map φ :  → X given by φ((αn )n ) =
1
fore
n=1 αn xn is continuous, and it has dense range since the linear span of the
xn , n ≥ 1, is dense in X. Since T ◦ φ = φ ◦ S, T is quasiconjugate to S, and
the conclusion follows from Proposition 1.40.  

Remark 8.10. We note that in the case that X is a Banach space, the mixing
operator T constructed above is of the form

T = I + K,

where K is a compact operator (see Exercise 5.2.4), and one may make K
arbitrarily small; see also Exercise 8.2.2.

Having answered the question of the existence of mixing operators one

might wonder if every infinite-dimensional separable Fréchet space admits a
chaotic operator. In fact, this is not even the case when we restrict ourselves
to Banach spaces, as will follow from a deep recent result in Banach space
theory.

Theorem 8.11 (Argyros–Haydon). Let K = R or C. Then there exists an

infinite-dimensional separable Banach space X over K on which all operators
are of the form
T = λI + K,
where λ ∈ K and K is a compact operator on X.

In view of Proposition 5.20 we have the following.

Theorem 8.12. Let K = R or C. Then there exists an infinite-dimensional

separable Banach space over K that supports no chaotic operator.

While we cannot always demand hypercyclic operators with a dense set of

periodic points, there are, nonetheless, always hypercyclic operators with an
infinite-dimensional closed subspace of such points.

Theorem 8.13. Every infinite-dimensional separable Fréchet space supports

a mixing, and therefore hypercyclic, operator with an infinite-dimensional
closed subspace of fixed points.

Proof. The proof is a variant of that of Theorem 8.9.

We first suppose that the Fréchet space X is not isomorphic to ω. Applying
Lemma 8.8 we again find sequences (xn )n in X and (x∗n )n in X ∗ with the
specified properties, and we define an operator T on X by
222 8 Existence of hypercyclic operators
∞

Tx = x + 2−n x, x∗2n xn , x ∈ X.
n=1

It follows, as in the proof of Theorem 8.9, that T is quasiconjugate to the

operator S on 1 given by
x2 , x∗2  x4 , x∗4  x6 , x∗6  
S((αn )n ) = α1 + α2 , α2 + α 4 , α3 + α6 , . . . .
2 22 23
By Proposition 8.5, S is mixing on 1 , and hence so is T on X. Moreover,
every vector in the infinite-dimensional closed subspace span{xn ; n ≥ 1 odd}
of X is a fixed point of T .
Finally, for the case when X = ω, we see directly from Proposition 8.5
that (αn )n → (αn + α2n )n defines a mixing operator on 1 and therefore, by
continuous and dense inclusion of 1 into ω, also a mixing operator on ω, and
it has an infinite-dimensional closed subspace of fixed points.  
In Chapter 7 we considered C0 -semigroups of operators on Banach spaces
X. In particular, every operator A : X → X defines a C0 -semigroup (Tt )t≥0
on X given by Tt = etA , and A is called the infinitesimal generator. In a way
that is analogous to the existence of mixing operators on Fréchet spaces, we
can establish the existence of mixing C0 -semigroups on Banach spaces.
Theorem 8.14. Every infinite-dimensional separable Banach space supports
a mixing, and therefore hypercyclic, C0 -semigroup (Tt )t≥0 .
Proof. As in the Fréchet case we find sequences (xn )n in X and (x∗n )n in X ∗
satisfying the hypotheses of Lemma 8.8, and we consider the operator A on
X given by
∞

Ax = 2−n x, x∗n+1 xn , x ∈ X.
n=1

Then A is quasiconjugate to a weighted shift Bw on 1 via the continuous

linear map φ : 1 → X given in the proof of Theorem 8.9. Consequently, eA is
quasiconjugate to eBw , and the latter is a mixing operator by Corollary 8.3.
It then follows from Propositions 1.40 and 7.21 that (etA )t≥0 is a mixing
C0 -semigroup on X.  
We also have an analogue of Theorem 8.12.
Theorem 8.15. Let K = R or C. Then there exists an infinite-dimensional
separable Banach space over K that supports no chaotic C0 -semigroup.
Proof. Let X be one of the spaces constructed by Argyros and Haydon (see
Theorem 8.11), and suppose that (Tt )t≥0 is a chaotic C0 -semigroup on X. By
the Conejero–Müller–Peris theorem, every operator Tt , t > 0, is hypercyclic.
Now, if K = C, then Lemma 5.19 implies that σ(Tt ) is a singleton contained
in T, for every t > 0. We deduce from the point spectral mapping theorem
8.3 Density of hypercyclic operators 223

for semigroups (see Appendix B) that etσp (A) contains at most one element,
for every t > 0, where A is the infinitesimal generator of (Tt )t≥0 . This is
only possible if σp (A) itself contains at most one element, which contradicts
Proposition 7.18. In the case when K = R we consider the complexification
of the semigroup, which, by Corollary 7.24, is again a chaotic semigroup
consisting of compact perturbations of multiples of the identity, and we argue
as before. 

8.3 Density of hypercyclic operators

For vectors we have the phenomenon that as soon as one hypercyclic vec-
tor (for a given operator) exists there is then automatically a large supply
of them. Could something similar be true for operators? That is, does the
existence of one hypercyclic operator on a given space imply that there must
then be many of them, in a certain sense?
The most natural interpretation of this question leads to a negative answer.
Indeed, no operator T on a Banach space X with T  ≤ 1 can be hypercyclic,
because all of its orbits are bounded. Thus the set of hypercyclic operators
on X is not dense in the space L(X) of all operators, when we endow it with
the operator norm topology. One can even show that in this topology the set
of hypercyclic operators is nowhere dense.
But there is another well-known, weaker topology on L(X), the strong op-
erator topology, abbreviated SOT; it is the topology of pointwise convergence
of operators. We will show in this section that the set of hypercyclic operators
is indeed dense in this topology.
In order to prepare the ground, let X be any Fréchet space and L(X) the
space of operators on X. The strong operator topology is then defined in the
following way: for T ∈ L(X), a base of neighbourhoods is given by

Ux1 ,...,xn (T, ε) = {S ∈ L(X) ; T xk − Sxk  < ε for k = 1, . . . , n},

where x1 , . . . , xn , n ≥ 1, is an arbitrary collection of linearly independent

vectors of X,  ·  is an F-norm defining the topology of X, and ε > 0. Thus,
a sequence (more generally, a net) (Tα )α is SOT-convergent to an operator
T if and only if, for every x ∈ X, Tα x → T x.
Starting from a single hypercyclic operator T on a given Fréchet space
X it is very easy to find additional ones: if A : X → X is an invertible
operator then A−1 T A is conjugate to T and therefore also hypercyclic. The
fundamental, and perhaps surprising result in this context says that, under
a very weak condition on an arbitrary operator T , the similarity orbit of T ,

S(T ) = {A−1 T A ; A : X → X invertible},

is SOT-dense in L(X).
224 8 Existence of hypercyclic operators

Proposition 8.16. Let X be an infinite-dimensional Fréchet space. Suppose

that, for any n ≥ 1, there are n vectors x1 , . . . , xn in X such that

x1 , . . . , xn , T x1 , . . . , T xn

are linearly independent. Then S(T ) is SOT-dense in L(X).

For the proof we need the following.

Lemma 8.17. Let x1 , . . . , xn and y1 , . . . , yn be two sets of linearly indepen-

dent vectors from X. Then there exists an invertible operator A of X such
that Axk = yk , k = 1, . . . , n.

Proof. Let M = span{x1 , . . . , xn , y1 , . . . , yn }, which is a space of dimension

m ≤ 2n. We can find vectors xn+1 , . . . , xm and yn+1 , . . . , ym such that both
x1 , . . . , xm and y1 , . . . , ym are bases of M . Since finite-dimensional subspaces
are closed, the Hahn–Banach theorem implies the existence of functionals
x∗k ∈ X ∗ such that xj , x∗k  = δj,k for j, k = 1, . . . , m.
We then define A : X → X by

m 
m
Ax = x, x∗k yk + x − x, x∗k xk ,
k=1 k=1

which is clearly continuous. Since Axk = yk for k = 1,. . . , m, A : M →

M is a bijection. Moreover, A is the identity on M  := k=1 ker x∗k . Since,
m

algebraically, X = M ⊕ M  , A is a bijection (on X). By the inverse mapping

theorem, A−1 is continuous so that A is invertible.  

Proof of Proposition 8.16. In order to show that S(T ) is SOT-dense we

fix an operator S ∈ L(X), linearly independent vectors x1 , . . . , xn , n ≥ 1,
of X, and ε > 0. We then define inductively vectors y1 , . . . , yn such that
x1 , . . . , xn , y1 , . . . , yn are linearly independent and Sxk − yk  < ε, k =
1, . . . , n; in fact, if x1 , . . . , xn , y1 , . . . , yk−1 , Sxk are linearly independent we
take yk = Sxk , and if not we choose a suitable small perturbation of Sxk .
On the other hand, by hypothesis, there are z1 , . . . , zn in X such that
z1 , . . . , zn , T z1 , . . . , T zn are linearly independent. By Lemma 8.17 there is
an invertible operator A on X such that Axk = zk and Ayk = T zk for
k = 1, . . . , n. Thus A−1 T Axk = yk , so that

Sxk − A−1 T Axk  < ε

for k = 1, . . . , n. In other words, A−1 T A ∈ Ux1 ,...,xn (S, ε), as desired. 



We are now in a position to prove the announced density result for hyper-
cyclic operators.

Theorem 8.18. Let X be an infinite-dimensional separable Fréchet space.

Then the set of hypercyclic operators on X is SOT-dense in L(X).
8.4 Existence of hypercyclic operators with prescribed orbits 225

Proof. We know from the Ansari–Bernal theorem that there exists a hyper-
cyclic operator T on X. By conjugacy, the similarity orbit S(T ) consists en-
tirely of hypercyclic operators. Thus the result follows from Proposition 8.16
once we know that, for any n ≥ 1, there are n vectors x1 , . . . , xn in X such
that x1 , . . . , xn , T x1 , . . . , T xn are linearly independent. But these are easily
found: the orbit of any hypercyclic vector x for T is linearly independent and
hence so is

x, T 2 x, T 4 x, . . . , T 2n−2 x, T x, T 3 x, T 5 x, . . . , T 2n−1 x;

see Proposition 2.60. 



The proof tells us much more: the same result will be true for any
nonempty class of hypercyclic operators that is invariant under conjugacies.
In particular, we obtain the following.

Theorem 8.19. Let X be an infinite-dimensional separable Fréchet space.

(a) The set of mixing operators on X is SOT-dense in L(X).
(b) The set of chaotic operators on X is either empty or SOT-dense in
L(X); in particular, it is SOT-dense if X is a Hilbert space.
(c) The set of mixing operators on X with an infinite-dimensional closed
subspace of fixed points is SOT-dense in L(X).

8.4 Existence of hypercyclic operators with prescribed

orbits

There is another sense in which one can measure the richness of the set of
hypercyclic operators. By Proposition 2.60 we know that a dense orbit is nec-
essarily a linearly independent set. Therefore one may wonder if, conversely,
every countable dense linearly independent set arises as an orbit of an oper-
ator; this operator is then automatically hypercyclic. We will show here that
the answer is positive, at least in the setting of Banach spaces. The idea of
the proof is not dissimilar to the approach in the last section: we construct
an invertible operator that transforms a known dense orbit into an arbitrary
countable dense linearly independent set.
We first need a property of perturbations of invertible operators.

Lemma 8.20. Let X be a Banach space, A an invertible operator on X. If B

is an operator on X such that A − B < 1/A−1  then B is also invertible.
Moreover, in that case,

A−1 2 A − B
A−1 − B −1  ≤ .
1 − A−1  A − B
226 8 Existence of hypercyclic operators

Proof.
∞ First, if T is an operator on X with T  < 1 then the series C :=
n
n=0 T converges in L(X) under the operator norm, so that C defines an
operator on X. Moreover, since

(I − T )C = (I + T + T 2 + . . .) − (T + T 2 + . . .) = I,

and similarly C(I − T ) = I, the operator I − T is invertible with inverse C.

Now, if A − B < 1/A−1  then I − A−1 B = A−1 (A − B) ≤
A−1  A − B < 1, so that A−1 B = I − (I − A−1 B) is invertible and
therefore also B. In addition,
∞

A−1 − B −1 = (I − B −1 A)A−1 = − (A−1 (A − B))n A−1 .
n=1

Applying the operator norm and the sum formula of the geometric series
gives the desired estimate on A−1 − B −1 . 

The next step is to prescribe values of an invertible operator.
Lemma 8.21. Let X be an infinite-dimensional Banach space and X0 and Y0
dense subsets of X. Let E and F be subspaces of X with dim(E) = dim(F ) <
∞ and A an invertible operator on X such that A(E) = F .
Then, for any x0 ∈/ E, y0 ∈
/ F and ε > 0, there exists an invertible operator
B on X such that B|E = A|E , Bx0 ∈ Y0 , B −1 y0 ∈ X0 and A − B < ε.
Proof. Since E is closed as a finite-dimensional subspace, the Hahn–Banach
theorem implies that there exists x∗ ∈ X ∗ such that x, x∗  = 0 for all x ∈ E
and x0 , x∗  = 1. By density of Y0 there is some y1 ∈ Y0 with x∗ y1 −
Ax0  < min(ε/2, 1/A−1 ); perturbing y1 if necessary we can assume that it
does not belong to F ⊕ span{y0 }. Then we define the operator C by

Cx = Ax + x, x∗ (y1 − Ax0 ), x ∈ X.

Consequently, C|E = A|E , Cx0 = y1 and A − C ≤ x∗ y1 − Ax0  <

min(ε/2, 1/A−1 ). By Lemma 8.20, C is invertible.
Now we repeat the procedure, with F ⊕ span{y1 }, E, X0 , y0 , x0 , and
C −1 taking the roles of E, F , Y0 , x0 , y0 , and A, respectively; note that
y0 ∈
/ F ⊕ span{y1 }. Thus, for η > 0, there is some x1 ∈ X0 and an invertible
operator D on X such that D|F = C −1 |F , Dy1 = C −1 y1 = x0 , Dy0 = x1
and C −1 − D < η; moreover, by Lemma 8.20, if η is small enough then
C − D −1  < ε/2. Then B := D−1 is the desired invertible operator with
A − B ≤ A − C + C − D−1  < ε.  
This lemma implies that any two dense linearly independent sequences are
isomorphically equivalent, in the following sense.
Lemma 8.22. Let X be a Banach space, let X0 = {xn ; n ∈ N} and Y0 =
{yn ; n ∈ N} be dense linearly independent sets in X, and ε > 0. Then there
exists an invertible operator A on X such that A(X0 ) = Y0 and I − A < ε.
8.4 Existence of hypercyclic operators with prescribed orbits 227

Proof. We will show by induction that there are finite sets Vn ⊂ X0 , Wn ⊂

Y0 and invertible operators An on X such that, for any n ≥ 1, Vn−1 ⊂
Vn , {x1 , . . . , xn } ⊂ Vn , Wn−1 ⊂ Wn , {y1 , . . . , yn } ⊂ Wn , An (Vn ) = Wn ,
An |Vn−1 = An−1 |Vn−1 and An − An−1  < 2εn , where A0 = I and V0 = W0 =
∅.
For n = 1 we apply Lemma 8.21 to E = F = {0} and A = I to obtain an
invertible operator A1 with A1 − I < 2ε such that there are p1 , q1 ≥ 1 with
A1 x1 = yp1 and A1 xq1 = y1 . We then set V1 = {x1 , xq1 } and W1 = {y1 , yp1 };
note that both sets may be singletons.
Suppose that V1 , . . . , Vn−1 , W1 , . . . , Wn−1 and A1 , . . . , An−1 have been de-
fined, n ≥ 2. Since card(Vn−1 ) = card(Wn−1 ), we can apply Lemma 8.21 to
E = span Vn−1 , F = span Wn−1 and A = An−1 ; for x0 we choose the vector
xkn ∈ V of smallest index that does not belong to E, and similarly a vector
ymn for y0 . Then we obtain an invertible operator An and pn , qn ≥ 1 such
that An |E = An−1 |E , An xkn = ypn , An xqn = ymn , and An − An−1  < 2εn .
We set Vn = Vn−1 ∪ {xkn , xqn } and Wn = Wn−1 ∪ {ymn , ypn }. Using the
linear independence of X0 and Y0 , one deduces easily from the construction
and the induction hypothesis that {x1 , . . . , xn } ⊂ Vn , {y1 , . . . , yn } ⊂ Wn and
An (Vn ) = Wn . ∞
Having obtained the operators An we note that n=1 An − An−1  < ∞.
Thus
∞
A := I + (An − An−1 )
n=1
∞
defines an operator on X. Moreover, I − A ≤ n=1 An − An−1  < ε;
choosing ε < 1, Lemma 8.20 implies that A is invertible. Moreover, since
A = limn→∞ An we deduce easily that A(X0 ) = Y0 . 
We will need the following variant of this lemma.
Lemma 8.23. Let X be a Banach space, let X0 = {xn ; n ∈ N} and Y0 =
{yn ; n ∈ N} be dense linearly independent sets in X, B an invertible operator
with Bx1 = y1 , and ε > 0. Then there exists an invertible operator A on X
such that A(X0 ) = Y0 , Ax1 = y1 and B − A < ε.
Proof. The proof is identical to that of the previous lemma if one starts with
A1 = B, V1 = {x1 }, W1 = {y1 }, and then proceeds with the induction.  
We can now show that dense orbits can be prescribed.
Theorem 8.24. Let X be a Banach space, and let {xn ; n ∈ N} be a dense
linearly independent set in X. Then there exists an operator T on X, neces-
sarily hypercyclic, such that orb(x1 , T ) = {xn ; n ∈ N}.
Proof. Since X is an infinite-dimensional separable Banach space, the Ansari–
Bernal theorem implies that there is a hypercyclic operator S on X. Applying
Lemma 8.21 to E = F = {0}, X0 = {xn ; n ∈ N}, Y0 = HC(S) and A = I
we obtain an invertible operator B with y1 := Bx1 ∈ HC(S).
228 8 Existence of hypercyclic operators

By Proposition 2.60, orb(y1 , S) =: {yn ; n ∈ N} is linearly independent.

Now applying Lemma 8.23 to X0 = {xn ; n ∈ N} and Y0 = {yn ; n ∈ N} we
obtain an invertible operator A on X with Ax1 = y1 and A(X0 ) = Y0 . By
conjugacy, the operator T = A−1 SA is hypercyclic, and we have that

orb(x1 , T ) = {T n x1 ; n ∈ N0 } = A−1 {S n y1 ; n ∈ N0 } = A−1 (Y0 ) = X0 ,

which had to be shown. 



A more careful analysis of the proof shows that, for every ε > 0, the
operator T can be of the form T = I + K, where K is a compact operator
with K < ε; see Exercise 8.4.3.

Exercises

Exercise 8.1.1. Given m, n ∈ N with m ≥ n, verify the identity

⎛ ⎞
1
m!
1
(m+1)! ··· 1
(m+n)!
⎜ 1 1
··· 1 ⎟ n
⎜ (m−1)! m! (m+n−1)! ⎟ (n − k)!
det ⎜ .. .. .. .. ⎟= .
⎝ . . . . ⎠ k=0 (m + k)!
1 1
(m−n)! (m−n+1)!
··· 1
m!
∞
Exercise 8.1.2. An entire function γ(z) = γ z n is called an admissible compar-
n=0 n
ison function provided that γn > 0 for n ≥ 0 and the sequence ((n + 1)γn+1 /γn )n≥0 is
bounded. Each admissible comparison function induces Banach spaces of entire functions
 
∞

∞ 
E p (γ) = f ∈ H(C) ; f (z) = an z n , f pp,γ = γn−p |an |p < ∞ , 1 ≤ p < ∞;
n=0 n=0

see also Exercise 4.2.4. Show the following:

(i) any function f ∈ E p (γ) is of exponential type τ := lim supn→∞ (n + 1)γn+1 /γn ,
that is, for every ε > 0 there is some M > 0 such that |f (z)| ≤ M e(τ +ε)|z| for
z ∈ C;
(ii) D : f → f  defines an operator on E p (γ);
(iii) deduce from Theorem 8.1 that the translation operators Ta f (z) = f (z + a), a ∈ C,
a = 0, are mixing on E p (γ);
(iv) conversely, show that the result of (iii) implies Theorem 8.1.
(Hint: Show that aD on E p (γ) is conjugate to B on p (v) for a suitable sequence of
weights v = (vn )n , and observe that Ta = eaD .)

Exercise 8.1.3. Use the previous exercise to prove Theorem 4.23. (Hint: Note that
rn ≤ Mn φ(r); then find γn such that any f ∈ E 1 (γ) satisfies the required growth
condition, and apply the hypercyclic comparison principle.)
v
Exercise 8.1.4. Let v = (vn )n be a positive weight sequence with supn∈N vn+1 n
< ∞.
Show directly that I + B is mixing on p (v), 1 ≤ p < ∞, or on c0 (v). (Hint: Following
the proof of Theorem 8.1, use the same argument substituting the matrix W by a matrix
Wn such that limn Wn = W .)
Exercises 229

Exercise 8.1.5. Deduce Theorem 8.1 from Theorem 8.2.

Exercise 8.1.6. An operator T on an infinite-dimensional Banach space !∞X is called a
generalized backward shift if dim ker T = 1 and the generalized kernel n=0 ker T n is
dense in X. Show that there is a sequence (en )n of nonzero vectors in X such that
T e1 = 0, T en+1 = en , n ≥ 1, and span{en ; n ∈ N} is dense in X. Deduce that T is an
extended backward shift. (Hint: For Xn = ker T n , show that Xn−1  Xn , and use the
rank-nullity theorem to show inductively that T (Xn ) = Xn−1 and dim Xn = n.)
Exercise 8.1.7. Let A be an operator on a Banach space X for which there is a sequence
(en )n of nonzero vectors such that Ae1 = 0, Aen+1 = en , n ≥ 1, and span{en ; n ∈ N} is
dense in X. Prove that I +A is mixing on X. (Hint: Construct a suitable quasiconjugacy
to use Theorem 8.2.)
Exercise 8.1.8. Let A be an operatoron a Banach space X that satisfies the hypothesis
∞
of the previous exercise. Let ϕ(z) = a z n be a nonconstant formal power series

n=0 n
∞
with |a0 | = 1 such that T := ϕ(A) = a An converges pointwise on X; by the
n=0 n
Banach–Steinhaus theorem, T is then an operator on
∞
X. Show that T is mixing. (Hint:
One may assume that a0 = 1. Then T − I = an An with k ≥ 1, ak = 0. For
n=k 
n
j = 1, . . . , k, take fj := ej , and for n ≥ 1 let fk+n := w e
j=1 n,j k+j
, where the wn,j
have to be determined by induction to satisfy (T − I)fk+n = fn . Now, for j = 1, . . . , k
define Xj := span{fj+nk ; n ≥ 0}. Then each Xj is (T −I)-invariant, and one may apply
the previous exercise to each operator Aj := (T − I)|Xj on Xj . To finish, construct a
commutative diagram
k
(I+Aj )
X1 ⊕ · · · ⊕ Xk −−−−−−−−−→ X1 ⊕ · · · ⊕ Xk
j=1

⏐ ⏐
⏐ ⏐
φ  φ 
T =ϕ(A)
X −−−−−−→ X
and apply Proposition 1.40.)
Exercise 8.1.9. Prove Theorem 8.6. (Hint: By Proposition 2.37 it suffices to show that
for any x ∈ X, ε > 0, and for any sufficiently large n there are z, z  ∈ X (depending on
n) such that x − z !< ε and T n z < ε, z   < ε and x − T n z   < ε. It suffices to
∞
consider x = 0 from n=0 (ker T n ∩ ran T n ). Choose m ≥ 1 minimal and y ∈ X such
that T m x = 0 and x = T m y; then construct ek , k = 1, . . . , 2m, such that T e1 = 0,
T ek+1 = ek for k ≥ 1, and em = x. By minimality of m, the ek are linearly independent.
Now, for T = eA , use the argument in the proof of Theorem 8.1 to show that there
is some C > 0 such that, for any n ≥ 1, there is z ∈ X (depending on n) of the
2m 2m 2m nj−k
form z = em + z e that satisfies T n z =
k=m+1 k k k=m+1
( j=k (j−k)! zj )ek and
|zk | ≤ Cnm−k , k = m + 1, . . . , 2m. Deduce that, for large n, x − z < ε and T n z < ε.
Construct z  in a similar way. For the proof of T = I +A replace W by Wn as in Exercise
8.1.4.)
Exercise 8.2.1. Show that Lemma 8.8 fails for X = ω.
Exercise 8.2.2. Let T be an operator on a Banach space X. For n ≥ 1 the nth approx-
imation number αn (T ) is defined as

αn (T ) = inf{T − F  ; F ∈ L(X), dim(ran F ) ≤ n}.

By Exercise 5.2.4, if αn (T ) → 0 as n → ∞ then T is compact. Let ε > 0, and let εn > 0,

n ≥ 1, be such that εn → 0 as n → ∞. Show that, on any infinite-dimensional separable
230 8 Existence of hypercyclic operators

Banach space X, there exists a mixing operator of the form T = I + K, where K is

a compact operator with K < ε and αn (K) ≤ εn for any n ≥ 1. (Hint: Modify the
proof of Theorem 8.9.)

Exercise 8.2.3. Show the following:

(i) every infinite-dimensional separable Banach space supports an invertible mixing
operator;
(ii) the operator constructed in the proof of Theorem 8.9 is not chaotic.
(Hint: See the previous exercise, Lemma 8.20 and Proposition 5.20.)

Exercise 8.2.4. Let X be an infinite-dimensional separable Fréchet space. Show that

there are operators T1 , T2 on X that are disjoint hypercyclic; see Exercise 3.4.1. (Hint:
First note that when x, y = 0 then there is an invertible operator A such that Ax = y;
indeed, if x, y are linearly independent then find x∗ , y ∗ ∈ X ∗ with x, x∗  = 1, y, y ∗  =
1, y, x∗  = 0, x, y ∗  = 0, and set Az = z + z, y ∗ − x∗ (x − y). Now, let T be a mixing
operator on X and (x1 , x2 ) hypercyclic for T ⊕ T . Let x = 0, and A1 , A2 invertible
operators such that Aj xj = x, j = 1, 2. Then consider Tj = Aj T A−1 j , j = 1, 2.)

Exercise 8.2.5. Let X be a separable Banach space and M a complemented subspace

of X of infinite codimension; that is, M is a closed subspace for which there is an infinite-
dimensional closed subspace M  of X such that, algebraically, X = M ⊕ M  . Show that
there exists a hypercyclic operator T on X that is the identity on M . (Hint: Apply
Lemma 8.8 separately to both subspaces and follow the proof of Theorem 8.13.)

Exercise 8.2.6. Give an alternative proof of Theorem 8.14 as follows. Let T be the
mixing operator on an infinite-dimensional separable Banach space X provided by the
proof of Theorem 8.9 of the form T = I + S, where we can even have that S < 1. Then
find an operator A such that eA = T to deduce that (etA )t≥0 is a mixing C0 -semigroup
on X. (Hint: Use Exercise 7.1.5.)

Exercise 8.3.1. Even though every hypercyclic operator on 2 has norm bigger than
one, Theorem 8.18 tells us that one can approximate the zero operator in the SOT-
topology by such operators. Show that, concretely, if B is the backward shift, then
n B , n ≥ 1, are hypercyclic operators that converge to 0 in SOT as n → ∞.
n+1 n

Exercise 8.3.2. Show that the class of all operators on an infinite-dimensional separable
Fréchet space that have no nontrivial closed invariant subset is either empty or SOT-
dense.

Exercise 8.3.3. Show that on any infinite-dimensional separable Fréchet space that is
not isomorphic to ω the set of hypercyclic non-chaotic operators is SOT-dense. (Hint:
Exercise 8.2.3.)

Exercise 8.3.4. Let X be an infinite-dimensional separable Fréchet space. Show the

following:
(i) the operators in the similarity orbit S(T ) of an operator T on X either have no
common hypercyclic vector, or else every nonzero vector in X is hypercyclic for
every operator in S(T );
(ii) there is no vector that is hypercyclic for every hypercyclic operator on X.

Exercise 8.3.5. Let X be an infinite-dimensional separable Fréchet space and x ∈ X,

x = 0. Show that the set of hypercyclic operators on X having x as hypercyclic vector
is SOT-dense in L(X). (Hint: Consider {A−1 T A ; A invertible and Ax ∈ HC(T )}.)
Sources and comments 231

Exercise 8.3.6. Let X be a complex Banach space with the properties of Theorem
8.11. Show that if an operator T on X is the sum of two hypercyclic operators then it
is of the form T = λI + K, where |λ| ≤ 2 and K is a compact operator. In particular,
not every operator on X can be written as the sum of two hypercyclic operators. (Hint:
Lemma 5.19.)

Exercise 8.4.1. Let X be a Banach space and M a dense subspace of X of countably

infinite dimension. Show that for any nonzero vector x ∈ M there exists an operator
T on X such that M = span orb(x, T ); moreover, x may be hypercyclic for T . (Hint:
Starting with x, construct a dense linearly independent sequence in M , complete to an
algebraic basis of M , and apply Theorem 8.24.)

Exercise 8.4.2. Show that every normed space of countably infinite dimension sup-
ports an operator without nontrivial closed invariant subsets. (Hint: Apply the previous
exercise to the completion of X.)

Exercise 8.4.3. Show that, for any ε > 0, the operator T in Theorem 8.24 can be of
the form T = I + K, where K is a compact operator with K < ε. (Hint: Remark 8.10
and Exercise 5.2.3.)

Exercise 8.4.4. Show directly that the operator C defined in the proof of Lemma 8.21
is invertible if and only if it is injective, and if and only if A−1 y1 − x0 , x∗  = −1. In
that case, calculate the inverse.

Exercise 8.4.5. Show that Theorem 8.24 is false in ω. Indeed, there exists a dense
linearly independent sequence in ω that cannot be the orbit of any operator on ω. To
see this, let X1 = {xn ; n ∈ N} be a dense linearly independent set in ω such that, for
n ≥ 1, xn = (xn,k )k is a finite sequence with xn,1 = 0. We consider X2 = {yn ; n ∈ N}
defined by yn,k = 0 if k ≤ n, and yn,k = nk if k > n. Then X0 = X1 ∪ X2 is dense
and linearly independent in ω, but there is no operator T on ω such that X0 coincides
with an orbit under T . (Hint: Otherwise there would be infinitely many elements in X2
whose image under T belonged to X1 ; derive that then the composition of T with the
projection onto the first coordinate would not be continuous.)

Exercise 8.4.6. Prove the following extension of Theorem 8.24 to C0 -semigroups. For
any dense linearly independent set {xn ; n ∈ N} in a Banach space X, there is a C0 -
semigroup (Tt )t≥0 on X such that orb(x1 , T1 ) = {xn ; n ∈ N}. (Hint: Combine the
proof of Theorem 8.24 with Exercises 8.2.6 and 8.4.3.)

Sources and comments

Section 8.1. The main result of this section comes in three equivalent forms. Salas [274]
showed that any perturbation of the identity by a weighted backward shift is hypercyclic
on p or c0 ; see Corollary 8.3. Following his arguments, Desch, Schappacher and Webb
[131] showed that the exponential of the backward shift is hypercyclic on the correspond-
ing weighted spaces; see Theorem 8.1. Some years earlier, Chan and Shapiro [106] had
obtained the hypercyclicity of the translation operator Ta on Hilbert spaces of entire
functions of slow growth; see Exercise 8.1.2. The fact that the equivalence of these results
was only noticed much later is, perhaps, due to the fact that the authors worked in very
different contexts. While Salas regarded iterates of a function of the backward shift,
Desch, Schappacher and Webb were interested in C0 -semigroups generated by shifts,
and Chan and Shapiro studied translation operators on spaces of entire functions. In
232 8 Existence of hypercyclic operators

addition, the proof by Chan and Shapiro was very different from the one by Salas by
using, in a crucial way, tools from complex analysis. Incidentally, the quasiconjugacies
between I + B and eB were first observed in Martínez and Peris [232].
León and Montes [220] showed that Salas’ operators satisfy the Hypercyclicity Cri-
terion, while Grivaux [172] showed that they are even mixing.
Proposition 8.5 is due to Grivaux [171]. Theorem 8.6 was obtained by Grivaux and
Shkarin [176, 286]; see also Bayart and Matheron [44]. Hypercyclicity of functions ϕ(T )
of rather general operators T have been studied by Herzog and Schmoeger [200], Miller
and Miller [239], Bermúdez and Miller [53], Martínez and Peris [232], and Müller [248].

Section 8.2. The existence of hypercyclic operators on any infinite-dimensional sepa-

rable Banach space was independently obtained by Ansari [10] and Bernal [54], solving
a problem from Rolewicz [268]. Later, it was generalized to Fréchet spaces by Bonet
and Peris [85]. The construction heavily depends on the hypercyclicity of perturbations
of the identity by weighted shifts proved by Salas [274]. Grivaux [172] noticed that the
operators of Ansari and Bernal are even mixing. It is also fair to say that, on Banach
spaces, these operators are perturbations of the identity by the supercyclic operators con-
structed earlier by Herzog [197]. Lemma 8.7 was obtained by Metafune and Moscatelli
[238].
Theorem 8.11 is due to Argyros and Haydon [12]. Bonet, Martínez and Peris [82]
obtained Theorem 8.12; the deep result by Argyros and Haydon allowed us to simplify
the proof. Theorem 8.13 is due to Grivaux [171]. De la Rosa, Frerick, Grivaux and
Peris [127] showed that on every infinite-dimensional complex Fréchet space with an
unconditional basis there is a chaotic operator, and that there is an infinite-dimensional
real Banach space with an unconditional basis without chaotic operators.
Theorems 8.14 and 8.15 are due to Bermúdez and Bonilla with Conejero and Peris
[49] and with Martinón [50].
Further existence results are due to Grivaux [172] (every infinite-dimensional separa-
ble Banach space supports a mixing operator that fails Kitai’s criterion), Shkarin [285]
(every infinite-dimensional separable Banach space supports an operator that satisfies
Kitai’s criterion), Grivaux and Shkarin [176, 286] (every infinite-dimensional separable
Fréchet space not isomorphic to ω supports a hypercyclic non-mixing operator; on ω,
every hypercyclic operator is mixing) and to Salas [276], see also Grivaux and Shkarin
[176, 286] (every infinite-dimensional separable Banach space with separable dual sup-
ports a hypercyclic operator whose adjoint is also hypercyclic).
Hájek and Vivi [190] have applied the Ansari–Bernal theorem to limit sets of solu-
tions of ordinary differential equations in Banach spaces.

Section 8.3. Chan [100] showed that the set of hypercyclic operators on an infinite-
dimensional separable Hilbert space is SOT-dense in the space of all operators, which was
extended by Bès and Chan [68] to Fréchet spaces. The simplified proofs using Proposition
8.16 by Hadwin, Nordgren, Radjavi, and Rosenthal [189] are due to Bès and Chan [67]
and Prǎjiturǎ [260]. In the same vein many similar results are possible; see Bès and Chan
[67].
The fact that hypercyclic operators are nowhere dense under the operator norm
topology was observed by Wu [302], who even showed that the set of cyclic operators
has this property. Herrero [194] has obtained a spectral description of the operator norm
closure of the set of hypercyclic operators on an infinite-dimensional separable complex
Hilbert space; see also Müller [248]. Herrero then deduced in [195] that the chaotic
operators are norm dense in the set of hypercyclic operators.
As the set of hypercyclic operators is not dense in the operator norm topology one
might wonder if its linear hull is dense. First, Bès and Chan [67] proved even more: on an
infinite-dimensional separable complex Hilbert space the set of sums of two hypercyclic
Sources and comments 233

(or even chaotic) operators is dense in the operator norm. This motivated the following
deep result.

Theorem 8.25 (Grivaux [170]). Let X be an infinite-dimensional separable complex

Hilbert space. Then every operator is the sum of two hypercyclic operators.

In fact, both operators can even be chosen to be chaotic. But Grivaux also showed
that, even for hypercyclicity, the result cannot be extended to all Banach spaces; see
Exercise 8.3.6.

Section 8.4. The problem of whether every linearly independent sequence in a separable
Banach space is contained in the orbit of a hypercyclic vector was posed by Halperin,
Kitai and Rosenthal [192], who had obtained the result for Hilbert spaces. Theorem
8.24, which solves the problem, is due to Grivaux [169]. Albanese [5] has extended the
theorem to Fréchet spaces with a continuous norm.

Exercises. Exercise 8.1.2 generalizes Exercise 4.2.4 for Birkhoff’s operators. It was
proved (with the exception of (iv)) by Chan and Shapiro [106] with completely different
techniques based on complex analysis. Exercise 8.1.3 was observed by Chan and Shapiro
[106]. Exercise 8.1.4 essentially asks for the original proof of Corollary 8.3 by Salas [274].
The notion of a generalized backward shift was introduced by Godefroy and Shapiro
[165], where one also finds Exercise 8.1.6. Exercises 8.1.7 and 8.1.8 are from Martínez
and Peris [229], and [230, 232], respectively. Exercise 8.2.1 is taken from Bonet and Peris
[85], Exercise 8.2.2 generalizes a result by Chan and Shapiro [106]. The result of Exercise
8.2.4 is due to Bès, Martin, Peris, Salas and Shkarin [70, 277, 291]. Exercise 8.2.6 is taken
from Bernal and Grosse-Erdmann [63], Exercise 8.3.1 from Chan [100], and Exercises
8.3.4 and 8.3.5 from Chan and Sanders [103]. Concerning Exercise 8.3.3, De la Rosa,
Frerick, Grivaux and Peris [127] have shown that also on ω there are hypercyclic non-
chaotic operators, so that the result of the exercise extends to all infinite-dimensional
separable Fréchet spaces. Exercise 8.3.6 is due to Grivaux [170]; the use of the Argyros–
Haydon spaces simplifies the proof. Exercises 8.4.1, 8.4.2, 8.4.3 are taken from Grivaux
[169], and Exercises 8.4.4, 8.4.5 and 8.4.6 from Albanese [5], Bonet, Frerick, Peris and
Wengenroth [81] and Bernal and Grosse-Erdmann [63], respectively.
Chapter 9
Frequently hypercyclic operators

The theory of linear dynamical systems has its roots in topological dynamics.
But there is also a parallel theory of measurable dynamics, which is better
known under the name of ergodic theory. In this chapter we show how con-
cepts and results from that theory lead to a deepened understanding of linear
dynamics. More specifically, we will see how the celebrated Birkhoff ergodic
theorem suggests an interesting and rather strong variant of hypercyclicity,
that of frequently hypercyclic operators. We point out that, while ergodic
theory has turned out to be a most powerful tool in linear dynamics, we will
use it here only for motivating the new concept.
Having introduced frequently hypercyclic operators, we then derive a Fre-
quent Hypercyclicity Criterion and an eigenvalue criterion that allow us to
show that, quite surprisingly, many of the hypercyclic operators met so far
are in fact frequently hypercyclic. In the final section we revisit several of the
structural properties of hypercyclicity within the new framework.

9.1 Frequently recurrent orbits

Let T be an operator on a separable Fréchet space X. In order to look at

T from the point of view of ergodic theory we need to have a probability
measure μ on X. Since we are in a topological situation it is natural to
assume that μ is defined on the Borel σ-algebra B(X), that is, the smallest
σ-algebra containing the open subsets of X; the elements of B(X) are called
the Borel sets of X. Since T is continuous, it is then also measurable. We
assume that T satisfies the minimum requirement in ergodic theory, namely,
that it is μ-invariant, that is, μ(T −1 (A)) = μ(A) for every Borel set A.
Of course, μ-invariance alone does not yet give us interesting dynamics
since, for example, the identity operator is automatically μ-invariant for any
measure μ. This changes when we inject ergodicity: one way of defining this
notion is by demanding that, for any Borel sets A and B with μ(A) > 0

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 235

DOI 10.1007/978-1-4471-2170-1_9, © Springer-Verlag London Limited 2011
236 9 Frequently hypercyclic operators

and μ(B) > 0, there is some n ∈ N0 such that μ(T −n (A) ∩ B) > 0. This
notion is not only formally similar to topological transitivity. Suppose that
the measure μ has the additional property that μ(U ) > 0 for any nonempty
open set U ; μ is then said to be of full (topological) support. Under this
assumption, ergodicity obviously implies topological transitivity.
But there is an added bonus in the form of the Birkhoff ergodic theorem.
It tells us that if T is ergodic with respect to μ then, for any μ-integrable
function f on X, its time average with respect to T coincides with its space
average; more precisely we have that

1 
N
f (T n x) → f dμ, for μ-almost all x ∈ X, (9.1)
N + 1 n=0 X

as N → ∞. This then implies an interesting topological property for T .

Indeed, since X is separable, its topology has a countable base (Uk )k . When
we apply (9.1) to the indicator functions 1Uk , k ≥ 1, the left-hand side turns
out to be

1 
N
card{0 ≤ n ≤ N ; T n x ∈ Uk }
1Uk (T n x) = ,
N + 1 n=0 N +1

while the right-hand side is simply X 1Uk dμ = μ(Uk ) > 0, where we have
assumed again that μ is of full support. Thus there are subsets Ak ⊂ X,
k ≥ 1, of full measure such that, for any x ∈ Ak ,

card{0 ≤ n ≤ N ; T n x ∈ Uk }
lim > 0.
N →∞ N +1

Since every nonempty open set contains some Uk and since k≥1 Ak has full
measure we obtain that, for μ-almost all x ∈ X and every nonempty open
subset U of X,

card{0 ≤ n ≤ N ; T n x ∈ U }
lim inf > 0.
N →∞ N +1
What we have found here is that, under the mentioned assumptions, the
operator T has a property that is much stronger than hypercyclicity. There
must even be an x ∈ X whose orbit meets every nonempty open set very
often, in the sense given above. Let us recall here the following.
Definition 9.1. The lower density of a subset A ⊂ N0 is defined as

card{0 ≤ n ≤ N ; n ∈ A}
dens(A) = lim inf .
N →∞ N +1

Our discussion so far leads us to the following concept.

9.1 Frequently recurrent orbits 237

Definition 9.2. An operator T on a Fréchet space X is called frequently

hypercyclic if there is some x ∈ X such that, for any nonempty open subset
U of X,
dens {n ∈ N0 ; T n x ∈ U } > 0.
In this case, x is called a frequently hypercyclic vector for T . The set of
frequently hypercyclic vectors for T is denoted by F HC(T ).
The orbit of a frequently hypercyclic vector is therefore, in the specified
sense, frequently recurrent. Obviously, frequent hypercyclicity is a stronger
notion than hypercyclicity.
There is an equivalent formulation of frequent hypercyclicity that nicely
differentiates it from hypercyclicity. Let A be a subset of N0 ; if (nk )k≥1 is the
increasing sequence of integers forming A and nk ≤ N < nk+1 then

k card{0 ≤ n ≤ N ; n ∈ A} k
≤ ≤ ,
nk+1 N +1 nk

which implies that dens (A) = lim inf k→∞ nkk . Thus A has positive lower
density if and only if ( nkk )k is bounded; in other words, if nk = O(k).
Proposition 9.3. A vector x ∈ X is frequently hypercyclic for T if and only
if, for any nonempty open subset U of X, there is a strictly increasing se-
quence (nk )k of positive integers such that

T nk x ∈ U for all k ∈ N, and nk = O(k).

By contrast, T is hypercyclic if and only if the same is true for some

(nk )k , not necessarily of order O(k). This seems to indicate that our new
notion requires much more than mere hypercyclicity.
We have the usual behaviour under quasiconjugacies, which can be proved
as in Proposition 1.19.
Proposition 9.4. Frequent hypercyclicity is preserved under quasiconjugacy.
Our first task will be to show that frequently hypercyclic operators exist.
We saw above that an operator T on a separable Fréchet space X is frequently
hypercyclic if one can find a Borel probability measure μ of full support
on X with respect to which T is ergodic. However, in order to keep our
introduction to frequent hypercyclicity simple we will not pursue this circle of
ideas any further. Instead, we will favour a constructive approach to frequent
hypercyclicity.
So, what does it take for a vector x to be frequently hypercyclic for an
operator T ? Let · denote an F-norm defining the topology of X, and let
(yl )l be a dense sequence in X. Then there are subsets A(l, ν), l, ν ≥ 1, of N0
of positive lower density such that, for any n ∈ A(l, ν),
1
T n x − yl < .
ν
238 9 Frequently hypercyclic operators

Moreover, if yl
= yk then the sets A(l, ν) and A(k, μ) are disjoint if ν and μ
are big. In fact, in the sequel we will need the existence of sets A(l, ν) with
a stronger separation property.

Lemma 9.5. There exist pairwise disjoint subsets A(l, ν), l, ν ≥ 1, of N0 of

positive lower density such that, for any n ∈ A(l, ν) and m ∈ A(k, μ), we
have that n ≥ ν and

|n − m| ≥ ν + μ if n
= m.

Proof. We start by partitioning N in a very natural fashion by using the

dyadic representation
∞

n= aj 2j =: (a0 , a1 , a2 , . . .)
j=0

of any positive integer n. We define I(l, ν), l, ν ≥ 1, as the set of all n ∈ N

whose dyadic representation has the form

n = (0, . . . , 0, 1, . . . , 1, 0, ∗)

with l − 1 leading zeros, followed by ν ones, then one zero, followed by an

arbitrary tail. It is clear that the sets I(l, ν) form a partition of N, but they
do not satisfy the required separation property. To achieve this we let δk = ν
if k ∈ I(l, ν) for some l ≥ 1, and we define


k−1
nk = 2 δ i + δk , k ≥ 1,
i=1

which is a strictly increasing sequence. We claim that

A(l, ν) = {nk ; k ∈ I(l, ν)}, l, ν ≥ 1

has the desired properties. First, these sets are pairwise disjoint. Moreover,
if nk ∈ A(l, ν) then nk ≥ δk = ν; and if nj ∈ A(l, ν), nm ∈ A(k, μ) with
nj
= nm , where we can assume that j > m, then


j−1
nj − nm = δm + 2 δi + δj ≥ μ + ν.
i=m+1

It remains to show that each set A(l, ν) has positive lower density. We
begin by proving that there is some M > 0 such that

nk ≤ M k, k ≥ 1. (9.2)
9.1 Frequently recurrent orbits 239

It suffices to do this for k = 2N , N ≥ 1, because we then have for 2N −1 ≤

k < 2N that
nk ≤ n2N ≤ M 2N ≤ 2M k.
Thus let k = 2N . A simple but tedious enumeration shows that, if l + ν ≤
N + 2, then I(l, ν) contains at most 2N +2−l−ν elements that do not exceed
2N , and none if l + ν > N + 2. Hence we have that

   ν 
N
2 
n2 N ≤ 2 δi ≤ 2 2N +2−l−ν ν ≤ 8 2N ,
2l+ν
i=1 l+ν≤N +2 l,ν≥1

so that (9.2) holds for some M > 0.

Now let l, ν ≥ 1. Let (kj )j be the increasing sequence of elements of I(l, ν).
Since the latter set has positive lower density, the argument leading up to
Proposition 9.3 shows that there is some constant K > 0 such that

kj ≤ Kj, j ≥ 1.

It then follows that A(l, ν) = {nkj ; j ≥ 1} and

nkj ≤ M kj ≤ M Kj, j ≥ 1.

Hence each set A(l, ν) has positive lower density. 

This result allows us to obtain a first example of a frequently hypercyclic

operator.

Example 9.6. (Birkhoff’s operators) The translation operators Ta : f →

f (· + a), a
= 0, on the space H(C) of entire functions are frequently hyper-
cyclic. By Proposition 9.4 and Example 4.26 it suffices to consider a = 1.
Thus, let A(l, ν), l, ν ≥ 1, be subsets of N0 as given by Lemma 9.5, and
let (Pl )l be a dense sequence
 of polynomials. Let (nk )k be the increasing
sequence of elements of l,ν≥1 A(l, ν). If nk ∈ A(l, ν) then we define Bk as
the closed ball around nk of radius rk := ν/2, and on this ball we consider
the function gk := Pl (z − nk ); see Figure 9.1. It follows from the lemma that

Fig. 9.1 Constructing Birkhoff frequently hypercyclic functions

240 9 Frequently hypercyclic operators

the Bk are pairwise disjoint. We now apply Runge’s theorem recursively. We

start with f1 = g1 . If entire functions f1 , . . . , fk , k ≥ 1, have been constructed
then we consider the function that is defined as fk on |z| ≤ nk + rk and as
gk+1 on Bk+1 . Let εk > 0 be numbers that will be specified later. By Runge’s
approximation theorem there is an entire function fk+1 such that

sup |fk+1 (z) − fk (z)| < εk and sup |fk+1 (z) − gk+1 (z)| < εk .
|z|≤nk +rk z∈Bk+1

If ∞k=1 εk < ∞ then it follows from the first inequality and the fact that
nk → ∞ that
∞

f (z) := f1 (z) + (fk+1 (z) − fk (z)) = lim fk (z)
k→∞
k=1

defines an entire function. Moreover we have with ε0 = 0 that

∞

sup |f (z) − gk (z)| ≤ sup |fk (z) − gk (z)| + sup |fj+1 (z) − fj (z)|
z∈Bk z∈Bk z∈Bk
j=k
∞

≤ εj ,
j=k−1

that is,
∞

sup |f (z) − Pl (z − nk )| ≤ εj
|z−nk |≤ν/2 j=k−1

k ∈ A(l, ν). It is easy to see that we can choose the εj in such a way
for n
that ∞ j=k−1 εj < ν whenever nk ∈ A(l, ν). We therefore have that
1

1
sup |T1nk f (z) − Pl (z)| <
|z|≤ν/2 ν

for nk ∈ A(l, ν). Since the sets {g ∈ H(C) ; sup|z|≤ν/2 |g(z) − Pl (z)| < ν1 },
l, ν ≥ 1, form a basis of the topology of H(C) and since each set A(l, ν) has
positive lower density, it follows that T1 is frequently hypercyclic.

It is of interest to compare the new, and strong, form of hypercyclicity

with other strong forms such as weak mixing, mixing and chaos. We start
here by showing that every frequently hypercyclic operator is weakly mixing.
For the proof we need a property of sets of positive lower density. For any
subset A of N0 its difference set is defined as

A − A = {n − m ; n, m ∈ A, n ≥ m};
9.1 Frequently recurrent orbits 241

it should be noted that we consider here only nonnegative differences. We

recall that a subset B of N0 is called syndetic if its complement does not
contain intervals of arbitrary length; one also says that it has bounded gaps.
Theorem 9.7 (Erdős–Sárközy). Let A ⊂ N0 be a set of positive lower
density. Then the difference set A − A is syndetic.
Proof. Suppose that the difference set D of A is not syndetic. In particu-
lar, there exists some n1 ∈ / D. Moreover, since N0 \ D contains intervals
of arbitrary length, there is some n2 ∈ / D such that also n2 + n1 ∈ / D.
Hence {n1 , n2 , n1 + n2 } ⊂ N0 \ D. Similarly, there is some n3 ∈ / D such that
n3 +n1 , n3 +n2 ∈/ D, which implies that {n1 , n2 , n3 , n1 +n2 , n1 +n3 , n2 +n3 } ⊂
N0 \ D. Continuing in this way we obtain a sequence (nk )k in N0 such that
any finite sum of elements in the sequence belongs to N0 \ D.
1
We now fix a positive integer m such that dens(A) > m , and we consider
the sets
Ak = A + (n1 + . . . + nk ), k ∈ N.
Since each set Ak also has lower density larger than 1
m
, there is some N ≥ 1
such that, for any k ≤ m,
N +1
card{n ≤ N ; n ∈ Ak } > .
m
If the Ak , k = 1, . . . , m, were pairwise disjoint, we would have that
N +1
card{n ≤ N ; n ∈ A1 ∪ . . . ∪ Am } > m = N + 1,
m
which is impossible. Hence there are j < k with Aj ∩ Ak
= ∅, which implies
that
nj+1 + . . . + nk ∈ A − A = D.
This contradicts the construction of the nk . 

With this we can prove the announced result.
Theorem 9.8. Any frequently hypercyclic operator on a Fréchet space is
weakly mixing.
Proof. Let T be a frequently hypercyclic operator on a Fréchet space X. We
want to show that the condition of Theorem 2.47 is satisfied. Thus, let W be
a 0-neighbourhood and U and V nonempty open subsets of X.
First, since T is hypercyclic and therefore topologically transitive, there is
some n0 ≥ 0 such that T n0 (U ) ∩ W
= ∅. By continuity there is a nonempty
open subset U0 of U such that T n0 (U0 ) ⊂ W . Now let x be an arbitrary
frequently hypercyclic vector for T . Then there is a set A ⊂ N0 of positive
lower density such that

T n x ∈ U0 for any n ∈ A.
242 9 Frequently hypercyclic operators

For m, n ∈ A, m ≥ n, we then find that

T n0 +m−n (T n x) = T n0 (T m x) ∈ W.

We thus have that

n0 + (A − A) ⊂ N (U0 , W ) ⊂ N (U, W ).

It follows from Theorem 9.7 that N (U, W ) is syndetic.

Secondly, by continuity and linearity, each set T −k (W ) is a 0-neigh-
bourhood. Thus, given any m ≥ 1, there is a 0-neighbourhood W0 such
that T k (W0 ) ⊂ W for k = 1, . . . , m. Again by topological transitivity there is
some K > m and some y ∈ W0 such that T K y ∈ V and hence, for 1 ≤ k ≤ m,

T K−k (T k y) ∈ T K−k (W ) ∩ V.

This shows that, for any m ≥ 1, N (W, V ) contains m consecutive integers.

Our two conclusions imply that N (U, W ) ∩ N (W, V )
= ∅, so that, by
Theorem 2.47, T is weakly mixing. 

9.2 The Frequent Hypercyclicity Criterion

In order to obtain further examples of frequently hypercyclic operators, we

derive here a sufficient condition for frequent hypercyclicity that resembles
the Hypercyclicity Criterion. Its proof is inspired by Kitai’s constructive ap-
proach to that criterion; see the alternative proof of Theorem 3.12. However,
at a crucial point we have to depart from that proof: since we require ap-
proximation on sets of positive lower density we can no longer define the kj
inductively. On the other hand, fixing the kj in advance is no option either
as they will necessarily depend on the chosen operator. Lemma 9.5 provides
us with exactly the right tool for the construction.
We refer to Appendix A for the notion of unconditionally convergent series.

Theorem 9.9 (Frequent Hypercyclicity Criterion). Let T be an oper-

ator on a separable Fréchet space X. If there is a dense subset X0 of X and
a map S : X0 → X0 such that, for any x ∈ X0 ,
∞

(i) T n x converges unconditionally,
n=0
∞
(ii) S n x converges unconditionally,
n=0
(iii) T Sx = x,
then T is frequently hypercyclic.
9.2 The Frequent Hypercyclicity Criterion 243

Proof. Since X is separable we can choose a sequence (yj )j from X0 that

is dense in X. Let · denote an F-norm that defines the topology of X.
Conditions (i) and (ii) imply that there are Nl ∈ N, l ≥ 1, such that, for any
j ≤ l and any finite set F ⊂ {Nl , Nl + 1, Nl + 2, . . .} we have that



1

T n yj
< l , (9.3)
l2
n∈F



1

S n yj
< l . (9.4)
l2
n∈F

Now let A(l, ν), l, ν ≥ 1, be subsets of N0 as given by Lemma 9.5. We set

∞

A= A(l, Nl )
l=1

and
z n = yl if n ∈ A(l, Nl ).
We then consider 
x= S n zn . (9.5)
n∈A

First we want to verify that this series converges unconditionally. Let us

fix l ≥ 1. For any finite set F ⊂ N0 we have that

 ∞
  
l  ∞
 
S n zn = S n yj = S n yj + S n yj .
n∈A j=1 n∈A(j,Nj ) j=1 n∈A(j,Nj ) j=l+1 n∈A(j,Nj )
n∈F n∈F n∈F n∈F

It follows from (9.4) that, for j ≤ l and F ⊂ {Nl , Nl + 1, Nl + 2, . . .} finite,



1

S n yj
< l ;
l2
n∈A(j,Nj )
n∈F

moreover, since n ≥ Nj for any n ∈ A(j, Nj ) by Lemma 9.5, we also have by

(9.4) that, for any j ≥ 1 and any finite set F ,



1 1

S n yj
< j ≤ j .
j2 2
n∈A(j,Nj )
n∈F

Altogether we have that, for any finite set F ⊂ {Nl , Nl + 1, Nl + 2, . . .},


 l ∞



1 1 2

S n
z n
< l
+ j
= l.
j=1
l2 2 2
n∈A j=l+1
n∈F
244 9 Frequently hypercyclic operators

Since l was arbitrary we have proved that the series (9.5) converges uncon-
ditionally.
We now show that x is frequently hypercyclic for T . To this end, fix l ≥ 1.
Then, for n ∈ A(l, Nl ),
 
T n x − yl = T n S k zk + T n S k zk + T n S n zn − yl .
k∈A k∈A
k<n k>n

For the second sum we have, for any m ≥ n, using condition (iii),

 
l  ∞
 
T n S k zk = S k−n yj + S k−n yj .
k∈A j=1 k∈A(j,Nj ) j=l+1 k∈A(j,Nj )
n<k≤m n<k≤m n<k≤m

Note that, by Lemma 9.5, k − n ≥ Nl in the first sum and k − n ≥ Nj in the

second sum. Therefore, the same argument as above shows that


 l ∞



1 1 2

T n S k zk
< l
+ j
= l,
l2 2 2
k∈A j=1 j=l+1
n<k≤m

hence



2

T n S k zk
≤ l .
2
k∈A
k>n

In the same way, but using (9.3) instead of (9.4), we obtain that also



2

T n S k zk
≤ l .
2
k∈A
k<n

Finally, since n ∈ A(l, Nl ) we have that

T n S n zn = yl .

Altogether we find that for all n ∈ A(l, Nl )

4
T n x − yl ≤ .
2l
Since the yl form a dense set in X and since each set A(l, Nl ) is of positive
lower density we conclude that x is frequently hypercyclic for T . 

Remark 9.10. For a later application we note that the same proof works when
we replace conditions (ii) and (iii) by the following:
9.2 The Frequent Hypercyclicity Criterion 245

For any x ∈ X0 there is a sequence (un )n≥0 in X with u0 = x such that

∞
n=0 un converges unconditionally and T uk = uk−n if n ≤ k.
n

Let us also note here that the Frequent Hypercyclicity Criterion not only
implies frequent hypercyclicity but also two other strong forms of hypercyclic-
ity.

Proposition 9.11. An operator on a separable Fréchet space that satisfies

the Frequent Hypercyclicity Criterion is also chaotic and mixing.

Proof. The mixing property follows immediately from Kitai’s criterion.

As for chaos, we have from conditions (i) and (ii) that, for any x ∈ X0 and
N ≥ 1,
∞
 ∞

yx,N := S jN x + x + T jN x
j=1 j=1

converges in X. Moreover, by condition (iii), T N yx,N = yx,N , and by (i)

and (ii) we have that yx,N → x as N → ∞. Since X0 is dense, the yx,N
therefore form a dense set of periodic points for T . Knowing already that T
is hypercyclic we deduce that T is even chaotic. 

In the last section we saw that Birkhoff’s operators are frequently hyper-
cyclic. The Frequent Hypercyclicity Criterion allows us to show that also
the other two classical hypercyclic operators, the operators of MacLane and
Rolewicz, are in fact frequently hypercyclic; see also Exercise 9.2.2.

Example 9.12. (MacLane’s operator) The differentiation operator D on

H(C) is frequently hypercyclic. To see this we proceed as in Example
z 3.7.
Let X0 be the set of polynomials and S the operator Sf (z) = 0 f (ζ) dζ.
Condition (i) of the Frequent Hypercyclicity Criterion is satisfied since
any finite series converges unconditionally, and (iii) is trivial.
∞ Forn (ii)k we
need
∞ only consider the monomials, for which we find that n=0 S (z ) =
1
k! n=0 (k+n)! z k+n , which converges uniformly and unconditionally on any
compact set.

We study Rolewicz’s operators in the broader context of general weighted

shifts. We will use the notation and terminology of Section 4.1. In particular,
we consider weighted (backward) shifts

Bw : (x1 , x2 , x3 , . . .) → (w2 x2 , w3 x3 , w4 x4 , . . .),

where w = (wn )n is a weight sequence. By en , n ≥ 1, we denote the canonical

unit sequences.

Proposition 9.13. Let Bw be a weighted shift on a Fréchet sequence space

X in which span{en ; n ≥ 1} is dense. If the series
246 9 Frequently hypercyclic operators
∞ 
 n −1
wν en
n=1 ν=1

converges unconditionally in X then Bw is frequently hypercyclic.

Proof. We apply the Frequent Hypercyclicity Criterion. We choose X0 as the

set of finite sequences, which is dense by assumption, and for S we consider
the weighted forward shift (x1 , x2 , x3 , . . .) → (0, x1 /w2 , x2 /w3 , . . .). Then con-
dition (i) holds because any finite series converges unconditionally, and condi-
tion (iii) is obvious. By linearity, we need to confirm (ii) only for the sequences
ek , k ≥ 1. But then
∞
 ∞
 ek+n k  ∞  k+n
 −1
S n ek = = wν wν ek+n ,
n=0
w
n=0 k+1
· · · wk+n ν=1 n=0 ν=1

which converges unconditionally by hypothesis.



In particular, by Theorem 4.8 we have the following.

Corollary 9.14. On a Fréchet sequence space X in which (en )n is an un-

conditional basis, every chaotic weighted shift is frequently hypercyclic.

The result covers some interesting special cases.

Example 9.15. (Rolewicz’s operators) For λ ∈ K we consider the multiples

T = λB of the shift operator B with |λ| > 1. Then T is frequently hypercyclic
on any Fréchet sequence space on which it is defined, in which (en )n is an
unconditional basis and that contains the sequence (1/λn )n . This includes,
in particular, the spaces p , 1 ≤ p < ∞, and c0 .

Example 9.16. (a) We follow Example 4.9(b) and consider weighted shifts
T = Bw on H(C), or rather its corresponding sequence space. Since the
sequences en , n ≥ 0, correspond
∞ to the monomials z n , Bw turns out to be
frequently hypercyclic if n=1 ( ν=1 wν )−1 z n converges unconditionally in
n

H(C), which is equivalent to


n 1/n
lim |wν | = ∞.
n→∞
ν=1

Since the differentiation operator D corresponds to the weights wn = n, we

also get a new proof that D is frequently
hypercyclic.
∞
(b) In the space ω = KN , every series n=1 an en converges uncondition-
ally. As a consequence, every weighted shift is frequently hypercyclic on ω.

One is still far away from a characterization of frequently hypercyclic

weighted shifts. We complement here the sufficient condition derived above
by a necessary condition.
9.2 The Frequent Hypercyclicity Criterion 247

Proposition 9.17. Let Bw be a weighted shift on a Fréchet sequence space

X in which (en )n is an unconditional basis. If Bw is frequently hypercyclic
then there exists a subset A ⊂ N0 of positive lower density such that


n −1
wν en converges.
n∈A ν=1

Proof. Let x be a frequently hypercyclic vector for T = Bw . Since

T n x = (w2 w3 · · · wn+1 xn+1 , . . .)

and since the projection onto the first coordinate is continuous on X, there
is a set B ⊂ N0 of positive lower density such that, for any n ∈ B,

|w2 w3 · · · wn+1 xn+1 − 2| < 1,

hence
1
|xn+1 | > .
|w2 w3 · · · wn+1 |
∞
Together with the unconditional convergence of n=1 xn en this implies that
 1
en+1
w1 w2 · · · wn+1
n∈B

converges; see Theorem A.16. This proves the claim for A = {n + 1 ; n ∈ B}.


We note that this condition is not, in general, a sufficient condition; see
Exercise 9.2.5.
While at the outset it was not even clear if frequently hypercyclic operators
exist, we have now actually seen that all the classical hypercyclic operators
and many others have this strong form of hypercyclicity. Although it is not to
be expected that hypercyclicity and frequent hypercyclicity coincide, we are
also in the position to give an example that differentiates the two concepts.
Example 9.18. On X = 2 we consider the weighted shift Bw with weights
wn = ( n+1
n )
1/2
. It follows from Example 4.9(a) that Bw is hypercyclic and
even mixing. However, if Bw were frequently hypercyclic then by the previous
proposition we could find a set A = {nk ; k ≥ 1} of positive lower density
∞
such that k=1 nk1+1 < ∞, which is impossible since nk = O(k); see the
discussion before Proposition 9.3. Note that Bw is conjugate to the shift
operator B on the Bergman space A2 ; see Example 4.4(b).
This example takes us back to the problem of comparing frequent hyper-
cyclicity with other forms of hypercyclicity. We saw in Theorem 9.8 that
every frequently hypercyclic operator is weakly mixing. Some other implica-
tions have turned out to be false. We have just seen that the mixing property
248 9 Frequently hypercyclic operators

does not imply frequent hypercyclicity. But there is also a frequently hyper-
cyclic operator on c0 that is neither mixing nor chaotic. In particular, by
Proposition 9.11, this operator does not satisfy the Frequent Hypercyclicity
Criterion. The construction of this example goes well beyond the scope of
this book.
The reader may have noticed that in frequent hypercyclicity so far we have
not made use of the Baire category theorem. This is unlike the situation in
hypercyclicity where the existence of a hypercyclic vector was deduced from
the fact that, in the sense of Baire category, there must be many of them;
see the proof of the Birkhoff transitivity theorem. In fact, this procedure is
ruled out in frequent hypercyclicity because, in general, the set F HC(T ) of
frequently hypercyclic vectors for an operator T is only of first Baire category.

Proposition 9.19. Let T be an operator on a Fréchet space X. If there is a

dense set X0 such that T n x → 0 for all x ∈ X0 then F HC(T ) is of first Baire
category. This is true, in particular, for all operators satisfying the Frequent
Hypercyclicity Criterion.

Proof. Let · be an F-norm defining the topology of X, and choose δ > 0

such that {x ∈ X ; x > δ} is nonempty. Then every frequently hypercyclic
vector for T belongs to the set

E := {x ∈ X ; dens{n ∈ N0 ; T n x ≥ δ} > 0}.

We have that  
E= Ek,M ,
k≥1 M ≥1

where
  
Ek,M = x ∈ X ; card{n ≤ N ; T n x ≥ δ} ≥ N +1
k .
N ≥M

The continuity of T implies that the complement of Ek,M ,

  
X \ Ek,M = x ∈ X ; card{n ≤ N ; T n x < δ} > (N + 1)(1 − k1 ) ,
N ≥M

is open, and it contains the dense set X0 . Hence each set Ek,M is nowhere
dense, so that E is of first Baire category. 

Thus one cannot argue as in the case of hypercyclicity (see Proposition

2.52), that any vector in the underlying space is the sum of two frequently
hypercyclic vectors. Indeed, there are frequently hypercyclic operators T for
which X
= F HC(T )+F HC(T ); see Exercise 9.2.6. On the other hand, there
are operators T for which the set F HC(T ) is sufficiently large to ensure that
X = F HC(T ) + F HC(T ); see Exercise 9.1.4.
9.3 An eigenvalue criterion for frequent hypercyclicity 249

We end this section with another interesting phenomenon. Chapter 11 will

be devoted to the question of whether an uncountable family of hypercyclic
operators on a given space can have a common hypercyclic vector. The answer
is positive, for example, for the Rolewicz operators λB, λ > 1, on any of the
spaces X = p , 1 ≤ p < ∞, or c0 ; see Example 11.11. The corresponding
result is false, however, for frequent hypercyclicity.

Example 9.20. Let X be one of the spaces p , 1 ≤ p < ∞, or c0 . Then the

Rolewicz operators λB, λ > 1, on X have no common frequently hypercyclic
vector. Indeed, suppose that x was such a vector. By the proof of Proposition
9.17 it then follows that, for any λ > 1,

δλ := dens{n ∈ N0 ; |λn xn+1 − 2| < 1} > 0.

Since there are uncountably many λ, one can find a finite subset, λ1 < λ2 <
. . . < λK say, such that
K
δλk > 2.
k=1
λ
Let ρ = min1≤k<K λk+1
k
,and choose M ∈ N such that ρM ≥ 3. We then have
for N sufficiently large that, for any k = 1, . . . , K,

card{M ≤ n ≤ N ; |λnk xn+1 − 2| < 1} ≥ 12 δλk N.

Since K 1
k=1 2 δλk N > N , the corresponding sets cannot be pairwise disjoint.
Hence there are 1 ≤ k < l ≤ K and n ≥ M such that |λnk xn+1 − 2| < 1 and
|λnl xn+1 − 2| < 1. Thus, λnk |xn+1 | > 1 and λnl |xn+1 | < 3, which implies that
ρM ≤ (λl /λk )n < 3, a contradiction.

9.3 An eigenvalue criterion for frequent hypercyclicity

By the Godefroy–Shapiro criterion, eigenvalues inside and outside the unit

disk with many associated eigenvectors are useful for proving an operator to
be hypercyclic. Additional eigenvectors to certain unimodular eigenvalues are
responsible for chaos. We recall that an eigenvalue λ is called unimodular if
|λ| = 1.
In this section we will see that, rather surprisingly, a large supply of eigen-
vectors to unimodular eigenvalues by itself may lead to hypercyclicity, and in
some cases to frequent hypercyclicity. Let us only mention that the correct
interpretation of largeness in this context was again motivated by ergodic
theoretic considerations.
Suppose for the moment that T is an operator on a complex Fréchet space
X whose eigenspaces to unimodular eigenvalues all have dimension at most
one. One can then define an eigenvector field E : T → X so that, for any
250 9 Frequently hypercyclic operators

λ ∈ T, E(λ) is either an eigenvector to the eigenvalue λ, or 0. Since we want

a large supply of eigenvectors we would demand that span{E(λ) ; λ ∈ T}
is dense in X, in which case E is called spanning. In order to capture the
situation where eigenspaces are higher-dimensional one has to allow for a
collection of eigenvector fields. In the sequel, J is a nonempty index set.

Definition 9.21. Let T be an operator on a complex Fréchet space X. Then

a collection of functions Ej : T → X, j ∈ J, is called a spanning eigenvector
field associated to unimodular eigenvalues if Ej (λ) ∈ ker(λI − T ) for any
λ ∈ T, j ∈ J, and

span{Ej (λ) ; λ ∈ T, j ∈ J} is dense in X.

In addition, the vector field is said to be continuous (or C 2 ) if each function

Ej : T → X, j ∈ J, is continuous (or C 2 , respectively).

As usual, a function E : T → X, is called C 2 if it is twice continuously

differentiable, where differentiation is defined as in the scalar-valued case.
We now have the announced eigenvalue criterion.

Theorem 9.22. Let T be an operator on a complex separable Fréchet space.

(a) If T has a spanning continuous eigenvector field associated to unimod-
ular eigenvalues then it is mixing and chaotic.
(b) If T has a spanning C 2 -eigenvector field associated to unimodular
eigenvalues then it is frequently hypercyclic.

The proof is similar to that of Theorem 7.32. We will need the Riemann
integral  2π
f (t) dt
0

for a continuous function f : [0, 2π] → X; see Appendix A for details and
basic properties.

Lemma 9.23. Let X be a complex Fréchet space and f : [0, 2π] → X a

continuous function.  2π
(a) (Riemann–Lebesgue lemma) Then 0 eint f (t) dt → 0 as n →
±∞.
(b) If f is twice continuously differentiable with f (0) = f (2π) and
 2π int ∞  2π
f (0) = f  (2π) then ∞

n=0 0 e f (t) dt and n=0 0 e−int f (t) dt converge
unconditionally.

Proof. Let (pk )k be an increasing sequence of seminorms defining the topol-

ogy of X.
(a) As in the proof of Lemma 7.31 one shows that, for any k ≥ 1,
 2π
pk ( 0 eint f (t) dt) → 0 as n → ±∞. This implies the claim.
(b) Upon integrating by parts twice we obtain that, for any k ≥ 1, n
= 0,
9.3 An eigenvalue criterion for frequent hypercyclicity 251
 2π    2π   2π
1 1
pk eint f (t) dt = pk − 2 eint f  (t) dt ≤ 2 pk (f  (t)) dt,
0 n 0 n 0

which implies the claim. 

We are now in a position to prove the eigenvalue criterion.

Proof of Theorem 9.22. (a) Let (Ej )j∈J be the given eigenvector field of T .
Since each Ej : T → X is continuous the integrals
 2π
xk,j := eikt Ej (eit ) dt ∈ X, k ∈ Z, j ∈ J,
0

are defined. In order to apply Kitai’s criterion we set

X0 = Y0 = span{xk,j ; k ∈ Z, j ∈ J}.

We will use the Hahn–Banach theorem to show that this set is dense. Thus,
let x∗ be a continuous linear functional on X so that, for all k ∈ Z, j ∈ J,
 2π
xk,j , x∗  = eikt Ej (eit ), x∗  dt = 0.
0

The functions t → Ej (eit ), x∗  are continuous and therefore belong to

L2 [0, 2π]. Since ( √12π eikt )k∈Z is an orthonormal basis in this Hilbert space,
we deduce that, by continuity,

Ej (eit ), x∗  = 0 for all t ∈ [0, 2π], j ∈ J.

Hence x∗ vanishes on the set span{Ej (λ) ; λ ∈ T, j ∈ J}, which is dense by

assumption, so that x∗ itself must vanish. Thus X0 = Y0 is dense.
Now, since each Ej (λ) is in the eigenspace of λ we have for any k ∈ Z and
j ∈ J that
 2π  2π
T n xk,j = eikt T n Ej (eit ) dt = ei(k+n)t Ej (eit ) dt → 0
0 0

as n → ∞, as a result of the Riemann–Lebesgue lemma. By linearity, we

conclude that Tn x → 0 for all x ∈ X0 .
It would seem natural to define the mapping S : Y0 → Y0 by
 2π  2π
xk,j = ikt
e Ej (e ) dt →it
ei(k−1)t Ej (eit ) dt = xk−1,j ,
0 0

followed by linear extension to Y0 . Since this may lead to a conflict if the

xk,j are not linearly independent we apply, instead, the variant, Exercise
3.1.1, of Kitai’s criterion. Thus, for any y ∈ Y0 , we consider a representation
m
y = l=1 al xkl ,jl and define
252 9 Frequently hypercyclic operators


m
un = al xkl −n,jl , n ≥ 0.
l=1

We then have T n un = y and, again by the Riemann–Lebesgue lemma, that

xkl −n,j → 0 as n → ∞, so that un → 0. We can therefore conclude that T is
mixing.
Moreover, by continuity of the eigenvector field, also

span{Ej (λ) ; j ∈ J, λ = eαπi for some α ∈ Q}

is dense in X, and each vector in this span is a periodic point for T . Conse-
quently, T is chaotic.
(b) The proof follows the same lines, this time using Lemma 9.23(b) and
the Frequent Hypercyclicity Criterion in the form of Remark 9.10. 

The eigenvalue criterion provides a new proof that the three classical hy-
percyclic operators are even frequently hypercyclic.

Example 9.24. (Rolewicz’ operators) We consider the Rolewicz operators

T = μB, |μ| > 1, on one of the complex spaces X = p , 1 ≤ p < ∞, or c0 .
Then
E : T → X, λ → (λn /μn )n
is an eigenvector field associated to unimodular eigenvalues. An elementary
but tedious calculation shows that the field is C 2 (see Exercise 9.3.2), while
the spanning property was proved in Example 3.2.

Concerning MacLane’s and Birkhoff’s operators we will show a much more

general result, namely that Theorem 4.21 by Godefroy and Shapiro also holds
for frequent hypercyclicity.

Theorem 9.25. Suppose that T : H(C) → H(C), T

= λI, is an operator
that commutes with D, that is,

T D = DT.

Then T is frequently hypercyclic.

Proof. Following the proof of Theorem 4.21 we can write T = ϕ(D) with a
nonconstant entire function ϕ of exponential type, which also implies that ev-
ery function eλ (z) = eλz , λ ∈ C, is an eigenvector of T to the eigenvalue ϕ(λ).
Since ϕ(C) is connected and dense (see Appendix A), there is a point z ∈ C
with w := ϕ(z) ∈ T; and since ϕ(C) is open and the zeros of ϕ are isolated
points we can also achieve that ϕ (z)
= 0. Thus ϕ maps a neighbourhood of
z conformally onto a neighbourhood U of w; let ψ be the inverse map, which
is holomorphic. Fix a nontrivial closed subarc γ ⊂ U of T containing w and a
C 2 -function f : T → C with f (w)
= 0 that vanishes outside γ. It follows that
E : T → H(C) with E(λ) = f (λ)eψ(λ) if λ ∈ γ and E(λ) = 0, else, defines an
9.3 An eigenvalue criterion for frequent hypercyclicity 253

eigenvector field associated to unimodular eigenvalues for T . It was shown in

the proof of Lemma 2.34 that the function C → H(C), λ → eλ is, in fact,
infinitely differentiable, so that E is a C 2 -field. Finally, E is spanning by
Lemma 2.34. Now the eigenvalue criterion for frequent hypercyclicity implies
the result. 

As in the case of hypercyclicity one may ask how slowly a frequently hy-
percyclic entire function can grow at infinity. The eigenvalue criterion al-
lows us to deduce corresponding results for any operator T = ϕ(D); see
Exercise 9.3.3. Here we consider only the special case of Birkhoff’s operators
Ta f (z) = f (z + a), a
= 0. The theorem of Duyos-Ruiz tells us that corre-
sponding hypercyclic functions can grow arbitrarily slowly. This is no longer
true in the frequent context.

Theorem 9.26. Let a

= 0.
(a) Let ε > 0. Then there exists an entire function f that is frequently
hypercyclic for Ta and that satisfies

|f (z)| ≤ M eεr for |z| = r > 0

with some M > 0.

(b) Let ε : R+ → R+ be a function with lim inf r→∞ ε(r) = 0. Then there
is no entire function f that is frequently hypercyclic for Ta and that satisfies

|f (z)| ≤ M eε(r)r for all |z| = r > 0 sufficiently large

with some M > 0.

Proof. (a) This result follows from a general growth result for all operators
that commute with D (see Exercise 9.3.3) because Ta = eaD and eaz = 1 for
z = 0.
(b) We will assume that a = 1; see Example 4.26. Suppose, on the contrary,
that f is a frequently hypercyclic entire function with the stated growth
condition; by adding a constant, if necessary, we can assume that f (0) = 1.
Then there is a strictly increasing sequence (nk )k of positive integers with
nk = O(k) such that, for any k ≥ 1,
1 1
|f (z + nk ) − z| < for |z| ≤ .
2 2
Thus, by Rouché’s theorem (see Appendix A), f has a zero in |z − nk | < 12 . If
N (r) denotes the number of zeros of f in |z| < r, counting multiplicity, then

N (nk + 1) ≥ k, k ≥ 1.

On the other hand, it follows from Jensen’s formula (see Theorem A.23) and
the growth assumption on f that
254 9 Frequently hypercyclic operators

N (r) log 2 ≤ log M + ε(2r)2r

for r sufficiently large.

Now let rν → ∞ be such that ε(2rν ) → 0 as ν → ∞. For sufficiently large
ν choose kν such that nkν + 1 ≤ rν ≤ nkν +1 . Altogether we conclude that

kν N (nkν + 1) N (rν ) log M + ε(2rν )2rν

≤ ≤ ≤ → 0,
nkν +1 nkν +1 rν rν log 2
nkν +1 kν nkν +1
hence that kν +1 = kν +1 kν → ∞, which is a contradiction. 

9.4 Structural properties

In the previous chapters we derived various structural properties of hyper-

cyclicity. In this section we revisit several of them in the context of frequent
hypercyclicity.
We begin by looking at the main results of Chapter 6. Ansari’s theorem
says that every power T p , p ≥ 1, of a hypercyclic operator is again hyper-
cyclic; in fact, T and T p have the same hypercyclic vectors. For frequent
hypercyclicity we have the corresponding property, but its proof relies on
very different techniques than Ansari’s theorem.
Theorem 9.27. Let T be an operator on a Fréchet space. Then, for any
p ∈ N, F HC(T ) = F HC(T p ). In particular, if T is frequently hypercyclic
then so is every power T p .
Proof. Since every orbit orb(x, T p ) is obtained from the orbit orb(x, T ) by
retaining only the powers T np x, n ≥ 0, it is clear that every frequently
hypercyclic vector for T p is also frequently hypercyclic for T .
Conversely, let x ∈ X be a frequently hypercyclic vector for T and p ≥ 1. In
order to show that x is also frequently hypercyclic for T p we fix a nonempty
open subset U of X. Since the sequence (kp − 1)k≥1 is syndetic, we can
deduce from Theorems 9.8 and 1.54 that there is some m1 ≥ 0 of the form
m1 = k1 p − 1 such that U1 := U ∩ T −m1 (U )
= ∅. For the same reason, there
is some m2 ≥ 0 of the form m2 = k2 p−2 such that U2 := U1 ∩T −m2 (U1 )
= ∅.
Proceeding inductively we find, for j = 1, . . . , p − 1, integers mj ≥ 0 of the
form mj = kj p − j such that Uj := Uj−1 ∩ T −mj (Uj−1 )
= ∅, where U0 := U .
Moreover we set k0 = 0.
Now let V = Up−1 , which clearly satisfies V ⊂ U and T kj p−j (V ) ⊂ U , for
j = 0, 1, . . . , p − 1. Since x is frequently hypercyclic there is a subset A ⊂ N0
of positive lower density such that T n x ∈ V for all n ∈ A. We then define the
function f : N0 → N0 by f (n) = n−j p + kj if n = j (mod p), j = 0, . . . , p − 1;
note that this is well defined.
We finally set B = f (A). It is easy to show that dens(B) ≥ dens(A) > 0;
see Exercise 9.4.1. Moreover, if m ∈ B, then m = n−j p + kj for some n ∈ A
9.4 Structural properties 255

with n = j (mod p), and hence

(T p )m x = T n−j+kj p x = T kj p−j (T n x) ∈ T kj p−j (V ) ⊂ U.

This proves that x is frequently hypercyclic for T p . 

We saw in Section 6.3 how Ansari’s theorem follows from the fact that if
the union of the orbits of finitely many vectors is dense then one of these
orbits must already be dense. The corresponding result fails for frequent
hypercyclicity.

Example 9.28. We consider the Rolewicz operator T = 2B on 1 . We claim

that there are two vectors v, w ∈ 1 such that, for any nonempty open subset
U ⊂ 1 ,
dens{n ∈ N0 ; T n v ∈ U or T n w ∈ U } > 0, (9.6)
but neither v nor w is frequently hypercyclic for T .
To see this, let (yj )j be a dense sequence in 1 consisting of finite sequences.
The proof of the Frequent Hypercyclicity Criterion, with S being half the
forward shift, then constructs a frequently hypercyclic vector x for T given
by 
x= S n zn ,
n∈A

where A is the union of certain pairwise disjoint sets A(l, Nl ), l ≥ 1, of positive

lower density, and zn = yl for n ∈ A(l, Nl ). For an increasing sequence
(mk )k≥0 of positive integers with m0 = 0, that will be determined later, we
split the set A into two subsets

B = {n ∈ A ; ∃ k ≥ 0 : m2k ≤ n < m2k+1 },

C = {n ∈ A ; ∃ k ≥ 0 : m2k+1 ≤ n < m2k+2 }.

The proof of the Frequent Hypercyclicity Criterion then shows that the series
 
v := S n zn , w := S n zn
n∈B n∈C

converge and that, for any n ∈ A(l, Nl ),

4 4
T n v − yl ≤ or T n w − yl ≤
2l 2l
depending on whether n ∈ B or n ∈ C. This implies that the joint orbits of
v and w are frequently recurrent, in the sense of (9.6).
On the other hand, let ln be the length of the finite sequence zn , n ∈ A.
Let k ≥ 0. If n ∈ N0 satisfies

M2k+1 := max (ν + lν ) ≤ n < m2k+2

ν∈B,ν<m2k+1
256 9 Frequently hypercyclic operators

then the sequence T n v starts with a 0 and hence T n v ∈

/ U , where U = {x ∈
1 ; x − e1 < 1}. Now, if we choose the m2k+2 , k ≥ 1, such that

M2k+1 1
≤
m2k+2 k

then dens{n ≥ 0 ; T n v ∈ U } = 0, which shows that v is not frequently

hypercyclic for T . Imposing, in addition, a similar condition on the m2k+1
one can also achieve that w is not frequently hypercyclic for T .
For a variant of the Bourdon–Feldman theorem for frequent hypercyclicity
see Exercise 9.4.4.
We next turn to the results of Section 6.4. For this we need to define
frequent hypercyclicity for C0 -semigroups. The lower density of a measurable
subset A ⊂ R+ is given by

λ{t ∈ [0, T ] ; t ∈ A}
dens(A) := lim inf ,
T →∞ T
where λ denotes the Lebesgue measure.
Definition 9.29. A C0 -semigroup (Tt )t≥0 on a Banach space X is called
frequently hypercyclic if there is a vector x ∈ X such that, for any nonempty
open subset U of X,

dens {t ∈ R+ ; Tt x ∈ U } > 0.

In this case, x is called a frequently hypercyclic vector for (Tt )t≥0 .

As before we will treat the problems of unimodular multiples and of dis-
cretizations of semigroups within the common framework of semigroup ac-
tions; see Section 6.4. We recall that if T is an operator on a complex Fréchet
space X then
Ψ (n, t) = e2πti T n , n ∈ N0 , t ≥ 0, (9.7)
defines a semigroup action. Similarly, if (Tt )t≥0 is a C0 -semigroup on a Banach
space X then
Ψ (n, t) = Tt , n ∈ N0 , t ≥ 0, (9.8)
defines a semigroup action. In both cases, properties (α) and (β) of Section
6.4 are satisfied.
We then also need a concept of frequent hypercyclicity for semigroup ac-
tions. The natural notion of lower density on G = N0 × R+ is given by

1 N
dens(A) := lim inf λ{t ∈ [0, N ] ; (n, t) ∈ A},
N →∞ N (N + 1) n=0

where A ⊂ G is such that {t ≥ 0 ; (n, t) ∈ A} is measurable for each n ≥ 0;

λ denotes the Lebesgue measure.
9.4 Structural properties 257

Definition 9.30. A semigroup action Ψ : G → L(X) is called frequently

hypercyclic if there is some x ∈ X such that, for any nonempty open subset
U of X,
dens {g ∈ G ; Ψ (g)x ∈ U } > 0.
In this case, x is called a frequently hypercyclic vector for Ψ .

Now, frequent hypercyclicity of some operator Ψ (n, t), n, t > 0, implies

frequent hypercyclicity of Ψ .

Proposition 9.31. Let Ψ be a semigroup action on a Fréchet space X sat-

isfying property (α). If x ∈ X is frequently hypercyclic for some operator
Ψ (n, t), n, t > 0, then it is frequently hypercyclic for Ψ .

Proof. Let U be a nonempty open subset of X. Since Ψ (0, 0) = I and Ψ is

continuous, there is a nonempty open subset V of U and some η > 0 such
that Ψ (0, s)V ⊂ U if 0 ≤ s < η. By assumption, there is some (n, t) ∈ G
such that dens(A) = δ > 0, where A = {k ∈ N0 ; Ψ (n, t)k x ∈ V }. Now,
if k ∈ A and 0 ≤ s < η then Ψ (kn, kt + s)x = Ψ (0, s)Ψ (n, t)k x ∈ U . In
view of property (α), if Ψ (1, 0) = I then also Ψ (kn + m, kt + s)x ∈ U for
m ∈ Z with kn + m ≥ 0; if Ψ (0, 1) = I then Ψ (kn, kt + s + m)x ∈ U for
m ∈ Z with kt + s + m ≥ 0. In both cases a simple count reveals that
dens{(k, s) ∈ G ; Ψ (k, s)x ∈ U } ≥ ηδ/ max(n, t) > 0. 

Our main aim is to prove the converse statement. The following will be
crucial.

Lemma 9.32. Let Ψ be a semigroup action on an infinite-dimensional Fré-

chet space X satisfying properties (α) and (β). If x ∈ X is frequently hyper-
cyclic for Ψ then, for any k ∈ N and any nonempty open subset U of X, we
have that
 ∞ 
dens (n, t) ∈ G ; Ψ (n, t)x ∈ U, t ∈ m=1 [m − k1 , m[ > 0.

Proof. We fix k ∈ N and a nonempty open subset U of X. For j = 1, . . . , k

we define the sets ∞

Ij = [m − kj , m − j−1
k [.
m=1

By Theorem 6.10, x is hypercyclic for Ψ (1, 1). It follows from property (β)
that, for any t ≥ 0, also Ψ (0, t)x is hypercyclic for Ψ (1, 1). Therefore there
are nj ∈ N0 , j = 1, . . . , k, such that

k )x ∈ U.
j−1
Ψ (nj , nj + k )x = Ψ (1, 1)nj Ψ (0, j−1

By continuity there is a neighbourhood V of x such that

Ψ (nj , nj + j−1
k )(V ) ⊂ U, j = 1, . . . , k.
258 9 Frequently hypercyclic operators

Let N0 = max(n1 , . . . , nk ) + 1. It follows from frequent hypercyclicity of x

for Ψ that there are δ > 0 and N1 ≥ N0 such that, if N ≥ N1 , then

1 N
λ{t ∈ [0, N ] ; Ψ (n, t)x ∈ V } ≥ δ.
N (N + 1) n=0

We now fix N ≥ N1 . Since the Ij , j = 1, . . . , k, form a partition of R+ ,

there is some j such that


N
1
N
λ{t ∈ [0, N ] ; Ψ (n, t)x ∈ V, t ∈ Ij } ≥ λ{t ∈ [0, N ] ; Ψ (n, t)x ∈ V }.
n=0
k n=0

We fix such a j. If 0 ≤ n ≤ N , 0 ≤ t ≤ N , t ∈ Ij , and Ψ (n, t)x ∈ V then

ν := n + nj ≤ 2N , τ := t + nj + j−1
k ≤ 2N , τ ∈ I1 , and
      
Ψ (ν, τ )x = Ψ nj , nj + j−1
k Ψ n, t x ∈ Ψ nj , nj + j−1
k V ⊂ U.

We conclude that

2N
λ{τ ∈ [0, 2N ] ; Ψ (ν, τ )x ∈ U, τ ∈ I1 }
ν=0

1
N
≥ λ{t ∈ [0, N ] ; Ψ (n, t)x ∈ V },
k n=0

so that

1 
2N
δ
λ{τ ∈ [0, 2N ] ; Ψ (ν, τ )x ∈ U, τ ∈ I1 } ≥ .
2N (2N + 1) ν=0 4k

Since N ≥ N1 was arbitrary, the claim follows. 

We can now prove the analogue of Theorem 6.10 for frequent hypercyclic-
ity.

Theorem 9.33. Let Ψ be a semigroup action on an infinite-dimensional

Fréchet space X satisfying properties (α) and (β). If x ∈ X is frequently
hypercyclic for Ψ then it is frequently hypercyclic for every operator Ψ (1, t),
t > 0.

Proof. We first prove the case when t = 1. Thus, let U be a nonempty

open subset of X. Since Ψ (0, 0) = I, continuity of Ψ implies that there is a
nonempty open subset V of U and some η > 0 such that Ψ (0, s)V ⊂ U if
0 ≤ s < η. Let k ∈ N be such that k1 < η. Then, by Lemma 9.32, there are
δ > 0 and N0 ∈ N such that, for any N ≥ N0 ,
9.4 Structural properties 259


N
∞ 
r := λ{t ∈ [0, N ] ; Ψ (n, t)x ∈ V, t ∈ m=1 [m − k1 , m[ ≥ N (N + 1)δ.
n=0

Now, if Ψ (n, t)x ∈ V and t ∈ [m− k1 , m[ then Ψ (n, m)x = Ψ (0, m−t)Ψ (n, t)x ∈
U . Thus, for

p := card{(n, m) ; 0 ≤ n, m ≤ N, Ψ (n, m)x ∈ U }

we have that p k1 ≥ r.
Next, let Ψ (1, 1)n x ∈ U . We distinguish the two cases described by (α). If
Ψ (1, 0) = I then Ψ (m, n)x = Ψ (n, n)x = Ψ (1, 1)n x ∈ U for any m ∈ Z; and
if Ψ (0, 1) = I then Ψ (n, m)x = Ψ (n, n)x ∈ U for any m ∈ Z. Thus, for

q := card{0 ≤ n ≤ N ; Ψ (1, 1)n x ∈ U }

we have that p = (N + 1)q.

Altogether we find that, for any N ≥ N0 ,

card{0 ≤ n ≤ N ; Ψ (1, 1)n x ∈ U } p kr kN

= ≥ ≥ δ.
N +1 (N + 1)2 (N + 1)2 N +1

Hence x is frequently hypercyclic for Ψ (1, 1).

Now, if t > 0 is arbitrary then we rescale the semigroup action as in the
proof of Theorem 6.10. It is then not difficult to see, using property (α),
that x is also frequently hypercyclic for Ψ and thus frequently hypercyclic for
Ψ(1, 1) = Ψ (1, t). 

If we combine Theorem 9.33 with Theorem 9.27, noting that ψ(n, t) =

Ψ (1, t/n)n , we obtain the announced converse of Proposition 9.31.

Corollary 9.34. Let Ψ be a semigroup action on a Fréchet space X satisfying

properties (α) and (β). If x ∈ X is frequently hypercyclic for Ψ then it is
frequently hypercyclic for every operator Ψ (n, t), n, t > 0.

Proposition 9.31 and Theorem 9.33, applied to the semigroup action (9.7),
immediately imply a version of the León–Müller theorem for frequent hyper-
cyclicity.

Theorem 9.35. Let T be an operator on a complex Fréchet space and λ ∈ C

with |λ| = 1. Then T and λT have the same frequently hypercyclic vectors,
that is, F HC(T ) = F HC(λT ).

Similarly, applying Theorem 9.33 to the semigroup action (9.8) yields an

analogue of the Conejero–Müller–Peris theorem.

Theorem 9.36. Let (Tt )t≥0 be a C0 -semigroup on a Banach space X. If

x ∈ X is frequently hypercyclic for (Tt )t≥0 , then it is frequently hypercyclic
for every operator Tt , t > 0.
260 9 Frequently hypercyclic operators

Apart from being interesting in its own right, Theorem 9.35 has an impor-
tant application. We saw in Chapter 5 that the spectrum of a hypercyclic op-
erator on a complex Banach space has the property that each of its connected
components meets the unit circle; this is the content of Kitai’s theorem. In
particular, the spectrum cannot have isolated points outside the unit circle.
We will now show that the spectrum of frequently hypercyclic operators, just
like that of chaotic operators (see Proposition 5.7), cannot even have isolated
points on the unit circle.
We start with a crucial lemma whose proof uses complex analysis in a very
clever way.
Lemma 9.37. Let T be an operator on a real Fréchet space X. Let x ∈ X
and x∗ ∈ X ∗ with x, x∗ 
= 0 be such that

|(T − I)n x, x∗ |1/n → 0

as n → ∞. Then x is not frequently hypercyclic for T .

Proof. First, we may assume that x, x∗  = 1. Suppose that x is frequently
hypercyclic for T . Then (T n x, x∗ )n≥0 is dense in R. Thus there must be
some n ≥ 0 such that T n x, x∗  ≤ 0 and T n+1 x, x∗  > 0. Then, for α > 0
sufficiently small, T n x − αx, x∗  < 0 and T n+1 x − αT x, x∗  > 0, so that
the open set

U = {y ∈ X ; y, x∗  < 0 and T y, x∗  > 0}

is nonempty.
We now consider the series
∞
 z(z − 1) · · · (z − k + 1)
f (z) = (T − I)k x, x∗  , z ∈ C,
k!
k=0

where we regard the quotient as 1 if k = 0. We claim that this defines an

entire function. Indeed, it follows from the assumption that, for any ε ∈ ]0, 1[,
there is some M > 0 such that

|(T − I)n x, x∗ | ≤ M εn , n ≥ 0,

so that, for any R > 0 and |z| ≤ R,

∞
  z(z − 1) · · · (z − k + 1) 
 
|(T − I)k x, x∗ | 
k!
k=0
∞
 R(R + 1) · · · (R + k − 1)
≤M εk
k!
k=0
∞
 −R 
M
=M (−ε)k = < ∞,
k (1 − ε)R
k=0
9.4 Structural properties 261

where we have used the binomial theorem. Moreover, setting η = − log(1−ε),

this inequality implies that

|f (z)| ≤ M eη|z| , z ∈ C.

In addition, f (0) = x, x∗  = 1. It follows from Jensen’s formula (see Theorem

A.23) that if N (r) denotes the number of zeros of f in |z| < r, counting
multiplicity, then

N (r) log 2 ≤ log M + 2rη, r > 0. (9.9)

On the other hand we have that for n ∈ N0 ,


n
n(n − 1) · · · (n − k + 1)
f (n) = (T − I)k x, x∗ 
k!
k=0
n  
 
n
= (T − I)k x, x∗ = T n x, x∗ .
k
k=0

Thus, if T n x ∈ U then f (n) < 0 and f (n + 1) > 0, so that f , being real on

the real axis, has a zero in the interval ]n, n + 1[. It follows with (9.9) that

card{0 ≤ n ≤ m ; T n x ∈ U } N (m + 1) log M + 2(m + 1)η 2η

≤ ≤ →
m+1 m+1 (m + 1) log 2 log 2

as m → ∞. Since η > 0 is arbitrary, we deduce that x is not frequently

hypercyclic. 

As an immediate consequence we have the following. Recall that an oper-
ator T on a Banach space is called quasinilpotent if T n 1/n → 0 as n → ∞.
Lemma 9.38. Let T be an operator on a Banach space X of the form T =
λI +S with |λ| = 1 and S quasinilpotent. Then T is not frequently hypercyclic.
Proof. By Theorem 9.35 we may assume that λ = 1, so that (T −I)n 1/n →
0 as n → ∞. Moreover, we can regard X as a real Banach space and T − I
as a (real-linear) operator on X. We then have that, for any x ∈ X and any
(real-linear) continuous linear functional x∗ on X,
 1/n
|(T − I)n x, x∗ |1/n ≤ (T − I)n x x∗ → 0.

By the previous lemma, T cannot be frequently hypercyclic on X; note that

this notion does not depend on the scalar field. 

We can now prove the mentioned spectral property of frequently hyper-
cyclic operators.
Theorem 9.39. Let T be a frequently hypercyclic operator on a complex Ba-
nach space. Then its spectrum σ(T ) has no isolated points.
262 9 Frequently hypercyclic operators

Proof. Suppose that λ ∈ C is an isolated point of the spectrum. Then σ(T )

can be partitioned into some closed subset and the singleton {λ}. By the Riesz
decomposition theorem (see Appendix B) there are nontrivial T -invariant
closed subspaces M1 and M2 of X such that X = M1 ⊕ M2 and σ(T |M2 ) =
{λ}. By Exercise 2.2.8 and Proposition 9.4, T |M2 is frequently hypercyclic.
By Kitai’s theorem we have that |λ| = 1, and the spectral radius formula (see
Appendix B) implies that T |M2 = λI + S with a quasinilpotent operator S.
This contradicts Lemma 9.38. 

Lemma 9.38 has another application. First, combining it with Lemma 5.19
yields the following.

Proposition 9.40. No compact perturbation of a multiple of the identity on

a Banach space is frequently hypercyclic.

We can then apply the Argyros–Haydon theorem; see Theorem 8.11.

Corollary 9.41. Let K = R or C. Then there exists an infinite-dimensional

separable Banach space over K that supports no frequently hypercyclic oper-
ator.

With this we end our introduction to frequent hypercyclicity.

Exercises

Exercise 9.1.1. Show that the Herrero–Bourdon theorem also holds for frequent hyper-
cyclicity. In particular, every frequently hypercyclic operator on a Fréchet space admits
a dense T -invariant subspace consisting, except for 0, of frequently hypercyclic vectors.

Exercise 9.1.2. Using Lemma 9.5, show that every weighted shift is frequently hyper-
cyclic on the space ω = KN .

Exercise 9.1.3. Show that every frequently hypercyclic operator on a Fréchet space is
topologically ergodic; see Exercise 1.5.6. Deduce that if T is a frequently hypercyclic
operator on a Banach space then its adjoint T ∗ cannot be frequently hypercyclic. (Hint:
Exercise 2.5.5, Remark 4.17.)

Exercise 9.1.4. Show that every entire function is the sum of two functions that are
frequently hypercyclic for the translation operator T1 f (z) = f (z + 1). (Hint: Use a
variant of the construction in Example 9.6.)

Exercise 9.2.1. Let Tn : X → Y , n ≥ 0, be operators between separable Fréchet spaces

X and Y . The definition of frequent hypercyclicity for the sequence (Tn )n≥0 is obvious.
Prove the following version of the Frequent Hypercyclicity Criterion for (Tn )n . For the
notion of uniformly unconditionally convergent series see Definition 11.7 below.
If there is a dense subset Y0 of Y and maps Sn : Y0 → X, n ≥ 0, such that, for any
y ∈ Y0 ,

m

(i) Tm Sm−n y converges unconditionally in Y , uniformly for m ≥ 0,

n=0
Exercises 263


∞

(ii) Tm Sm+n y converges unconditionally in Y , uniformly for m ≥ 0,

n=0
∞

(iii) Sn y converges unconditionally in X,

n=0
(iv) Tn Sn y → y, as n → ∞,
then (Tn )n is frequently hypercyclic.
Note that, in (i), the finite sums can be understood as infinite series by adding 0
terms.

Exercise 9.2.2. Use the Frequent Hypercyclicity Criterion to give a new proof that
Birkhoff’s operators are frequently hypercyclic. (Hint: Example 3.8.)

Exercise 9.2.3. Formulate and prove an analogue of Proposition 9.13 for weighted bi-
lateral shifts.

Exercise 9.2.4. Let Bw be a frequently hypercyclic weighted shift on p , 1 ≤ p < ∞.

Show that, for any ε > 0, there exists a subset A ⊂ N0 of positive lower density such
that, for any m ∈ A,
 1
< ε.
|w2 w3 · · · wn−m+1 |p
n∈A
n>m

(Hint: Proceed as in the proof of Proposition 9.17 and consider the coordinates of index
n − m + 1 in Bwm
x − e1 .)

Exercise 9.2.5. Let Nj = 2j 2 −2j +1, j ≥ 1. Define wn as ( n+1

n ) for Nj ≤ n < Nj +j,
2

−2/j
as (Nj + j) for Nj + j ≤ n < Nj + 2j, as 1 for Nj + 2j ≤ n < Nj + 3j, and as
(Nj+1 )2/j for Nj + 3j ≤ n < Nj + 4j = Nj+1 . Show that Bw is a weighted shift on p ,
1 ≤ p < ∞, that satisfies the condition given in Proposition 9.17 but not the condition
in the previous exercise. Thus, the condition in Proposition 9.17 does not characterize
frequent hypercyclicity of weighted shifts on p . (Hint: Use the result by Erdős and
Sárközy.)

Exercise 9.2.6. The aim of this exercise is to show that not every vector x ∈ 1 is the
sum of two frequently hypercyclic vectors for the Rolewicz operator 2B. Suppose that
x = y+z with y and z frequently hypercyclic. Then there is an increasing sequence (mk )k
of positive integers such that T mk y < 1 for k ≥ 1; further let dens{n ∈ N0 ; T n z <
1} =: 2δ > 0. Deduce that there are positive integers nk such that T nk z < 1 and
δmk ≤ nk ≤ mk , k ≥ 1 sufficiently large, and hence that T mk x ≤ 1 + 2(1−δ)mk .
Finally find some x ∈ 1 that fails this inequality for any δ > 0 and any increasing
sequence (mk )k of positive integers.

Exercise 9.2.7. Generalize Example 9.20: let T be an operator on a Fréchet space

X and Λ ⊂ ]0, ∞[ an uncountable set such that λT is frequently hypercyclic for any
λ ∈ Λ. Show that these operators have no common frequently hypercyclic vector. (Hint:
Consider U = {x ∈ X ; | x, x∗
− 2| < 1}.)

Exercise 9.3.1. Let Lp (T), 1 ≤ p < ∞, be the space of all complex-valued functions
 2π
f on T such that f p := ( 0 |f (eit )|p dt)1/p < ∞. Show that T f (λ) = λf (λ) −

(1,λ)
f (ζ) dζ defines a mixing and chaotic operator on Lp (T), where (λ1 , λ2 ) denotes the
positively oriented arc from λ1 to λ2 . (Hint: Consider the indicator functions f = 1(λ,1) .)
264 9 Frequently hypercyclic operators

Exercise 9.3.2. Let X be one of the complex spaces p , 1 ≤ p < ∞, or c0 . Show that
the map D → X, λ → (λn )n , is infinitely differentiable. Deduce that also the maps
D → H 2 , λ → kλ (see Proposition 4.38) and Dτ → Eτ2 , λ → eλ (see Exercise 4.2.4) are
infinitely differentiable.

Exercise 9.3.3. Let ϕ be a nonconstant entire function of exponential type and A =

min{|z| ; z ∈ C, |ϕ(z)| = 1}. Show that, for any ε > 0, there is an entire function f that
is frequently hypercyclic for ϕ(D) such that

|f (z)| ≤ M e(A+ε)r for |z| = r > 0

with some M > 0. (Hint: Combine the ideas of Exercise 4.2.4 and the proof of Theorem
9.25.)

Exercise 9.3.4. Let D be the differentiation operator on H(C). Let φ : ]0, ∞[→ [1, ∞[
be a function with φ(r) → ∞ as r → ∞. Show that there exists an entire function f
that is frequently hypercyclic for D and that satisfies

|f (z)| ≤ M φ(r)er for |z| = r > 0

with some M > 0. (Hint: Look at the proof of Theorem 4.22, using the Frequent Hy-
percyclicity Criterion in the version of Exercise 9.2.1.)

Exercise 9.3.5. Let ϕ be a nonconstant bounded holomorphic function on D and let

Mϕ∗ be the corresponding adjoint multiplication operator on H 2 ; see Section 4.4. Show
that Mϕ∗ is frequently hypercyclic if and only if it is hypercyclic, that is, if ϕ(D) ∩ T = ∅.
(Hint: Look at the proofs of Theorems 4.42 and 9.25.)

Exercise 9.4.1. In the proof of Theorem 9.27, show that dens(B) ≥ dens(A).

Exercise 9.4.2. Let T be a frequently hypercyclic operator on a Fréchet space X. Show

that then T p ⊕ T q is hypercyclic on X ⊕ X for any p, q ∈ N. (Hint: Exercises 2.5.5 and
9.1.3.)

Exercise 9.4.3. Let T be a topologically ergodic operator on a separable Fréchet space;

see Exercise 1.5.6. Show that T p is then also topologically ergodic for any p ≥ 1. (Hint:
Follow the proof of Theorem 9.27, using Exercise 2.5.5; see Exercise 6.1.5 for an alter-
native proof.)

Exercise 9.4.4. Let T be an operator on a separable Fréchet space X. Suppose that

there is a vector x ∈ X and a nonempty open subset U of X such that dens{n ∈
N0 ; T n x ∈ V } > 0 for all nonempty open subsets V of U . Show that x is frequently
hypercyclic for T . (Hint: Use the Bourdon–Feldman theorem.)

Exercise 9.4.5. Let T be an operator on a (real or complex) Banach space X. Show that
if there is some x∗ ∈ X ∗ , x∗ = 0, and some λ with |λ| = 1 such that (λI−T ∗ )n x∗ 1/n →
0 as n → ∞ then T is not frequently hypercyclic.

Sources and comments

Section 9.1. Frequently hypercyclic operators were introduced by Bayart and Grivaux
[38], [40]. The idea of using ergodic theory to obtain the dynamical properties of linear
operators seems to be due to Rudnicki [272] and Flytzanis [152, 153]. Bayart and Grivaux
Sources and comments 265

[40] obtained Lemma 9.5 (see also Bonilla and Grosse-Erdmann [87]) as well as the
frequent hypercyclicity of the Birkhoff operators. The theorem of Erdős and Sárközy
can be found in [296]. Theorem 9.8 is due to Grosse-Erdmann and Peris [185]; Bayart
and Matheron [45] show that this result is essentially optimal.
For an introduction to ergodic theory we refer to Walters [300].

Section 9.2. The Frequent Hypercyclicity Criterion was obtained by Bayart and Gri-
vaux [38, 40]; the form given here is due to Bonilla and Grosse-Erdmann [87]. Grivaux
[173] also provided a probabilistic version of it. Proposition 9.11 is due to Bonilla and
Grosse-Erdmann [87]. The remaining results in this section can essentially be found in
Bayart and Grivaux [40]; see also Bonilla and Grosse-Erdmann [87]. The latter paper also
contains further conditions under which the set F HC(T ) of frequently hypercyclic op-
erators is of first Baire category, or when F HC(T ) + F HC(T ) does or does not coincide
with the full space.
Bayart and Grivaux [41] constructed a weighted shift on c0 that is frequently hy-
percyclic, but neither chaotic nor mixing; this also shows that not every frequently
hypercyclic operator satisfies the Frequent Hypercyclicity Criterion, and that Proposi-
tion 9.13 does not characterize frequently hypercyclic weighted shifts on c0 . Badea and
Grivaux [19] found operators on a Hilbert space that are frequently hypercyclic and
chaotic but not mixing.
It remains an open problem whether every chaotic operator is frequently hypercyclic,
and to find a characterization of frequently hypercyclic weighted shifts, even on 2 or on
c0 .

Section 9.3. The proof of Theorem 9.22 follows Bayart and Grivaux [38]; see also [39].
Theorems 9.25 and 9.26 are due to Blasco, Bonilla and Grosse-Erdmann [86, 76]; these
authors also show that the operators of differentiation and translation on the space of
harmonic functions on RN are frequently hypercyclic, and they obtain some related
growth results.
In order to keep the presentation simple we have imposed rather strong assumptions
on the eigenvector fields. A much deeper analysis leads to one of the most striking results
in linear dynamics.
To be more specific, an operator T on a complex separable Banach space X is said
to have a perfectly spanning set of eigenvectors associated to unimodular eigenvalues if
one of the following two equivalent conditions holds:
(i) there exists an atomless probability measure σ on T such that, for any measurable
set A ⊂ T with σ(A) = 1, span{ker(λI − T ) ; λ ∈ A} is dense in X;
(ii) for any countable set D ⊂ T, span{ker(λI − T ) ; λ ∈ T \ D} is dense in X.
These conditions were first introduced by Flytzanis [152, 153]. Their equivalence was
shown by Grivaux [174], who also obtained the following fundamental principle.

Theorem 9.42. Any operator on a complex separable Banach space with a perfectly
spanning set of eigenvectors associated to unimodular eigenvalues is frequently hyper-
cyclic.

When the underlying space is even a Hilbert space then one can show that there exists
a Borel probability measure of full support on X with respect to which T is ergodic (see
Bayart and Grivaux [40]); as explained in Section 9.1, this immediately implies that T
is frequently hypercyclic. The measure can even be a so-called Gaussian measure. A
similar result for nuclear Fréchet spaces is due to Grosse-Erdmann [182]. For surveys
on the application of ergodic theory to linear dynamics we refer to Godefroy [164] and
Grosse-Erdmann [182]. A detailed treatment can be found in Bayart and Matheron [44].
Bayart and Grivaux [40, 41] have applied their results to various operators. In par-
ticular they have shown that if ϕ is an automorphism of the unit disk D then the
266 9 Frequently hypercyclic operators

corresponding composition operator Cϕ (see Section 4.5) is frequently hypercyclic on

the Hardy space H 2 if and only if it is hypercyclic, that is, if and only if ϕ is parabolic
or hyperbolic.
The example of Bayart and Grivaux [41] of a frequently hypercyclic weighted shift on
c0 , mentioned above, has no unimodular eigenvalues, so that the approach chosen in this
section is not always possible. Moreover, their operator does not possess any invariant
Gaussian measure of full support.

Section 9.4. Theorem 9.27 is due to Bayart and Grivaux [40], whose proof uses Ansari’s
theorem. The alternative proof given in Grosse-Erdmann and Peris [185] contains an
error; in fact, Example 9.28 contradicts Theorem 1.4 in that paper. The proof given
here is due to Grosse-Erdmann and Peris [186].
Theorem 9.33 provides a new common approach to Theorems 9.35 and Theorem 9.36
that were previously obtained by Bayart and Matheron [44] and by Conejero, Müller
and Peris [110], respectively. The remainder of the section, including Theorem 9.39
and Corollary 9.41, is due to Shkarin [287]. Grivaux [174] has recently shown that the
necessary spectral conditions of Theorems 5.6 and 9.39 actually characterize spectra of
frequently hypercyclic operators on Hilbert spaces.

Theorem 9.43. Let K ⊂ C be a nonempty compact set. There exists a frequently hy-
percyclic operator T on a complex Hilbert space such that σ(T ) = K if and only if K
has no isolated points and each of its connected components meets the unit circle.

Further interesting results on frequent hypercyclicity include the facts that every op-
erator on an infinite-dimensional complex separable Hilbert space is the sum of two
frequently hypercyclic operators (Bayart and Grivaux [40]) and that every infinite-
dimensional complex Fréchet space with an unconditional basis supports a frequently
hypercyclic and chaotic operator (De la Rosa, Frerick, Grivaux, and Peris [127]).
Many questions concerning frequently hypercyclic operators remain open. For exam-
ple (see Bayart and Grivaux [40]), whether the frequent hypercyclicity of an operator T
is inherited by its direct sum T ⊕ T ; and whether it is inherited by its inverse T −1 , if it
exists.

Exercises. Exercise 9.1.1 is taken from Bayart and Grivaux [40], Exercises 9.1.4, 9.2.1
and 9.2.6 from Bonilla and Grosse-Erdmann [87], and Exercises 9.2.4 and 9.2.5 from
Grosse-Erdmann and Peris [185]. For Exercise 9.3.1 we refer to Bayart and Grivaux
[39], for Exercise 9.3.3 to Bonilla and Grosse-Erdmann [86]. Exercise 9.3.4 is taken from
Blasco, Bonilla and Grosse-Erdmann [76] who also show that, in the converse direction,
given any function φ : R+ → R+ with limr→∞ φ(r) = 0 there r
is no entire function f
that is frequently hypercyclic for D such that |f (z)| ≤ φ(r) re1/4 for |z| = r sufficiently
large. Exercise 9.3.5 is taken from Bayart and Grivaux [40], Exercise 9.4.2 from Costakis
and Ruzsa [122], Exercise 9.4.4 from Grosse-Erdmann and Peris [185], and Exercise 9.4.5
from Shkarin [287].
Chapter 10
Hypercyclic subspaces

By the Herrero–Bourdon theorem, every hypercyclic operator admits a dense

subspace in which every nonzero vector is hypercyclic. In this chapter we ask
for a large space of hypercyclic vectors in a different sense: does a given hy-
percyclic operator admit a closed and infinite-dimensional subspace in which
every nonzero vector is hypercyclic? We will see that in many cases the an-
swer is positive. However, in contrast to the problem of dense subspaces of
hypercyclic vectors, we will also find counterexamples. This makes the notion
considered here particularly interesting.
We point out that the two meanings of largeness are almost incompatible.
The only dense and closed subspace is the whole space itself. And while
there do exist operators for which every nonzero vector is hypercyclic, Read’s
operator being one of them, such examples are extremely rare and difficult to
construct. Indeed, the commonly known hypercyclic operators all have large
supplies of non-hypercyclic vectors.
Since several of the proofs in this chapter are technically more demanding
in the Fréchet space setting, we present them first in Banach (or Hilbert)
spaces; in the final section we then supply the proofs in the general case.

10.1 Operators with hypercyclic subspaces

In the present context the following terminology has been generally accepted.

Definition 10.1. Let T be an operator on a separable Fréchet space X. Then

a hypercyclic subspace for T is an infinite-dimensional closed subspace M of
X so that every nonzero vector in M is hypercyclic for T .

In this section we derive a useful sufficient condition for an operator to

have a hypercyclic subspace.
Thus let T be a hypercyclic operator on X. Which additional assumption
do we need in order for T to support a hypercyclic subspace? The following

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 267

DOI 10.1007/978-1-4471-2170-1_10, © Springer-Verlag London Limited 2011
268 10 Hypercyclic subspaces

idea is at least plausible. Suppose that there is an infinite-dimensional closed

subspace M0 of X so that T n x → 0 for every x ∈ M0 , and suppose that
M0 has a basis (en )n . By the density of hypercyclic vectors we can find,
arbitrarily closely to each en , a hypercyclic vector fn . Now, if the fn are
sufficiently close to the en then one may hope that the fn , in turn, form a
basis in their closed linear span M , and that each nonzero vector in M is also
hypercyclic. This strategy does indeed work, at least in certain spaces.

Theorem 10.2 (Montes). Let X be a separable Fréchet space with a con-

tinuous norm, and let T be an operator on X. Suppose that there exists an
increasing sequence (nk )k of positive integers such that
(i) T satisfies the Hypercyclicity Criterion for (nk )k ,
(ii) there exists an infinite-dimensional closed subspace M0 of X such that
T nk x → 0 for all x ∈ M0 .
Then T has a hypercyclic subspace.

We have already met the notion of a continuous norm; see Lemma 8.7.

Example 10.3. Of course, the norm in any Banach space is continuous. The
Fréchet space H(C) of entire functions has a continuous norm. One may take,
for example, f  = sup|z|≤1 |f (z)|, f ∈ H(C). By contrast, the space ω = KN
of all sequences (see Example 2.2) has no continuous norm: it would have to
be dominated by a multiple of some seminorm pn (x) = sup1≤k≤n |xk | (see
Exercise 2.1.7) and would thus assign the value 0 to some nonzero vector.

Remark 10.4. While Montes’ theorem is usually applied in the form stated
above, we will see that it remains true if condition (ii) is replaced by the
following weaker condition:
(ii ) there exists an infinite-dimensional closed subspace M0 of X such that
(T nk x)k converges for all x ∈ M0 .

We will give two proofs of this result. The first one is straightforward but
slightly technical, and it uses the notion of a basic sequence. The second
one, which we describe in the next section, provides a very interesting link
between the hypercyclicity of an operator T and the hypercyclicity of the
corresponding left-multiplication operator LT : S → T S.
Let us turn to the first proof. As is the case for several results in this
chapter, the proof of Montes’ theorem is considerably more transparent when
X is a Banach space. We will therefore restrict ourselves here to these spaces.
The proof in the general case will be given in Section 10.5.

Definition 10.5. A (finite or infinite) sequence (en )n in a Banach space X

is called a basic sequence if it is a basis in its closed linear span.

We have that (en )n is a basic sequence if and only if every vector x ∈

M := span{en ; n ≥ 1} has a unique representation
10.1 Operators with hypercyclic subspaces 269

x= an en
n≥1

with scalars an ∈ K, n ≥ 1. One then defines the coefficient functionals

e∗n : M → K, x → an

for n ≥ 1. Since M is a Banach space these functionals are continuous; see

Appendix A. By e∗n , n ≥ 1, we denote the norm of e∗n on M .
We will now show that a sufficiently small perturbation of a basic sequence
remains basic.
Lemma 10.6. Let (en )n be a basic sequence in a Banach space X with coef-
ficient functionals e∗n , n ≥ 1. If (fn )n is a sequence in X with

e∗n  en − fn  = δ < 1,
n≥1

then (fn )n is also a basic

 sequence. 
Moreover, a series n≥1 an en converges if and only if n≥1 an fn does,
and
1
fn∗  ≤ e∗ , n ≥ 1.
1−δ n
Proof. Let M denote the closed linear span of the en . We then consider the
operator  
T : M → X, x = an en → a n fn ,
n≥1 n≥1
 ∗
that is, T x = n≥1 x, en fn . In order to see that this is well defined, we note
that for 1 ≤ n ≤ m,
m   m   m 
     
 a k fk  ≤  ak (fk − ek ) +  ak ek 
k=n k=n k=n

m  m 
∗  
≤ ek fk − ek  x +  ak ek , (10.1)
k=n k=n

and the right-hand side tends to 0 as m ≥ n → ∞. Thus T x exists and, by

(10.1),
T x ≤ δx + x = (1 + δ)x, x ∈ M, (10.2)
so that T is continuous.
Moreover, the second triangle inequality tells us that, for any x ∈ M ,
     
     
T x =  a n fn  ≥  an en  −  an (en − fn )
n≥1 n≥1 n≥1
 
≥ x − e∗n  fn − en  x
n≥1
270 10 Hypercyclic subspaces

≥ x − δx = (1 − δ)x. (10.3)

Now, (10.2) and (10.3) imply that T is an isomorphism of M onto M  :=


ran T . Thus, (fn )n = (T e
n )n , is a basis in M and therefore 
a basic sequence
in X. Moreover, a series n≥1 an en converges if and  only if n≥1 an fn does.
Finally, we have by (10.3) that for any y = T x = n≥1 an fn ∈ M  and n ≥ 1

1
|y, fn∗ | = |an | = |x, e∗n | ≤ e∗n x ≤ e∗ y,
1−δ n

which implies that fn∗  ≤ 1 ∗

1−δ en . 

As a second tool we need the following fundamental result.

Theorem 10.7 (Mazur). Every infinite-dimensional Banach space contains

a basic sequence.

We prepare its proof by a lemma.

Lemma 10.8. Let X be an infinite-dimensional Banach space, E a finite-

dimensional subspace of X and ε > 0. Then there exists an x ∈ X with
x = 1 such that, for any λ ∈ K and y ∈ E,

y ≤ (1 + ε)λx + y.

Since we will obtain a more general result later (see Lemma 10.39), we
omit the proof.

Proof of Theorem
 10.7. First, let (εn )n be a sequence of positive numbers
such that ∞n=1 (1 + εn ) ≤ 2. Choose e1 ∈ X with e1  = 1. By Lemma 10.8
we can inductively construct vectors e2 , e3 , . . . of norm 1 such that, for all
n ≥ 1 and a1 , . . . , an+1 ∈ K,
   n+1 

n
  
 ak ek  ≤ (1 + εn ) ak ek .
k=1 k=1

We then have for any a1 , . . . , an ∈ K and any m ≤ n that

   m+1 

m
  
 ak ek  ≤ (1 + εm ) ak e k 
k=1 k=1
n 
 
≤ (1 + εm )(1 + εm+1 ) . . . (1 + εn−1 ) ak ek 
k=1
n 
 
≤ 2 ak ek , (10.4)
k=1

which also implies that, for k ≤ n,

10.1 Operators with hypercyclic subspaces 271

   k−1   
 
k n
   
|ak | = ak ek  ≤  aj ej  +  aj ej  ≤ 4 aj ej . (10.5)
j=1 j=1 j=1

We claim that (en )n is a basic sequence in X. Indeed, let M be the closed

linear span of these vectors, and let x ∈ M . Then there are vectors


Nν
xν = aν,k ek
k=1

that converge to x as ν → ∞. We let aν,k = 0 for k > Nν . It then follows

from (10.5) that, for k ≥ 1,

|aν,k − aμ,k | ≤ 4xν − xμ  → 0

as μ, ν → ∞, so that ak := limν→∞ aν,k exists for all k ≥ 1. Using (10.4) we

deduce that, for any n ≥ 1,
n 
n  n 
n 
   
 aν,k ek − ak ek  = lim  aν,k ek − aμ,k ek 
μ→∞
k=1 k=1 k=1 k=1
≤ 2 lim sup xν − xμ  = 2xν − x
μ→∞

and therefore, whenever n ≥ Nν ,

 
n  Nν 
n 
   
x − ak ek  ≤ x − xν  +  aν,k ek − ak ek  ≤ 3xν − x,
k=1 k=1 k=1
∞
so that x = k=1 ak ek . This proves the claim.


We are now in a position to prove Theorem 10.2 in a special case.
Proof of Theorem 10.2 for Banach spaces. The proof will be divided into
three steps. To simplify notation we perform the proof in the case when
(nk )k is the full sequence of positive integers. The general case follows in
exactly the same way.

Step 1. By Mazur’s theorem, since M0 is an infinite-dimensional Banach space

in the induced topology, there exists a basic sequence (en )n in X of elements
from M0 .

Step 2. We show that (en )n can be perturbed into a basic sequence (fn )n of
hypercyclic vectors. To this end, let Kn = max(1, e∗n ). Further let X0 and
Y0 be the dense subsets of X appearing in the Hypercyclicity Criterion and
let (yn )n be a sequence in Y0 that is dense in X. We claim that there then
exist vectors xj,k ∈ X0 and positive integers n(j, k) such that (n(j, k))k≥1 is
increasing for each j ≥ 1 and such that, for all j, k, j  , k ≥ 1,
272 10 Hypercyclic subspaces

1
xj,k  ≤ , (10.6)
2j+k+1 Kj
1
T n(j,k) xj,k − yk  ≤ k , (10.7)
2
n(j  ,k ) 1
T xj,k  ≤ j+k+k if (j  , k  ) = (j, k), (10.8)
2 Kj
1
T n(j,k) ej  ≤ k . (10.9)
2
This is easily seen by induction with respect to the strict order < on N×N that
is defined by (1, 1) < (1, 2) < (2, 1) < (1, 3) < (2, 2) < (3, 1) < (1, 4) < . . ..
The existence of xj,k and n(j, k) satisfying (10.6), (10.7) and (10.9) follows
from the assumptions on Y0 in the Hypercyclicity Criterion and the fact that
ej ∈ M0 ; note that xj,k ∈ X0 can be achieved because X0 is dense in X.
Condition (10.8) can be rewritten as
1 
,k ) 1
T n(j,k) xj  ,k  ≤ and T n(j xj,k  ≤
2j  +k +k Kj  2j+k+k Kj

if (j  , k ) < (j, k); this condition can therefore also be ensured when we use
the fact that each xj  ,k belongs to X0 and that T is continuous.
We now define, for any j ≥ 1,
∞

fj = ej + xj,k .
k=1

By (10.6) these series converge, and simple calculations using (10.6)–(10.9)

show that we have, for j, k, j  ≥ 1,
1
ej − fj  ≤ , (10.10)
2j+1 Kj
3
T n(j,k) fj − yk  ≤
, (10.11)
2k
 1
T n(j ,k) (ej − fj ) ≤ j+k if j  = j. (10.12)
2 Kj

By (10.10) we have that ∞ ∗
j=1 ej  ej − fj  < 1, whence (fn )n is a basic
sequence by Lemma 10.6.

Step 3. We claim that the closed linear span M of the fn , n ≥ 1, is the

desired hypercyclic subspace. Thus let z ∈ M , z = 0; we need to show that z
is hypercyclic for T . Since (fn )n is a basis of M we can write
∞

z= a j fj .
j=1
10.1 Operators with hypercyclic subspaces 273

Since z = 0, one of the coefficients, am say, must be nonzero. We can assume

that am = 1 because any nonzero multiple of a hypercyclic vector remains
hypercyclic. Moreover, by Lemma 10.6,

w := aj ej
j=m

exists, and w belongs to M0 .

We then have by (10.11) and (10.12) that

T n(m,k) z − yk 
 
 
≤ T n(m,k) fm − yk  +  aj T n(m,k) (fj − ej ) + T n(m,k) w
j=m
3 
≤ k + |aj |T n(m,k) (fj − ej ) + T n(m,k) w
2
j=m
3  1
≤ k + e∗j w j+k + T n(m,k) w
2 2 Kj
j=m
3 1
≤ + k w + T n(m,k) w → 0
2k 2
as k → ∞ because w ∈ M0 . Since the yk , k ≥ 1, are dense in X, z is
hypercyclic for T .


Remark 10.9. In order to see that Montes’ theorem remains true under con-
dition (ii ) of Remark 10.4 one need only weaken (10.9) to
1
T n(j,k) ej − vj  ≤
2k
with certain vj ∈ X, j ≥ 1, so that (10.11) has to be replaced by

3
T n(j,k) fj − vj − yk  ≤ .
2k
In addition, there is some v ∈ X such that T n w → v. We can then conclude
as before that T n(m,k) z − vm − v − yk  → 0 as k → ∞. Since the vectors
yk + vm + v, k ≥ 1, form a dense set in X, we have again that z is hypercyclic.
Our first application of Montes’ theorem treats weighted shifts.
Example 10.10. Let w = (wn )n be a bounded weight sequence and Bw the
corresponding weighted backward shift on one of the spaces X = p , 1 ≤
p < ∞, or X = c0 ; see Section 4.1. By Example 4.9(a), Bw satisfies the
Hypercyclicity Criterion if (and only if)


n
sup |wν | = ∞.
n≥1 ν=1
274 10 Hypercyclic subspaces

We claim that Bw has a hypercyclic subspace if, in addition,


n
sup lim sup |wν+k | < ∞.
n≥1 k→∞ ν=1

For example, if wn = ( n+1 n ) , n ≥ 1, α > 0, then Bw has a hypercyclic

α

subspace. In particular, the backward shift on the Bergman space does; see
Example 4.9(a).
In order to show the claim, let C denote the latter supremum, and suppose
that the Hypercyclicity Criterion is satisfied for the sequence (nk )k . Setting
m1 = n1 , we can find a subsequence (mk )k of (nk )k such that, for k ≥ 1,


mk
|wν+μ | ≤ C + 1 for μ ≥ mk+1 − mk .
ν=1

This implies that, for j > k ≥ 1,


mj
|wν | ≤ C + 1. (10.13)
ν=mj −mk +1

If en , n ≥ 1, denote the unit sequences then


∞

M0 := ak emk ; (ak )k ∈ X
k=1
∞
is an infinite-dimensional closed subspace of X. For any x = k=1 ak emk ∈
M0 we have, using (10.13), that
∞    
mj  
   
T mk
x =  aj T mk e mj  =  aj wν emj −mk 
j=1 j>k ν=mj −mk +1
 
 
≤ (C + 1) aj emj −mk  → 0
j>k

as k → ∞. Montes’ theorem then implies the claim.

For the application to Birkhoff’s operators we derive a simple but useful

consequence from Montes’ theorem.

Corollary 10.11. Let X be a separable Fréchet space with a continuous

norm, and let T be an operator on X that satisfies the Hypercyclicity Crite-
rion. If ker(λI − T ) is infinite-dimensional for some λ with |λ| < 1, then T
has a hypercyclic subspace.

Proof. By continuity of T , M0 := ker(λI − T ) is a closed subspace, and for

any x ∈ M0 we have that T n x = λn x → 0.

10.1 Operators with hypercyclic subspaces 275

We remark that this result remains true for any λ with |λ| ≤ 1; see Exercise
10.1.3.

Example 10.12. (Birkhoff’s operators) The Birkhoff operators Ta : f →

f ( · + a), a = 0, on H(C) have hypercyclic subspaces. This follows from the
fact that the linearly independent entire functions
1  2kπi 
z→ exp z , k ∈ Z,
2z/a a
belong to ker( 12 I − Ta ). The remaining hypotheses of Corollary 10.11 are also
satisfied; see Examples 2.38 and 10.3.
More generally, if ϕ is an entire function of exponential type that is not
a polynomial then the operator ϕ(D) on H(C) defined in Proposition 4.19
has a hypercyclic subspace. Indeed, by the proof of Theorem 4.21 we have
that ϕ(D)eμ = ϕ(μ)eμ , where eμ (z) = eμz . The big Picard theorem tells us
that for any λ ∈ C with at most one exception the equation ϕ(μ) = λ has
infinitely many solutions. This implies that for such a λ, ker(λI − ϕ(D)) is
infinite-dimensional, and we can choose |λ| < 1.

As our last example in this section we consider MacLane’s operator.

Example 10.13. (MacLane’s operator) The operator of differentiation on

H(C) has a hypercyclic subspace. Since D satisfies the Hypercyclicity Crite-
rion for the full sequence (see Example 3.7), it suffices to exhibit an infinite-
dimensional closed subspace M0 of H(C) on which suitable powers of D tend
to 0.
To start with, let us note that, for any n ≥ 1, there is some Cn > 0 such
that
xn ≤ 2x for all x ≥ Cn . (10.14)
We choose a strictly increasing sequence of positive integers (nk )k with n1 ≥ 1
such that, for all k ≥ 1,
nk+1 ≥ Cnk .
For j ≥ k + 1 we then have that nj ≥ nk+1 ≥ Cnk and hence, by (10.14),

nnj k ≤ 2nj for j ≥ k + 1. (10.15)

Now let M0 be the closed subspace of H(C) of all entire functions f of the
form
∞

f (z) = ak z nk −1 .
k=1

We claim that
Dnk f → 0 in H(C) as k → ∞.
Indeed, let R ≥ 1. Then we have that
276 10 Hypercyclic subspaces
 
∞ 
 
sup |Dnk f (z)| = sup  aj Dnk z nj −1 
|z|≤R |z|≤R j=k+1

∞
≤ |aj | (nj − 1) · · · (nj − nk )Rnj −nk −1
j=k+1
∞
≤ |aj | nnj k Rnj
j=k+1
∞
≤ |aj |(2R)nj → 0
j=k+1

as k → ∞, where in the last inequality we have used (10.15). This had to be

shown.

Further examples will be considered in Section 10.4.

10.2 Hypercyclic left-multiplication operators

In this section we outline an alternative approach to hypercyclic subspaces

that is also interesting in its own right. In order to emphasize ideas over tech-
nical details we first restrict ourselves again to Banach spaces and postpone
the general proofs to Section 10.5.
Thus, let T be an operator on a separable Banach space X. The principal
idea, a priori unrelated to the topic of this chapter, is to study the dynamical
properties of the operator T by way of its induced left-multiplication operator

LT : L(X) → L(X), LT S = T S,

on the space L(X) of operators on X. The inequality

LT S = T S ≤ ST , S ∈ L(X),

shows that LT indeed defines an operator on L(X) when the latter is endowed
with the operator norm topology.
This new approach of studying the operator T is quite promising. For if
we can describe properties of an orbit

{(LT )n S ; n ≥ 0} = {T n S ; n ≥ 0} ⊂ L(X) (10.16)

of some fixed operator S under LT then we can deduce properties of the orbit

{T n Sx ; n ≥ 0} ⊂ X (10.17)
10.2 Hypercyclic left-multiplication operators 277

of Sx under T , for an arbitrary x ∈ X. In particular if the orbit (10.16) is

dense in L(X) then any orbit (10.17) is dense in X, unless for x = 0, giving
us automatically a large set of hypercyclic vectors for T .
This striking idea, however, meets with an unexpected obstacle: for an
orbit in L(X) to be dense the space of operators would have to be separable,
which is hardly ever the case under the operator norm topology; see Exercise
10.2.1 (but also Exercise 10.2.8). The only hope is to weaken the topology of
L(X). In Section 8.3 the strong operator topology (SOT) has already come
to our rescue in another case where the operator norm topology proved to be
too strong. It will save us again because L(X) is indeed separable under this
weaker topology. Here, and in the sequel, topological notions will be prefixed
by SOT if they refer to the strong operator topology.

Proposition 10.14. Let X be a separable Banach space. Then L(X) is SOT-

separable.

For the proof we need a related result concerning the dual X ∗ of X. Note
that the dual of a separable Banach space need not be separable under the
usual operator norm; a simple example is provided by the sequence space
1 whose dual is ∞ . Again we obtain separability under a weaker topology.
Just like the strong operator topology on L(X), the weak-∗-topology on X ∗
is defined as the topology of pointwise convergence on X. A base of neigh-
bourhoods of an element x∗ ∈ X ∗ is given by

Ux1 ,...,xn (x∗ , ε) = {y ∗ ∈ X ∗ ; |xk , x∗  − xk , y ∗ | < ε for k = 1, . . . , n},

where x1 , . . . , xn , n ≥ 1, is an arbitrary collection of linearly independent

vectors of X and ε > 0. Thus, a sequence (more generally, a net) (x∗α )α is
weak-∗-convergent to x∗ if and only if, for every x ∈ X, x, x∗α  → x, x∗ .
For later use we consider general normed spaces.

Lemma 10.15. Let X be a separable normed space. Then X ∗ is weak-∗-

separable.

Proof. Let E be a countable dense subset of X. We then consider the set

of all continuous linear functionals on X of the following form: there are
linearly independent elements e1 , . . . , em ∈ E and numbers q1 , . . . , qm in Q
or in Q + iQ such that
 m
x∗ = qj e∗j ,
j=1

where e∗1 , . . . , e∗m are the coordinate functionals corresponding to the (finite)
basic sequence (ej )j=1,...,m (see Definition 10.5 and the subsequent discus-
sion); we assume that the e∗j are extended continuously to all of X, with
preservation of the norm.
We show that this countable set of functionals is weak-∗-dense in X ∗ . To
see this, let x∗ ∈ X ∗ , let x1 , . . . , xm , m ≥ 1, be linearly independent vectors
278 10 Hypercyclic subspaces

of X and ε > 0. Let x∗1 , . . . , x∗m be the coordinate functionals corresponding

to the (finite) basic sequence (xj )j=1,...,m , extended to all of X, and let M =
maxj=1,...,m x∗j . We fix rational numbers q1 , . . . , qm ∈ Q(+iQ) such that,
for j = 1, . . . , m,
ε
|xj , x∗  − qj | < .
2
Let e1 , . . . , em be vectors from E such that, for j = 1, . . . , m,
 1 ε 
ej − xj  ≤ min , m .
2mM 4M ( k=1 |qk | + 1)

By Lemma 10.6, applied to the Banach space span{x1 , . . . , xm , e1 , . . . , em },

we have that (ej )j=1,...,m is a basic sequence in X with e∗j  ≤ 2M , j =
1, . . . , m.
Let
m
y∗ = qj e∗j
j=1
m ∗
be the corresponding continuous linear functional. Since j=1 qj ek , ej  = qk ,
we have for k = 1, . . . , m that
m 
m 
 
|xk , x∗  − xk , y ∗ | ≤ |xk , x∗  − qk | +  qj ek , e∗j  − qj xk , e∗j 
j=1 j=1

ε 
m
< + |qj |e∗j ek − xk 
2 j=1

ε ε  m
≤ + 2M m |qj | ≤ ε,
2 4M ( j=1 |qj | + 1) j=1

as had to be shown.

Proof of Proposition 10.14. Let E be a countable dense subset of X. By

Lemma 10.15 there is a countable weak-∗-dense subset Φ of X ∗ . We claim
that the countable set

m

F = FΦ,E =  · , yj∗ ej ; yj∗ ∈ Φ, ej ∈ E, 1 ≤ j ≤ m

j=1

of operators on X is SOT-dense in L(X). Indeed, let T ∈ L(X), let

x1 , . . . , xm , m ≥ 1, be linearly independent vectors of X and ε > 0. De-
note by x∗1 , . . . , x∗m the coordinate functionals corresponding to the basic
sequence (xj )j=1,...,m , extended to X. Then, by density of E and weak-∗-
density of Φ, there are vectors e1 , . . . , em ∈ E and y1∗ , . . . , ym
∗
∈ Φ such that
for j, k = 1, . . . , m,
10.2 Hypercyclic left-multiplication operators 279
ε ε
T xk − ek  < , (xj , x∗k  − xj , yk∗ ) ek  < .
2 2m
m ∗
For S = j=1  · , yj ej ∈ F we then have that, for k = 1, . . . , m,

T xk − Sxk  ≤ T xk − ek  + (1 − xk , yk∗ )ek  + xk , yj∗ ej  < ε;
j=k

note that xj , x∗k  = δj,k . This had to be shown.

The SOT-separability of L(X) allows us to speak of SOT-dense orbits.

However, since L(X) is not a Fréchet space under the strong operator topol-
ogy, we are no longer in the framework of Chapter 2. Thus we need to define
the following.

Definition 10.16. Let T be an operator on a Banach space X. Then the left-

multiplication operator LT is called SOT-hypercyclic if there is some operator
S ∈ L(X) whose orbit {T n S ; n ≥ 0} under LT is SOT-dense in L(X); such
an operator S is then called SOT-hypercyclic for LT .

Remark 10.17. Under the strong operator topology, L(X) is a topological

vector space and LT is an operator on L(X). The dynamical properties of
operators on topological vector spaces will be treated in detail in Chapter 12.

The following explains our interest in SOT-hypercyclicity of left-multipli-

cation operators.

Proposition 10.18. Let T be an operator on a separable Banach space X.

If S ∈ L(X) is SOT-hypercyclic for LT then Sx is a hypercyclic vector for T
for any x ∈ X, x = 0.

Proof. By the Hahn–Banach theorem there is some x∗ ∈ X ∗ with x, x∗  = 1.

Let y ∈ X be arbitrary. Then the operator R =  · , x∗ y maps x to y. Since S
is SOT-hypercyclic there is an increasing sequence (nk )k of positive integers
such that T nk Sx = (LT )nk Sx → Rx = y. Hence Sx is hypercyclic for T .


Remark 10.19. As an immediate consequence we have the following: if T itself

is SOT-hypercyclic for LT , or if some surjective operator S is hypercyclic for
LT then every nonzero vector x ∈ X is hypercyclic for T .

The main result of this section provides us with a close link between the
hypercyclicity properties of the operator T and those of the operator LT .

Theorem 10.20. Let T be an operator on a separable Banach space X. Then

the following assertions are equivalent:
(i) T satisfies the Hypercyclicity Criterion;
(ii) LT is SOT-hypercyclic.
280 10 Hypercyclic subspaces

In order to prepare the proof we fix a countable weak-∗-dense subset Φ of

X ∗ . Then we define

K = KΦ = span{ · , y ∗ x ; y ∗ ∈ Φ, x ∈ X} ⊂ L(X),

where the closure is taken in the operator norm topology. Moreover, since

 · , y ∗ x ≤ y ∗ x, y ∗ ∈ X ∗ , x ∈ X, (10.18)

the countable set F of operators considered in the proof of Proposition 10.14

is dense in K. Consequently, K is a separable Banach space. Finally, LT
obviously maps F, and therefore also K, into K; thus LT defines an operator
on K.
Proof of Theorem 10.20. (i)=⇒(ii). We first note that, by the Hypercyclicity
Criterion, any n-fold direct sum T ⊕ . . . ⊕ T of T is topologically transitive;
see Remark 3.13(b).
We want to show that LT is topologically transitive as an operator on K.
Thus, let U and V be nonempty open subsets of K. By definition there are
m ≥ 1, y1∗ , . . . , ym
∗
∈ Φ as well as x1 , . . . , xm , z1 , . . . , zm ∈ X such that


m 
m
 · , yj∗ xj ∈ U and  · , yj∗ zj ∈ V ;
j=1 j=1

note that we can have the same elements y1∗ , . . . , ym ∗

∈ Φ here by taking some
xj and zj as zero. By (10.18) there are open neighbourhoods U1 × · · · × Um
of (x1 , . . . , xm ) and V1 × · · · × Vm of (z1 , . . . , zm ) such that, if (x1 , . . . , xm ) ∈
U1 × · · · × Um and (z1 , . . . , zm
) ∈ V1 × · · · × Vm , then


m 
m
 · , yj∗ xj ∈ U and  · , yj∗ zj ∈ V.
j=1 j=1

Applying the topological transitivity of the m-fold direct sum T ⊕ . . . ⊕

T we find some (x1 , . . . , xm ) ∈ U1 × · · · × Um and some n ≥ 1 such that
(T n x1 , . . . , T n xm ) ∈ V1 × · · · × Vm . This implies that


m 
m
S :=  · , yj∗ xj ∈ U and (LT )n S = T n S =  · , yj∗ T n xj ∈ V,
j=1 j=1

which proves our initial claim.

By the Birkhoff transitivity theorem, the operator LT is therefore hyper-
cyclic on K, that is, there is an operator S ∈ K whose orbit {(LT )n S ; n ≥ 0}
is dense in K with respect to the operator norm topology. But then the orbit
is also SOT-dense in K; and since K contains F , which is SOT-dense in L(X)
by the proof of Proposition 10.14, the orbit of S under LT is SOT-dense in
L(X). This shows that LT is SOT-hypercyclic on L(X).
10.2 Hypercyclic left-multiplication operators 281

(ii)=⇒(i). By the Bès–Peris theorem it suffices to show that T ⊕ T is

hypercyclic. Let S be an operator that is SOT-hypercyclic for LT . Since
X cannot be one-dimensional, there are two linearly independent vectors
x1 , x2 ∈ X, and we denote by x∗1 , x∗2 the coordinate functionals corresponding
to the basic sequence (x1 , x2 ), extended to X. Let (y1 , y2 ) ∈ X ⊕ X be
arbitrary. Then the operator R :=  · , x∗1 y1 +  · , x∗2 y2 maps x1 to y1 , x2
to y2 . By SOT-hypercyclicity of S there is an increasing sequence (nk )k of
positive integers such that T nk Sx1 → Rx1 = y1 and T nk Sx2 → Rx2 = y2 .
This shows that (Sx1 , Sx2 ) is hypercyclic for T ⊕ T .


The theorem allows us to deduce Montes’ theorem for Banach spaces in

the case when (nk ) is the full sequence (n). However, when we do not want to
restrict the sequence (nk )k we need to generalize Theorem 10.20 to sequences
(Tn )n≥0 of operators on a Banach space X.

Definition 10.21. Let (Tn )n≥0 be a sequence of operators on a Banach space

X. Then the sequence (LTn )n of left-multiplication operators is called SOT-
hypercyclic if there is some operator S ∈ L(X) whose orbit {Tn S ; n ≥ 0}
under (LTn )n is SOT-dense in L(X); such an operator S is then called SOT-
hypercyclic for (LTn )n .

Then exactly as in the proof of Theorem 10.20, using Theorems 3.24 and
3.25, we obtain the following generalization.

Theorem 10.22. Let (Tn )n be a commuting sequence of operators on a sep-

arable Banach space X with dim X ≥ 2. Then the following assertions are
equivalent:
(i) (Tn )n satisfies the Hypercyclicity Criterion;
(ii) (LTn )n is SOT-hypercyclic.

From this we can deduce Montes’ theorem for Banach spaces.

Second proof of Theorem 10.2 for Banach spaces. We suppose that T satis-
fies the Hypercyclicity Criterion for an increasing sequence (nk )k of positive
integers. Then also the sequence (T nk )k satisfies the Hypercyclicity Criterion;
see Theorem 3.24. By the preceding theorem, the sequence (LT nk )k admits
an SOT-hypercyclic vector S ∈ L(X). Since any nonzero multiple of an SOT-
hypercyclic vector is SOT-hypercyclic we may assume that S = 12 . Then,
for any x ∈ X,
1
(I + S)x ≥ x − Sx ≥ x.
2
Hence I + S is an injective operator, and it defines an isomorphism of X onto
its range ran(I + S).
Now let M0 be an infinite-dimensional closed subspace of X so that
T nk x → 0 for all x ∈ M0 . By the above, M := (I + S)M0 is an infinite-
dimensional closed subspace of X. We claim that every nonzero vector x ∈ M
282 10 Hypercyclic subspaces

is hypercyclic for T . Indeed, there is some y ∈ M0 , y = 0, such that

x = (I + S)y. Then

T nk x = T nk y + T nk Sy, k ≥ 1,

is a dense sequence in X because T nk y → 0 and Sy is hypercyclic for (T nk )k ;

Proposition 10.18 also holds for sequences of operators.


For the deduction of Montes’ theorem in the Fréchet space setting we refer
to Section 10.5.

10.3 Operators without hypercyclic subspaces

In this section we want to show that not all hypercyclic operators have hy-
percyclic subspaces. To this end we derive a useful necessary condition for
the existence of a hypercyclic subspace in which the behaviour of the iterates
of T on subspaces of finite codimension plays an essential role. We recall that
a subspace M of a vector space X is said to be of finite codimension if there
is a finite-dimensional subspace E of X such that X = M + E. We collect
here some rather obvious properties, whose proof we leave to the reader; see
Exercise 10.3.1.
Lemma 10.23. Let X be a vector space over K. Then we have the following:
(i) a subspace M is of finite codimension if and only if there is some n ≥ 1
and a linear map u : X → Kn such that M = ker u; n
(ii) if M1 , . . . , Mn are subspaces of finite codimension then so is k=1 Mk ;
(iii) if M is a subspace of finite codimension and T : X → X is a linear
map then T −1 (M ) is of finite codimension;
(iv) if M is an infinite-dimensional subspace and L is a subspace of finite
codimension then M ∩ L is infinite-dimensional.
We turn to the announced condition that prevents an operator from having
hypercyclic subspaces. We first state a version for Banach spaces; later on we
will present a more technical result for general Fréchet spaces.
Theorem 10.24. Let T be an operator on a Banach space X. Suppose that
there are subspaces Mn ⊂ X, n ≥ 1, of finite codimension and positive num-
bers Cn , n ≥ 1, with Cn → ∞ as n → ∞ such that

T n x ≥ Cn x for any x ∈ Mn , n ≥ 1.

Then T does not possess any hypercyclic subspace.

Since the proof is more transparent for Hilbert spaces we consider this
case first; the general case will be treated later within the context of Fréchet
spaces.
10.3 Operators without hypercyclic subspaces 283

Proof of Theorem 10.24 for Hilbert spaces. Thus, let X be a Hilbert space.
We assume that T satisfies the hypotheses of the theorem and that M is
an infinite-dimensional closed subspace of X; we need to show then that M
contains a vector x = 0 that is not hypercyclic for T . First, it follows from
the assumption that there is an increasing sequence (kn )n of positive integers
such that, for n ≥ 2,

Cj ≥ n3 for kn−1 < j ≤ kn .

We claim that there are points xn ∈ X and closed subspaces Ln ⊂ X of finite

codimension such that, for n ≥ 1 and 1 ≤ j ≤ kn ,
1
x n ∈ M ∩ Mj , xn  = , (10.19)
n2
T j x1 , . . . , T j xn−1 ⊥ Ln (n ≥ 2), (10.20)
T xn ∈ L1 ∩ . . . ∩ Ln .
j
(10.21)

We show existence by induction on n ≥ 1. For n = 1 one can take any vector

x1 ∈ M ∩ M1 ∩ . . . ∩ Mk1 of norm 1 and L1 = X. Now, if x1 , . . . , xn−1 and
L1 , . . . , Ln−1 , n ≥ 2, have been constructed then we let

En = span{T j x1 , . . . , T j xn−1 ; 1 ≤ j ≤ kn } and Ln = En⊥ .

Clearly, Ln is a closed subspace of finite codimension such that (10.20) holds.

Moreover, by Lemma 10.23 we can find a nonzero vector
 
xn ∈ M ∩ Mj ∩ T −j (L1 ∩ . . . ∩ Ln ) ,
1≤j≤kn

which we can assume to have norm n12 ; hence also (10.19) and (10.21) are
satisfied.
Having constructed the xn and Ln , n ≥ 1, we set
∞

x= xn ;
n=1

by (10.19), the series converges in X, and x belongs to M because M is a

closed subspace. Now let n ≥ 2 and kn−1 < j ≤ kn . Then


n−1 ∞

j j j
T x= T xν + T xn + T j xν .
ν=1 ν=n+1

It follows from (10.21) that, for ν ≥ n + 1,

T j xν ∈ Ln+1 .
284 10 Hypercyclic subspaces
∞
Since Ln+1 is closed, ν=n+1 T j xν belongs to Ln+1 and hence is orthogonal
n
to ν=1 T j xν by (10.20). By Pythagoras’ theorem we have for any orthogonal
vectors y, z ∈ X that y + z = (y2 + z2 )1/2 ≥ y. In our present
situation we therefore have that
 n−1 
 j 
T j x ≥  T xν + T j xn .
ν=1

n−1
Moreover, by (10.20) and (10.21) we have that ν=1 T j xν is orthogonal to
T j xn . Again by Pythagoras’ theorem this yields that

T j x ≥ T j xn ,

and therefore, using the fact that xn ∈ Mj and kn−1 < j ≤ kn ,

n3
T j x ≥ Cj xn  ≥ = n.
n2
Thus T j x → ∞ as j → ∞, which shows that x ∈ M , x = 0, is not a
hypercyclic vector for T , as desired.

For Banach spaces this proof breaks down because we no longer have a
notion of orthogonality. In Section 10.5 we will prove the following general-
ization of Theorem 10.24, which then also contains the Banach space case.

Theorem 10.25. Let T be an operator on a Fréchet space X with defining

increasing sequence (pn )n of seminorms. Suppose that there are subspaces
Mn ⊂ X, n ≥ 1, of finite codimension, positive numbers Cn , n ≥ 1, with
Cn → ∞ as n → ∞, and some N ≥ 1 such that
(i) pN (x) > 0 for every hypercyclic vector x for T ;
(ii) pN (T n x) ≥ Cn pn (x) for any x ∈ Mn , n ≥ 1.
Then T does not possess any hypercyclic subspace.

We remark that the additional assumption that pN (x) > 0 for any hy-
percyclic vector is automatically satisfied if the defining seminorms can be
chosen to be norms, that is, if the space admits a continuous norm.
Our first application uses the criterion in the case of Banach spaces.

Example 10.26. (Rolewicz’s operators) We consider Rolewicz’s operators

T = λB, |λ| > 1, on X = p , 1 ≤ p < ∞, or c0 . Theorem 10.24 implies
immediately that they do not have hypercyclic subspaces. Indeed, for the
subspaces Mn = {(xk )k ; x1 = x2 = . . . = xn = 0}, n ≥ 1, of finite
codimension we have that T n (xk )k  = λn (xk+n )k  = |λ|n (xk )k  for all
(xk )k ∈ Mn . Since |λ|n → ∞ as n → ∞ the claim follows.
More generally, let T = Bw be a weighted backward shift on one of these
spaces X, with w = (wn )n a bounded weight sequence; see Section 4.1. In
this case we have for n ≥ 1 that
10.3 Operators without hypercyclic subspaces 285

(Bw )n (xk )k = (w2 · · · wn+1 xn+1 , w3 · · · wn+2 xn+2 , w4 · · · wn+3 xn+3 , . . .);

hence, taking the same spaces Mn as above, we obtain that

 
n 
(Bw )n x ≥ inf |wν+k | x for x ∈ Mn .
k≥1
ν=1

Consequently, by Theorem 10.24, if


n
lim inf |wν+k | = ∞,
n→∞ k≥1
ν=1

then Bw has no hypercyclic subspace.

We note that, strangely enough, this condition implies that Bw is mixing;
see Example 4.9(a).

Theorem 10.25 can also be applied in a different direction. In Montes’ the-

orem, which gives a sufficient condition for an operator to have a hypercyclic
subspace, we have assumed that the space possesses a continuous norm. This
is not just a technical requirement, as the following example shows.

Example 10.27. There is an even mixing operator on a Fréchet space (with-

out a continuous norm) that satisfies condition (ii) of Theorem 10.2 for the
full sequence (n) but that does not have a hypercyclic subspace. Indeed, we
consider the sequence space over Z,
∞


X = (xn )n∈Z ; |xn | < ∞ ,

n=0

endowed with the seminorms

∞

pn (x) = |xk |, n = 1, 2, 3, . . . .
k=−n

Clearly, X is isomorphic to KN ⊕ 1 and therefore a separable Fréchet space.

On X we consider the multiple T = 2B of the bilateral backward shift, that
is,
T (xn )n∈Z = 2(xn+1 )n∈Z .
Then Kitai’s criterion, applied to the finite sequences and the inverse of T as
S, shows that T is mixing (or one might apply Theorem 4.13). Moreover,

M0 = {(xn )n∈Z ; xn = 0 for n ≥ 0}

is an infinite-dimensional closed subspace of X on which T n tends pointwise

to 0 as n → ∞. Thus T satisfies conditions (i) and (ii) of Theorem 10.2 for the
full sequence (n). But we will now show that T has no hypercyclic subspace.
286 10 Hypercyclic subspaces

For this we will apply Theorem 10.25. First, it is clear that every hyper-
cyclic sequence x satisfies p1 (x) > 0. For Mn we consider the subspaces of
finite codimension

Mn = {(xk )k∈Z ; xk = 0 for −n ≤ k ≤ n}, n ≥ 1.

We then have for x ∈ Mn , n ≥ 1, that

p1 (T n x) = 2n pn (x).

Therefore, T has no hypercyclic subspace.

10.4 Further operators with hypercyclic subspaces

We first show that operators with hypercyclic subspaces exist in a large class
of Fréchet spaces.

Theorem 10.28. Let X be an infinite-dimensional separable Fréchet space

with a continuous norm. Then there exists a mixing operator on X that pos-
sesses a hypercyclic subspace.

Proof. By Theorem 8.13 there is a mixing operator T on X with an infinite-

dimensional closed subspace M0 of fixed points. Thus, T satisfies the Hy-
percyclicity Criterion, and every x ∈ M0 trivially has a converging orbit. It
follows from Montes’ theorem, taking into account Remark 10.4, that T has
a hypercyclic subspace.


The usual techniques of Chapter 8 then also show that, in the setting of the
theorem, the mixing operators with a hypercyclic subspace are SOT-dense in
the space of operators on X; see Exercise 10.4.1.
Montes’ theorem provides us with a powerful tool for verifying that an
operator has hypercyclic subspaces. But it leaves us with the task of finding
an infinite-dimensional closed subspace on which (sub)orbits converge to 0. In
many cases, coming up with such a subspace is by no means easy or obvious.
In the remainder of this section we address this problem in two particular
cases. We will see that once again the notion of a basic sequence is crucial.
Our first result is in the spirit of Theorem 10.24.

Theorem 10.29. Let T be an operator on a separable Banach space X. Sup-

pose that there exists an increasing sequence (nk )k of positive integers such
that
(i) T satisfies the Hypercyclicity Criterion for (nk )k ,
(ii) there is a decreasing sequence (Mk )k of infinite-dimensional closed sub-
spaces of X such that supk≥1 T nk |Mk  < ∞.
Then T has a hypercyclic subspace.
10.4 Further operators with hypercyclic subspaces 287

Proof. For simplicity we will perform the proof in the case when (nk )k is the
full sequence; the general case follows in the same way.
By (i) there is a dense subset X0 of X such that, for all x ∈ X0 , T n x → 0 as
n → ∞. By (ii) there is some C > 0 such that T n |Mn  ≤ C for n ≥ 1. And by
a slight strengthening of Mazur’s theorem (see Exercise 10.4.4), there exists
a basic sequence (en )n in X with en ∈ Mn for n ≥ 1; let Kn = max(1, e∗n ).
By density of X0 and continuity of T we can then find fn ∈ X0 , n ≥ 1,
such that
1
T n ej − T n fj  < j , j ≥ 1, n = 0, 1, . . . , j. (10.22)
2 Kj
Since the fn belong to X0 one can construct inductively an increasing se-
quence (nk )k of positive integers such that, for all k ≥ 1,
1
T nk fj  ≤ , j = 1, . . . , nk−1 , (10.23)
2j+k Kj

where n0 = 1. Since also (enk )k is a basic sequence and

∞
 ∞
 1
e∗nk enk − fnk  < Knk ≤ 1,
2nk K nk
k=1 k=1

we have by Lemma 10.6 that (fnk )k is a basic sequence. Let M0 be its closed
linear span.
nk
∞the powers T of T tend pointwise to 
We now claim that 0 on M0 . Indeed,
∞
let x ∈ M0 , x = k=1 ak fnk . By Lemma 10.6, also z := k=1 ak enk con-
verges. Moreover, since the subspaces Mk are closed and (Mk )k is decreasing,
we have that
∞
aj enj ∈ Mnk , k ≥ 1. (10.24)
j=k

We then deduce from (10.22), (10.23) and (10.24) that


k−1  
∞   ∞
 
T nk x ≤ |aj |T nk fnj  + T nk aj enj  + |aj |T nk (fnj − enj )
j=1 j=k j=k


k−1 
∞   ∞
1   1
≤ e∗nj z nj +k + C aj enj  + e∗nj z
j=1
2 K nj 2nj Knj
j=k j=k
∞ 
1   2
≤ k z + C  aj enj  + n z → 0
2 2 k
j=k

as k → ∞, which proves the claim.

An application of Montes’ theorem finishes the proof.


As an application one obtains a more direct verification of Example 10.10;
see Exercise 10.4.2.
288 10 Hypercyclic subspaces

We use the theorem here to obtain a new, large class of operators with
hypercyclic subspaces. For this we need the following.
Lemma 10.30. Let K be a compact operator on a Banach space X. Then
there is a closed subspace M of finite codimension such that K|M  ≤ 12 .
Proof. Since the image under K of the closed unit sphere is relatively compact
in X there are finitely many points xk ∈ X with xk  = 1 for k = 1, . . . , N ,
such that, for any x ∈ X with x = 1 there is some k with Kxk −Kx ≤ 14 .
By the Hahn–Banach theorem (see Appendix A), there are continuous linear
functionals yk∗ , k = 1, . . . , N , on X such that yk∗ (Kxk ) = Kxk  and yk∗  =
1. By Lemma 10.23,

N
M= ker(yk∗ ◦ K)
k=1

is a closed subspace of finite codimension.

Now let x ∈ M with x = 1, and let k be such that Kxk − Kx ≤ 14 .
Then
1
Kxk  = yk∗ (Kxk ) = yk∗ (Kxk − Kx) ≤ yk∗  Kxk − Kx ≤
4
and therefore
1
Kx ≤ Kx − Kxk  + Kxk  ≤ .
2
This implies that, for all x ∈ M , Kx ≤ 12 x, and the claim follows.


We can now present the announced class of operators with hypercyclic
subspaces.
Corollary 10.31. Let T be an operator on a separable Banach space of the
form
T = U + K,
where U  ≤ 1 and K is compact. If T satisfies the Hypercyclicity Criterion
then it has a hypercyclic subspace.
Proof. For n ≥ 1 we have that

T n = (U + K)n = U n + Kn

with compact operators Kn ; see Exercise 5.2.3. It then follows from Lemma
10.30 that there are closed subspaces Mn of finite codimension such that
Kn |Mn  ≤ 12 , hence

3
T n |Mn  = (U n + Kn )|Mn  ≤ .
2
Since X must be infinite dimensional, an application of Theorem 10.29, us-
ing the decreasing sequence (M1 ∩ . . . ∩ Mn )n of infinite-dimensional closed
subspaces, yields the result.


10.4 Further operators with hypercyclic subspaces 289

Example 10.32. (a) Let X be one of the spaces p , 1 ≤ p < ∞, or c0 , w =

(wn )n a weight sequence with limn→∞ wn = 0, and Bw the corresponding
weighted shift. Then T = I + Bw has a hypercyclic subspace. This follows
from the corollary with Corollary 8.3 and Exercise 5.2.10.
(b) In the Banach space setting the mixing operator constructed in the
proof of Theorem 8.9 is a compact perturbation of the identity; see also
Exercise 8.2.2. By the corollary it therefore has a hypercyclic subspace. This
provides a new proof of Theorem 10.28 for Banach spaces.
We finish this section with another example where basic sequences help us
to verify the subspace condition in Montes’ theorem. By Example 10.12 we
know already that Birkhoff’s translation operators, f → f (· + a) on H(C),
have hypercyclic subspaces. The proof was based on the largeness of some
eigenspace. We will give here another proof in a more general setting. As in
Section 4.3 we consider general composition operators Cϕ : f → f ◦ ϕ on
arbitrary domains Ω in C with automorphisms ϕ of Ω.
Proposition 10.33. Let Ω be a domain in C. Then every hypercyclic com-
position operator Cϕ , ϕ ∈ Aut(Ω), has a hypercyclic subspace.
Proof. Let Cϕ be hypercyclic on H(Ω). By Proposition 4.31, Ω is either sim-
ply connected or infinitely connected. By Remark 4.35 and Theorem 3.25,
there is a subsequence (mn )n such that Cϕ satisfies the Hypercyclicity Cri-
terion for the sequence (mn )n and such that (ϕmn )n is a run-away sequence.
By Montes’ theorem it then suffices to construct an infinite-dimensional
closed subspace M0 of H(Ω) and a subsequence (nk )k of (mn )n such that
(Cϕ )nk g → 0 as k → ∞, for all g ∈ M0 .
Since Ω is conformally equivalent to any domain rΩ + z0 , r > 0, z0 ∈ C,
we can assume by Proposition 4.25 that Ω contains the closed unit disk D.
Let (Kn )n be an exhaustion of Ω by compact sets with K1 = D. Then we
construct inductively Ω-convex compact subsets Lk and a subsequence (nk )k
of (mn )n such that, for k ≥ 1,

ϕnk (Lk ) ∩ Lk = ∅, ϕnk (Lk ) ∪ Lk is Ω-convex, (10.25)

Kk ∪ Lk−1 ∪ ϕnk−1 (Lk−1 ) ⊂ Lk , (10.26)

where L0 = ∅ and n0 = 0; in the simply connected case this follows imme-

diately from the run-away property, in the infinitely connected case we use
Lemma 4.33.
Since the unit circle T is contained in Ω, we can consider functions f ∈
H(Ω) as functions in the Hilbert space
 2π

L2 (T) = f : T → C ; |f (eit )|2 dt < ∞

0
 2π
with norm f 2 = ( 2π1
0
|f (eit )|2 dt)1/2 ; the embedding is obviously con-
tinuous. In particular, the functions en (z) = z n , n ≥ 1, form an orthonormal
290 10 Hypercyclic subspaces

system in L2 (T) and are therefore a basic sequence there; the norms of the
corresponding coefficient functionals are 1.
By (10.25) we can apply Runge’s theorem to obtain, for any j ≥ 1, func-
tions fj,k ∈ H(Ω) such that

1 1
sup |ej (z) − fj,1 (z)| < , sup |fj,1 (z)| < (10.27)
z∈L1 2j+2 z∈ϕn1 (L 1)
2j+2

and, for k ≥ 2,
1 1
sup |fj,k−1 (z) − fj,k (z)| < , sup |fj,k (z)| < . (10.28)
z∈Lk 2j+k+1 z∈ϕnk (Lk ) 2j+k+1

Since, by (10.26), Lk ⊂ Lk+1 for all k ≥ 1, (10.28) implies that

∞
 ∞

fj := fj,1 + (fj,k − fj,k−1 ) = fj,m + (fj,k − fj,k−1 )
k=2 k=m+1

converges uniformly on any Lm and hence on any Km , m ≥ 1, so that fj ∈

H(Ω), j ≥ 1.
By (10.26) we also have that ϕnk (Lk ) ⊂ Lk+1 ⊂ Lk+2 ⊂ . . . for all k ≥ 1.
By (10.27) and (10.28) we then have, for any m ≥ 1,
∞

sup |fj (z)| ≤ sup |fj,m (z)| + sup |fj,k (z) − fj,k−1 (z)|
z∈ϕnm (Lm ) z∈ϕnm (Lm ) k=m+1
z∈Lk
∞
 1 1
≤ = . (10.29)
2j+k+1 2j+m
k=m

In the same way we have, using that T ⊂ D ⊂ L1 ,

∞
 1 1
ej − fj 2 ≤ sup |ej (z) − fj (z)| ≤ = j+1
z∈L1 2j+k+1 2
k=1
∞
and therefore j=1 ej − fj 2 ≤ 12 . It follows from Lemma 10.6 that (fn )n
is a basic sequence in L2 (T) with fn∗  ≤ 2.
We now claim that the closed linear span M0 of the fj in H(Ω) is the de-
sired subspace; note that it is infinite dimensional because the fj are linearly
independent. Thus, let g ∈ M0 ; then there are linear combinations


Nν
gν = aν,j fj
j=1

that converge to g in H(Ω) and then also in L2 (T). Fix m ≥ 1; then we have
for k ≥ m that Lm ⊂ Lk and therefore, using (10.29),
10.5 The Fréchet space setting 291


Nν
sup |(Cϕ )nk g(z)| ≤ sup |(Cϕ )nk (g − gν )(z)| + |aν,j | sup |fj (ϕnk (z))|
z∈Lm z∈Lm j=1 z∈Lk
∞
 1
≤ sup |(g − gν )(ϕnk (z))| + fj∗  gν 2
z∈Lm 2j+k
j=1
2
≤ sup |(g − gν )(w)| + gν 2 .
w∈ϕnk (Lm ) 2k

Letting ν → ∞ we obtain that, for any m ≥ 1 and k ≥ m,

2
sup |(Cϕ )nk g(z)| ≤ g2 .
z∈Lm 2k

Since Lm contains Km we obtain that (Cϕ )nk g → 0 in H(Ω) for any g ∈ M0 ,

which had to be shown.


10.5 The Fréchet space setting

In this section we discuss the extensions of the main results of the previous
sections to Fréchet spaces.

Montes’ theorem. We start with the proof of Montes’ theorem.

Proof of Theorem 10.2. As in the Banach space case we assume that (nk )k
is the full sequence of positive integers.

Step 1. In order to construct a basic sequence (en )n again, we first need to

specify a Banach space in which it lives. Thus, let | · | denote a continuous
norm on X. Then (X, | · |) is a normed space and as such has a comple-
 | · |), which is a Banach space that contains X densely. Let M
tion (X, 0 be
 
the closure of M0 in X. By Mazur’s theorem, (M0 , | · |) contains a basic
sequence (en )n . Approximating each en sufficiently closely from inside M0 ,
Lemma 10.6 tells us that we can find a sequence (en )n from M0 that is a
basic sequence in X.

Step 2. We will again perturb (en )n into a basic sequence (fn )n of hypercyclic
vectors. To do this, let Kn = max(1, |e∗n |), n ≥ 1, where |e∗n | is the norm
of the coefficient functionals e∗n on the space (X,  | · |). Furthermore, let
 ·  denote an F-norm defining the topology of X; see Section 2.1. As in
the Banach space case there is then a dense sequence (yn )n in X, vectors
xj,k ∈ X and positive integers n(j, k) such that (n(j, k))k≥1 is increasing for
each j ≥ 1 and such that, for all j, k, j  , k ≥ 1,
292 10 Hypercyclic subspaces

1
max(xj,k , |xj,k |) ≤ , (10.30)
2j+k+1 Kj
1
T n(j,k) xj,k − yk  ≤ k ,
2
n(j  ,k ) 1
T xj,k  ≤ j+k+k if (j  , k ) = (j, k),
2 Kj
1
T n(j,k) ej  ≤ k .
2
 Defining
In (10.30) we have used the continuity of the inclusion of X into X.
∞

fj = ej + xj,k , j ≥ 1,
k=1

we obtain as before that these series converge in X and that, for j, k, j  ≥ 1,

1
max(ej − fj , |ej − fj |) ≤ , (10.31)
2j+1 Kj
3
T n(j,k) fj − yk  ≤ k , (10.32)
2
n(j  ,k) 1
T (ej − fj ) ≤ j+k if j  = j. (10.33)
2 Kj
∞
By (10.31), |e∗j | |ej − fj | ≤
j=1
1
2 < 1. Lemma 10.6 implies that (fn )n
is a basic sequence in X  with

|fj∗ | ≤ 2Kj , j ≥ 1. (10.34)

Step 3. We now define the desired hypercyclic subspace. We denote by M the

closed linear span (in X) of the fn , n ≥ 1. Note that (fn )n is not necessarily
a basis in M , but since the fn are linearly independent, M is an infinite-
dimensional closed subspace of X. It remains to show that every nonzero
vector in M is hypercyclic for T .
We first derive an estimate for finite linear combinations

N
z= a j fj
j=1

with am = 1 for some m. To this end we define


w := aj ej .
j≤N, j=m

Let k ≥ 1. Then by (2.3), (10.32), (10.33) and (10.34) we find that

10.5 The Fréchet space setting 293
 n(m,k) 
T z − yk 
  
 
≤ T n(m,k) fm − yk  +  aj T n(m,k) (fj − ej ) + T n(m,k) w
j≤N,j=m
3 
≤ + (|aj | + 1)T n(m,k) (fj − ej ) + T n(m,k) w
2k
j≤N, j=m
∞
4 1
≤ + |e∗j | |w| + T n(m,k) w
2k j=1
2j+k K j

4 1
≤ + k |w| + T n(m,k) w. (10.35)
2k 2
Now let z ∈ M , z = 0. We want to show that z is hypercyclic for T . As z
 we have a
also belongs to the closed linear span of the fn when taken in X,
representation
∞
z= a j fj
j=1

 As in the Banach space case we can assume some am

with convergence in X.
to be 1. By assumption there are vectors


Nν
zν := aν,j fj
j=1

converging to z in X. Since we also have convergence in X  and the coefficient

functionals are continuous, we have that aν,j → aj for all j ≥ 1. In particular,
aν,m → am , and we can assume without loss of generality that aν,m = 1 for
all ν. Let us also consider the vectors

wν := aν,j ej ∈ M0 .
j≤Nν , j=m

Setting aν,j = 0 for j > Nν , we find for ν, μ ≥ 1,


(wν − zν ) − (wμ − zμ ) ≤ (aν,j − aμ,j )(ej − fj ).
j=m

Since (aν,j )ν converges for all j ≥ 1 and since, by (2.3), (10.31) and (10.34),

(aν,j − aμ,j )(ej − fj ) ≤ (|aν,j − aμ,j | + 1)ej − fj 

1
≤ |fj∗ | |zν − zμ | ej − fj  +
2j+1
|zν − zμ | 1 C
≤ + j+1 ≤ j
2j 2 2
294 10 Hypercyclic subspaces

with some C > 0, the dominated convergence theorem implies that (wν −zν )ν
is a Cauchy sequence in X; hence (wν )ν converges to some vector w, which
necessarily belongs to M0 .
By our previous argument, zν and wν satisfy (10.35) for any ν ≥ 1. Letting
ν → ∞ we have by continuity that, for any k ≥ 1,
4 1
T n(m,k) z − yk  ≤ k
+ k |w| + T n(m,k) w.
2 2
Since w ∈ M0 we obtain that T n(m,k) z − yk  → 0 as k → ∞, which implies
that z is hypercyclic.

Remark 10.34. Similar modifications to those in the Banach space case yield
that condition (ii ) of Remark 10.4 suffices in Montes’ theorem.

Left-multiplication operators. Next we turn to the generalization of The-

orems 10.20 and 10.22 to separable Fréchet spaces X with a continuous norm
| · |. A crucial role in the proof will be played by operators from X under
the norm topology induced by | · | to X under its original topology. For
brevity we will denote (X, | · |) by X|·| .

Theorem 10.35. Let (Tn )n be a commuting sequence of operators on a sep-

arable Fréchet space X with a continuous norm | · | and dim X ≥ 2. Then
the following assertions are equivalent:
(i) (Tn )n satisfies the Hypercyclicity Criterion;
(ii) (LTn )n is SOT-hypercyclic.

Note that the strong operator topology on L(X) is defined in Section 8.3,
and the definition of SOT-hypercyclicity is the same as in Definitions 10.16
and 10.21. Proposition 10.18 extends as well, with unchanged proof.
We endow the space L(X|·| , X) of operators X|·| → X with its natu-
ral operator topology. More precisely, if (pn )n is an increasing sequence of
seminorms defining the topology of X then we set

Sn = sup pn (Sx), n ≥ 1. (10.36)

|x|≤1

It is a standard exercise to show that this defines a Fréchet space topology

on L(X|·| , X) and to show that the left-multiplication operator LT induced
by an operator T on X is an operator on L(X|·| , X).
It is now not difficult to generalize the various preliminary results of Sec-
tion 10.2 to the present setting. First, when we apply Lemma 10.15 to the
separable normed space X|·| we obtain a countable set Φ of continuous lin-
ear functionals on X|·| that is weak-∗-dense in (X|·| )∗ . Note that every
continuous linear functional on X|·| is also a continuous linear functional on
X in its original topology.
10.5 The Fréchet space setting 295

Lemma 10.36. Let X be a separable Fréchet space with a continuous norm

| · |. Let Φ be a countable set of continuous linear functionals on X|·| that
is weak-∗-dense in (X|·| )∗ , and let E be a dense subset of X. We define


m

F = FΦ,E =  · , yj∗ ej ; yj∗ ∈ Φ, ej ∈ E, 1 ≤ j ≤ m ,

j=1

K = KΦ = span{ · , y ∗ x ; y ∗ ∈ Φ, x ∈ X},

where the closure is taken in L(X|·| , X). Then:

(i) F is SOT-dense in L(X);
(ii) L(X) is SOT-separable;
(iii) every S ∈ K is a compact operator S : X|·| → X;
(iv) K is a separable Fréchet space, and F is dense in K;
(v) for any operator T on X, LT is an operator on K.

Proof. (i) Suppose that  ·  is an F-norm that defines the topology of X. Let
T ∈ L(X), let x1 , . . . , xm , m ≥ 1, be linearly independent vectors of X, and
ε > 0. The coordinate functionals x∗1 , . . . , x∗m corresponding to the basic se-
quence (xj )j=1,...,m are continuous with respect to | · | on span{x1 , . . . , xm }
(because all norms on a finite-dimensional space are equivalent) and hence
have continuous linear extensions to X|·| . One then concludes exactly as
in the proof of Proposition 10.14 that there is some S ∈ F such that
T xk − Sxk  < ε for k = 1, . . . , m.
(ii) Since X is separable we can choose E to be countable. Then F is
countable, and the assertion follows from (i).
(iii) The proof of this assertion can be given as in Exercise 5.2.4.
(iv) Since, for any y ∗ ∈ Φ and x ∈ X,

 · , y ∗ xn = pn (x) sup |y, y ∗ |, n ≥ 1, (10.37)

|y|≤1

and since E is dense in X, F is dense in K. Moreover, if E is countable then

so is F, which makes K separable.
(v) Clearly, LT is an operator on L(X|·| , X) that maps F into K; by (iv)
it then also maps K into K.


Lemma 10.37. Under the assumptions of Theorem 10.35 and Lemma 10.36,
the following assertions are equivalent:
(i) (Tn )n satisfies the Hypercyclicity Criterion;
(ii) (LTn )n is hypercyclic on K;
(iii) (LTn )n is SOT-hypercyclic.
Moreover, if S ∈ K is hypercyclic for (LTn )n on K, then it is SOT-hypercyclic
for (LTn )n on L(X).

Proof. The proof follows the same lines as that of Theorem 10.20, using
(10.37), Theorem 3.24 and Theorem 3.25.

296 10 Hypercyclic subspaces

This lemma proves Theorem 10.35.

Lemma 10.38. Under the assumptions of Theorem 10.35 and Lemma 10.36,
let S ∈ K be hypercyclic for (LTn )n on K. Then there is some λ = 0 such
that |λSx| ≤ 12 |x| for all x ∈ X. If M0 is an infinite-dimensional closed
subspace of X such that Tn x → 0 for all x ∈ M0 then (I + λS)M0 is a
hypercyclic subspace for (Tn )n .

Proof. Since S : X|·| → X is continuous, S is also a continuous operator on

X|·| . Hence there is some λ = 0 such that |λSx| ≤ 12 |x| for all x ∈ X,
which implies that
1
|(I + λS)x| ≥ |x| − |λSx| ≥ |x|, (10.38)
2
so that I + λS is an injective operator.
Moreover, under the usual topology, I + λS maps closed sets onto closed
sets. Indeed, let A ⊂ X be closed, yn = (I + λS)xn , xn ∈ A, and yn → y in
X. Since (yn )n also converges in X|·| , it follows from (10.38) that (xn )n is
bounded in X|·| . By compactness of S, a subsequence of (Sxn )n converges
in X. Since xn = yn − λSxn , a subsequence of (xn )n converges, to x ∈ X say.
Then x ∈ A and y = (I + λS)x, so that (I + λS)(A) is closed.
Now let M0 be an infinite-dimensional closed subspace of X so that Tn x →
0 for all x ∈ M0 . Then M := (I + λS)M0 is an infinite-dimensional closed
subspace of X. Let x = (I + λS)y with y ∈ M0 , y = 0. Then

Tn x = Tn y + λTn Sy, n ≥ 1,

is a dense sequence in X because Sy is hypercyclic for (Tn )n by the extension

of Proposition 10.18. Hence M is a hypercyclic subspace for (Tn )n .


The previous two lemmas, applied to the sequence (T nk )k , imply Montes’

theorem in full generality.

Operators without hypercyclic subspaces. We give here the proof of

Theorem 10.25, which will also prove Theorem 10.24 for general Banach
spaces. For this we first need a substitute for the notion of orthogonality
in Hilbert spaces. We refer to Exercise 2.1.7 for a characterization of contin-
uous seminorms.

Lemma 10.39. Let X be a Fréchet space, E a finite-dimensional subspace

of X, p a continuous seminorm on X and ε > 0. Then there exists a closed
subspace L of finite codimension such that, for any x ∈ L and y ∈ E,
 p(x)p(y) 
p(x + y) ≥ max , . (10.39)
2+ε 1+ε
10.5 The Fréchet space setting 297

Fig. 10.1 A substitute for orthogonality

Proof. We first note that ker p is a subspace of X. There is then a subspace

F of E such that E = (ker p ∩ E) ⊕ F algebraically. Clearly, p is a norm on
F . On the finite-dimensional space F all norms define the same topology; see
Appendix A. Hence the unit sphere with respect to p in F , being bounded
and closed, is compact. Consequently there are finitely many points yk ∈ F
with p(yk ) = 1, k = 1, . . . , N , such that, for any y ∈ F with p(y) = 1, there
is some k with p(yk − y) ≤ 1+ε ε
. More generally, if y ∈ E with p(y) = 1 then
y = u + v with p(u) = 0 and v ∈ F , p(v) = 1. Hence there is some k with
p(yk − y) = p(yk − v) ≤ 1+ε ε
.
By the Hahn–Banach theorem (see Appendix A), there are linear function-
als yk∗ , k = 1, . . . , N , on X such that yk , yk∗  = p(yk ) = 1 and |y, yk∗ | ≤ p(y)
for all y ∈ X. By continuity of p, these functionals are continuous. By Lemma
10.23,
N
L= ker yk∗
k=1

is a closed subspace of finite codimension.

Now let x ∈ L and y ∈ E; see Figure 10.1. If p(y) = 0 then p(x + y) =
p(x) and the claim is trivially true. Hence let p(y) = 0. Then y  := y/p(y)
satisfies p(y  ) = 1, so that there is some k with p(yk − y  ) ≤ 1+ε
ε
. Setting

x = x/p(y) ∈ L we have that
ε ε 1
p(x +y  ) ≥ p(yk +x )−p(yk −y  ) ≥ |yk +x , yk∗ |− = 1− = ,
1+ε 1+ε 1+ε
and therefore
p(y)
p(x + y) ≥ .
1+ε
In addition,

(2 + ε)p(x + y) = p(x + y) + (1 + ε)p(x + y) ≥ p(x + y) + p(y) ≥ p(x).

298 10 Hypercyclic subspaces

This proves the claim. 

The similarities between this proof and that of Lemma 10.30 are evident.
In Exercise 10.5.4 we ask the reader to prove a common generalization of the
two lemmas in the case of Banach spaces.
We can now prove the generalization of Theorem 10.24 to Fréchet spaces.
Proof of Theorem 10.25. Suppose that M is a hypercyclic subspace for T .
Then, exactly as in the proof of Theorem 10.24, there is an increasing se-
quence (kn )n of positive integers such that Cj ≥ n3 for kn−1 < j ≤ kn , n ≥ 2,
and with Lemma 10.39 we can construct points xn in X, finite-dimensional
subspaces En of X and corresponding closed subspaces Ln of finite codimen-
sion satisfying (10.39) for all x ∈ Ln , y ∈ En , with p = pN , such that, for
n ≥ 1 and 1 ≤ j ≤ kn ,
1
pn (xn ) = (n ≥ N ), xn ∈ M ∩ Mj , (10.40)
n2
T j x1 , . . . , T j xn−1 ∈ En (n ≥ 2),
T j xn ∈ L1 ∩ . . . ∩ Ln ,

where we set E1 = {0} and L1 = X. Note that pn (xn ) = n12 , n ≥ N , is

possible because, by assumption, pn (x) ≥ pN (x) > 0 for every hypercyclic
vector x. Defining
∞
x= xn ,
n=1

it follows from (10.40) that the series converges in X, and we have that x ∈ M .
Once more following the proof of Theorem 10.24, replacing orthogonality by
condition (10.39), we find that, for n ≥ N and kn−1 < j ≤ kn ,

1  n  pN (T j xn ) Cj pj (xn )
pN (T j x) ≥ pN T j xν ≥ ≥
1+ε ν=1
(1 + ε)(2 + ε) (1 + ε)(2 + ε)
n3 pn (xn ) n
≥ = →∞
(1 + ε)(2 + ε) (1 + ε)(2 + ε)

as n → ∞; we have used here that xn ∈ Mj and j ≥ kn−1 + 1 ≥ n. This is a

contradiction because every nonzero element of M is hypercyclic.

Exercises

Exercise 10.1.1. Show that the following operators have hypercyclic subspaces:
(i) the operator T on C0 (R+ ) given by T f (x) = λf (x + a), a > 0, λ > 1 (see Exercise
2.2.1);
(ii) the operator T on Lpv (R+ ), 1 ≤ p < ∞, given by T f (x) = f (x + 1), where v is an
admissible weight function with lim inf x→∞ v(x) = 0; see Exercise 2.2.4.
Exercises 299

Compare with Examples 10.10 and 10.26.

Exercise 10.1.2. Show that a Fréchet space has a continuous norm if and only if its
topology can be defined by an increasing sequence (pn )n of norms.

Exercise 10.1.3. Let X be a separable Fréchet space with a continuous norm, and let T
be an operator on X that satisfies the Hypercyclicity Criterion. Show that if ker(λI − T )
is infinite-dimensional for some λ with |λ| ≤ 1, then T has a hypercyclic subspace. (Hint:
Use Remark 10.4 and the León–Müller theorem.)

Exercise 10.1.4. Let B wn be a weighted backward shift on H(C); see Example 4.9(b).
Show that if limn→∞ ( ν=1 |wν |)1/n = ∞ and limn→∞ |wn |1/n = 1 then Bw has a
hypercyclic subspace. See also Exercise 10.3.3.

Exercise 10.1.5. Let T be an operator on a separable Fréchet space. Show the follow-
ing:
(i) T and T n , n ≥ 1, have the same hypercyclic subspaces;
(ii) if T is surjective and M = X is a hypercyclic subspace then Mn := T −n (M ),
n ≥ 0, are hypercyclic subspaces of T ; moreover, Mn = Mk for n = k;
(iii) if T is bijective and M = X is a hypercyclic subspace then Mn := T −n (M ), n ∈ Z,
are hypercyclic subspaces of T ; moreover, Mn = Mk for n = k.
(Hint: Ansari’s theorem.)

Exercise 10.2.1. Let X be a Banach space with a basis (en )n .

(a) Suppose that the basis is unconditional. Show thatL(X) is not  separable in
the operator norm topology. (Hint: Consider the operators a e →
k∈N k k
a e ,
k∈A k k
A ⊂ N.)
(b) Give a short  nis SOT-separable. (Hint: For T ∈ L(X) consider
proof that L(X)
∞
Pn T Pn , where Pn : k=1
ak ek → a e .)
k=1 k k

Exercise 10.2.2. Let X be a Banach space with separable dual X ∗ . Show that X itself
is separable. (Hint: If {x∗n ; n ≥ 1} is dense in X ∗ , choose xn ∈ X with xn = 1,
|x∗n (xn )| ≥ 12 x∗n ; show that span{xn ; n ≥ 1} is dense in X.)

Exercise 10.2.3. Let T be an operator on a separable Banach space X such that LT

is SOT-hypercyclic. Show the following:
(i) the set of SOT-hypercyclic operators for LT is SOT-dense in L(X);
(ii) the set of SOT-hypercyclic operators for LT is not dense in L(X) under the operator
norm topology, unless all nonzero vectors are hypercyclic for T .
(Hint: Generalize Proposition 1.15; see Remark 10.19, Exercise 5.1.4.)

Exercise 10.2.4. Let X be a Banach separable space. Show that an operator T on X

is chaotic if and only if LT is SOT-chaotic, that is, LT is SOT-hypercyclic and has an
SOT-dense set of periodic points in L(X).

Exercise 10.2.5. A Banach space X is said to have the approximation property if, for
every Banach space Y , for every compact operator T : Y → X and for every ε > 0 there
is a finite-rank operator F : Y → X such that T − F < ε; see Exercise 5.2.2. Let X
be a Banach space with the approximation property and K(X) the space of compact
∗
operators on X. Let Φ ⊂ X mbe norm-dense and E ⊂ X be dense. Show that then
the operators of the form k=1

· , yk∗ xk , y1∗ , . . . , ym
∗
∈ Φ, x1 , . . . , xm ∈ E, m ≥ 1,
constitute a dense subset of K(X) under the operator norm topology. Deduce that if
X has the approximation property and X ∗ is separable then K(X) is separable. (Hint:
Exercise 10.2.2.)
300 10 Hypercyclic subspaces

Exercise 10.2.6. Let X be a Banach space with the approximation property such that
X ∗ is separable. By Exercises 5.2.4 and 10.2.5, the space K(X) of compact operators is
a separable Banach space under the operator norm topology. Show that LT : K(X) →
K(X), S → T S, is a well-defined operator on K(X). Show moreover that T satisfies the
Hypercyclicity Criterion if and only if LT is hypercyclic on K(X).

Exercise 10.2.7. Let X be a Banach space and T an operator on X. Then the right-
multiplication operator RT on L(X) is defined as RT S = ST for S ∈ L(X).
(a) Let xj ∈ X, yj ∈ X ∗ , j = 1, . . . , m. Show that RT ( j=1
· , yj∗ xj ) =
m
m
j=1

· , T ∗ yj∗ xj .
(b) Suppose that X and X ∗ are separable. Show that if T ∗ : X ∗ → X ∗ satisfies the
Hypercyclicity Criterion then RT is SOT-hypercyclic, that is, there is some S ∈ L(X)
such that {ST n ; n ≥ 0} is SOT-dense in L(X).

Exercise 10.2.8. By Theorem 8.11 there exists an infinite-dimensional separable Ba-

nach space X on which every operator is of the form T = λI + K with λ ∈ K and
K a compact operator. In fact, the space constructed by Argyros and Haydon has the
approximation property and a separable dual. For such a Banach space X, show the
following:
(i) the space L(X) of operators on X is separable under the operator norm topology;
(ii) the functional u : L(X) → K, λI + K → λ is well defined and continuous;
(iii) the functional u is an eigenvector of (LT )∗ for any operator T on X.
Deduce that, even for this Banach space X with separable operator space L(X), no left-
multiplication operator LT is hypercyclic on L(X) with respect to the operator norm.
(Hint: Exercises 5.2.3, 5.2.4 and 10.2.5.)

Exercise 10.2.9. Verify the contents of Remark 10.4 for Banach spaces using the ap-
proach of this section.

Exercise 10.3.1. Prove Lemma 10.23.

Exercise 10.3.2. Let Bw be a hypercyclic weighted backward shift on one of the spaces
X = p , 1 ≤ p < ∞, or c0 . Show the following:
(i) if lim inf n→∞ |wn | > 1, then Bw has no hypercyclic subspace;
(ii) if (|wn |)n is decreasing, then Bw has a hypercyclic subspace if and only if
limn→∞ |wn | = 1;
(iii) if (|wn |)n is increasing, then Bw is even mixing, but it has no hypercyclic subspace.

Exercise 10.3.3. Let Bw be a weighted backward shift on H(C); see Exercise 10.1.4.
n→∞
Show that if lim |wn |1/n = ∞ then Bw has no hypercyclic subspace. (Hint: The
∞ ∞
seminorms pn ( k=0 ak z k ) = k=0
|ak |nk , n ≥ 1, define the topology of H(C).)

Exercise 10.3.4. Use Example 10.27 to show that the existence of hypercyclic sub-
spaces is not preserved under quasiconjugacies.

Exercise 10.4.1. Using the methods of Section 8.3 show that in every Fréchet space X,
the set of operators having hypercyclic subspaces is either empty or SOT-dense in L(X).
Moreover, if X is an infinite-dimensional separable Fréchet space with a continuous norm
then the set of mixing operators having hypercyclic subspaces is SOT-dense in L(X).

Exercise 10.4.2. Verify Example 10.10 using Theorem 10.29.

Exercise 10.4.3. Let Bw be a hypercyclic weighted backward shift on X = p (Z),

1 ≤ p < ∞, or X = c0 (Z). Suppose that there is some N ∈ Z such that |wn | ≤ 1 for all
n ≤ N . Show that Bw has a hypercyclic subspace.
Sources and comments 301

Exercise 10.4.4. Prove the following strengthening of Mazur’s theorem. Let X be a

Banach space and Mn , n ≥ 1, infinite-dimensional closed subspaces of X. Then there
exists a basic sequence (en )n in X with en ∈ Mn for all n ≥ 1. (Hint: Use Lemma
10.39.)

Exercise 10.4.5. Show the following variant of Corollary 10.31. Let T be an operator
on a separable Banach space of the form T = S + K where ker(λI − S) is infinite
dimensional for some λ with |λ| ≤ 1 and K is compact. If T satisfies the Hypercyclicity
Criterion then it has a hypercyclic subspace.

Exercise 10.5.1. Let X be a separable Fréchet space with a continuous norm and T
an operator on X that satisfies the Hypercyclicity Criterion. Let K be the space defined
in Lemma 10.36.
(a) Show that the set of operators from K with dense range is a dense Gδ -set.
(b) Show that LT has an SOT-hypercyclic vector S that is a dense range operator.
(c) Deduce that there is a dense subspace of X consisting, except for zero, of hyper-
cyclic vectors for T . Compare with the Herrero–Bourdon theorem.

Exercise 10.5.2. Let Y be a separable normed space and X a separable Fréchet space.
The space L(Y, X) of operators from Y to X turns into a Fréchet space under seminorms
as in (10.36). For an operator T on X define the operator LT on L(Y, X) by LT S = T S.
The definitions of the strong operator topology on L(Y, X) and SOT-hypercyclicity of
LT are obvious. Show that the following assertions are equivalent when dim Y ≥ 2:
(i) T satisfies the Hypercyclicity Criterion;
(ii) LT is SOT-hypercyclic.
In that case, LT has an SOT-hypercyclic vector S that defines a compact operator
S : Y → X. Deduce Theorem 10.35 for iterates of an operator.

Exercise 10.5.3. Let T be the operator of Example 10.27. Show that it is even chaotic,
but that the left-multiplication operator LT is not SOT-hypercyclic. Hence, in Theorem
10.35, the assumption of existence of a continuous norm cannot be dropped, even for
iterates of an operator. (Hint: Let S ∈ L(X). If P denotes the canonical projection of
X onto 1 , then show that there exists some N ≥ 1 such that P Se−k = 0 for all k ≥ N
and deduce that Se−N is not hypercyclic for T .)

Exercise 10.5.4. Let X and Y be Banach spaces, T : X → Y an operator, E a subspace

of X such that T |E : E → Y is compact, and ε > 0. Then there is a closed subspace L
of X of finite codimension such that for any x ∈ L and y ∈ E

T x + T y − T y ≥ −ε y ,

and for any x ∈ L ∩ E

T x ≤ ε x .
Deduce Lemma 10.39 (for a Banach space X and p the norm on X) and Lemma 10.30.

Sources and comments

In view of the fact that every hypercyclic operator possesses a dense subspace all of
whose nonzero vectors are hypercyclic, it seemed natural to reserve the term “hyper-
cyclic subspace” for the more interesting case of infinite-dimensional closed subspaces,
as was first suggested by Chan and Taylor [107]. An example of an operator for which
302 10 Hypercyclic subspaces

every nonzero vector is due to Read [266]; see Theorem 2.64.

Section 10.1. For Banach spaces, Montes’ theorem is due to Montes [240]; see also
González, León and Montes [167]. Our proof is taken from Bonilla and Grosse-Erdmann
[88]; for a similar proof see León and Müller [223]. For Fréchet spaces, the result is due
independently to Bonet, Martínez and Peris [84] who use a tensor product technique,
and to Petersson [258] who adapts Montes’ technique. Bès and Conejero [69] have stud-
ied hypercyclic subspaces of operators on a Fréchet space without a continuous norm;
they have shown that every nonconstant polynomial p(B) of the backward shift has a
hypercyclic subspace on ω = KN .
For more on basic sequences in Banach spaces we refer to Diestel [133].
In the Banach space setting, González, León and Montes [167] (see also León and
Montes [221] and León and Müller [223]) have improved Montes’ theorem to a charac-
terization under the assumption that the operator satisfies the Hypercyclicity Criterion,
or equivalently, that it is weakly mixing.

Theorem 10.40 (González–León–Montes). Let T be a weakly mixing operator on a

separable Banach space X. Then the following assertions are equivalent:
(i) T has a hypercyclic subspace;
(ii) there exists an increasing sequence (nk )k of positive integers and an infinite-
dimensional closed subspace M0 of X such that T nk x → 0 for all x ∈ M0 ;
(iii) there exists an increasing sequence (nk )k of positive integers and an infinite-
dimensional closed subspace M1 of X such that supk≥1 T nk |M1 < ∞;
(iv) the essential spectrum σe (T ) of T meets the closed unit disk D.

It is important to note that the sequence (nk )k in (ii) need not be related to the
sequence in the Hypercyclicity Criterion. One way of defining the essential spectrum of
/ σe (T ) if and only if λI − T is a Fredholm operator, that is,
an operator T is that λ ∈
λI − T has finite-dimensional kernel and finite-codimensional closed range.
As an application, González, León and Montes [167], [221] identify the operators with
hypercyclic subspaces among various classes of operators. For example, a hypercyclic
weighted backward shift Bw on 2 has a hypercyclic subspace if and only if
 
n 1/n
lim inf |wν+k | ≤ 1,
n→∞ k≥0
ν=1

while any hypercyclic bilateral weighted backward shift on 2 (Z) has a hypercyclic sub-
space. A hypercyclic adjoint multiplier Mϕ∗ on the Hardy space H 2 , induced by an injec-
tive bounded holomorphic function ϕ on D (see Section 4.4), has a hypercyclic subspace
if and only if the boundary of ϕ(D) meets the closed unit disk. And for any automor-
phism ϕ of D the composition operator Cϕ on H 2 (see Section 4.5) has a hypercyclic
subspace whenever it is hypercyclic; see also Montes [240].
Remark 10.4 is due to Bernal [57]; in the case of Banach spaces it also follows from
the sufficiency of condition (iii) in Theorem 10.40 and the Banach–Steinhaus theorem.
Corollary 10.11 and Example 10.12 are due to Petersson [258]; hypercyclic subspaces for
Birkhoff’s operators were found earlier by Bernal and Montes [65]; see the remarks on
Proposition 10.33 below. The fact that MacLane’s operator has a hypercyclic subspace
was only recently proved by Shkarin [290]; he also notes that its essential spectrum is
empty, so that Theorem 10.40 breaks down for Fréchet spaces.
Frequently hypercyclic subspaces are introduced and studied in Bonilla and Grosse-
Erdmann [88].

Section 10.2. The alternative approach to hypercyclic subspaces via left-multiplication

operators is due to Chan [99], who considered Hilbert spaces. He also introduced the
Sources and comments 303

notion of SOT-hypercyclicity and observed Proposition 10.18. Chan and Taylor [107]
and Montes and Romero [242] extended Chan’s investigation to Banach spaces. In all
three papers, the implication (i)=⇒(ii) in Theorem 10.20 is obtained by a construction.
Martínez and Peris [231] show this implication by a tensor product technique; they also
prove the converse implication as well as Theorem 10.22. The proof of Theorem 10.20
given here follows Martínez and Peris [231] but avoids the language of tensor products;
see also Aron, Bès, León and Peris [14]. Remark 10.19 was made in Montes and Romero
[242].
For more on the weak-∗-topology we refer to Diestel [133].

Section 10.3. Montes [240] not only proved his sufficient condition for the existence
of hypercyclic subspaces, he was also the first to come up with operators without such
a subspace, namely Rolewicz’s operators; see Example 10.26. Theorem 10.24 is due to
León and Müller [223]; its extension to Fréchet spaces, Theorem 10.25, seems to be new.
Example 10.27 provides a new proof of a result by Bonet, Martínez and Peris [84].

Section 10.4. Theorem 10.28 was obtained independently by Bernal [57] and Petersson
[258]; our proof follows that of Bernal. For Banach spaces the result was obtained earlier
by León and Montes [220]. The problem of whether Theorem 10.28 remains true for
all infinite-dimensional separable Fréchet spaces remains open; in the case of X = ω, a
positive answer was given by Bès and Conejero [69], as already mentioned.
Theorem 10.29 is due to León and Müller [223], but it is also implicitly contained in
León and Montes [220]. Corollary 10.31 is due to León and Montes [220] who deduced
Theorem 10.28 for Banach spaces from it. In a related result, Petersson [258] has shown
that any weakly mixing nuclear perturbation of the identity on a separable Fréchet space
with a continuous norm has a hypercyclic subspace; he used it to deduce Theorem 10.28.
The result with which we close this section, Proposition 10.33, is due to Bernal and
Montes [65]; historically, it was, in fact, the first result on hypercyclic subspaces after
Read’s theorem mentioned above.

Section 10.5. The proof of Montes’ theorem via basic sequences in a containing Banach
space, given here, is from Bonilla and Grosse-Erdmann [88].
Theorem 10.35 is due to Bonet, Martínez and Peris [84] who refine the tensor product
technique of Martínez and Peris [231]. The proof given here uses similar ideas without
relying on the theory of tensor products.
For Lemma 10.39 in the case of Banach spaces we refer to Müller [246].

Exercises. Exercises 10.1.3 and 10.1.5 are taken from Petersson [258], Exercise 10.2.3
from Chan [99], Exercise 10.2.4 from Martínez and Peris [231], Exercises 10.2.6 and 10.2.7
from Bonet, Martínez and Peris [84]. As for Exercise 10.2.8, the space constructed by
Argyros and Haydon [12] has the approximation property because it has a basis, and its
dual is separable because it is isomorphic to 1 . For Exercise 10.4.1 we refer to Bès and
Chan [100], [67], [68] and Bernal [57]. Exercises 10.4.5 and 10.5.1 are taken from León
and Müller [223], and Exercises 10.5.2 and 10.5.3 from Bonet, Martínez and Peris [84].
Chapter 11
Common hypercyclic vectors

It is an immediate consequence of the Baire category theorem that whenever

we have two, three, or even countably many hypercyclic operators on a given
Fréchet space then there exists a vector that is simultaneously hypercyclic for
each of these operators; such a vector is called a common hypercyclic vector.
This is a purely structural property and has nothing to do with the nature
of the operators under consideration.
However, the problem comes to life again when we allow uncountable fami-
lies of operators. We show in this chapter that even some natural uncountable
families do not admit common hypercyclic vectors. The main result will pro-
vide a sufficient condition for the existence of common hypercyclic vectors.
This criterion will then be applied to the main families of hypercyclic op-
erators. In the final section we will again take up the topic of the previous
chapter by studying the existence of common hypercyclic subspaces.

11.1 The Common Hypercyclicity Criterion

We first fix our terminology.

Definition 11.1. Let (Tλ )λ∈Λ be a family of hypercyclic operators on a sep-

arable Fréchet space X. Then a vector x ∈ X is called a common hypercyclic
vector for this family if it is hypercyclic for each operator Tλ , λ ∈ Λ.

In other words, the common hypercyclic vectors are exactly the elements
of 
HC(Tλ ).
λ∈Λ

In view of the Baire category theorem, the following is an immediate conse-

quence of Theorem 2.19.

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 305

DOI 10.1007/978-1-4471-2170-1_11, © Springer-Verlag London Limited 2011
306 11 Common hypercyclic vectors

Proposition 11.2. If Λ is countable, then the set of common hypercyclic

vectors for (Tλ )λ∈Λ is a dense Gδ -set, and in particular, nonempty.

The situation changes dramatically when we allow uncountable families of

operators. The following example shows that in this case there may not exist
any common hypercyclic vector; see also Exercise 11.1.1.

Example 11.3. Let B be the backward shift operator on 2 . We consider the

operators
Tλ,μ := λB ⊕ μB, λ, μ > 1,
on 2 ⊕ 2 . We know that the Rolewicz operators λB, λ > 1, are mixing (see
Example 2.38), so that each operator Tλ,μ is hypercyclic by Proposition 1.42.
But they do not possess a common hypercyclic vector. Indeed, suppose that
(x, y) is such a vector. Then, in particular, for every λ, μ > 1, the pairs of
real numbers (λn B n x, μn B n y), n ≥ 1, form a dense set in R2+ . Let us fix
a, b > 0 and δ > 1. Then, for any λ, μ > 1, there is some n ≥ 1 such that

δ −1 a < λn B n x < δa and δ −1 b < μn B n y < δb. (11.1)

We set cn = n1 log(a B n x−1 ) and dn = n1 log(b B n y−1 ). Taking loga-

rithms in (11.1) we find that, for any λ, μ > 1 there is some n ≥ 1 such
that

− n1 log δ < log λ − cn < 1

n log δ and − 1
n log δ < log μ − dn < 1
n log δ.

Hence the squares of side length n2 log δ centred at (cn , dn ), n ≥ 1, cover

]0, ∞[2 . But this is impossible because
the total area of these squares is finite
∞
due to the convergence of the series n=1 n12 .

On the other hand, some uncountable families do have common hypercyclic

vectors. Two results that we have already met provide us with a multitude
of examples. By the León–Müller theorem, each family (λT )λ∈T has common
hypercyclic vectors as soon as T is a hypercyclic operator on a complex
Fréchet space. And by the Conejero–Müller–Peris theorem, if a C0 -semigroup
(Tt )t≥0 of operators on a Banach space is hypercyclic then the operators
Tt , t > 0, have common hypercyclic vectors. However, it is important to
realize that these results should be understood as structural properties of
hypercyclicity; they do not represent interesting examples of the phenomenon
of common hypercyclicity. Indeed, in both cases the operators in the family
have exactly the same hypercyclic vectors, which will not be the case in
general; see Exercise 11.1.4.
Suppose now that an uncountable family (Tλ )λ∈Λ of hypercyclic operators
does not fall into one of these two categories. How would we prove the exis-
tence of a common hypercyclic vector? In view of uncountability, one might
expect that we have to do without the Baire category theorem. This is true,
but only partly so. Baire’s theorem can still help us to simplify the search, and
11.1 The Common Hypercyclicity Criterion 307

it serves to show that in many cases the set of common hypercyclic vectors
is a dense Gδ -set.
To this end we suppose that the index set Λ is a metric space. Then a
family (Tλ )λ∈Λ of operators on a Fréchet space X is called continuous if, for
any x ∈ X, the map
Λ → X, λ → Tλ x
is continuous. In the present setting, this implies joint continuity with respect
to λ and x.

Proposition 11.4. Let Λ be a metric space and (Tλ )λ∈Λ a continuous family
of operators on a Fréchet space X. Then the map (λ, x) → Tλ x is continuous
on Λ × X.

Proof. Let xn → x in X and λn → λ in Λ. Then

Tλ x − Tλn xn = (Tλ − Tλn )x + Tλn (x − xn ).

Since (Tλn z)n converges for any z ∈ X, the Banach–Steinhaus theorem (The-
orem A.10) implies that (Tλn )n is equicontinuous. It then follows easily that
Tλn xn → Tλ x as n → ∞. 

We need to add one more assumption, namely that Λ is σ-compact, that

is, a countable union of compact sets.

Theorem 11.5. Let X be a separable Fréchet space, Λ a σ-compact metric

space and (Tλ )λ∈Λ a continuous family of operators on X.
(a) The set of common hypercyclic vectors is a Gδ -set.
(b) The following assertions are equivalent:
(i) the set of common hypercyclic vectors is a dense Gδ -set;
(ii) for any compact set K ⊂ Λ and for any nonempty open subsets U
and V of X there is some x ∈ U such that, for any λ ∈ K, there
is some n ≥ 0 such that Tλn x ∈ V .

Proof. (a) By separability, the topology of X has a countable base (Vk )k . Let
(Km )m be a sequence of compact sets whose union is Λ. We then have that
  
HC(Tλ ) = E(Km , Vk ),
λ∈Λ m≥1 k≥1

where we define

E(K, V ) = {x ∈ X ; for all λ ∈ K there is n ≥ 0 such that Tλn x ∈ V }

for any compact subset K ⊂ Λ and any nonempty open subset V ⊂ X.

We show that each set E(K, V ) is open. Let x0 ∈ E(K, V ). Then, for any
λ ∈ K, there is some nλ ≥ 0 such that Tλnλ x0 ∈ V . By Proposition 11.4, the
continuity of (Tμ )μ∈Λ implies the continuity of each family (Tμnλ )μ∈Λ , so that
308 11 Common hypercyclic vectors

there exist open neighbourhoods Oλ of λ and Uλ of x0 such that, for any

μ ∈ Oλ and x ∈ Uλ , Tμnλ x ∈ V . Since the sets Oλ , λ ∈ K, form an open cover
of the compact set K there exists a finite subcover Oλ1 , . . . , OλJ of K. Then

U0 := Jj=1 Uλj is a neighbourhood of x0 , and if x ∈ U0 and λ ∈ K then
nλ
there is some j such that λ ∈ Oλj , hence Tλ j x ∈ V , so that x ∈ E(K, V ).
This shows that E(K, V ) is open, which implies (a).
(b) It now follows from the Baire category theorem that the set of common
hypercyclic vectors is dense if and only if each set E(K, V ) is dense, that is,
if it meets any nonempty open set U . But this is precisely condition (ii). 
Note that if Λ is a singleton, then the result reduces to the Birkhoff tran-
sitivity theorem; see Theorem 2.19.
Remark 11.6. (a) If, under the conditions of the theorem, one of the operators,
Tλ0 say, commutes with all the others then the set of common hypercyclic
vectors is either empty or a dense Gδ -set. Indeed, if x is a common hypercyclic
vector then the vectors Tλn0 x, n ≥ 0, form a dense set of common hypercyclic
vectors; see Exercise 2.6.3. See also Exercise 11.1.5.
(b) Condition (ii) is deceivingly simple. As the proof shows, the follow-
ing, formally stronger, condition also characterizes when the set of common
hypercyclic vectors is a dense Gδ -set:
(ii ) for any compact set K ⊂ Λ and for any nonempty open subsets U and
V of X there are x ∈ U and n1 , . . . , nJ ≥ 0 such that, for any λ ∈ K
n
there is some j such that Tλ j x ∈ V ;
see Exercise 11.1.6. In all the existence proofs in this chapter we will in fact
implicitly verify condition (ii ) rather than condition (ii).
As for ordinary hypercyclicity, the Baire category theorem simplifies the
task of finding common hypercyclic vectors considerably. But, as for frequent
hypercyclicity, the conditions (ii) or (ii ) still demand the construction of a
vector x having several approximation properties simultaneously.
The following criterion, the main result of this chapter, provides a sufficient
condition that allows us to construct such a vector. It turns out that the
criterion has some similarities with the Frequent Hypercyclicity Criterion;
see Theorem 9.9 and Exercise 9.2.1.
Again the notion of unconditionally convergent series enters the picture,
but convergence should be uniform over a family of series. As usual we will
work with an F-norm  ·  that induces the topology of the given Fréchet
space.
∞
Definition 11.7. Let X be a Fréchet space. Then the series n=1 xλ,n , λ ∈
Λ, in X are said to be uniformly unconditionally convergent if, for any ε > 0,
there is some N ∈ N such that for any finite set F ⊂ {N, N + 1, N + 2, . . .}
we have that  
 
 xλ,n  < ε for all λ ∈ Λ.
n∈F
11.1 The Common Hypercyclicity Criterion 309

Example 11.8. The following are sufficient∞ conditions for uniform uncondi-
series n=1 xλ,n , λ ∈ Λ:
tional convergence of a family of 
(i) there is a convergent series ∞n=1 cn of positive numbers such that

xλ,n  ≤ cn for all n ≥ 1, λ ∈ Λ;

(ii) for any k ≥ 1 and ε > 0 there is some N ∈ N such that

∞

pk (xλ,n ) < ε for all λ ∈ Λ,
n=N

where (pk )k is an increasing sequence of seminorms defining the topol-

ogy of X; ∞
(iii) there are en ∈ X, aλ,n ∈ K and cn > 0 such that n=1 cn en converges
unconditionally and, for all n ≥ 1 and λ ∈ Λ,

xλ,n = aλ,n en and |aλ,n | ≤ cn .

Here, condition (ii) is sufficient because {x ∈ X ; x < ε} is a neighbourhood

of 0 and thus contains a set of the form {x ∈ X ; pk (x) < δ}, k ≥ 1, δ > 0.
And condition (iii) is sufficient in view of condition (vi) in Theorem A.16.

We can now formulate the announced sufficient condition for common

hypercyclicity. In order to increase readability we will use the notation

Tλ = T (λ) and Tλn = T n (λ).

Let us also note that the criterion requires a one-dimensional parameter set,
a restriction that we will discuss later.

Theorem 11.9 (Common Hypercyclicity Criterion). Let Λ ⊂ R be an

interval and (Tλ )λ∈Λ = (T (λ))λ∈Λ a continuous family of operators on a
separable Fréchet space X. Suppose that, for any compact subinterval K ⊂ Λ,
there is a dense subset X0 of X and maps Sn (λ) : X0 → X, n ≥ 0, λ ∈ K,
such that, for any x ∈ X0 ,
m
(i) T m (λ)Sm−n (μn )x converges unconditionally, uniformly for m ≥ 0
n=0
and λ ≥ μ0 ≥ . . . ≥ μm from K;
∞
(ii) T m (λ)Sm+n (μn )x converges unconditionally, uniformly for m ≥ 0
n=0
and λ ≤ μ0 ≤ μ1 ≤ . . . from K;
(iii) for any ε > 0 there is some δ > 0 such that, for any n ≥ 1, λ, μ ∈ K,

if 0≤μ−λ< δ
n then T n (λ)Sn (μ)x − x < ε;

(iv) T n (λ)x → 0 as n → ∞, uniformly for λ ∈ K.

310 11 Common hypercyclic vectors

Then the set of common hypercyclic vectors of the family (Tλ )λ∈Λ is a dense
Gδ -set, and in particular, nonempty.
In (i), we consider the finite sums as infinite series by adding 0 terms.
Proof. We will verify that the characterizing condition of Theorem 11.5 holds.
Thus let K ⊂ Λ be a compact set, which we can assume to be a subinterval
K = [a, b], and let U and V be nonempty open subsets of X. Then there are
points x0 , y0 ∈ X0 and some ε > 0 such that, whenever x − x0  < ε and
y − y0  < ε then x ∈ U and y ∈ V .
We can deduce from conditions (i), (ii) and (iv) that there is some N ≥ 0
such that, for any finite set F ⊂ {N, N + 1, N + 2, . . .}, we have that
   ε
 
 T m (λ)Sm−n (μn )y0  < for m ≥ 0, λ ≥ μ0 ≥ . . . ≥ μm , (11.2)
4
n∈F,n≤m
  ε
 
 T m (λ)Sm+n (μn )y0  < for m ≥ 0, λ ≤ μ0 ≤ μ1 ≤ . . . , (11.3)
4
n∈F
ε
T n (λ)x0  < for n ≥ N , λ ∈ K. (11.4)
4

Moreover, by condition (iii) there is some δ > 0 such that

ε
T n (λ)Sn (μ)y0 − y0  < for n ≥ 1, 0 ≤ μ − λ < δ
n. (11.5)
4
By the divergence of the harmonic series there is some J ≥ 1 such that


J−1
δ J
δ
a+ ≤b<a+ .
ν=1
2νN ν=1
2νN

We then set


j
δ
μ0 = a, μj = a + (0 < j < J), μJ = b,
ν=1
2νN

so that a = μ0 ≤ μ1 ≤ . . . ≤ μJ = b.
After these preparations we define

x = x0 + SN (μ1 )y0 + S2N (μ2 )y0 + . . . + SJN (μJ )y0

and claim that it is a vector as required in condition (ii) of Theorem 11.5(b).

Indeed, setting m = 0 and λ = a in (11.3) we find that

x − x0  = SN (μ1 )y0 + S2N (μ2 )y0 + . . . + SJN (μJ )y0  < ε,

so that x ∈ U . Moreover, let λ ∈ K. Then there is some j with 1 ≤ j ≤ J

such that μj−1 ≤ λ ≤ μj . Since 0 ≤ μj − λ ≤ 2jN
δ δ
< jN , we conclude that
11.1 The Common Hypercyclicity Criterion 311


j−1
T jN (λ)x − y0 = T jN (λ)x0 + T jN (λ)SνN (μν )y0
ν=1

  
J
+ T jN
(λ)SjN (μj )y0 − y0 + T jN (λ)SνN (μν )y0
ν=j+1


j−1
jN
=T (λ)x0 + T jN (λ)SjN −αN (μj−α )y0
α=1

  
J−j
+ T jN
(λ)SjN (μj )y0 − y0 + T jN (λ)SjN +αN (μj+α )y0
α=1
ε ε ε ε
< + + + = ε,
4 4 4 4
where we have applied, in turn, (11.4), (11.2), (11.5) and (11.3). This implies
that
T jN (λ)x ∈ V,
which proves the claim.

Remark 11.10. (a) In applications one often has that S0 (λ)x = x for all λ ∈ Λ
and x ∈ X0 . In such a case condition (iv) can be dropped because it follows
from condition (i) by considering the index n = m. Incidentally, condition
(iii) implies that T n (λ)Sn (λ)x = x for all n ≥ 1, λ ∈ K and x ∈ X0 .
(b) The conditions in the Common Hypercyclicity Criterion are rather
strong. They imply that any operator Tλ , λ ∈ Λ, is frequently hypercyclic
and mixing; see Exercise 9.2.1 and Remark 3.13(a). Moreover, if we can take,
for any λ ∈ Λ, Sn (λ) = S n (λ), n ≥ 0, with some map S(λ) : X0 → X0 , then
each operator Tλ , λ ∈ Λ, is also chaotic; see Proposition 9.11.
(c) Condition (i) in particular is quite restrictive; see the discussion before
Example 11.18.
(d) It is obvious from the proof that condition (iii) can be relaxed. Clearly,
∞onlyδ property of the sequence (δ/n)n that we needed was that the series
the
n=1 nN diverges for any N ≥ 1. Thus, condition (iii) can be weakened to
the following:
(iii ) for any ε 
> 0 there is a decreasing sequence (δn )n of positive numbers
∞
such that n=1 δn diverges and such that, for all n ≥ 1, λ, μ ∈ K,

if 0 ≤ μ − λ < δn T n (λ)Sn (μ)x − x < ε.

then
∞
The fact that (δn )n is decreasing ensures that also n=1 δnN diverges for
any N ≥ 1. See Exercise 11.3.1 for an application. As a concrete example,
one may replace δ/n by δ/(n log n) in condition (iii).
In many situations one faces families with a complex or (at least) two-
dimensional real parameter set of positive Lebesgue measure. When one tries
to extend the Common Hypercyclicity Criterion to such a setting, condition
312 11 Common hypercyclic vectors

(iii) seems to create a serious problem. The property that we needed in the
proof was that, essentially, the balls of radius δ/n about the μn cover the
compact part K of the parameter set. To have something similar in a complex√
setting, say, would require choosing larger balls, for example of radius δ/ n;
see also the argument in Example 11.3. But in this weakened form condition
(iii) can usually no longer be satisfied.
Let us now see the Common Hypercyclicity Criterion at work.

Example 11.11. (Rolewicz’s operators) The multiples λB, |λ| > 1, of the
backward shift on any of the spaces p , 1 ≤ p < ∞, or c0 , share a common
hypercyclic vector, and the set of common hypercyclic vectors is a dense
Gδ -set.
Thanks to the León–Müller theorem, for any fixed λ > 1, the operators
zλB, |z| = 1, have the same hypercyclic vectors. Hence it suffices to prove
the claim for Λ = ]1, ∞[.
In order to apply the Common Hypercyclicity Criterion, let K = [a, b]
with 1 < a < b. As is usual for backward shifts, we set S(λ) = λ1 F , with F
the forward shift (see Example 3.6), and we let Sn (λ) = S n (λ), n ≥ 0. For
X0 we choose the set of finite sequences.
It then suffices to verify conditions (i)–(iii) for the canonical unit sequences
x = ek , k ≥ 1; see Remark 11.10(a). For any finite set G ⊂ {N, N + 1, . . .}
we have that
  λm
T m (λ)Sm−n (μn )ek = n
m−n B ek ,
n∈G,n≤m n∈G,n≤m
μn

which will vanish, irrespective of m, λ, and the μn , whenever N ≥ k. This

shows condition (i). Next, for any m ≥ 0 and λ ≤ μ0 ≤ μ1 ≤ . . . in [a, b] we
have that 0 ≤ λm /μm+n
n ≤ 1/an and hence
    λm   ∞ 
   n   1 
 T m (λ)Sm+n (μn )ek  =  m+n F e k  ≤  e k+n  → 0
μn an
n∈G n∈G n=N

as N → ∞. This shows condition (ii). And finally,

λ n

T n (λ)Sn (μ)ek − ek  = − 1 .
μ

Therefore, if 0 ≤ μ − λ < δ
n and λ, μ ∈ K, then

λ n  1
λ δ δ
0≤1− =n tn−1 dt ≤ n 1 − < ≤ , (11.6)
μ λ/μ μ μ a

so that also condition (iii) holds with δ = aε.

It is obvious from the argument that this result allows a far-reaching gen-
eralization. We will return to this in the next section.
11.2 Common hypercyclic vectors for multiples of an operator 313

11.2 Common hypercyclic vectors for multiples of an

operator

One of the simplest parametrized families of operators is given by multiples

λT, λ ∈ K,

of a given operator, which always constitutes a continuous family. We start

with a simple observation that is an immediate consequence of Remark
11.6(a) and Exercise 11.1.5.

Proposition 11.12. Let T be an operator on a separable Fréchet space, Λ ⊂

K a σ-compact set. If the set

HC(λT )
λ∈Λ

of common hypercyclic vectors is nonempty then it is a dense Gδ -set and it

contains a dense subspace, except for 0.

Moreover, we know from the León–Müller theorem that any operator λT

shares its hypercyclic vectors with each of the operators zλT , |z| = 1. Thus
we only have to search for common hypercyclic vectors when λ varies inside
the positive real numbers.
We will consider here the special case of operators T whose generalized
kernel ∞

ker T n
n=0

forms a dense set in X; we have already encountered this notion in Corollary

2.49. For example, if T = Bw is a unilateral weighted backward shift on
a sequence space X, then its generalized kernel is simply the set of finite
sequences, which often forms a dense set.
In this context we have the following general result.

Theorem 11.13. Let T be an operator on a separable Fréchet space X with

dense generalized kernel X0 and λ0 ≥ 0. Suppose that there is a map S :
X0 → X such that T Sx = x and
1 n
S x→0
λn
for all x ∈ X0 , λ > λ0 . Then the set of common hypercyclic vectors for the
operators λT , |λ| > λ0 , is a dense Gδ -set, and in particular, nonempty.

Proof. We first note that, for any x ∈ X0 , T n Sx = T n−1 x = 0 if n is

sufficiently large, so that S maps X0 into itself. Therefore the maps S n are
defined on X0 . Now, by the theorem of León and Müller it suffices to show
314 11 Common hypercyclic vectors

that the operators λT , λ > λ0 , have a common hypercyclic vector. To this

end we apply the Common Hypercyclicity Criterion with Λ = ]λ0 , ∞[, where
we define
1
Sn (λ)x = n S n x, x ∈ X0 , λ ∈ Λ.
λ
Now let b > a > λ0 . Then we have for x ∈ X0 and λ, μ0 , . . . , μm ∈ [a, b]
that
m m
λm n
T m (λ)Sm−n (μn )x = T x,
n=0
μm−n
n=0 n

which implies condition (i) of the Common Hypercyclicity Criterion because

T n x = 0 for all sufficiently large n.
Moreover, for any x ∈ X0 and λ ≤ μ0 ≤ μ1 ≤ . . . from [a, b] we have that
∞
 ∞
λm n
T m (λ)Sm+n (μn )x = S x.
n=0
μm+n
n=0 n

Let (pk )k be an increasing sequence of seminorms defining the topology of

X, and let λ0 < c < a. Since
λm 1 1
0≤ ≤ n ≤ n,
μm+n
n μ n a

we then have for any k ≥ 1,

∞

c n
1 n 
∞
   
pk T (λ)Sm+n (μn )x ≤
m
pk n S x < ∞,
n=0 n=0
a c

which implies condition (ii); see Example 11.8.

Next, for any ε > 0 and x ∈ X there is some ε > 0 such that cx < ε
whenever |c| < ε. Now, if 0 ≤ μ − λ < aε/n, then |( μλ )n − 1| < ε (see (11.6)),
and hence 

λ n  
 
T n (λ)Sn (μ)x − x =  − 1 x < ε,
μ
which implies condition (iii).
Finally, condition (iv) is trivial. 

Remark 11.14. For some applications it is useful to allow arbitrary dense

∞
subsets X0 of the generalized kernel n=0 ker T n of T ; see Exercises 11.2.3–
11.2.5. In that case we assume the existence of maps Sn : X0 → X, n ≥ 0,
such that, for all x ∈ X0 , T n Sn x = x, T m Sm+n x = Sn x for m, n ≥ 0, and
λn Sn x → 0, λ > λ0 . Then the result remains true with virtually the same
1

proof.

In Banach spaces the theorem leads to a very general result under rather
weak hypotheses.
11.2 Common hypercyclic vectors for multiples of an operator 315

Corollary 11.15. Let T be a surjective operator on a separable Banach space

X with dense generalized kernel. Then there exists some λ0 > 0 such that the
set of common hypercyclic vectors for the operators λT , |λ| > λ0 , is a dense
Gδ -set, and in particular, nonempty.

Proof. By the open mapping theorem, the image of the open unit ball under
T contains the open ball of radius 2ε for some ε > 0. Hence for any y ∈ X
with y = ε there is some x ∈ X with x < 1, which we call Sy, such that
T x = T Sy = y. For general y ∈ X, y = 0, we define

y
ε 
Sy = S y ,
ε y

with S0 = 0. Then S is well defined, and for any y ∈ X we have that T Sy = y

and Sy ≤ y/ε. Hence
 1 
  1
 n S n y  ≤ n n y → 0
λ λ ε
for any y ∈ X and λ > 1/ε. We then conclude with Theorem 11.13. 

In situations where Theorem 11.13 can be applied, the task that remains
is that of determining a possibly small value of λ0 . We start with a general-
ization of Example 11.11; see Section 4.1 for the general framework of this
result.

Example 11.16. Let X be a Fréchet sequence space in which (en )n is a basis

and suppose that the (unweighted) backward shift B is an operator on X. If
λ0 ≥ 0 is such that the sequence (λ−n )n belongs to X for any λ > λ0 then
the multiples λB, |λ| > λ0 , share a hypercyclic vector. In fact, the set of
common hypercyclic vectors is a dense Gδ -set.
This follows from Theorem 11.13. Indeed, the generalized kernel of B is the
set offinite sequences, which is dense, and the condition (λ−n )n ∈ X implies
that ∞ n=1 λ
−n
en converges and hence that λ−n F n ek = λk λ−(n+k) en+k → 0
for all k ≥ 1, where F is the forward shift.

Example 11.17. (MacLane’s operator) Let D be the differentiation oper-

ator on the space H(C) of entire functions.
(a) The family (λD)λ=0 has a common hypercyclic vector. Indeed, the
generalized kernel X0 of D consists of the polynomials, which are  zdense in
H(C). The map S : X0 → H(C) is defined, as usual, by Sf (z) = 0 f (ζ) dζ,
λk k!
z ∈ C. Since, for every λ = 0 and k ≥ 0, λ1n S n (z k ) = (k+n)! (z/λ)k+n → 0 as
n → ∞, uniformly on compact sets, Theorem 11.13 implies the claim.
(b) Let T = μI +D, where μ ∈ C. Then (λT )λ=0 has a common hypercyclic
vector. In this case, we consider the set X0 of functions f (z) = e−μz p(z),
where p is a polynomial; then X0 is dense, and a subset of the general-
ized kernel of T (in fact, it is the generalized kernel). Moreover, Sf (z) =
316 11 Common hypercyclic vectors
z
e−μz 0 eμζ f (ζ) dζ defines a map S : X0 → X0 for which the maps S n ,
n ≥ 1, satisfy the assumptions of Remark 11.14; in particular, an argument
as in (a) shows that, for any f ∈ X0 and λ = 0, λ1n S n f → 0 as n → ∞,
uniformly on compact sets. Again, this implies the claim.

Having treated Rolewicz’s and MacLane’s operators it is natural to study

common hypercyclic vectors for the multiples of Birkhoff’s operators Ta on
H(C) given by Ta f (z) = f (z + a), a = 0. However, since Ta is injective its
generalized kernel is trivial so that we cannot apply the methods of this sec-
tion. What is worse, even reverting directly to the Common Hypercyclicity
Criterion does not work, at least when using the obvious choice for the maps
Sn (λ), namely Sn (λ)f (z) = λ1n f (z − na). Indeed, condition (i) is then not
satisfied; see Exercise 11.2.7. Thus we need to rely on what seems to be the
canonical tool when dealing with Birkhoff’s operators, the Runge approxima-
tion theorem; see Example 2.20.

Example 11.18. (Birkhoff’s operators, I) Let a ∈ C, a = 0. We claim that

the multiples of the corresponding translation operator,

λTa f (z) = λf (z + a), λ ∈ C \ {0},

have a common hypercyclic vector on H(C). As before, by the León–Müller

theorem, it suffices to consider real positive parameters λ. By Proposition
4.25 we may take a = 1.
We will show that condition (ii) of Theorem 11.5(b) is satisfied. Thus,
let K = [b, c], 0 < b < c, and let U and V be nonempty open subsets
of H(C). Then there are functions f ∈ U and g ∈ V , ε > 0, and some
N ∈ N such that an entire function h belongs to U (or to V ) whenever
sup|z|≤N |f (z) − h(z)| < ε (or sup|z|≤N |g(z) − h(z)| < ε, respectively).
Let M = max|z|≤N |g(z)| and δ = 2M εb
. There is then some J ≥ 1 such that


J−1
δ J
δ
b+ ≤c<b+ .
ν=1
3νN ν=1
3νN

We set

j
δ
λ0 = b, λj = b + (0 < j < J), λJ = c
ν=1
3νN
and nj = 3jN for j = 1, . . . , J.
Applying Runge’s theorem to the union of the closed balls of radius N
around 0, n1 , . . . , nJ (see Figure 11.1), we obtain an entire function h such
that
sup |f (z) − h(z)| < ε
|z|≤N

and, for j = 1, . . . , J,
11.3 Further examples 317

Fig. 11.1 Constructing common Birkhoff hypercyclic functions I

−nj ε −nj
sup |λj g(z − nj ) − h(z)| < λ .
|z−nj |≤N 2 j

By the first inequality we have that h ∈ U .

Moreover, let λ ∈ K. Then there is some j with 1 ≤ j ≤ J such that
λj−1 ≤ λ ≤ λj . Thus the second inequality implies that, for |z| ≤ N ,

λ nj
λ nj
g(z) − λnj h(z + nj )
|g(z) − λnj h(z + nj )| ≤ 1 − |g(z)| + j
λj λj

λ nj
ε
< 1 − M + .
λj 2

Using the estimate (11.6) and the fact that 0 ≤ λj − λ ≤ δ

nj we obtain that

δ ε
sup |g(z) − (λT1 )nj h(z)| < M + = ε,
|z|≤N b 2

so that (λT1 )nj h ∈ V , which proves the claim.

11.3 Further examples

As a first example of a family that is not made up of multiples of a given

operator we turn to families of weighted shifts. Following the notation and
terminology of Section 4.1, we will consider weighted shifts Bw(λ) , where the
weight sequence w(λ) = (wn (λ))n depends on a real parameter λ. Moreover,
we recall that a function f : [a, b] → R is called Lipschitz continuous if there
is a constant L ≥ 0 such that, for any x, y ∈ [a, b],

|f (x) − f (y)| ≤ L|x − y|.

The minimal constant L is called the Lipschitz constant of f .

Proposition 11.19. Let X be a Fréchet sequence space in which (en )n is an

unconditional basis. Let Λ ⊂ R be an interval and wn : Λ → R, n ≥ 1, be
strictly positive functions such that, for each λ ∈ Λ, Bw(λ) is an operator on
X. Suppose that
318 11 Common hypercyclic vectors

(i) each function wn , n ≥ 1, is increasing;

(ii) for any compact subinterval K ⊂ Λ, the functions log wn , n ≥ 1, are
Lipschitz continuous on K with uniformly bounded Lipschitz constants;
(iii) for any λ ∈ Λ the series
∞

 n −1
wν (λ) en
n=1 ν=1

converges in X.
Then the set of common hypercyclic vectors for the family (Bw(λ) )λ∈Λ is a
dense Gδ -set, and in particular, nonempty.

Proof. We apply the Common Hypercyclicity Criterion. Let  ·  denote an

F-norm defining the topology of X.
First, let b ∈ Λ and x ∈X. By assumption, the functions wn , n ≥ 1,
∞
are increasing and the series n=1 wn+1 (b)xn+1 en converges unconditionally.
Hence,
 by Theorem A.16, for any ε > 0 there is some N ≥ 1 such that
 n≥N wn+1 (λ)xn+1 en  < ε for any λ ∈ Λ, λ ≤ b. Now the continuity of
the wn implies that (Bw(λ) )λ∈Λ is a continuous family.
In order to verify the main conditions of the criterion, let K = [a, b] ⊂ Λ.
As usual for weighted shifts we take X0 as the space of finite sequences, which
is dense in X, and we define Sn (λ) : X0 → X by
1
Sn (λ)ek = ek+n , n ≥ 0, k ≥ 1,
wk+1 (λ) · · · wk+n (λ)

and linear extension to X0 . In view of Remark 11.10(a) we need only verify

conditions (i)–(iii) of the Common Hypercyclicity Criterion.
Condition (i) of the criterion holds because, for any k ≥ 1 and λ ∈ Λ,
m
Bw(λ) Sm−n (μ)ek = 0 for n ≥ k.

As for condition (ii) we have for any k ≥ 1 that

m wk+1+n (λ) · · · wk+m+n (λ)

Bw(λ) Sm+n (μn )ek = ek+n . (11.7)
wk+1 (μn ) · · · wk+m+n (μn )

Using the fact that each function wn is increasing we obtain, whenever μn , λ ∈

K with μn ≥ λ,

wk+1+n (λ) · · · wk+m+n (λ) 1

0≤ ≤ . (11.8)
wk+1 (μn ) · · · wk+m+n (μn ) wk+1 (a) · · · wk+n (a)

It follows from assumption (iii) and the unconditionality of the basis that
∞
 1
ek+n
n=1
wk+1 (a) · · · wk+n (a)
11.3 Further examples 319

converges unconditionally. By Example 11.8(iii), (11.7) and (11.8) then imply

condition (ii) of the Common Hypercyclicity Criterion.
Next we find that, for any k ≥ 1,

w 
k+1 (λ) · · · wk+n (λ)
n
Bw(λ) Sn (μ)ek − ek = − 1 ek .
wk+1 (μ) · · · wk+n (μ)

Choose δ > 0 such that (et − 1)ek  < ε whenever |t| < δ. By assumption
(ii) there is some L > 0 such that, for all ν ≥ 1 and λ, μ ∈ K,

| log wν (λ) − log wν (μ)| ≤ L|λ − μ|.

Thus, if n ≥ 1 and λ, μ ∈ K with 0 ≤ μ − λ < δ

Ln then

w  
k+n
k+1 (λ) · · · wk+n (λ)
log ≤ | log wν (λ) − log wν (μ)|
wk+1 (μ) · · · wk+n (μ)
ν=k+1

≤ nL(μ − λ) < δ,

so that Bw(λ)
n
Sn (μ)ek − ek  < ε. Hence also condition (iii) of the Common
Hypercyclicity Criterion holds, which concludes the proof. 

We apply the proposition in some known and some new situations.

Example 11.20. (a) Let X be one of the spaces p , 1 ≤ p < ∞, or c0 . Then

the multiples λB, λ > 1, have a common hypercyclic vector. Indeed, the
functions wn (λ) = λ, n ≥ 1, satisfy all the assumptions of the proposition on
Λ = ]1, ∞[. This confirms Example 11.11.
(b) On the same spaces as in (a) we consider the weighted shifts Bw(λ)
with wn (λ) = 1 + nλ , λ > 0, n ≥ 1. Since 1 + nλ ≥ n+1 n for λ ≥ 1 and
1 + n ≥ ( n ) for 0 < λ ≤ 1 we have that w1 (λ) · · · wn (λ) ≥ (n + 1)min(1,λ)
λ n+1 λ

for λ > 0. Thus, assumption (iii) is satisfied for λ > 1/p for X = p , and
for all λ > 0 for X = c0 . Since the other assumptions hold as well, we have
common hypercyclic vectors for the respective families of operators.
(c) Consider the multiples λD, λ > 0, of the differentiation operator D on
the space H(C) of entire functions. By Example 4.9(b), D can be regarded
as a weighted shift with weights wn = n, n ≥ 1, on a suitable sequence space.
The proposition then implies easily that the operators λD, λ > 0, have a
common hypercyclic vector, confirming Example 11.17(a).

We return to Birkhoff’s operators. Instead of the multiples of a single

operator we consider the family of all hypercyclic Birkhoff operators.

Example 11.21. (Birkhoff’s operators, II) We claim that the translation

operators
Ta f (z) = f (z + a), a ∈ C \ {0},
320 11 Common hypercyclic vectors

have a common hypercyclic vector on H(C). Since this is a continuous family

of commuting operators, the set of common hypercyclic vectors is then a
dense Gδ -set.
First, for every θ ∈ [0, 2π], the family (Taeiθ )a≥0 is a C0 -semigroup. One
may easily verify that the proof of the Conejero–Müller–Peris theorem re-
mains valid on Fréchet spaces. Hence the operators Taeiθ , a > 0, have the
same hypercyclic vectors. Consequently it suffices to show that the operators
Teiθ , θ ∈ [0, 2π], have a common hypercyclic vector.
The proof is very similar to the one given in Example 11.18. We will
show again that condition (ii) in Theorem 11.5(b) is satisfied, this time for
K = [0, 2π]. Thus, let U, V ⊂ H(C) be nonempty and open, let f ∈ U ,
g ∈ V , and let N ∈ N and ε > 0 be such that h ∈ H(C) belongs to U (or
to V ) whenever sup|z|≤N |f (z) − h(z)| < ε (or sup|z|≤N |g(z) − h(z)| < ε,
respectively).
By the continuity of g there is some δ > 0 such that |g(z) − g(ζ)| < 2ε
whenever |z| ≤ N , |ζ| ≤ 54 N and |z − ζ| ≤ δ. We can assume that δ ≤ N4 .
Then there is some J ≥ 1 such that


J−1
δ  δ
J
≤ 2π < .
ν=1
3νN ν=1
3νN

We set

j
δ
θ0 = 0, θj = (0 < j < J), θJ = 2π
ν=1
3νN
and nj = 3jN for j = 1, . . . , J.

Fig. 11.2 Constructing common Birkhoff hypercyclic functions II

Applying Runge’s theorem to the union of the closed balls of radius N

around 0 and of radius 54 N around n1 eiθ1 , . . . , nJ eiθJ (see Figure 11.2), we
obtain an entire function h such that
11.4 Common hypercyclic subspaces 321

sup |f (z) − h(z)| < ε

|z|≤N

and, for j = 1, . . . , J,

sup g(z − nj eiθj ) − h(z) < ε .
|z−nj eiθj |≤ 5 N
2
4

By the first inequality we have that h ∈ U .

Moreover, let θ ∈ [0, 2π]. Then there exists some j with 1 ≤ j ≤ J such
that θj−1 ≤ θ ≤ θj . Hence |θ − θj | ≤ 3jN
δ
and

iθ
nj e − nj eiθj ≤ nj |θ − θj | ≤ δ ≤ 1 N.
4
Now, if |z − nj eiθ | ≤ N then

z − nj eiθj ≤ z − nj eiθ + nj eiθ − nj eiθj ≤ N + δ ≤ 5 N.
4
Consequently we have that
   
sup g(z) − h z + nj eiθ ≤ sup g(z) − g z + nj eiθ − nj eiθj
|z|≤N |z|≤N
 
+ sup g z − nj eiθj − h(z)
|z−nj eiθ |≤N

≤ sup |g(z) − g(ζ)|

|z|≤N,|ζ|≤ 54 N,|z−ζ|≤δ
 
+ sup g z − nj eiθj − h(z)
|z−nj eiθj |≤ 54 N
ε ε
< + = ε.
2 2
This shows that (Teiθ )nj h ∈ V , which implies the claim.

11.4 Common hypercyclic subspaces

In this section we combine the ideas of this chapter with those of Chapter 10.
Having established criteria for the existence of common hypercyclic vectors
one may wonder if a family of operators even has a common hypercyclic
subspace. Our first result might already come as a surprise: not even two
operators with hypercyclic subspaces need to share a hypercyclic subspace.

Example 11.22. We consider the weights wn = n+1n and vn = 2, n ≥ 1,

and the corresponding weighted shifts Bw and Bv on the Hilbert space 2 .
Then the direct sums T1 = Bw ⊕ Bv and T2 = Bv ⊕ Bw on the Hilbert
322 11 Common hypercyclic vectors

space 2 ⊕ 2 satisfy the conditions of Montes’ theorem and therefore have

hypercyclic subspaces; see Exercise 11.4.1.
Let us suppose that T1 and T2 have a common hypercyclic subspace
M ⊂ 2 ⊕ 2 . We can construct inductively an orthonormal system (en )n
of vectors en = (xn , yn ) ∈ M , n ≥ 1, such that (xn )n and (yn )n are or-
thogonal sequences in 2 . Indeed, for e1 we choose a unit vector in M .
Now let ek = (xk , yk ), k = 1, . . . , n, be chosen. Then the linear map
M → K2n , z = (x, y) → (x, x1 , y, y1 , . . . , x, xn , y, yn ) must have non-
trivial kernel because M is infinite dimensional. Hence there is a unit vector
en+1 = (xn+1 , yn+1 ) in M such that (xk )1≤k≤n+1 and (yk )1≤k≤n+1 are or-
thogonal, so that (ek )1≤k≤n+1 is orthonormal.
The sequence (en )n then satisfies xn 2 + yn 2 = 1 for n ≥ 1. By passing
to a subsequence and interchanging x and y if necessary, we can assume
that xn 2 ≥ 12 for all n ≥ 1. Then (xn /x n )n is an orthonormal basis
∞
for L := span{xn ; n ≥ 1} in  2
. Let x = n=1 an xn /x n∞ ∈ L, x = 0.
∞ ∞
Since n=1 (|an |/xn )2 ≤ 2 n=1 |an |2 < ∞, then z := n=1 an /xn  en
converges in M , and z = 0. Therefore z is hypercyclic for T2 , which implies
that x is hypercyclic for Bv . But then L is a hypercyclic subspace for Bv ,
which contradicts Example 10.26.
Thus, even in the case of finitely or countably many operators, one is lead
to look for a sufficient condition on a family of operators to share a hypercyclic
subspace. The second approach to Montes’ theorem via left-multiplication
operators, paired with the Baire category theorem, yields the following quite
natural result.
Theorem 11.23. Let X be a separable Fréchet space with a continuous norm,
and let Tj , j ≥ 1, be operators on X. If they satisfy the hypotheses of Montes’
theorem for the same infinite-dimensional closed subspace M0 of X then they
have a common hypercyclic subspace.
Proof. Let K be the separable Fréchet space defined in Lemma 10.36. By
assumption, for any j ≥ 1, there exists an increasing sequence (nj,k )k of
positive integers such that Tj satisfies the Hypercyclicity Criterion for (nj,k )k
n
and Tj j,k x → 0 as k → ∞, for all x ∈ M0 . It follows from Lemma 10.37 that
the sequences (LT nj,k )k , j ≥ 1, are hypercyclic on K and therefore, by the
j
Baire category theorem, have a common hypercyclic vector S ∈ K. Then, by
Lemma 10.38, there is some λ = 0 such that (I + λS)M0 is a hypercyclic
subspace for each operator Tj , j ≥ 1. 

Example 11.24. Let Ω be a domain in C. By Proposition 10.33, every hyper-
cyclic composition operator Cϕ , ϕ ∈ Aut(Ω), possesses a hypercyclic sub-
space. We want to show here that any two of them, Cϕ and Cψ , ϕ, ψ ∈
Aut(Ω), possess a common hypercyclic subspace. In particular, every hy-
percyclic operator Cϕ has a common hypercyclic subspace with its inverse
operator (Cϕ )−1 = Cϕ−1 ; note that the hypercyclicity of the inverse operator
is automatic by Proposition 2.23.
11.4 Common hypercyclic subspaces 323

In fact, we need only modify the argument of Proposition 10.33 slightly.

As in that proof there are subsequences (mn )n and (mn )n such that Cϕ and
Cψ satisfy the Hypercyclicity Criterion for the sequence (mn )n and (mn )n ,

respectively, and such that (ϕmn )n and (ψ mn )n are run-away sequences. As-
suming again that Ω contains D and fixing an exhaustion (Kn )n of Ω by
compact sets with K1 = D, we can construct inductively Ω-convex compact
subsets Lk , and a strictly increasing sequence (nk )k of positive integers with
n2k−1 ∈ {mn ; n ≥ 1} and n2k ∈ {mn ; n ≥ 1} such that, for k ≥ 1,

ϕn2k−1 (L2k−1 ) ∩ L2k−1 = ∅, ϕn2k−1 (L2k−1 ) ∪ L2k−1 is Ω-convex,

ψ n2k (L2k ) ∩ L2k = ∅, ψ n2k (L2k ) ∪ L2k is Ω-convex,
K2k−1 ∪ L2k−2 ∪ ψ n2k−2 (L2k−2 ) ⊂ L2k−1 ,
K2k ∪ L2k−1 ∪ ϕn2k−1 (L2k−1 ) ⊂ L2k ,

where L0 = ∅ and n0 = 0.
Denoting by en the functions en (z) = z n , n ≥ 1, and applying successively
Runge’s theorem we obtain, for any j ≥ 1, functions fj,k ∈ H(Ω), k ≥ 1,
such that
1 1
sup |ej (z) − fj,1 (z)| < , sup |fj,k (z) − fj,k+1 (z)| < ,
z∈L1 2j+2 z∈Lk+1 2j+k+2
1 1
sup |fj,2k−1 (z)| < , sup |fj,2k (z)| < .
z∈ϕn2k−1 (L2k−1 ) 2j+2k z∈ψn2k (L2k ) 2j+2k+1

It can now be shown exactly as in the proof of Proposition 10.33 that

∞

fj := fj,1 + (fj,k+1 − fj,k ), j ≥ 1,
k=1

converges in H(Ω), that the closed linear span M0 of the functions fj , j ≥ 1,

is infinite dimensional and that (Cϕ )n2k−1 g → 0 and (Cψ )n2k g → 0 in H(Ω)
for any g ∈ M0 . Since Cϕ and Cψ satisfy the Hypercyclicity Criterion for
(n2k−1 )k and (n2k )k , respectively, Theorem 11.23 implies that Cϕ and Cψ
have a common hypercyclic subspace.
We now turn to uncountable families. Not surprisingly, in this case we
have to impose stronger assumptions in order to obtain common hypercyclic
subspaces. Again the approach using left-multiplication operators, this time
paired with the Common Hypercyclicity Criterion, leads to the desired result.
Theorem 11.25. Let Λ ⊂ R be an interval and (Tλ )λ∈Λ a continuous family
of operators on a separable Fréchet space X with a continuous norm. Suppose
that
(i) the family (Tλ )λ∈Λ satisfies the Common Hypercyclicity Criterion;
(ii) there exists an infinite-dimensional closed subspace M0 of X such that
Tλn x → 0 for all x ∈ M0 , λ ∈ Λ.
324 11 Common hypercyclic vectors

Then the family (Tλ )λ∈Λ has a common hypercyclic subspace.

Proof. We will use the terminology of Section 10.5. Let | · | be a continuous
norm on X, let Φ be a countable set of continuous linear functionals on X|·|
that is weak-∗-dense in (X|·| )∗ , and let K = KΦ be the closed subspace of
L(X|·| , X) defined in Lemma 10.36. Then K is a separable Fréchet space
under the seminorms

Sk = sup pk (Sx), k ≥ 1,

|x|≤1

where (pk )k is an increasing sequence of seminorms defining the topology of

X. Also, for every λ ∈ Λ, the left-multiplication operator LTλ is an operator
on K. We claim that the family (LTλ )λ∈Λ satisfies the Common Hypercyclicity
Criterion on K.
We leave the proof of continuity of the family to the reader; see Exercise
11.4.6. Now let K = [a, b] ⊂ Λ. By hypothesis there is a dense subset X0 of
X and maps Sn (λ) : X0 → X, n ≥ 0, λ ∈ K, such that, for any x ∈ X0 ,
conditions (i)–(iv) of Theorem 11.9 are satisfied. By Lemma 10.36, F = FΦ,X0
is dense in K. For any S ∈ F we choose a representation


l
S=  · , yj∗ xj , yj∗ ∈ Φ, xj ∈ X0 ,
j=1

and define, for any n ≥ 0, λ ∈ K,


l
Mn (λ)S =  · , yj∗ Sn (λ)xj ∈ K.
j=1

We then have that, for k ≥ 1, N ≥ 0 and any finite set F ⊂ {N, N + 1, . . .},

     
l 
   
 L m
Tλ M (μ
m−n n )S  =  Tλ
m
 · , yj∗ Sm−n (μn )xj 
k k
0≤n≤m 0≤n≤m j=1
n∈F n∈F


l  
= sup pk x, yj∗  Tλm Sm−n (μn )xj
|x|≤1 j=1 0≤n≤m
n∈F


l
 
≤ sup |x, yj∗ | pk Tλm Sm−n (μn )xj ,
j=1 |x|≤1 0≤n≤m
n∈F

and by assumption the right-hand side can be made arbitrarily small if N

is sufficiently large, uniformly for m ≥ 0 and λ ≥ μ0 ≥ . . . ≥ μm from K.
In the same way one can show also that conditions (ii)–(iv) of the Common
Hypercyclicity Criterion are satisfied for the family (LTλ )λ∈Λ . This implies
Exercises 325

that there is an operator S ∈ K that is hypercyclic for any operator LTλ on

K, λ ∈ Λ. Then, by Lemma 10.38, there is some μ = 0 such that (I + μS)M0
is a hypercyclic subspace for each operator Tλ , λ ∈ Λ. 

Combining this result with (the proof of) Theorem 11.13 we obtain the
following special case.

Theorem 11.26. Let X be a separable Fréchet space with a continuous norm,

let T be an operator on X with dense generalized kernel X0 and λ0 ≥ 0.
Suppose that
(i) there is a map S : X0 → X such that T Sx = x and λ1n S n x → 0 for all
x ∈ X0 , λ > λ0 ;
(ii) there exists an infinite-dimensional closed subspace M0 of X such that
λn T n x → 0 for all x ∈ M0 , λ > λ0 .
Then the family (λT )|λ|>λ0 has a common hypercyclic subspace.

Remark 11.27. Remark 11.14 applies here as well.

Example 11.28. Let X be one of the spaces p , 1 ≤ p < ∞, or c0 , and T the

operator on X defined by

T (x1 , x2 , x3 , . . .) = (x2 , x4 , x6 , . . .);

see also Proposition 8.5. Then the operators λT , |λ| > 1, have a common
hypercyclic subspace. To this end consider for X0 the set of finite sequences,
for Sn the maps Sn = S n with S(xn )n = (0, x1 , 0, x2 , 0, x3 , 0, . . .), x ∈ X, and
for M0 the subspace of all sequences x = (xn )n ∈ X whose even coordinates
vanish. Then the claim follows from Theorem 11.26 and Remark 11.27.

We end this section by mentioning a related problem, that of the existence

of a common dense subspace of hypercyclic vectors, excluding 0 as usual. By
the Herrero–Bourdon theorem, every hypercyclic operator has such a sub-
space. This easily implies that if a commuting family of operators has a com-
mon hypercyclic vector then it has a dense subspace of hypercyclic vectors in
common; see Exercise 11.1.5. In the noncommuting case we still get a positive
answer when the family satisfies the Common Hypercyclicity Criterion and
the underlying space has a continuous norm; see Exercise 11.4.10.

Exercises

Exercise 11.1.1. Consider the family of all weighted backward shifts Bw on 2 whose
weight sequence (wn )n only takes the values 1 and 2, with 2 appearing infinitely often.
Show that each of these operators is hypercyclic but that they do not possess a common
hypercyclic vector. (Hint: Construct, for any x ∈ 2 , a weight w such that w1 · · · wn |xn | ≤
1 for all sufficiently large n.)
326 11 Common hypercyclic vectors

Exercise 11.1.2. Let S and T be operators on a Banach space X and Λ ⊂ ]0, ∞[2 . Show
that if the family (λS ⊕ μT )(λ,μ)∈Λ of operators on X ⊕ X has a common hypercyclic
vector then Λ has two-dimensional Lebesgue measure 0. (Hint: First show as in Example
11.3 that the set {(log λ, log μ) ; (λ, μ) ∈ Λ} has Lebesgue measure 0.)

Exercise 11.1.3. Let 1 ≤ p < ∞, α > 0, v(t) = 1+|t| 1

α (t ∈ R), and b ∈ R. Show that
p
the following define families of operators on Lv (R) and can be extended to semigroups
of operators:
(i) Tλ f (x) = f (x + λ), λ > 0;
(ii) Tλ f (x) = f (λx), λ > 1;
(iii) Tλ f (x) = f (λ(x − b) + b), λ > 1.
Deduce that in each family the operators have the same sets of hypercyclic vectors.
(Hint: When suitable, replace λ with λ > 1 by et with t > 0.)

Exercise 11.1.4. Let B be the backward shift on one of the spaces p , 1 ≤ p < ∞, or c0 .
Show that, although the operators λB, λ > 1, have common hypercyclic vectors, no two
of them have the same hypercyclic vectors. More precisely, if λ, μ > 1 with λ = μ then
there is a vector that is λB-hypercyclic but not μB-hypercyclic. (Hint: Show that there
exists x ∈ HC(λB) with inf n≥1 μn |xn | > 0 (λ < μ) or supn≥1 μn |xn | < ∞ (λ > μ).)

Exercise 11.1.5. Let (Tλ )λ∈Λ be a family of hypercyclic operators on a separable

Fréchet space X. Show that, if some Tλ0 commutes with each operator Tλ , λ ∈ Λ,
then the set 
HC(Tλ )
λ∈Λ

of common hypercyclic vectors is either empty or it contains, except for 0, a dense

subspace.

Exercise 11.1.6. Let X be a separable Fréchet space, Λ a σ-compact metric space and
(Tλ )λ∈Λ a continuous family of operators on X. Show the following:
(i) if x ∈ X is a common hypercyclic vector then, for any compact subset K ⊂ Λ and
for any nonempty open subset V of X there are n1 , . . . , nJ ≥ 0 such that for any
n
λ ∈ K there is some j such that Tλ j x ∈ V ;
(ii) the set of common hypercyclic vectors is a dense Gδ -set if and only if, for any
compact set K ⊂ Λ and for any nonempty open subsets U and V of X, there are
x ∈ U and n1 , . . . , nJ ≥ 0 such that, for any λ ∈ K there is some j such that
n
Tλ j x ∈ V .
Moreover, show that the following condition is satisfied for any continuous family
(Tλ )λ∈Λ of hypercyclic operators and therefore does not characterize common hyper-
cyclicity:
(iii) for any compact set K ⊂ Λ and for any nonempty open subsets U and V of
X there are n1 , . . . , nJ ≥ 0 such that for any λ ∈ K there is some j such that
n
Tλ j (U ) ∩ V = ∅.

Exercise 11.1.7. (a) Let Λ ⊂ R be an interval and (T (λ))λ∈Λ a continuous family of

operators on a separable Fréchet space X. Suppose that, for any compact subinterval
K ⊂ Λ, there are dense subsets X0 , Y0 of X, an increasing sequence (nk )k of positive
integers, and maps Snk (λ) : Y0 → X, k ≥ 1, λ ∈ Λ, such that, for any x ∈ X0 , y ∈ Y0 ,
(i) T nk (λ)x → 0, uniformly for λ ∈ K;
(ii) Snk (λ)y → 0 for any λ ∈ K;
(iii) for any ε > 0 there is some δ > 0 such that, for any μ ∈ K and N ≥ 1 there is
some k ≥ N such that, if λ ∈ K with |λ − μ| < δ then T nk (λ)Snk (μ)y − y < ε.
Exercises 327

Show that the set of common hypercyclic vectors of the family (Tλ )λ∈Λ is a dense Gδ -
set, and in particular, nonempty. (Hint: Fix a partition K1 , . . . , Km of K of intervals
of length < δ; obtain U ⊃ U1 ⊃ . . . ⊃ Um and ν1 , . . . , νm such that T νj (Uj ) ⊂ V for
λ ∈ Kj .)
(b) Show that, for a singleton Λ, this criterion reduces to (a weak form of) the
Hypercyclicity Criterion.
(c) Show that the criterion cannot be applied in the canonical way to the family of
Rolewicz operators.

Exercise 11.2.1. Show that the family (λ(B ⊕ B))|λ|>1 = (λB ⊕ λB)|λ|>1 of operators
on p ⊕ p , 1 ≤ p < ∞, has a common hypercyclic vector. Compare the result with
Exercise 11.1.2.

Exercise 11.2.2. Let Bw be a weighted shift on one of the spaces p , 1 ≤ p < ∞, or c0 ,

where w = (wn )n is a bounded weight sequence; see Section 4.1. Furthermore, let λ0 > 0
be such that λBw is a mixing operator for any λ > λ0 . Show then that (λBw )|λ|>λ0 has
a common hypercyclic vector. Formulate λ0 in terms of the weights.

Exercise 11.2.3. Let D be the differentiation operator on H(C) and p a nonconstant

polynomial. By using the following steps, show that (λp(D))λ=0 has a common hyper-
cyclic vector on H(C).
(a) Fix a root μ ∈ C of p. Then the set X0 of functions f (z) = eμz q(z), q a polynomial,
is dense in H(C).
∞ ∞ n!an
(b) If f (z) = eμz q(z) = a z n belongs to X0 then f(z) =
n=0 n n=0 z n+1
is
holomorphic for |z| > |μ|. (Hint: Look at the proof of Lemma 4.18.)
(c) Fix ρ > |μ| such that p has no roots on |z| = ρ. Then, for any f ∈ X0 and n ≥ 0,
 f(ζ) dζ defines an entire function with p(D)n S f = f .
Sn f (z) = 1
2πi |ζ|=ρ
ezζ p(ζ) n n
(d) Prove the claim by letting ρ → ∞.

Exercise 11.2.4. Let μ ∈ C with |μ| < 1. Consider the adjoint multiplier T = Mϕ∗ on
the Hardy space H 2 given by ϕ(z) = z − μ; see Section 4.4.
zn
n+1 , n ≥ 0. Show that X0 := span{gn ; n ≥ 0} is dense in H
2
(a) Let gn (z) = (1−μz)
and that T gn = gn−1 with g−1 = 0. (Hint: Show that f, gn  is a multiple of f (n) (μ).)
(b) Deduce that the operators λMϕ∗ , |λ| > 1−|μ|
1
, have common hypercyclic vectors.
(c) Deduce also that the operators λ(μI + B) on 2 , |λ| > 1
1−|μ| , have common
hypercyclic vectors, where B is the backward shift.

Exercise 11.2.5. Let ϕ be a nonconstant complex polynomial all of whose roots lie in D.
Show that the operators λMϕ∗ on H 2 , |λ| > max|z|=1 |ϕ(z)|
1
, have common hypercyclic
vectors. (Hint: Fix a root μ of ϕ; consider the set X0 of Exercise 11.2.4 and define
Sn : X0 → H 2 by Sn = Mψn , where ψn (z) = (1/ϕ∗ (1/z))n ; see Proposition 4.41. Use
Proposition 4.40.)

Exercise 11.2.6. Let ϕ be a nonconstant bounded holomorphic function on D and

Mϕ∗ the corresponding adjoint multiplier on H 2 . Using Theorem 4.42, show that λMϕ∗
is hypercyclic if and only if inf z∈D |ϕ(z)|−1 < |λ| < supz∈D |ϕ(z)|−1 , where 1/0 =
∞. Deduce that, for a certain ϕ, Theorem 11.13 cannot be used to obtain common
hypercyclic vectors.

Exercise 11.2.7. Let T be an operator on a Fréchet space X with a continuous norm,

and set T (λ) = λT , λ > 0. Let X0 be a dense subset of X and Sn : X0 → X be maps
such that T n Sn x = x for all x ∈ X0 , n ≥ 0. Set Sn (λ) = λ1n Sn . Show that if condition
(i) of the Common Hypercyclicity Criterion is satisfied for some set K = [a, b] with
0 < a < b then T has a dense generalized kernel.
328 11 Common hypercyclic vectors

Exercise 11.2.8. The proof of the existence of entire functions that are hypercyclic for
all multiples λTa , λ > 0, of the Birkhoff operator Ta (see Example 11.18), can easily be
turned into a purely constructive proof. Explain.

Exercise 11.3.1. Show that Proposition 11.19 remains true when assumption (ii) is
weakened as follows: for any compact subinterval K ⊂ Λ, thefunctionsn log wn−1are
∞
Lipschitz continuous on K with Lipschitz constant Ln such that n=1
( L ) =
k=1 k
∞. Deduce that the weighted shifts Bw(λ) , λ > 0, with wn (λ) = nλ , n ≥ 1, have a
common hypercyclic vector on any of the spaces p , 1 ≤ p < ∞, or c0 . (Hint: Use
Remark 11.10(d).)

Exercise 11.3.2. Proposition 11.19 remains true if each function wn , n ≥ 1, is decreas-

ing. More generally, it remains true if the interval I can be partitioned into countably
many subintervals on which either all functions wn are decreasing or all functions wn
are increasing. Explain.

Exercise 11.3.3. Let Λ ⊂ R be an interval. Let wn : Λ → R, n ≥ 1, be strictly

positive functions such that, for any compact subinterval K ⊂ Λ, the functions log wn
are Lipschitz continuous on K with uniformly bounded Lipschitz constant. Show that
the family (Bw(λ) )λ∈Λ of weighted backward shifts has a common hypercyclic vector on
ω = KN and that the set of common hypercyclic vectors is a dense Gδ -set.

Exercise 11.3.4. Use the results of this section (and the León–Müller theorem) to
confirm the claim of Exercise 11.2.2.

Exercise 11.4.1. Show that the operators T1 and T2 of Example 11.22 have hypercyclic
subspaces.

Exercise 11.4.2. Explain how the construction in Example 11.24 can be modified in
order to show that any countable family of hypercyclic composition operators Cϕj ,
ϕj ∈ Aut(Ω), j ≥ 1, on a domain Ω ⊂ C possesses a common hypercyclic subspace.

Exercise 11.4.3. Let B be the family of all weighted backward shifts Bw on X = p ,

1 ≤ p < ∞, or on X = c0 , such that


n

n

lim |wν | = ∞ and sup lim sup |wν+k | < ∞.

n→∞ n≥1 k→∞
ν=1 ν=1

Show that any countable family of operators in B has a common hypercyclic subspace.

Exercise 11.4.4. Let Tj , j ≥ 1, be operators on a separable Banach space X. Suppose

that there exist increasing sequences (nj,k )k of positive integers such that
(i) for j ≥ 1, Tj satisfies the Hypercyclicity Criterion for (nj,k )k ,
(ii) there is a decreasing sequence (Mk )k of infinite-dimensional closed subspaces of X
n
such that supk≥1,j≤k Tj j,k |Mk  < ∞.
Then the operators Tj , j ≥ 1, have a common hypercyclic subspace. (Hint: Modify the
proof of Theorem 10.29 using Exercise 3.4.9.)

Exercise 11.4.5. Let Tj , j ≥ 1, be weakly mixing operators on a separable Banach

space that are of the form Tj = Uj + Kj , Uj  ≤ 1 and Kj compact, j ≥ 1. Show
that these operators have a common hypercyclic subspace. (Hint: Modify the proof of
Corollary 10.31 using the previous exercise.)

Exercise 11.4.6. Show that, in the setting of Theorem 11.25 and its proof, (LTλ )λ∈Λ
is a continuous family on K. (Hint: Restrict the λ to a compact subset; show that the
corresponding Tλ are equicontinuous on X; use that F is dense in K).
Sources and comments 329

Exercise 11.4.7. Let X be one of the spaces p (Z), 1 ≤ p < ∞, or c0 (Z). Show that
the bilateral weighted backward shifts Bw(λ) , λ > 1, with weights

w(λ) = (. . . , λ1 , λ1 , λ1 , 2, 2, 2, . . .)

have a common hypercyclic subspace.

Exercise 11.4.8. Generalize the result of Example 11.28 to arbitrary operators Dw

given by Dw (xn )n = (w
2 x2 , w4 x4 , w6 x6 , . . .) such that (w2n )n is bounded and, for any
n
m ≥ 1 odd, lim inf n→∞ k=1 wm2k > 0.

Exercise 11.4.9. Let a > 0, and let T be the operator on C0 (R+ ) given by T f (x) =
f (x + a); see Exercise 10.1.1. Show that the operators λT , λ > 1, have a common
hypercyclic subspace.

Exercise 11.4.10. Let Λ ⊂ R be an interval and (Tλ )λ∈Λ a continuous family of op-
erators on a separable Fréchet space X with a continuous norm. Show that if (Tλ )λ∈Λ
satisfies the Common Hypercyclicity Criterion then its set of common hypercyclic vec-
tors contains, except for 0, a dense subspace. (Hint: Use the ideas of Exercise 10.5.1.)

Sources and comments

Section 11.1. The problem of the existence of common hypercyclic vectors for un-
countable families was first raised by Godefroy and Shapiro [165] when they asked if
the differential operators ϕ(D) on H(C) with nonconstant ϕ (see Section 4.2) share a
hypercyclic vector. However, this question was largely ignored, so that it was Salas [275]
who initiated the present study into common hypercyclicity by asking if the Rolewicz
operators λB, |λ| > 1, have a common hypercyclic vector. For real parameters, a positive
answer was given by Abakumov and Gordon [1] and, independently, by Peris [252]; the
general case is due to Costakis and Sambarino [123]; see Example 11.11.
Example 11.3 is due to Borichev (see [1, 42]). Theorem 11.5(a) is due to Saint Ray-
mond (see [1, 33]); the proof given here is taken from Costakis and Sambarino [123, p.
304]. Assertion (b) was obtained by Shkarin [288], while its equivalent formulation given
in Remark 11.6(b) is due to Chan and Sanders [102]. Remark 11.6(a) was observed by
Bayart [33].
The Common Hypercyclicity Criterion, under stronger assumptions, is due to Costakis
and Sambarino [123]. The present form was inspired by the Frequent Hypercyclicity Cri-
terion. An alternative criterion was found by Bayart and Matheron [42]. They use it to
show that, for the family of all automorphisms ϕ of D having 1 as an attractive fixed
point, the corresponding composition operators Cϕ on H 2 (see Section 4.5) share hyper-
cyclic vectors; see also Bayart and Grivaux [39]. By Bayart [33], this result breaks down
when the attractive fixed points are allowed to cover a set of positive Lebesgue measure
on T.
We have already commented on the difficulty of extending the criterion to more than
one-dimensional families of operators. Shkarin [288] confirmed that there is an intrinsic
obstruction to common hypercyclicity for higher-dimensional families by proving the
following remarkably general result.

Theorem 11.29. Let T be an operator on a complex Fréchet space and Λ ⊂ C × R+

be such that the family (λI + μT )(λ,μ)∈Λ has a common hypercyclic vector. Then Λ has
three-dimensional Lebesgue measure 0.
330 11 Common hypercyclic vectors

For example, if D denotes the differentiation operator on H(C), then the operators
λI + μD, λ ∈ C, μ > 0, cannot have a common hypercyclic vector. This finally solved
the Godefroy–Shapiro problem in the negative.

Section 11.2. Proposition 11.12 is due to Bayart [33]. Theorem 11.13, Corollary 11.15
and Example 11.16 are due to Bayart and Matheron [42]; see also Bayart [33]. Remark
11.14 was made by Costakis and Mavroudis [120] and Bernal [59].
Common hypercyclic vectors for the multiples of MacLane’s operator (see Example
11.17) were found by Costakis and Sambarino [123]. This was extended to multiples
of certain operators ϕ(D) by Costakis and Mavroudis [120] and Bernal [59]. Finally,
Shkarin [288] proved that for any nonconstant entire function ϕ of exponential type the
differential operators λϕ(D), λ = 0, share hypercyclic vectors. This then also includes
Example 11.18.
Bayart [33] and Gallardo and Partington [160] have obtained common hypercyclic
vectors for multiples λMϕ∗ of certain adjoint multipliers on H 2 . Shkarin [288] proved
that, for any bounded holomorphic function ϕ, all hypercyclic multiples λMϕ∗ share
hypercyclic vectors.
The two mentioned results by Shkarin [288] are based on a powerful sufficient condi-
tion for common hypercyclicity for multiples of operators.

Section 11.3. Proposition 11.19 is a special case of a theorem by Bayart and Matheron
[42]. The first result in this direction is due to Costakis and Sambarino [123]; see Example
11.20(b). These authors also found Example 11.21; in fact, they gave a direct proof that
for parameters along rays starting from zero the sets of hypercyclic vectors are the same.
Chan and Sanders [102] showed that between any two hypercyclic weighted shifts
on 2 there is a continuous path of weighted shifts that shares hypercyclic vectors; and
there is another such path without any common hypercyclic vector.
Shkarin [288] studied common hypercyclicity for genuinely two- or higher-dimensional
families, that is, where one cannot reduce the number of dimensions by the León–
Müller theorem or the Conejero–Müller–Peris theorem. He was the first to obtain a
genuinely higher-dimensional family with common hypercyclic vectors: the multiples of
the Birkhoff operators f → λf (z + a), λ, a ∈ C \ {0}, on H(C) (note that one can reduce
the two complex dimensions to two real ones).

Section 11.4. Theorem 11.23 is due to Aron, Bès, León and Peris [14] where Example
11.22 can also be found. Theorem 11.25 improves on the corresponding result in Bayart
[35], Theorem 11.26 seems to be new. Common hypercyclic subspaces for some sequence
of operators on ω = KN were studied by Bès and Conejero [69].
As for the problem mentioned at the end of the section we recall that by a remarkable
result of Grivaux [169], any countable family of hypercyclic operators on a Banach space
has a common dense subspace of hypercyclic vectors, except for 0. See also Exercise
3.4.10.

Exercises. Exercise 11.1.1 is from Bayart and Matheron [42], Exercise 11.1.2 is due to
Borichev (see [42]), Exercise 11.1.5 is taken from Grivaux [169], and Exercises 11.1.6
and 11.1.7 from Chan and Sanders [102]. Exercise 11.2.1 is mentioned in Bayart and
Matheron [42], while Exercise 11.2.3 follows Costakis and Mavroudis [120] and Bernal
[59]. Exercises 11.3.1 and 11.3.2 are taken from Bayart and Matheron [42], Exercise
11.4.5 from Aron, Bès, León and Peris [14], and Exercise 11.4.10 partially improves a
result by Bayart [34].
Chapter 12
Linear dynamics in topological vector
spaces

So far, we have been working with operators on Banach or Fréchet spaces.

One of the main reasons was that we then had the Baire category theorem
at our disposal, which is a basic tool in hypercyclicity.
We have made one exception. In Chapter 10, the left-multiplication opera-
tors that we needed were defined on the space L(X) with the strong operator
topology, which is not a Fréchet space unless X is finite dimensional. But
even there, in the final analysis, we worked in a separable Fréchet space K of
operators on X.
Dealing with more general spaces in which Baire category arguments can-
not be applied makes life certainly more difficult for hypercyclicity; but there
are several dynamical properties, like mixing or weak mixing, where the previ-
ous arguments extend, essentially unchanged, to arbitrary topological vector
spaces. Also, several interesting and natural operators are defined on non-
Fréchet topological vector spaces, which is a good motivation to study linear
dynamics in a wider context. This is the purpose of this chapter.

12.1 Topological vector spaces

A topological vector space is a vector space X over the scalar field K =

R or C endowed with a Hausdorff topology such that addition and scalar
multiplication,

+ : X × X → X, (x, y) → x + y,
· : K × X → X, (λ, x) → λx,

are continuous maps. We recall that a topology is Hausdorff if any two distinct
points in the space have disjoint neighbourhoods.

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 331

DOI 10.1007/978-1-4471-2170-1_12, © Springer-Verlag London Limited 2011
332 12 Linear dynamics in topological vector spaces

Many arguments in Banach and Fréchet spaces use the triangle inequality
of the norm or the seminorms. In general topological vector spaces, such
arguments are replaced by operations with 0-neighbourhoods.
A subset A of a vector space X is called balanced if λA ⊂ A whenever
λ ∈ K, |λ| ≤ 1.
Proposition 12.1. Let X be a topological vector space.
(a) A set U is a neighbourhood of a point x ∈ X if and only if there is a
0-neighbourhood W such that

x + W ⊂ U.

(b) Let W be a 0-neighbourhood. For any λ, μ ∈ K there is a 0-neighbour-

hood W1 such that
λW1 + μW1 ⊂ W.
In particular, there is a 0-neighbourhood W1 such that

W 1 + W1 ⊂ W and W1 − W1 ⊂ W.

(c) If W is a 0-neighbourhood and M > 0, then there is a 0-neighbourhood

W1 ⊂ W such that λW1 ⊂ W for every λ ∈ K with |λ| ≤ M . In particular,
every 0-neighbourhood contains a balanced 0-neighbourhood.
Proof. Properties (a) and (b) are easy consequences of the continuity of the
vector operations. For property (c), given W and M , by the continuity of
scalar multiplication we can find ε > 0 and a 0-neighbourhood W0 such
that λW0 ⊂ W for every λ ∈ K with |λ| < ε. Let δ = ε/(M + 1), and
consider W1 = δW0 , which is a 0-neighbourhood since multiplication by a
fixed nonzero scalar is a homeomorphism
 of X. If |λ| ≤ M then λW1 =
(λδ)W0 ⊂ W . As a consequence, |λ|≥1 λW is a balanced 0-neighbourhood
contained in W .  
Let us apply the proposition to obtain some basic facts.
Proposition 12.2. Let X be a topological vector space.
(a) If A is an arbitrary subset of X and U an open set then A + U is open.
(b) For any 0-neighbourhood W there is a 0-neighbourhood W1 such that

W1 ⊂ W ;

in particular, every 0-neighbourhood contains a closed 0-neighbourhood.

Proof. (a) Let x = y + z, y ∈ A, z ∈ U . Since U is open there is a 0-
neighbourhood W such that z + W ⊂ U . Then x + W = y + (z + W ) ⊂ A + U ,
so that x is an interior point of A + U . This proves the claim.
(b) By Proposition 12.1(b) there is a 0-neighbourhood W1 such that W1 −
W1 ⊂ W . Let x ∈ W1 . Then (x + W1 ) ∩ W1 = ∅, hence x ∈ W1 − W1 ⊂ W .


12.1 Topological vector spaces 333

In view of Propositions 12.1(c) and 12.2(b), the set of all closed and
balanced 0-neighbourhoods of a topological vector space X is a base of 0-
neighbourhoods in X, which will be denoted by U0 (X).
There are some classes of topological vector spaces that deserve special
consideration.
To start with, any finite-dimensional topological vector space X is isomor-
phic to KN , for some N ≥ 0, where K is the scalar field of X; see Exer-
cise 12.1.2.
If a topological vector space X admits a countable base (Wn )n of 0-
neighbourhoods, then there is a translation-invariant metric d on X gen-
erating the topology of X. If, moreover, (X, d) is complete, then X is called
an F-space. Metrizable topological vector spaces are, thus, exactly the topo-
logical vector spaces admitting a countable base of 0-neighbourhoods, and
the completion X  of a metrizable topological vector space is an F-space. See
also the related discussion in Section 2.1.
A topological vector space X whose topology is defined by a family of
seminorms is called a locally convex space; that is, X is locally convex if
there is a family (pα )α∈A of seminorms on X such that a subset W of X is
a 0-neighbourhood if and only if there are α1 , . . . , αn ∈ A, n ≥ 1, and ε > 0
such that
{x ∈ X ; pαk (x) < ε for k = 1, . . . , n} ⊂ W.
Fréchet spaces are, precisely, the locally convex F-spaces.
A subset A of a vector space X is called absolutely convex if, for any
x1 , x2 ∈ A and λ1 , λ2 ∈ K with |λ1 | + |λ2 | ≤ 1, the absolutely convex com-
bination λ1 x1 + λ2 x2 belongs to A. An easy observation nis that, if A ⊂ X is
absolutely convex, xk ∈ A, λk ∈ K, k = 1, . . . , n, with k=1 |λk | ≤ 1, then


n
λk xk ∈ A.
k=1

Also, if p is a seminorm on X and M ≥ 0, then the set A = {x ∈ X ; p(x) ≤

M } is absolutely convex. Conversely, if A ⊂ X is an absolutely convex set,
then the associated gauge of A, also called its Minkowski functional, is defined
as
pA (x) = inf{λ > 0 ; x ∈ λA}, x ∈ span A.
One can verify that pA is a seminorm on span A; see Exercise 12.1.5. There-
fore, a topological vector space X is locally convex if, and only if, it has a
base of 0-neighbourhoods (Wα )α∈A consisting of absolutely convex sets.

Example 12.3. In Sections 8.3 and 10.2 we considered the space L(X) of op-
erators on a Fréchet space X, endowed with the strong operator topology
(SOT). In this topology, a base of neighbourhoods of T ∈ L(X) is given by

Ux1 ,...,xn (T, ε) = {S ∈ L(X) ; T xk − Sxk  < ε for k = 1, . . . , n},

334 12 Linear dynamics in topological vector spaces

where x1 , . . . , xn , n ≥ 1, is an arbitrary collection of linearly independent

vectors of X,  ·  is an F-norm defining the topology of X, and ε > 0. We
immediately obtain that L(X) with the strong operator topology is a locally
convex space.

Example 12.4. Given 0 < p < 1 and a < b, let

  b 
Lp [a, b] = f : [a, b] → K ; f is measurable and |f (t)|p dt < ∞ .
a
b
We set Wn = {f ∈ Lp [a, b] ; a |f (t)|p dt < 1/n}, n ∈ N. Then the sequence
(Wn )n defines a base of 0-neighbourhoods and, by translation, a topology on
Lp [a, b] that makes it an F-space that is not locally convex; see Exercise 12.1.3.
Therefore, Lp [a, b] is not a Fréchet space.

Example 12.5. Let (Xn )n be an increasing sequence of Banach spaces such

∞ inclusion map in : Xn → Xn+1 , n ≥ 1, is continuous. We consider
that each
X = n=1 Xn . For each sequence δ = (δn )n of strictly positive numbers, let
∞ 
 n
Wδ = δk Bk ,
n=1 k=1

where Bk is the open unit ball of Xk , k ∈ N. The family of absolutely convex

sets {Wδ ; δ = (δn )n ∈ ]0, ∞[N } forms a base of 0-neighbourhoods for a locally
convex topology on X, called the inductive limit of (Xn )n ; see Exercise 12.1.7.

12.2 Hypercyclicity, topological transitivity, and linear

chaos

We are now in a position to study linear dynamics in its widest possible

framework, that of operators on arbitrary topological vector spaces.
In the sequel we will not always define a notion when its generalization
from the Fréchet space setting is evident. Still, we cannot help stating the
following.

Definition 12.6. An operator T on a topological vector space X is called

hypercyclic if there is some x ∈ X whose orbit under T is dense in X. In such
a case, x is called a hypercyclic vector for T . The set of hypercyclic vectors
is denoted by HC(T ).

Clearly, separability of a space is again a necessary condition for the ex-

istence of a hypercyclic operator. Moreover, any finite-dimensional topolog-
ical vector space is isomorphic to some KN and therefore cannot support
a hypercyclic operator. But unlike for the case of Fréchet spaces there are
12.2 Hypercyclicity, topological transitivity, and linear chaos 335

infinite-dimensional separable topological vector spaces that do not admit

any hypercyclic operator.

Example 12.7. We consider the space ϕ of finite sequences,

ϕ = {(xn )n ∈ KN ; there is some m ∈ N such that xn = 0 for all n > m}.

The space ϕ has a natural locally convex topology, which is the strongest one
that can be defined on it; it is generated by the family of norms
∞

xv = |xn |vn , x ∈ ϕ,
n=1

where v = (vn )n is an arbitrary sequence of strictly positive numbers. Now

let x ∈ ϕ be a hypercyclic vector for an operator T on ϕ. We set En = {x ∈
ϕ ; xk = 0 for k > n}, n ∈ N. Suppose that each En contains only a finite
number of elements of the orbit of x. We define F1 = orb(x, T ) ∩ (E1 \ {0})

∞ Fn = orb(x, T ) ∩ (En \ En−1 ), n > 1. Then each Fn is finite, orb(x, T ) =

and
n=1 Fn , and every element y ∈ Fn satisfies yn = 0. We can therefore define

1
vn = , n ∈ N,
min{|yn | ; y ∈ Fn }

if Fn is nonempty, and vn = 1 otherwise. Considering the sequence v = (vn )n

we then find that yv ≥ 1 for every y ∈ orb(x, T ), which contradicts the
hypercyclicity of x. Therefore, some En , n ≥ 1, must contain an infinite
number of elements of orb(x, T ), which is impossible because En is finite di-
mensional and the vectors in a dense orbit are linearly independent; note that
Proposition 2.60 continues to hold. Consequently, ϕ admits no hypercyclic
operator.

By the Birkhoff transitivity theorem, an operator on a separable Fréchet

space is hypercyclic if and only if it is topologically transitive. One implication
remains true since no topological vector space has isolated points.

Observation 12.8. Any hypercyclic operator on a topological vector space is

topologically transitive.

But the converse is no longer true, as the following example shows.

Example 12.9. We consider again the space X = ϕ of finite sequences, but

this time endowed with the topology inherited from 2 . Consider the multiple
of the backward shift operator T = 2B : X → X, B(xn )n = 2(xn+1 )n . Then
T is topologically transitive, and even mixing, because Rolewicz’s operator
is. On the other hand, T cannot be hypercyclic since the orbit of any vector
in X is finite.
336 12 Linear dynamics in topological vector spaces

Example 12.10. For a separable Banach space X, let us consider the space
L(X) of operators on X, endowed with the strong operator topology. In
Theorem 10.20 we proved that an operator T on X satisfies the Hypercyclicity
Criterion if, and only if, the left-multiplication operator LT : L(X) → L(X),
S → T S, is hypercyclic. This characterization provides a good collection of
hypercyclic operators on the non-metrizable locally convex space L(X).

Since hypercyclicity and topological transitivity no longer coincide, we

adopt Devaney’s original definition of chaos in the general setting.

Definition 12.11 (Linear chaos). An operator T on a topological vector

space X is said to be chaotic if it satisfies the following conditions:
(i) T is topologically transitive;
(ii) T has a dense set of periodic points.

We recall the useful result, Proposition 2.33, that the set Per(T ) of periodic
points for an operator T on a complex space X is given by

Per(T ) = span{x ∈ X ; T x = eαπi x for some α ∈ Q},

whose density can often be checked easily for a concrete operator T .

Example 12.12. Let T be a chaotic operator on a separable Banach space X.

Then there is a countable dense set E ⊂ X such that each element of E is a
periodic point for T . In the proof of Proposition 10.14 we showed that, if Φ
is a countable weak-∗-dense subset of X ∗ , then the countable set

m 
F = FΦ,E =  · , yj∗ ej ; yj∗ ∈ Φ, ej ∈ E, 1 ≤ j ≤ m
j=1

of operators on X is SOT-dense in L(X). We have that every element of

F is periodic for the left-multiplication operator LT on L(X). On the other
hand, hypercyclicity, and therefore topological transitivity, of LT follows from
Theorem 10.20 and the fact that chaotic operators satisfy the Hypercyclicity
Criterion; see Theorem 3.18. We thus conclude that LT is chaotic.

We next want to show that many fundamental results for hypercyclic op-
erators on Fréchet spaces extend to arbitrary topological vector spaces. For
this we need the notion of a quotient space.
Let X be a vector space and L ⊂ X a subspace. Defining x ∼ y if x−y ∈ L,
we obtain an equivalence relation on X. Let us denote by [x] = x + L the
equivalence class of x ∈ X, and by X/L the set of equivalence classes. Then
X/L inherits in a natural way a vector space structure, and we denote by
q : X → X/L, x → [x], the quotient map, which is linear and surjective.
If, now, X is a topological vector space and L ⊂ X is a closed subspace,
then X/L becomes a topological vector space, called the quotient space of X
modulo L, when endowed with the induced topology:
12.2 Hypercyclicity, topological transitivity, and linear chaos 337

 ⊂ X with q(U
U ⊂ X/L is open if and only if there is an open set U  ) = U.

The quotient map q is then a continuous and open map. The requirement
that L be closed is necessary for the Hausdorff property of X/L; see Exer-
cise 12.2.5.
The following result, which generalizes Bourdon’s theorem, is the key to
the announced extensions.

Lemma 12.13 (Wengenroth). Let T be an operator on topological vector

space X. If either
(i) T is topologically transitive, or
(ii) T has a somewhere dense orbit,
then, for any nonzero polynomial p, the operator p(T ) has dense range.

Proof. We will only show the complex case. The real case can be deduced in
a similar way after some minor considerations; see Exercise 12.2.6.
As in the proof of Bourdon’s theorem it suffices to show that T − λI has
dense range for every λ ∈ C. Let L = (T − λI)(X), which is a closed subspace
of X, and suppose that L = X. We then consider the quotient space X/L,
which is nontrivial, and the quotient map q : X → X/L. Since, for any x ∈ X,
q((T − λI)x) = 0 we have that q(T x) = λq(x). Hence the operator S on X/L
given by S[x] = λ[x] is quasiconjugate to T via q and therefore inherits the
stated properties from T ; see the following section.
Under assumption (i), S is topologically transitive. On the other hand, let
[x] ∈ X/L, [x] = 0. By the Hausdorff property there is an open neighbourhood
U of [x] and a balanced 0-neighbourhood W such that U ∩ W = ∅. Now, if
|λ| ≥ 1, then W ⊂ λn W , and therefore S n (U ) ∩ W = ∅ for all n ∈ N0 . And
if |λ| < 1, then λn W ⊂ W , and therefore S n (W ) ∩ U = ∅ for all n ∈ N0 .
Thus, for any λ ∈ C, S is not topologically transitive, a contradiction.
Under assumption (ii), S has a somewhere dense orbit {λn [x] ; n ∈ N0 }.
Then span{[x]} = span{[x]} = X/L (see Exercises 12.1.1(v) and 12.1.2), so
that X/L is isomorphic to C. But every orbit {λn z ; n ∈ N0 }, z ∈ C, is
nowhere dense, a contradiction.  

Now, looking back at the proofs of the following fundamental results we

see that they work unrestrictedly once one has Wengenroth’s lemma at hand.
They therefore hold for operators on all topological vector spaces.
Herrero–Bourdon theorem. Any hypercyclic operator admits a dense in-
variant subspace consisting, except for zero, of hypercyclic vectors.
Ansari’s theorem. Any power of a hypercyclic operator is hypercyclic.
Costakis–Peris theorem. Any multi-hypercyclic operator is hypercyclic.
Bourdon–Feldman theorem. Any somewhere dense orbit is (everywhere)
dense.
Indeed, each result holds in the more detailed form given in Chapters 2 and 6.
338 12 Linear dynamics in topological vector spaces

12.3 Dynamical transference principles

Working with general topological vector spaces instead of F-spaces sometimes

requires some abstract considerations; for example, instead of sequences,
balls, or the distance one needs to use notions like nets or neighbourhoods. In
addition, of course, one has to do without the Baire category theorem. There
do exist topological vector spaces beyond F-spaces in which Baire’s theorem
holds but they are rare.
In this section we want to discuss three techniques that allow us to transfer
dynamical properties from operators on F-spaces to operators on general
topological vector spaces.
The first technique is by now well known, that of quasiconjugacies. As
before, if X and Y are topological vector spaces then an operator T on X is
called quasiconjugate to an operator S on Y via a continuous map φ : Y → X
with dense range if T ◦ φ = φ ◦ S. Then the usual notions of linear dynamics
are preserved under quasiconjugacy: hypercyclicity, topological transitivity,
(weak) mixing, chaos, frequent hypercyclicity, etc. Moreover, if y ∈ Y is a
hypercyclic vector for S, then x := φ(y) is hypercyclic for T .
A particular case of quasiconjugacy that frequently occurs naturally is
when one can find a T -invariant dense subspace Y ⊂ X that carries its own,
not necessarily the induced, vector space topology such that the restriction
T |Y is an operator Y . If, in addition, the embedding Y → X is continuous,
then T is quasiconjugate to T |Y , so that T inherits dynamical properties from
T |Y . This is commonly known as the hypercyclic comparison principle; see
Exercise 2.2.6.
Now, if Y is, in particular, an F-space, then the results of the previous
chapters can be applied. We illustrate this by an example.
Example 12.14. Let (Xn )n be an increasing sequence of Banach spaces with
continuous inclusions, and let X be the inductive limit of (Xn )n ; see Example
12.5. Suppose that T is an operator on X such that, for some n ≥ 1, Xn is
dense in X, T (Xn ) ⊂ Xn and T |Xn is continuous and hypercyclic. Then, by
the comparison principle, T is hypercyclic.

As a particular case, let 1 < p ≤ ∞, and consider the space p− := q<p q .

∞
Obviously, p− = n=1 pn for any strictly increasing sequence (pn )n in ]1, p[
tending to p. A natural topology on p− is the corresponding inductive limit
topology. If λ ∈ K is any scalar with |λ| > 1, then the multiple T = λB of the
backward shift satisfies the above requirements and is therefore hypercyclic
on p− .
The second method is a kind of converse to the first technique. Of course,
if, for a given operator T , all operators S that are quasiconjugate to T are
hypercyclic then T itself must be hypercyclic; one may simply take S =
T . It is, however, remarkable that the result remains true when we only
admit operators S defined on F-spaces. In addition, the map φ defining the
quasiconjugacy may be required to be linear.
12.3 Dynamical transference principles 339

Proposition 12.15. Let T be an operator on a topological vector space X,

and x ∈ X.
(a) If every operator S, defined on an F-space and quasiconjugate to T
via a linear map, is hypercyclic (topologically transitive, weakly mixing, or
mixing), then the same holds for T .
(b) If for any operator S, defined on an F-space and quasiconjugate to
T via a linear map φ, φ(x) is (frequently) hypercyclic for S, then x is (fre-
quently) hypercyclic for T .
(c) If X is a locally convex space then it suffices in (a) and (b) to allow
operators S on Fréchet spaces.
Proof. We will only show assertion (a) for the mixing property; the remaining
cases follow similarly.
Let U, V ⊂ X be arbitrary nonempty open subsets. Let x1 ∈ U , x2 ∈ V ,
and choose W ∈ U0 (X) such that x1 +W ⊂ U and x2 +W ⊂ V . By continuity,
we obtain a decreasing sequence (Wn )n of closed balanced 0-neighbourhoods
∞ W1 = W , Wn+1 + Wn+1 ⊂ Wn , and T (Wn+1 ) ⊂ Wn , n ∈ N. Let
such that
L := n=1 Wn , which is easily seen to be a closed T -invariant subspace of
X. We set Y  = X/L, and endow it with the topology τ generated by the
family of neighbourhoods {y  + W n ; y  ∈ Y  , n ∈ N}, where W n is the

image of Wn under the quotient map q : X → Y , n ∈ N. It is routine to
verify that (Y  , τ ) is a topological vector space, which is metrizable since it
has a countable base of 0-neighbourhoods, and that the operator T induces
an operator S  : Y  → Y  that is quasiconjugate to T via q.
Now let Y be the completion of (Y  , τ ), which is an F-space, S : Y → Y
the extension of S  to the completion, and φ : X → Y the operator induced
by q, which has dense range. It is clear that S is quasiconjugate to T via
the linear map φ. It then follows from the assumption that S is a mixing
operator. Therefore, also S  is mixing, so that there is some N ∈ N0 such
that
n  
2 ∩ q(x2 ) + W

2 = ∅
q(T n (x1 + W2 )) ∩ q(x2 + W2 ) = (S  ) q(x1 ) + W

for every n ≥ N . This implies that

T n (U ) ∩ V ⊃ T n (x1 + W ) ∩ (x2 + W ) ⊃ T n (x1 + W2 ) ∩ (x2 + W2 + L) = ∅

for every n ≥ N , so that T is mixing.

In the case that X is a locally convex space, the 0-neighbourhoods Wn ,
n ∈ N, can be chosen to be absolutely convex, and Y is a Fréchet space. 

An application of this result yields the generalization of the León–Müller
theorem to arbitrary complex topological vector spaces.
Corollary 12.16. Let T be an operator on a complex topological vector space
X. Then, for any λ ∈ C with |λ| = 1, T and λT have the same hypercyclic
vectors, that is, HC(T ) = HC(λT ).
340 12 Linear dynamics in topological vector spaces

Proof. Let x ∈ HC(T ) and λ ∈ C with |λ| = 1. By Proposition 12.15 it

suffices to show that every operator S, defined on an arbitrary F-space Y and
quasiconjugate to λT via a linear map φ : X → Y , has φ(x) as a hypercyclic
vector. But under these assumptions, λ−1 S is quasiconjugate to T via φ, so
that φ(x) is a hypercyclic vector for λ−1 S. Since Y is an F-space, we can
apply Theorem 6.7; note that its proof also works in arbitrary F-spaces. We
then obtain that φ(x) is hypercyclic for S, as demanded.  

As another application of Proposition 12.15 we show that the space ϕ of

finite sequences supports a mixing operator, which is surprising in view of
the fact that ϕ does not admit any hypercyclic operator.

Example 12.17. We claim that the operator T = I + B on ϕ is mixing, where

B is the backward shift. Thus, let S : Y → Y be an operator on an arbitrary
Fréchet space Y that is quasiconjugate to T via a linear map φ : ϕ → Y . We
fix an increasing sequence of seminorms (pn )n on Y generating its topology;
by the continuity of φ there are strictly positive sequences v(n) = (vn,k )k ,
n ∈ N, such that pn (φ(x)) ≤ xv(n) , for any n ∈ N and x ∈ ϕ; see Example
12.7. Defining v = (vk )k by vk = maxn,m≤k vn,m , k ∈ N, a simple calculation
shows that there are constants Mn > 0 such that pn (φ(x)) ≤ Mn xv for
any n ∈ N, x ∈ ϕ. Since Y is complete, there is a continuous extension
φ : 1 (v) → Y , and S is quasiconjugate to I + B : 1 (v) → 1 (v) via φ. Since,
by Theorem 8.2, I + B is mixing on 1 (v), so is S on Y . Proposition 12.15
then implies that T is mixing on ϕ.

We turn to the third transference principle. For this we need a new concept.
A projective spectrum X of Fréchet spaces consists of a family (Xα )α∈I of
Fréchet spaces, where I is a directed index set, and operators α β : Xβ → X α
for α ≤ β, called the spectral maps, that satisfy β ◦ γ = γ and α
α β α
α = IXα ,
the identity on Xα , for any α ≤ β ≤ γ. The projective limit of X is defined
as   
proj X = (xα )α∈I ∈ Xα ; αβ xβ = xα for all α ≤ β ,
α∈I

endowed with the topology inherited from the product topology on α∈I Xα ;
in this way, proj X is a locally convex space. We denote by α : proj X → Xα
the projection onto the component with index α. It is not difficult to see that
the sets (α )−1 (Wα ), Wα ∈ U0 (Xα ), α ∈ I, form a base of 0-neighbourhoods
for the topology of proj X . We say that X is strongly reduced if for each α
there is a larger β such that αβ (Xβ ) is contained in the closure of  (proj X )
α

in Xα .
Now, a family (Tα )α∈I of operators Tα on Xα is called an endomorphism
of X if their elements commute with the spectral maps in the sense that, for
any α ≤ β, Tα ◦ α β = β ◦ Tβ . The projective limit of the endomorphism is
α

the operator T on proj X defined by T (xα )α∈I = (Tα xα )α∈I .

12.3 Dynamical transference principles 341

Proposition 12.18. Let X be a strongly reduced projective spectrum of

Fréchet spaces, (Tα )α∈I an endomorphism of X , and T its projective limit.
(a) If every Tα , α ∈ I, is topologically transitive (mixing, weakly mixing)
on Xα then T is topologically transitive (mixing, weakly mixing) on proj X .
(b) If x ∈ proj X is such that, for every α ∈ I, α x ∈ Xα is (frequently)
hypercyclic for Tα then x is (frequently) hypercyclic for T .

Proof. (a) We will only show the result for the mixing property. Let x, y ∈
X := proj X and W0 ∈ U0 (X) be given. Then there are α ∈ I and
W1 ∈ U0 (Xα ) with W0 ⊃ (α )−1 (W1 ), and there is some β ≥ α with
αβ Xβ ⊂  (X). For each W ∈ U0 (Xα ) we obtain that β (Xβ ) ⊂  (X) +
α α α
α −1
W , and thus Xβ ⊂  (X) + (β ) (W ). This means that the image of
β

β is dense in Xβ with respect to the vector space topology τ having

−1
{(αβ) (W ) ; W ∈ U0 (Xα )} as a base of 0-neighbourhoods. Moreover, Tβ
is continuous on (Xβ , τ ) since Tβ−1 ((α
β)
−1
(W )) = (α
β)
−1
(Tα−1 (W )) ∈ τ for
every W ∈ U0 (Xα ), and Tβ is mixing on (Xβ , τ ) since τ is coarser than the
original topology on Xβ .
Hence there is some N ∈ N0 such that

Un := (β x + (α
β)
−1
(W1 )) ∩ (Tβ−n (β y + (α
β)
−1
(W1 ))) = ∅

for all n ≥ N . Since Un is open with respect to τ , there are zn ∈ X such that
β zn ∈ Un for all n ≥ N . Then we have that α (zn − x) = α β ( zn −  x) ∈
β β

W1 , that is, zn ∈ x + W0 , and

β ( (T zn )) = β (Tβ ( zn )) ∈  y + W1 ,
α (T n zn ) = α β n α n β α

which gives that T n zn ∈ y + W0 for every n ≥ N . This proves that

(x + W0 ) ∩ (T −n (y + W0 )) = ∅,

for each n ≥ N , and therefore that T is mixing on X.

(b) Now let x ∈ X be such that, for every α ∈ I, α x ∈ Xα is hypercyclic
for Tα . Let y ∈ X and W0 ∈ U0 (X) be given. Then there are α ∈ I and
W1 ∈ U0 (Xα ) with W0 ⊃ (α )−1 (W1 ). It follows that there is some n ∈ N0
such that α T n x = Tαn α x ∈ α y + W1 , hence T n x − y ∈ (α )−1 (W1 ) ⊂ W0 .
This shows that x is hypercyclic for T . The same argument also shows the
claim for frequent hypercyclicity.  

We single out a particular case of this result. Let (Xn )n be a decreasing

sequence of Fréchet spaces such that each inclusion map in : Xn+1 → Xn ,
n ≥ 1, is continuous. For any n ≥ 1, let (pn,k )k be an increasing sequence
of
∞ seminorms defining the topology of Xn . We then consider the space X :=
n=1 Xn with the locally convex topology induced by the seminorms pk (x) :=
maxn≤k pn,k (x), x ∈ X. Then X is a Fréchet space, also called the projective
limit of (Xn )n .
342 12 Linear dynamics in topological vector spaces

Taking the inclusion maps as spectral maps, we see that (Xn )n is also
∞ and proj X is the space of constant sequences with
a projective spectrum,
entries from X = n=1 Xn . It is then clear that proj X is isomorphic to X.
Moreover, the projective spectrum is strongly reduced if, for example, X is
dense in each Xn , n ≥ 1.

Corollary 12.19. Let (Xn )n be a decreasingsequence of Fréchet spaces with

continuous inclusion maps such that X = ∞ n=1 Xn is dense in Xn for all
n ≥ 1. Let T : X → X be an operator that can be extended to an operator
Tn : Xn → Xn for any n ≥ 1.
(a) If every Tn , n ≥ 1, is topologically transitive (mixing, weakly mixing)
on Xn then T is topologically transitive (mixing, weakly mixing) on X.
(b) If x ∈ X is (frequently) hypercyclic for every Tn , n ≥ 1, then x is
(frequently) hypercyclic for T .

Proof. This follows directly from Proposition 12.18 because, by the assump-
tions, (Tn )n is an endomorphism of proj X and the projective limit of (Tn )n
on proj X turns into T via the identification of proj X with X.  

We note that, by quasiconjugacy, the conditions in the corollary are also

necessary.

Example 12.20. Let X = L∞− [0, 1] := p<∞ Lp [0, 1] be endowed with the
Fréchet space topology induced by the increasing sequence of norms (pn )n ,
where pn is the norm of Ln [0, 1], n ∈ N. Let C : L∞− [0, 1] → L∞− [0, 1] be
t
the Cesàro operator given by Cf (t) = 1t 0 f (s) ds. By Exercise 3.1.4, C is
mixing on Lp [0, 1] for any 1 < p < ∞. Hence C is mixing on L∞− [0, 1].

12.4 Mixing and weakly mixing operators

In this section we will convince ourselves that the central results of Sections
2.4 and 2.5 remain true in general topological vector spaces when we replace
the assumption of hypercyclicity there by topological transitivity. We will
omit the proofs when the arguments given in those sections translate directly
to the general situation.

Proposition 12.21. An operator T on a topological vector space X is mixing

if and only if, for any nonempty open set U ⊂ X and any 0-neighbourhood
W , the sets
N (U, W ) and N (W, U )
are cofinite.

Example 12.22. We pointed out in Example 12.17 that there are mixing op-
erators on ϕ. Another easy example of a mixing operator is T = λB, |λ| > 1,
12.4 Mixing and weakly mixing operators 343

on the inductive limit p− for 1 < p ≤ ∞; see Example 12.14. Also, if X is
an arbitrary topological vector space, then the product space X N , endowed
with the product topology, is a topological vector space, and the backward
shift B : X N → X N , (x1 , x2 , . . . ) → (x2 , x3 , . . . ) is a mixing operator. Thus,
if I is an infinite set then the space X I , endowed with the product topology,
is isomorphic to (X I )N and therefore admits a mixing operator.

We turn to the weak mixing property. As before, under mild additional

assumptions, topologically transitive operators turn out to be weakly mixing.
We start with a useful auxiliary result.

Lemma 12.23. Let T be a topologically transitive operator on a topological

vector space X. Then, for any nonempty open sets U and V in X and for
any 0-neighbourhood W , there is a nonempty open set U1 ⊂ U and a 0-
neighbourhood W1 ⊂ W such that

N (U1 , W1 ) ⊂ N (V, W ) and N (W1 , U1 ) ⊂ N (W, V ).

From this we can deduce the main result of this section.

Theorem 12.24. Let T be a topologically transitive operator on a topolog-

ical vector space X. If, for any nonempty open set U ⊂ X and any 0-
neighbourhood W , there is a continuous map S : X → X commuting with T
such that
S(U ) ∩ W = ∅ and S(W ) ∩ U = ∅,
then T is weakly mixing.

Recall that an operator T is flip transitive if, for any pair U, V ⊂ X of

nonempty open sets, N (U, V ) ∩ N (V, U ) = ∅. Thus we have in particular:

Corollary 12.25. Every flip transitive operator is weakly mixing.

Theorem 12.24 also implies that Theorem 2.47 extends to general topo-
logical vector spaces.

Theorem 12.26. An operator T on a topological vector space X is weakly

mixing if and only if, for any nonempty open sets U, V ⊂ X and any 0-
neighbourhood W ,
N (U, W ) ∩ N (W, V ) = ∅.

Another application of Theorem 12.24 provides us with a useful sufficient

condition for a topologically transitive operator to be weakly mixing; see
Exercise 12.2.4 for the notion of a bounded set.

Theorem 12.27. Let T be a topologically transitive operator on a topological

vector space X. If there exists a dense subset X0 of X such that the orbit of
each x ∈ X0 is bounded, then T is weakly mixing.
344 12 Linear dynamics in topological vector spaces

Proof. In order to adapt the proof of Theorem 2.48 one need only note that,
for any x ∈ X0 and any 0-neighbourhood W , there is some ε > 0 such that
εT n x ∈ W for all n ∈ N0 . 

∞
Recall that the generalized kernel of an operator T is given by n=0 ker T n .
Corollary 12.28. Let T be a topologically transitive operator on a topological
vector space X. If one of the following conditions is satisfied:
(i) T is chaotic;
(ii) T has a dense set of points for which the orbits converge;
(iii) T has dense generalized kernel;
then T is weakly mixing.
As a final application of Theorem 12.24 we will characterize weakly mixing
operators by the behaviour of multiples of iterates of T . For the notion of
topological transitivity for sequences of operators we refer to Section 1.6.
Theorem 12.29. Let T be an operator on a topological vector space X and
λ, μ ∈ K \ {0} with λ = μ. Then the following assertions are equivalent:
(i) T is weakly mixing;
(ii) for any M > δ > 0 and for any (λn )n with δ ≤ |λn | ≤ M , n ∈ N0 , the
sequence (λn T n )n is topologically transitive;
(iii) for any (λn )n with {λn ; n ∈ N0 } ⊂ {λ, μ}, the sequence (λn T n )n is
topologically transitive.
Proof. (i) =⇒ (ii). Given (λn )n with δ ≤ |λn | ≤ M , n ∈ N0 , we let U, V ⊂ X
be nonempty open sets. By Exercise 12.1.1 there are nonempty open sets U1
and V1 and a 0-neighbourhood W such that U1 +W ⊂ U and V1 +W ⊂ V . By
Proposition 12.1, if L = max( 1δ , M ), then there is a 0-neighbourhood W1 such
that λW1 ⊂ W , for any λ ∈ K with |λ| ≤ L. Now, since T is weakly mixing,
there are n ∈ N0 , u ∈ U1 and w ∈ W1 such that T n u ∈ W1 and T n w ∈ V1 .
Thus u + λ−1 n −1
n w ∈ U1 + W ⊂ U and λn T (u + λn w) = λn T u + T w ∈
n n

W + V1 ⊂ V . Hence (λn T )n is topologically transitive.

n

(ii) =⇒ (iii) is trivial.

(iii) =⇒ (i). By taking (λn )n to be a constant sequence, T is easily seen
to be topologically transitive. It therefore suffices to verify the hypothesis of
Theorem 12.24. Thus let W be a 0-neighbourhood and U ⊂ X a nonempty
open set. Let x ∈ U . Using the properties of Proposition 12.1 we can find
some M > 0 and an open neighbourhood U1 ⊂ U of x such that

M (λ − μ) λ μ
U1 ⊂ W, U1 − U1 ⊂ M −1 W, and U1 − U1 ⊂ U.
λ λ−μ λ−μ

Let α = M (λ − μ). The hypothesis implies that there is some n ∈ N0 such

that λT n (U1 ) ∩ αU1 = ∅ and μT n (U1 ) ∩ αU1 = ∅; otherwise there would
exist a sequence (λn )n with entries λ or μ such that λn T n (U1 ) ∩ αU1 = ∅
for all n ∈ N0 . Thus there are u1 , u2 ∈ U1 with T n (α−1 λu1 ) ∈ U1 and
T n (α−1 μu2 ) ∈ U1 . Then
12.5 Criteria for weak mixing, mixing and chaos 345

λ
α−1 λu1 ∈ U1 ⊂ W, T n (α−1 λu1 ) ∈ U,
M (λ − μ)

and

M α−1 (λu1 − μu2 ) ∈ U, T n (M α−1 (λu1 − μu2 )) ∈ M (U1 − U1 ) ⊂ W.

We then conclude with Theorem 12.24. 



12.5 Criteria for weak mixing, mixing and chaos

We extend here the criteria of Chapter 3 to general topological vector spaces.

Since the arguments given there can be adapted directly to the general situ-
ation we omit the proofs.
Following the same order, we start with the criterion based on a large
supply of eigenvectors.

Theorem 12.30 (Godefroy–Shapiro criterion). Let T be an operator on

a topological vector space X. Suppose that the subspaces

X0 := span{x ∈ X ; T x = λx for some λ ∈ K with |λ| < 1},

Y0 := span{x ∈ X ; T x = λx for some λ ∈ K with |λ| > 1}

are dense in X. Then T is mixing.
If, moreover, X is a complex space and also the subspace

Z0 := span{x ∈ X ; T x = eαπi x for some α ∈ Q}

is dense in X, then T is chaotic.

Kitai’s Criterion for mixing extends likewise.

Theorem 12.31 (Kitai’s criterion). Let T be an operator on a topological

vector space X. If there are dense subsets X0 , Y0 ⊂ X and a map S : Y0 → Y0
such that, for any x ∈ X0 , y ∈ Y0 ,
(i) T n x → 0,
(ii) S n y → 0,
(iii) T Sy = y,
then T is mixing.

Example 12.32. Let X = Lp [0, 1], 0 < p < 1, be the space of p-integrable
functions on [0, 1]; see Example 12.4. Let ϕ : [0, 1] → [0, 1] be the invertible
function given by ϕ(t) = t/2 if t ∈ [0, 1/2], and ϕ(t) = (3/2)t − 1/2 if
t ∈ ]1/2, 1]. We then consider the composition operator Cϕ : X → X, Cϕ f =
f ◦ ϕ. The set
346 12 Linear dynamics in topological vector spaces

X0 = Y0 = {f ∈ X ; f is continuous and f (0) = f (1) = 0},

is dense in X, and for S : Y0 → Y0 we choose the composition operator S =

Cϕ−1 , so that T Sf = f for all f ∈ Y0 . Moreover, it is easy to check that, if t ∈
[0, 1[, then limn→∞ ϕn (t) = 0, which implies that Cϕn f → 0 for every f ∈ X0
by the dominated convergence theorem. Analogously, limn→∞ (ϕ−1 )n (t) = 1
for all t ∈ ]0, 1], so that Cϕn−1 f → 0 for every f ∈ Y0 . An application of
Kitai’s criterion shows that T is mixing.

Finally, the Hypercyclicity Criterion turns out to be a weak mixing crite-

rion within the general framework.

Theorem 12.33. Let T be an operator on a topological vector space X. If

there are dense subsets X0 , Y0 ⊂ X, an increasing sequence (nk )k of positive
integers, and maps Snk : Y0 → X, k ≥ 1, such that, for any x ∈ X0 , y ∈ Y0 ,
(i) T nk x → 0,
(ii) Snk y → 0,
(iii) T nk Snk y → y,
then T is weakly mixing.

By Example 12.9 then, an operator satisfying this criterion need not be

hypercyclic. With this realization that the Hypercyclicity Criterion is a mis-
nomer we conclude the book.

Exercises

Exercise 12.1.1. Let X be a topological vector space. Prove the following assertions:
(i) if x ∈ X and W is a 0-neighbourhood, then there exists some M > 0 and a
neighbourhood U of x such that U ⊂ M W ;
(ii) if U is a nonempty open set, then there is a 0-neighbourhood W and a nonempty
open set U1 ⊂ U such that U1 + W ⊂ U ;
(iii) if U is a nonempty open set, λ, μ ∈ K with λ + μ = 0, and x ∈ U , then there is a
neighbourhood U1 ⊂ U of x such that λU1 + μU1 ⊂ (λ + μ)U ;
(iv) for any λ ∈ K \ {0} and y ∈ X, the operators Mλ : X → X, x → λx, and
Ty : X → X, x → x + y, are homeomorphisms;
(v) if A ⊂ X is somewhere dense in X then span A is dense in X.

Exercise 12.1.2. Show that every finite-dimensional topological vector space X over
the field K = R or C is isomorphic to KN , where N is the dimension of X. Deduce that
finite-dimensional subspaces of topological vector spaces are closed. Here, as usual, an
isomorphism between two topological vector spaces is, by definition, a linear homeomor-
phism.

Exercise 12.1.3. Given 0 < p < 1 and a < b, show that the vector space
  b 
X = Lp [a, b] = f : [a, b] → K ; f is measurable and |f (t)|p dt < ∞
a
Exercises 347

is an F-space if we endow it with the base of neighbourhoods (g + Wn )n , g ∈ X, where

b
Wn = {f ∈ X ; a |f (t)|p dt < 1/n}, n ∈ N. Prove that X is not a Fréchet space.

Exercise 12.1.4. (a) Let X be a vector space. Show that a subset A of X is absolutely
convex if and only if it is convex and balanced.
(b) Show that a topological vector space is locally convex if, and only if, it has a base
of 0-neighbourhoods consisting of convex sets.

Exercise 12.1.5. Given a vector space X, prove that, if A ⊂ X is an absolutely convex

set, the associated gauge of A,

pA (x) = inf{λ > 0 ; x ∈ λA}, x ∈ span A,

is a seminorm on span A.

Exercise 12.1.6. If X is an infinite-dimensional Fréchet space, then show that L(X)

endowed with the strong operator topology is not metrizable.

∞ such that each
Exercise 12.1.7. Let (Xn )n be an increasing sequence of Banach spaces
inclusion map in : Xn → Xn+1 , n ≥ 1, is continuous, and set X = n=1 Xn . Show that
the inductive limit topology defined in Example 12.5 is a locally convex topology on X.
Prove that it is not metrizable unless there is some m ∈ N such that Xn = Xm for all
n ≥ m.

Exercise 12.2.1. The weak topology on a Banach space X is the locally convex topology
on X defined by the seminorms x → |x, x∗ |, x∗ ∈ X ∗ ; that is, it is the topology of
pointwise convergence on X ∗ . An operator T on a Banach space X is called weakly
hypercyclic if it is hypercyclic on X endowed with the weak topology.
Let X = p , 1 ≤ p < ∞, or X = c0 . Show that a weighted backward shift Bw on X
is hypercyclic if and only if it is weakly hypercyclic.

Exercise 12.2.2. Construct a chaotic and mixing operator that is not hypercyclic.
(Hint: Enlarge the space in Example 12.9.)

Exercise 12.2.3. An operator T on a locally convex space X is called compact if there

exists some W ∈ U0 (X) such that T (W ) is compact. Show that no compact operator on
a locally convex space is hypercyclic, and that no compact perturbation of a multiple of
the identity is chaotic. (Hint: If W ∈ U0 (X) is absolutely convex and pW is the gauge
of W , then pW induces a norm on X/ ker pW . The completion of this normed space is
called the local Banach space XW . If W ∈ U0 (X) is absolutely convex such that T (W )
is compact, then consider the operator TW on XW induced by T .)

Exercise 12.2.4. A subset B of a topological vector space X is called bounded if, for any
W ∈ U0 (X), there is some M > 0 such that B ⊂ M W . If, in addition, B is absolutely
convex then XB := span B is a normed space when endowed with the gauge of B.
An operator T on X is called bounded if there is some U ∈ U0 (X) such that T (U ) is a
bounded subset of X. Show that a bounded operator T with dense range is hypercyclic
(or mixing, weakly mixing or chaotic) if and only if there is a bounded absolutely convex
set B ⊂ X such that XB is a T -invariant dense subspace of X and the induced operator
TB : XB → XB is hypercyclic (or mixing, weakly mixing or chaotic, respectively).
Moreover, if T is a bounded operator and λ ∈ K, show that there is some M > 0
such that M (λI + T ) is not hypercyclic.

Exercise 12.2.5. Let X be a topological vector space and L a subspace of X. Show that
X/L is Hausdorff if and only if L is closed. (Hint: Use some properties of Exercise 12.1.1.)
348 12 Linear dynamics in topological vector spaces

Exercise 12.2.6. Let X be a topological vector space over R, T an operator on X that

is topologically transitive or has a somewhere dense orbit, and p a nonzero polynomial
over R. Prove that p(T ) has dense range. (Hint: Define, as for Fréchet spaces, the com-
plexifications X  of X and T of T . As in the proof of Theorem 2.54 it suffices to show that
T −λI has dense range for all λ ∈ C. First case: proceed as in the proof of Theorem 2.54,
taking into account that T has the property that, for any U, V ⊂ X open and nonempty,
there is n ∈ N with Tn (U + iU ) ∩ (V + iV ) = ∅. Second case: there is x ∈ X such
that {T n x + iT m x ; n, m ≥ 0} is somewhere dense in X.  Applying the quotient map q,
deduce that X  can be identified with C. By continuity and openness of |q| : X → R+ ,
{|q(T n x)| ; n ≥ 0} is somewhere dense in R+ , but |q(T n x)| = |q(Tn x)| = |λ|n |q(x)|.)

Exercise 12.3.1. Given a compact subset K ⊂ C, a holomorphic germ on K is a

function that is defined and holomorphic on an open set U containing K. Let H(K) be
the space of holomorphic germs on K, and let A(K) be the Banach space of continuous
functions on K that are holomorphic on the interior of K, endowed with the sup-norm.
If (Kn )n is a 
decreasing sequence of compact sets such that the interior of each Kn
∞
contains K and n=1 Kn = K, then H(K) can be viewed as the inductive limit of the
increasing sequence (A(Kn ))n of Banach spaces. Prove that the differentiation operator
D is a well-defined operator on H(K), and that it is hypercyclic and chaotic if K is
connected with connected complement C \ K.

Exercise 12.3.2. Consider the weighted Banach space of holomorphic functions on the
unit disk,  
Hvn (D) := f ∈ H(D) ; f  := sup|f (z)|vn (z) < ∞ ,
z∈D

where vn (z) := (1 − |z|)n , n ∈ N; the inclusions Hvn (D) → Hvn+1 (D), n ≥ 1, are
continuous. The Korenblum space A−∞ is defined as the inductive limit of (Hvn (D))n .
Show that the differentiation operator D is a well-defined operator on A−∞ , and prove
that any finite-order differential operator on A−∞ that is not a multiple of the identity
is chaotic. In contrast, observe that D(Hvn (D)) ⊂ Hvn (D) for any n ∈ N, so that the
argument in Example 12.14 cannot be applied.

Exercise 12.3.3. Let w = (wn )n be a weight sequence and Bw the corresponding

weighted backward shift. Show that T := I + Bw is mixing on ϕ. (Hint: Show that T is
quasiconjugate to I + B on ϕ via a suitable diagonal operator.)

Exercise 12.3.4. Let X be a topological sequence space, that is, a topological vector
space X such that X ⊂ ω = KN with continuous inclusion. Suppose that ϕ is contained
and dense in X. If the weighted backward shift Bw is a well-defined operator on X,
prove that T := I + Bw is mixing on X.

Exercise 12.3.5. For this exercise we will need Young’s inequality: given any x ∈ p (Z)
and y ∈ q (Z), 1 ≤ p, q < 2, the convolution product
 
x ∗ y := xk yn−k ,
n∈Z
k∈Z

exists and belongs to r (Z), where 1/p + 1/q = 1/r + 1. Moreover, x ∗ yr ≤ xp yq .


We consider the Fréchet space 1+ (Z) = p>1 p (Z). The space 1+ is defined similarly.
(a) Show that, for any y ∈ 1+ (Z), the map y ∗ · : 1+ (Z) → 1+ (Z) given
∞ by x → y ∗x
defines an operator on 1+ (Z). Deduce that, for any function f (z) = α z n with
 ∞
n=0 n
(αn−1 )n ∈ 1+ , T = f (B) = n=0 n
α B n defines an operator on 1+ , where B is the
(unweighted) backward shift.
Sources and comments 349
∞
(b) Let λ ∈ K \ {0}, and suppose that f (z) = n=0 n
α z n with (αn−1 )n ∈ 1+ is
a nonconstant function such that there exist z1 , z2 ∈ K with |z1 |, |z2 | < 1 such that
|λf (z1 )| < 1 and |λf (z2 )| > 1. Show that λf (B) is hypercyclic on 1+ . If K = C and,
moreover, there exists z ∈ C with |z| < 1 such that |λf (z)| = 1, show that λf (B) is
chaotic on  1+
. As a consequence obtain that λT is chaotic on 1+ for every λ = 0,
∞
where T := n=1 n
B . (Hint: If |λ| < 1, then eλ := (λn )n is an eigenvector of f (B).)
1 n

Observe that T ( ) ⊂ p for any p ≥ 1, which shows that the above result cannot be
p

transferred from the Banach spaces defining the projective spectrum of 1+ .

Exercise 12.4.1. Let (Tn )n be a topologically transitive commuting sequence of oper-

ators on a topological vector space X such that, for any nonempty open set U ⊂ X and
any 0-neighbourhood W , there is a continuous map S : X → X commuting with all Tn ,
n ∈ N0 , such that
S(U ) ∩ U = ∅ and S(U ) ∩ W = ∅.
Show that (Tn )n is weakly mixing. (Hint: Given U, W , find U1 ⊂ U and W1 ⊂ W
with U1 − U1 ⊂ W , U1 − W1 ⊂ U . Apply the hypothesis and the 4-set trick to get that
N (U1 , U1 )∩N (U1 , W1 ) = ∅. Then deduce the result from the analogue of Theorem 12.24
for sequences of operators.)

Exercise 12.4.2. Inspired by Theorems 1.54 and 12.29, prove that an operator T on a
topological vector space X is weakly mixing if and only if, for any M > δ > 0, for any
(λn )n with δ ≤ |λn | ≤ M , n ∈ N0 , and for any syndetic sequence (nk )k , the sequence
(λk T nk )k is topologically transitive.

Exercise 12.5.1. Let ϕ : [0, 1] → [0, 1] be a continuous, surjective, and strictly increas-
ing function such that ϕ(t) = t for all t ∈ ]0, 1[. Show that the composition operator Cϕ
is mixing on Lp [0, 1] for any p > 0.

Exercise 12.5.2. Let T be an operator on a separable Banach space X satisfying the

Godefroy–Shapiro criterion (or the Hypercyclicity Criterion with respect to (nk )k ).
Prove that the left-multiplication operator LT on L(X), endowed with the strong op-
erator topology, satisfies the hypotheses of Theorem 12.30 (or the hypotheses of Theo-
rem 12.33 with respect to (nk )k , respectively).

Exercise 12.5.3. Let X be a Banach space with separable dual X ∗ and T an operator
on X whose adjoint T ∗ : X ∗ → X ∗ satisfies the Hypercyclicity Criterion (or is chaotic).
Show that the right-multiplication operator RT : L(X) → L(X), S → ST , is hypercyclic
and weakly mixing (or is hypercyclic and chaotic, respectively) on L(X), endowed with
the strong operator topology; see also Exercise 10.2.7.

Sources and comments

Section 12.1. All the basic results of this section can be found in the books by Meise and
Vogt [237] and Rudin [271]. For F-spaces we also refer to Kalton, Peck and Roberts [212].

Section 12.2. Dynamical properties of linear operators on topological vector spaces

beyond F-spaces were apparently first studied by Ansari [10]. The fact that the space
ϕ admits no hypercyclic operators was obtained by Bonet and Peris [85] and Grosse-
Erdmann [179]. The definition of chaos in general topological vector spaces was proposed
by Bonet [78], where one also finds Example 12.9. The crucial Lemma 12.13 is due to
Wengenroth [301].
350 12 Linear dynamics in topological vector spaces

We want to mention an interesting result on weakly topologically transitive operators,

that is, operators that are topologically transitive with respect to the weak topology; it
is due to Desch and Schappacher [130] (for Banach spaces) and Shkarin [286].

Theorem 12.34. An operator T on a complex locally convex space is weakly topologi-

cally transitive if and only if T ∗ has no eigenvalues.

Section 12.3. Proposition 12.15 is new, while Corollary 12.16 is due to Shkarin [284]
and Bayart and Matheron [44].
Example 12.17 and Proposition 12.18 are due to Bonet, Frerick, Peris and Wengen-
roth [81]. Since ϕ is thus a topological vector space without hypercyclic operators, but
that admits a mixing, and therefore topologically transitive, operator, one might wonder
if every infinite-dimensional topological vector space necessarily admits a topologically
transitive operator. This is not the case, as Bermúdez and Kalton [52] have shown: there
are (non-separable) Banach spaces, like ∞ , or L(2 ) with the operator norm, without
any topologically transitive operators. Further existence and nonexistence results con-
cerning hypercyclic or topologically transitive operators on locally convex spaces beyond
Fréchet spaces are due to Bonet and Peris [85], Bonet, Frerick, Peris and Wengenroth [81],
and Shkarin [286, 288, 293].
Example 12.20 solves a problem from León, Piqueras and Seoane [224].

Section 12.4. For the results of this section we refer to Grosse-Erdmann and Peris [187].
Similar investigations can be found in Bayart and Matheron [44], [45] and Moothathu
[245].

Section 12.5. Example 12.32 is from Grosse-Erdmann [179].

Exercises. Exercise 12.2.1 is taken from Chan and Sanders [101], where the authors
also show that there are weakly hypercyclic bilateral shifts that are not hypercyclic.
Exercise 12.2.2 is taken from Bonet [78], Exercise 12.2.3 from Bonet and Peris [85] and
Martínez and Peris [229]. The first part of Exercise 12.2.4 is from Bonet and Peris [85].
The second part is extracted from Bonet [80]; this paper studies the open problem of
the existence of non-normable Fréchet spaces X such that every operator on X is of the
form λI + T with T a bounded operator. It is also asked whether, for every infinite-
dimensional separable non-normable Fréchet space X, there exists an operator T on X
such that λT is hypercyclic for any λ = 0. Exercise 12.2.6 is taken from Wengenroth
[301]. For Exercise 12.3.2 we refer to Bonet [78], for Exercises 12.3.3 and 12.3.4 to Bonet,
Frerick, Peris and Wengenroth [81], and for Exercise 12.3.5 to Frerick and Peris [155].
Exercise 12.5.3 is taken from Bonet, Martínez and Peris [84].
Appendix A – Prerequisites

Throughout this book we suppose that the reader is familiar with metric
spaces, the basics of Hilbert and Banach space theory, and the fundamentals
of complex analysis. An introduction to the theory of Fréchet spaces and
their operators is given in Section 2.1. Some more advanced results of these
theories will be provided here, with suitable references.

Metric spaces

A good understanding of metric spaces and their topology is essential for

reading this book. A short introduction can be found in most texts on func-
tional analysis; for a more thorough account we refer to Shirali and Vasudeva
[282].
A point x in a metric space X is called isolated if some neighbourhood of
x contains no other point from X. A Gδ -set is the intersection of countably
many open sets. A set is of first Baire category if it is a countable union of
nowhere dense sets, otherwise of second Baire category.

Theorem A.1 (Baire category theorem). If X is a complete metric space

then the intersection of countably many dense open sets in X is dense in X.

For a proof see Rudin [270].

Banach spaces

We suppose that the reader has had a first introduction to Banach spaces,
as can be found, for example, in Chapter 5 of Rudin [270], Chapter III of
Conway [116], or Chapters 2–5 of Bollobás [77].

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 351

DOI 10.1007/978-1-4471-2170-1, © Springer-Verlag London Limited 2011
352 Appendix A

If X, Y are two normed spaces then L(X, Y ) denotes the space of (contin-
uous, linear) operators T : X → Y ; under the operator norm T → T  this
space turns into a Banach space whenever Y is a Banach space.
As a special case, L(X) = L(X, X) is the space of operators on X. The
operator norm then has the property that

ST  ≤ ST 

for S, T ∈ L(X). The identity operator is denoted by I. Two operators S and

T are said to commute if ST = T S.
If X is a normed space and Y is a Banach space then an operator T ∈
L(X, Y ) is called compact if the image of the closed unit ball is relatively
compact. Recall that a subset is said to be relatively compact if its closure
is compact. Thus, T is compact if and only if, for any sequence (xn )n in X
with xn  ≤ 1, n ≥ 1, the sequence (T xn )n has a convergent subsequence.
One then writes T ∈ K(X, Y ), with K(X) = K(X, X).
The dual X ∗ = L(X, K) of a normed space X is the space of all continuous
linear functionals on X. If x∗ ∈ X ∗ then we write

x∗ (x) = x, x∗ , x ∈ X.

The adjoint T ∗ : X ∗ → X ∗ of an operator T on X is defined by T ∗ x∗ = x∗ ◦T ,

that is,
x, T ∗ x∗  = T x, x∗ , x ∈ X, x∗ ∈ X ∗ .

Proposition A.2. Let S and T be operators on a Banach space X, and

λ ∈ K. Then
(i) I ∗ = I;
(ii) (S + T )∗ = S ∗ + T ∗ ;
(iii) (λT )∗ = λT ∗ ;
(iv) (ST )∗ = T ∗ S ∗ ;
(v) if T is invertible then (T ∗ )−1 = (T −1 )∗ ;
(vi) T ∗  = T .

Only the last assertion is not immediate; for a proof see [116, Chapter VI,
Proposition 1.4].

Example A.3. (a) The spaces p , 1 ≤ p < ∞, c0 and Lp [a, b] are recalled in
Example 2.4. The dual of p is given by q , where p1 + 1q = 1, in the sense
that the continuous linear functionals x∗ on p are precisely the maps of the
form
∞

x∗ (x) = x, x∗  = xn y n (A.1)
n=1

with y = (yn )n ∈ q , and we have that x∗  = y. In the same way the dual
of c0 is given by 1 .
Prerequisites 353

The dual of Lp [a, b], 1 ≤ p < ∞, is given by Lq [a, b] with 1p + 1q = 1, in the

sense that the continuous linear functionals x∗ on Lp [a, b] are precisely the
maps of the form
 b
x∗ (f ) = f, x∗  = f (t)g(t) dt (A.2)
a

with g ∈ Lq [a, b]; moreover, x∗  = g.

(b) Let v = (vn )n be a strictly positive sequence. Then the weighted spaces
p (v), 1 ≤ p < ∞, are defined as
 ∞
 
p
 (v) = (xn )n ; |xn |p vn < ∞ .
n=1

Under the representation (A.1), for p > 1, the dual of p (v) is given by q (w),
−q/p
where p1 + 1q = 1 and wn = vn , n ≥ 1; for p = 1, it is given by the space
of sequences (yn )n with supn≥1 |yn vn−1 | < ∞. This follows from the fact that
1/p
(xn )n → (xn vn )n defines an isometric isomorphism from p (v) to p .
(c) The analogous representation, via (A.2), holds for the duals of the
spaces Lpv (R+ ) introduced in Example 7.4; see also Exercise 2.1.6.

Of the four great principles in functional analysis, which will be stated

in the larger context of Fréchet spaces below, only the Banach–Steinhaus
theorem has a version that is specific to Banach spaces.

Theorem A.4 (Banach–Steinhaus theorem). Let X, Y be Banach spaces

and Tj : X → Y, j ∈ J, operators. If, for every x ∈ X, supj∈J Tj x < ∞,
then
sup Tj  < ∞.
j∈J

Further properties of Banach spaces will be treated in the context of

Fréchet spaces below.

Hilbert spaces

We suppose that the reader is familiar with the basic properties of Hilbert
spaces as can be found, for example, in Chapter 4 of Rudin [270], Chapter II
of Conway [116], or Chapter 9 of Bollobás [77].
Let H be a Hilbert space with inner product  · , · . We formulate two
consequences of the projection theorem.

Proposition A.5. A subspace M of H is dense if and only if only the zero

vector is orthogonal to M .
354 Appendix A

Theorem A.6 (Riesz representation theorem). Let y ∈ H. Then

x∗ (x) = x, y, x∈H

defines a continuous linear functional x∗ on H with x∗  = y. Conversely,

any continuous linear functional on H can be represented in this way, and
the vector y is uniquely determined by the functional.
Let T be an operator on H. As a consequence of the Riesz representation
theorem there exists a unique operator T ∗ : H → H such that, for all x, y ∈
H,
T x, y = x, T ∗ y.
It is called the (Hilbert space) adjoint of T .
Remark A.7. The Hilbert space adjoint T ∗ = T hil∗ is nothing but the (Ba-
nach space) adjoint T ∗ = T ban∗ of T as defined previously when we identify
the dual H ∗ of H with H itself via the Riesz representation theorem. More
precisely, if J : H ∗ → H denotes the Riesz representation of x∗ ∈ H ∗ by
y ∈ H, then T hil∗ = JT ban∗ J −1 . In that sense it is justified to use the no-
tation T ∗ for both adjoints, and it suggests using the notation  · , ·  for the
evaluation of continuous linear functionals in general Banach spaces.
We collect some useful properties.
Proposition A.8. Let S and T be operators on a Hilbert space H, and λ ∈
K. Then
(i) I ∗ = I;
(ii) (T ∗ )∗ = T ;
(iii) (S + T )∗ = S ∗ + T ∗ ;
(iv) (λT )∗ = λT ∗ ;
(v) (ST )∗ = T ∗ S ∗ ;
(vi) if T is invertible then (T ∗ )−1 = (T −1 )∗ ;
(vii) T ∗  = T .

Fréchet spaces

Section 2.1 gives a short introduction to the theory of Fréchet spaces. For
more detailed accounts we refer to Rudin [271] and Meise and Vogt [237]. We
collect here some further definitions and results that are used in this book;
their proofs can be found in the two books mentioned or in the references
provided.
The dual X ∗ of a Fréchet space X and the adjoint T ∗ of an operator T
on X are defined as in the case of Banach spaces. However we will not (need
to) address the problem of topologizing X ∗ , which is more delicate than for
Banach spaces; as a consequence we will consider T ∗ only as a linear map.
Prerequisites 355

The compactness of an operator T : X → Y from a normed space X into

a Fréchet space Y is defined as in the case of Banach spaces.
In a Banach space a set A is bounded if supx∈A x < ∞. In a Fréchet
space one cannot use the F-norm to the same effect because, for example, the
F-norm (2.2) is always bounded.

Definition A.9. Let X be a Fréchet space with defining increasing sequence

of seminorms (pn )n . Then a subset A of X is bounded if, for any n ≥ 1,

sup pn (x) < ∞.

x∈A

If X and Y are Fréchet spaces, then a family (Tj )j∈J of operators Tj :

X → Y , j ∈ J, is called equicontinuous if for any neighbourhood W of 0
in Y there is a neighbourhood V of 0 in X such that Tj (V ) ⊂ W for all
j ∈ J. In other terms, if (pn )n and (qn )n are defining increasing sequences of
seminorms on X and Y , respectively, then (Tj )j∈J is equicontinuous if and
only if, for any m ≥ 1, there are n ≥ 1 and M > 0 such that, for any j ∈ J,

qm (Tj x) ≤ M pn (x), x ∈ X.

We turn to the four great principles of functional analysis.

Theorem A.10 (Banach–Steinhaus theorem). Let X and Y be Fréchet

spaces and Tj : X → Y , j ∈ J, operators.
(a) If, for every x ∈ X, the set {Tj x : j ∈ J} is bounded in Y , then the
family (Tj )j∈J is equicontinuous.
(b) If J = N and, for every x ∈ X, the sequence (Tn x)n converges in Y ,
then T x = limn→∞ Tn x, x ∈ X, defines an operator T : X → Y .

Theorem A.11 (Open mapping theorem). Let X, Y be Fréchet spaces.

If T : X → Y is a surjective operator, then T is an open map, that is, for
any open set U ⊂ X, T (U ) is open in Y .

As a consequence, we have the following.

Corollary A.12 (Inverse mapping theorem). Let X, Y be Fréchet spaces.

If T : X → Y is a bijective operator, then T has a continuous linear inverse
T −1 .

Theorem A.13 (Closed graph theorem). Let X, Y be Fréchet spaces. If

a linear map T : X → Y has closed graph, that is, if xn → x in X and
T xn → y in Y implies that T x = y, then T is continuous.

Theorem A.14 (Hahn–Banach theorem). Let X be a vector space, M a

subspace of X, p a seminorm on X and u : M → K a linear functional such
that |u(x)| ≤ p(x) for all x ∈ M . Then u has a linear extension u
 to X such
that |
u(x)| ≤ p(x) for all x ∈ X.
356 Appendix A

We will mostly apply the Hahn–Banach theorem through one of the fol-
lowing corollaries:
(i) if p is a seminorm on X and x0 ∈ X then there exists a linear functional
u on X such that u(x0 ) = p(x0 ) and |u(x)| ≤ p(x) for all x ∈ X;
if X is a Fréchet space then
(ii) every continuous linear functional on a subspace of X extends to a con-
tinuous linear functional on X (with preservation of norm, if X is a
Banach space);
(iii) if M is a closed subspace of X and x ∈/ M then there exists a continuous
linear functional x∗ on X that vanishes on M with x, x∗  = 0;
(iv) a subspace M of X is dense in X if and only if every continuous linear
functional that vanishes on M also vanishes on X;
(v) for any x ∈ X, if x, x∗  = 0 for all x∗ ∈ X ∗ then x = 0.
A sequence (en )n in a Fréchet space X is called a basis if every x ∈ X has a
unique representation
∞
x= an e n
n=1

with scalars an ∈ K, n ≥ 1. The coefficient functionals

e∗n : X → K, x → an ,

are then continuous.

At several places in this book unconditional convergence plays a crucial role.
∞
Definition A.15. A series n=1 xn in a Fréchet space is called uncondition-
ally convergent if for any bijection π : N → N the series
∞

xπ(n)
n=1

converges.

There are several useful equivalent formulations; in the following result,

let  ·  denote an F-norm that induces the topology of X.

Theorem A.16. Let X be a Fréchet space. Then the following assertions are
equivalent:
∞
(i) n=1 xn is unconditionally convergent;
∞
(ii) for any 0-1-sequence (εn )n , n=1 εn xn converges;

(iii) for any bounded sequence (αn )n of scalars, ∞ n=1 αn xn converges;
(iv) for any ε > 0 there is some N ∈ N such that for any finite set F ⊂
{N, N + 1, N + 2, . . .} we have that
 
 
 xn  < ε;
n∈F
Prerequisites 357

∞any ε > 0 there is some N ∈ N such that for any 0-1-sequence (εn )n ,
(v) for
n=1 εn xn converges and
 
 
 εn xn  < ε;
n≥N

(vi) for any

∞ ε > 0 there is some N ∈ N such that whenever supn≥1 |αn | ≤ 1
then n=1 αn xn converges and
 
 
 αn xn  < ε.
n≥N

For a proof we refer to [269, 3.8.2 and p. 153] and

[213, 3.3.8 and 3.3.9].
Unconditional convergence of series of the form n∈Z xn is defined simi-
larly and has the corresponding properties.
A sequence (en )n in a Fréchet space X is called an unconditional basis if
it is a basis such that, for every x ∈ X, the representation
∞

x= an en ,
n=1

converges unconditionally.
In connection with eigenvalue criteria for hypercyclicity, but also for Ap-
pendix B on spectral theory, we need the Riemann integral for Fréchet space-
valued continuous functions. Its definition poses no particular problems and
follows exactly the same lines as in the real-valued case. Thus, if f : [a, b] → X
is a continuous function with values in a Fréchet space X, then its Riemann
integral is defined as
 b 
N −1
f (t) dt = lim f (tk )(tk+1 − tk ),
a k=0

where the limit is taken as max0≤k≤N −1 |tk+1 − tk | → 0. Apart from the

usual properties of linearity in f and additivity with respect to the domain
of integration we have that, for any continuous seminorm p on X, for the
norm  ·  on a Banach space X, for any operator T ∈ L(X, Y ), where Y is
another Fréchet space, and for any x∗ ∈ X ∗ that
 b
 b
p f (t) dt ≤ p(f (t)) dt,
a a
 b   b
 
 f (t) dt ≤ f (t) dt,
a a
 b  b
T f (t) dt = T f (t) dt,
a a
358 Appendix A
 b   b
f (t) dt, x∗ = f (t), x∗  dt.
a a

As usual, if X is a complex Fréchet space and Γ is a continuously differ-

entiable curve in C then one defines
  1
f (z) dz = f (γ(t))γ  (t) dt,
Γ 0

where γ : [0, 1] → Γ is a parametrization of the curve. This extends to

integrals over finite collections of such curves.

Finite-dimensional Fréchet spaces

Finite-dimensional Fréchet spaces play a particular role in the theory, for

several reasons.
First, every Fréchet space of dimension N ≥ 1 is isomorphic to KN under
the Euclidean topology. In particular, all Fréchet space topologies on finite-
dimensional vector spaces coincide.
Moreover, any subspace of finite dimension of a Fréchet space is closed
and therefore a Fréchet space in the induced topology.
Finally, a Fréchet space with a relatively compact neighbourhood of 0 is
finite dimensional.
Indeed, all these results hold in arbitrary topological vector spaces, which
are studied in Chapter 12; see Rudin [271].

Complex analysis

Finally, we assume familiarity with the basic concepts of complex analysis

as can be gained, for example, from Chapter 10 of Rudin [270] or Chapters
II–IV of Conway [115]. We collect here some important results.

Theorem A.17 (Liouville). Every bounded entire function is constant.

Theorem A.18 (Casorati–Weierstrass). For any punctured neighbour-

hood U of an essential singularity, possibly ∞, of a holomorphic function
f , f (U ) is dense in C.

As a consequence, every nonconstant entire function has dense range.

Theorem A.19 (Open mapping theorem). Any nonconstant holomor-

phic function f on a domain Ω is an open mapping, that is, for any open set
O ⊂ Ω, f (O) is open.
Prerequisites 359

Theorem A.20 (Rouché). Let f and g be holomorphic functions on a

neighbourhood of {z ∈ C ; |z − a| ≤ r}, where a ∈ C, r > 0, such that

|f (z) − g(z)| < |g(z)| for |z − a| = r.

Then f and g have the same number of zeros, counting multiplicity, for
|z − a| < r.
Let now D = {z ∈ C ; |z| < 1} be the open unit disk.
Theorem A.21 (Schwarz lemma). Let f : D → D be a holomorphic func-
tion with f (0) = 0. Then |f (z)| ≤ |z| for any z ∈ D. Moreover, if there is
some z ∈ D, z = 0, with |f (z)| = |z| then there is some a ∈ C with |a| = 1
such that f (z) = az, z ∈ D.
Theorem A.22 (Big Picard theorem). Let f be an entire function that is
not a polynomial. Then f takes every value in C, with at most one exception,
infinitely often.
The following is a consequence of Jensen’s formula; see Rudin [270, 15.20] or
Conway [115, Chapter XI, § 1].
Theorem A.23. Let f be an entire function, and let N (r) denote the number
of zeros of f in |z| < r, counting multiplicity. If f (0) = 1 and M (r) =
max{|f (z)| ; |z| = r} then

N (r) log 2 ≤ log M (2r).

The Runge approximation theorem is a crucial tool for several of our exam-
we denote the extended complex plane C ∪ {∞}.
ples. By C
Theorem A.24 (Runge’s theorem). Let K ⊂ C be a compact set and A a
\ K.
set that contains at least one point from each connected component of C
Let f be a function that is holomorphic on some neighbourhood of K, and
let ε > 0. Then there exists a rational function h with poles only at points
from A such that
sup |f (z) − h(z)| < ε.
z∈K

We note that the complement of a compact set is open and therefore has
at most a countable number of connected components. Thus, A may always
be chosen to be at most countable.
\ K is connected then A may be taken
Moreover, if K ⊂ Ω is such that C
to be {∞}, and the function h will a polynomial.
Corollary A.25. Let Ω ⊂ C be a domain.
(a) Let A ⊂ C contain at least one point from each bounded component of

C \ Ω. Then the rational functions with poles only in A form a dense set in
H(Ω).
360 Appendix A

(b) If Ω is simply connected, then the polynomials form a dense set in

H(Ω).

Next we discuss an important family of maps. A linear fractional transfor-

mation, also known as a Möbius transformation, is a map of the form
az + b
f (z) = ,
cz + d
with constants a, b, c, d ∈ C. In addition, one demands that ad − bc = 0,
which excludes the constant functions. Under obvious conventions for deal-
ing with the point at infinity, each linear fractional transformation defines a
bijection f : C → C, where C is the extended complex plane, and its inverse
is also a linear fractional transformation. Moreover, any composition of linear
fractional transformations is a linear fractional transformation.
Every linear fractional transformation maps each circle and each line onto
either a circle or a line. It has either one or two fixed points, unless it is the
identity function. And given any distinct points z1 , z2 , z3 ∈ C and distinct
points w1 , w2 , w3 ∈ C there is a (unique) linear fractional transformation that
maps zj to wj , j = 1, 2, 3.
For more information on the topic we refer to Ahlfors [3].
Finally, the notion of homotopy is only used once in the book, but then in a
crucial way; see Section 6.4. Let X and Y be metric spaces. Two maps f, g :
X → Y are called homotopic if there is a continuous map H : X × [0, 1] → Y ,
called a homotopy, such that H(x, 0) = f (x) and H(x, 1) = g(x), x ∈ X. The
map f is called homotopically trivial or null-homotopic if it is homotopic to
a constant map.
In particular, let T = {z ∈ C ; |z| = 1} be the unit circle in C. Then every
continuous map f : T → T has an index n ∈ Z. Indeed, f can be identified
with a closed curve that lies entirely in T, and the index is nothing but the
winding number of the curve with respect to the origin. Then f has index
n if and only if it is homotopic to the map z → z n , and it is homotopically
trivial if and only if it has index 0.
For more details we refer the reader to Chapter 10 of Rudin [270]; note
also Exercise 28 there.
Appendix B – Spectral theory

Throughout this appendix, T denotes an operator on a complex Banach space

X.

Basic spectral theory

In the finite-dimensional setting, λ ∈ C is an eigenvalue of an operator T

if and only if λI − T is not injective, or equivalently, not invertible. For
general complex Banach spaces, the latter is taken as the defining property
of elements in the spectrum.
Definition B.1. Let T be an operator on a complex Banach space X. The
spectrum σ(T ) of T is defined as

σ(T ) = {λ ∈ C ; λI − T is not invertible}.

The point spectrum σp (T ) of T is the set of eigenvalues of T .

We recall that an operator is said to be invertible if it is bijective and
its inverse is also continuous. However, by the inverse mapping theorem (see
Appendix A), the continuity of the inverse is automatic in our setting. There-
fore, λ ∈ C belongs to the spectrum of T if and only if either λI − T fails to
be injective, in which case λ is an eigenvalue, or if it fails to be surjective.
One obviously has the useful formula

σ(λI + μT ) = λ + μσ(T )

for any λ, μ ∈ C, which is a special case of the spectral mapping theorem; see
below.
Proposition B.2. The spectrum σ(T ) is a nonempty compact set. Moreover,
|λ| ≤ T  for any λ ∈ σ(T ).

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 361

DOI 10.1007/978-1-4471-2170-1, © Springer-Verlag London Limited 2011
362 Appendix B

The spectral radius of T is defined as r(T ) = supλ∈σ(T ) |λ|.

Theorem B.3 (Spectral radius formula). For the spectral radius we have
that
r(T ) = lim T n 1/n .
n→∞

The existence of the limit is part of the result. In particular, we have that
σ(T ) = {0} if and only if limn→∞ T n 1/n = 0.
For the adjoint T ∗ : X ∗ → X ∗ of T we refer to Appendix A.

Proposition B.4. We have that σ(T ∗ ) = σ(T ).

For a more detailed introduction to spectral theory we refer to Chapter 18

of Rudin [270] or Chapter 12 of Bollobás [77].

The Riesz–Dunford functional calculus

∞ n
If f (z) = n=0 an z is a holomorphicfunction on some disk {z ∈ C ; |z| < r}
∞ ∞
with r > T  then n=0 an T n  ≤ n=0 |an |T n < ∞, so that
∞

f (T ) = an T n
n=0

defines an operator on X; we apply this procedure, for example, in Section

4.4. If f is only holomorphic on a neighbourhood O of the spectrum σ(T ) of
T then one can still define an operator f (T ) by

1
f (T ) = f (λ)(λI − T )−1 dλ,
2πi Γ

where Γ is the union of finitely many continuously differentiable Jordan

curves in O that contains σ(T ) in its interior and C \ O in its exterior;
see Figure B.1. The integral is to be understood as an operator-valued Rie-
mann integral; see Appendix A. This definition of f (T ) goes back to Riesz
and Dunford and is therefore referred to as the Riesz–Dunford functional
calculus.

Theorem B.5. Let f and g be holomorphic functions on a neighbourhood of

σ(T ) and λ ∈ C. Then we have the following:
(i) (λf )(T ) = λ(f (T ));
(ii) (f + g)(T ) = f (T ) + g(T );
(iii) (f g)(T ) = f (T )g(T ); in particular, f (T ) and g(T ) commute;
(iv) f (T )∗ = f (T ∗ );
(v) if f (z) = 0 for all z ∈ σ(T ), then f (T ) is invertible and f (T )−1 =
(1/f )(T );
Spectral theory 363

Fig. B.1 Jordan curves for the definition of f (T )

N N
(vi) if p(z) = n=0 an z n is a polynomial then p(T ) = n=0 an T n ;
(vii) if f (z) = ∞ n
n=0 an z  is holomorphic on {z ∈ C ; |z| < r} for some
∞
r > T  then f (T ) = n=0 an T n .
The following is a central result of spectral theory.
Theorem B.6 (Spectral mapping theorem). Let f be a holomorphic
function on a neighbourhood of σ(T ). Then

σ(f (T )) = f (σ(T )).

There is also a version for the point spectrum. Since the result is more
difficult to find in the literature we give a proof here.
Theorem B.7 (Point spectral mapping theorem). Let f be a holomor-
phic function on an open neighbourhood O of σ(T ) that is not constant on
any connected component of O. Then

σp (f (T )) = f (σp (T )).

Proof. First suppose that λ ∈ σp (T ). Since the function z → f (λ) − f (z) has
a zero at λ, it factorizes as f (λ) − f (z) = (λ − z)g(z), where g is holomorphic
on some neighbourhood of σ(T ). Hence

f (λ)I − f (T ) = g(T )(λI − T ).

Since λI − T is non-injective, the same is true for f (λ)I − f (T ), so that

f (λ) ∈ σp (f (T )).
Conversely, let μ ∈ σp (f (T )). It follows from the assumption that, on σ(T ),
the function z → μ − f (z) only has a finite number of zeros λ1 , . . . , λn , so
364 Appendix B
n
that it factorizes as μ − f (z) = k=1 (λk − z)Nk g(z), where Nk ≥ 1 and g is
holomorphic on O and zero-free on σ(T ). Thus g(T ) is invertible and


n
μI − f (T ) = (λk I − T )Nk g(T ).
k=1

Since μI − f (T ) is non-injective, so is some λk I − T . Hence μ = f (λk ) ∈

f (σp (T )). 

We state an analogous result for semigroups of operators; see Section 7.1.
Theorem B.8 (Point spectral mapping theorem for semigroups). Let
(A, D(A)) be the generator of a C0 -semigroup (Tt )t≥0 defined on a complex
Banach space X. Then we have the following identities:

σp (Tt ) \ {0} = etσp (A) for t ≥ 0,

  
ker(λI − A) = ker eλt I − Tt for λ ∈ C,
t≥0
 2πni

ker(eλt I − Tt ) = span ker λ+ I −A for t > 0.

t
n∈Z

The Riesz–Dunford functional calculus is treated in detail in Chapter VII

of Conway [116] or in Radjavi and Rosenthal [264]. The point spectral map-
ping theorem is taken from Hille and Phillips [203, Theorem 5.12.2]. Its ver-
sion for C0 -semigroups can be found in the book by Engel and Nagel [143].

The Riesz decomposition theorem

We turn to a result that is crucial for the application of spectral theory to

linear dynamics. Since its proof is not so well known we will provide one here.
We recall that an invariant closed subspace is called nontrivial if it is different
from {0} and the whole space.
Theorem B.9 (Riesz decomposition theorem). Suppose that the spec-
trum σ(T ) splits into two disjoint nonempty closed subsets σ1 and σ2 :

σ(T ) = σ1 ∪ σ2 .

Then there are nontrivial T -invariant closed subspaces M1 and M2 of X such

that X = M1 ⊕ M2 ,

σ(T |M1 ) = σ1 and σ(T |M2 ) = σ2 .

It should be noted that, the spectrum being closed, a subset of σ(T ) is

closed in σ(T ) if and only if it is closed in C.
Spectral theory 365

Proof. Since σ(T ) is compact, the sets σ1 and σ2 are disjoint compact subsets
of C, so that there exist disjoint open neighbourhoods O1 of σ1 and O2 of σ2 .
The function f1 that takes the value 1 on O1 and the value 0 on O2 is then
holomorphic on a neighbourhood of σ(T ), as is the function f2 that takes the
value 0 on O1 and the value 1 on O2 . We can therefore define two operators
by
P1 = f1 (T ) and P2 = f2 (T ).
Since the functions f1 and f2 have the properties

f1 + f2 = 1, f1 f2 = f2 f1 = 0, f12 = f1 , f22 = f2 ,

an application of Theorem B.5 yields that

P1 + P2 = I, P1 P2 = P2 P1 = 0, P12 = P1 , P22 = P2 . (B.1)

We define
M1 = ran P1 and M2 = ran P2
and claim that these are the desired subspaces.
First, (B.1) implies that

x = P1 x + P2 x and P1 P2 x = P2 P1 x = 0 for all x ∈ X, (B.2)

and hence that X = M1 + M2 , M1 ⊂ ker P2 and M2 ⊂ ker P1 . Now if

x ∈ M1 ∩ M2 then P2 x = P1 x = 0 and hence x = 0 by (B.2); this shows that
X = M1 ⊕ M2 . Moreover, if x ∈ ker P2 then x = P1 x ∈ M1 by (B.2); this
shows that M1 = ker P2 is a closed subspace and x = P1 x for all x ∈ M1 .
The corresponding statements hold for M2 . Moreover, for x ∈ M1 , T x =
T P1 x = P1 T x ∈ M1 , so that M1 is T -invariant; note that P1 and T commute
because P1 is a function of T . As a consequence, T |M1 is an operator on M1 .
By symmetry the same holds for M2 . Finally, M1 = {0} and hence M2 = X.
Otherwise, ker P1 = M2 = X, so that f1 = 0 on σ(T ) (see Exercise 5.0.8),
which contradicts the fact that σ1 = ∅. By symmetry we also have that
M2 = {0} and hence M1 = X.
It remains to identify the spectra of T |M1 and T |M2 . Since the operator
λI − T leaves the spaces M1 and M2 invariant, it is invertible if and only if
(λI − T )|M1 and (λI − T )|M2 are; in other words,

σ(T ) = σ(T |M1 ) ∪ σ(T |M2 ). (B.3)

Moreover, let λ ∈
/ σ1 . Then the function

g(z) = (λ − z)f1 (z) + f2 (z)

is holomorphic on a neighbourhood of σ(T ) and nonzero on σ(T ). Hence the

operator
g(T ) = (λI − T )P1 + P2
366 Appendix B

is invertible. It follows from the properties of P1 and P2 that g(T )|M1 =

(λI − T )|M1 and g(T )|M2 = I|M2 . Since X is the direct sum of M1 and M2 ,
(λI − T )|M1 must then be invertible, hence λ ∈/ σ(T |M1 ).
Thus we have shown that σ(T |M1 ) ⊂ σ1 , and similarly σ(T |M2 ) ⊂ σ2 . But
since σ1 and σ2 form a partition of σ(T ), (B.3) implies that these inclusions
are in fact identities, which had to be shown.  

For the Riesz decomposition theorem we have followed Radjavi and Rosen-
thal [264].
References

1. E. Abakumov and J. Gordon, Common hypercyclic vectors for multiples of back-

ward shift, J. Funct. Anal. 200 (2003), 494–504.
2. Y.A. Abramovich and C.D. Aliprantis, An invitation to operator theory, American
Mathematical Society, Providence, RI, 2002.
3. L.V. Ahlfors, Complex analysis, third edition, McGraw-Hill, New York, 1978.
4. E. Akin, Recurrence in topological dynamics. Furstenberg families and Ellis actions,
Plenum Press, New York, 1997.
5. A.A. Albanese, Construction of operators with prescribed orbits in Fréchet spaces
with a continuous norm, Math. Scand., to appear.
6. M.P. Aldred and D.H. Armitage, Harmonic analogues of G.R. MacLane’s universal
functions, J. London Math. Soc. (2) 57 (1998), 148–156.
7. M.P. Aldred and D.H. Armitage, Harmonic analogues of G.R. MacLane’s universal
functions. II, J. Math. Anal. Appl. 220 (1998), 382–395.
8. T. Ando, Operators with a norm condition, Acta Sci. Math. (Szeged) 33 (1972),
169–178.
9. S.I. Ansari, Hypercyclic and cyclic vectors, J. Funct. Anal. 128 (1995), 374–383.
10. S.I. Ansari, Existence of hypercyclic operators on topological vector spaces, J.
Funct. Anal. 148 (1997), 384–390.
11. R. Arens, Dense inverse limit rings, Michigan Math. J. 5 (1958), 169–182.
12. S.A. Argyros and R.G. Haydon, A hereditarily indecomposable L∞ -space that
solves the scalar-plus-compact problem, Acta Math. 206 (2011), 1–54.
13. D.H. Armitage, Permissible growth rates for Birkhoff type universal harmonic func-
tions, J. Approx. Theory 136 (2005), 230–243.
14. R. Aron, J. Bès, F. León and A. Peris, Operators with common hypercyclic sub-
spaces, J. Operator Theory 54 (2005), 251–260.
15. R. Aron and D. Markose, On universal functions, J. Korean Math. Soc. 41 (2004),
65–76.
16. R.M. Aron, J.B. Seoane-Sepúlveda and A. Weber, Chaos on function spaces, Bull.
Austral. Math. Soc. 71 (2005), 411–415.
17. A. Atzmon, Power regular operators, Trans. Amer. Math. Soc. 347 (1995), 3101–
3109.
18. J. Auslander and J.A. Yorke, Interval maps, factors of maps, and chaos, Tôhoku
Math. J. (2) 32 (1980), 177–188.
19. C. Badea and S. Grivaux, Unimodular eigenvalues, uniformly distributed sequences
and linear dynamics, Adv. Math. 211 (2007), 766–793.
20. C. Badea, S. Grivaux and V. Müller, Multiples of hypercyclic operators, Proc.
Amer. Math. Soc. 137 (2009), 1397–1403.

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 367

DOI 10.1007/978-1-4471-2170-1, © Springer-Verlag London Limited 2011
368 References

21. C. Badea, S. Grivaux and V. Müller, Epsilon-hypercyclic operators, Ergodic Theory

Dynam. Systems 30 (2010), 1597–1606.
22. J. Banasiak, Chaos in some linear kinetic models, “WASCOM 2003”—12th Confer-
ence on Waves and Stability in Continuous Media (Proc. Conf., Villasimius, 2003),
32–37, World Sci. Publishing, River Edge, NJ, 2004.
23. J. Banasiak, Birth-and-death type systems with parameter and chaotic dynamics
of some linear kinetic models, Z. Anal. Anwendungen 24 (2005), 675–690.
24. J. Banasiak and M. Lachowicz, Chaos for a class of linear kinetic models, C. R.
Acad. Sci. Paris Sér. IIB Méc. 329 (2001), 439–444.
25. J. Banasiak and M. Lachowicz, Topological chaos for birth-and-death-type models
with proliferation, Math. Models Methods Appl. Sci. 12 (2002), 755–775.
26. J. Banasiak, M. Lachowicz and M. Moszyński, Semigroups for generalized birth-
and-death equations in lp spaces, Semigroup Forum 73 (2006), 175–193.
27. J. Banasiak, M. Lachowicz and M. Moszyński, Chaotic behavior of semigroups
related to the process of gene amplification–deamplification with cell proliferation,
Math. Biosci. 206 (2007), 200–215.
28. J. Banks, Regular periodic decompositions for topologically transitive maps, Er-
godic Theory Dynam. Systems 17 (1997), 505–529.
29. J. Banks, Topological mapping properties defined by digraphs, Discrete Contin.
Dynam. Systems 5 (1999), 83–92.
30. J. Banks, Chaos for induced hyperspace maps, Chaos Solitons Fractals 25 (2005),
681–685.
31. J. Banks, J. Brooks, G. Cairns, G. Davis and P. Stacey, On Devaney’s definition
of chaos, Amer. Math. Monthly 99 (1992), 332–334.
32. W. Bauer and K. Sigmund, Topological dynamics of transformations induced on
the space of probability measures, Monatsh. Math. 79 (1975), 81–92.
33. F. Bayart, Common hypercyclic vectors for composition operators, J. Operator
Theory 52 (2004), 353–370.
34. F. Bayart, Topological and algebraic genericity of divergence and universality, Stu-
dia Math. 167 (2005), 161–181.
35. F. Bayart, Common hypercyclic subspaces, Integral Equations Operator Theory 53
(2005), 467–476.
36. F. Bayart, Dynamics of holomorphic groups, Semigroup Forum 82 (2011), 229–241.
37. F. Bayart and T. Bermúdez, Semigroups of chaotic operators, Bull. Lond. Math.
Soc. 41 (2009), 823–830.
38. F. Bayart and S. Grivaux, Hypercyclicité: le rôle du spectre ponctuel unimodulaire,
C. R. Math. Acad. Sci. Paris 338 (2004), 703–708.
39. F. Bayart and S. Grivaux, Hypercyclicity and unimodular point spectrum, J. Funct.
Anal. 226 (2005), 281–300.
40. F. Bayart and S. Grivaux, Frequently hypercyclic operators, Trans. Amer. Math.
Soc. 358 (2006), 5083–5117.
41. F. Bayart and S. Grivaux, Invariant Gaussian measures for operators on Banach
spaces and linear dynamics, Proc. Lond. Math. Soc. (3) 94 (2007), 181–210.
42. F. Bayart and É. Matheron, How to get common universal vectors, Indiana Univ.
Math. J. 56 (2007), 553–580.
43. F. Bayart and É. Matheron, Hypercyclic operators failing the hypercyclicity crite-
rion on classical Banach spaces, J. Funct. Anal. 250 (2007), 426–441.
44. F. Bayart and É. Matheron, Dynamics of linear operators, Cambridge University
Press, Cambridge, 2009.
45. F. Bayart and É. Matheron, (Non-)weakly mixing operators and hypercyclicity
sets, Ann. Inst. Fourier (Grenoble) 59 (2009), 1–35.
46. B. Beauzamy, Un opérateur, sur l’espace de Hilbert, dont tous les polynômes sont
hypercycliques, C. R. Acad. Sci. Paris Sér. I Math. 303 (1986), 923–925.
47. B. Beauzamy, An operator on a separable Hilbert space with many hypercyclic
vectors, Studia Math. 87 (1987), 71–78.
References 369

48. B. Beauzamy, Introduction to operator theory and invariant subspaces, North-

Holland, Amsterdam, 1988.
49. T. Bermúdez, A. Bonilla, J.A. Conejero and A. Peris, Hypercyclic, topologically
mixing and chaotic semigroups on Banach spaces, Studia Math. 170 (2005), 57–75.
50. T. Bermúdez, A. Bonilla and A. Martinón, On the existence of chaotic and hy-
percyclic semigroups on Banach spaces, Proc. Amer. Math. Soc. 131 (2003), 2435–
2441.
51. T. Bermúdez, A. Bonilla and A. Peris, On hypercyclicity and supercyclicity criteria,
Bull. Austral. Math. Soc. 70 (2004), 45–54.
52. T. Bermúdez and N.J. Kalton, The range of operators on von Neumann algebras,
Proc. Amer. Math. Soc. 130 (2002), 1447–1455.
53. T. Bermúdez and V.G. Miller, On operators T such that f (T ) is hypercyclic,
Integral Equations Operator Theory 37 (2000), 332–340.
54. L. Bernal-González, On hypercyclic operators on Banach spaces, Proc. Amer.
Math. Soc. 127 (1999), 1003–1010.
55. L. Bernal-González, Hypercyclic sequences of differential and antidifferential oper-
ators, J. Approx. Theory 96 (1999), 323–337.
56. L. Bernal-González, Norms of hypercyclic sequences, Math. Pannon. 15 (2004),
221–230.
57. L. Bernal-González, Hypercyclic subspaces in Fréchet spaces, Proc. Amer. Math.
Soc. 134 (2006), 1955–1961.
58. L. Bernal-González, Disjoint hypercyclic operators, Studia Math. 182 (2007), 113–
131.
59. L. Bernal-González, Common hypercyclic functions for multiples of convolution
and non-convolution operators, Proc. Amer. Math. Soc. 137 (2009), 3787–3795.
60. L. Bernal-González and A. Bonilla, Exponential type of hypercyclic entire func-
tions, Arch. Math. (Basel) 78 (2002), 283–290.
61. L. Bernal-González and M.D.C. Calderón-Moreno, Dense linear manifolds of mon-
sters, J. Approx. Theory 119 (2002), 156–180.
62. L. Bernal-González and K.-G. Grosse-Erdmann, The hypercyclicity criterion for
sequences of operators, Studia Math. 157 (2003), 17–32.
63. L. Bernal-González and K.-G. Grosse-Erdmann, Existence and nonexistence of
hypercyclic semigroups, Proc Amer. Math. Soc. 135 (2007), 755–766.
64. L. Bernal-González and A. Montes-Rodríguez, Universal functions for composition
operators, Complex Variables Theory Appl. 27 (1995), 47–56.
65. L. Bernal González and A. Montes Rodríguez, Non-finite dimensional closed vec-
tor spaces of universal functions for composition operators, J. Approx. Theory 82
(1995), 375–391.
66. J.P. Bès, Invariant manifolds of hypercyclic vectors for the real scalar case, Proc.
Amer. Math. Soc. 127 (1999), 1801–1804.
67. J. Bès and K.C. Chan, Approximation by chaotic operators and by conjugate
classes, J. Math. Anal. Appl. 284 (2003), 206–212.
68. J. Bès and K.C. Chan, Denseness of hypercyclic operators on a Fréchet space,
Houston J. Math. 29 (2003), 195–206.
69. J. Bès and J.A. Conejero, Hypercyclic subspaces in omega, J. Math. Anal. Appl.
316 (2006), 16–23.
70. J. Bès, Ö. Martin, A. Peris and S. Shkarin, Disjoint mixing operators, preprint.
71. J. Bès and A. Peris, Hereditarily hypercyclic operators, J. Funct. Anal. 167 (1999),
94–112.
72. J. Bès and A. Peris, Disjointness in hypercyclicity, J. Math. Anal. Appl. 336 (2007),
297–315.
73. J.J. Betancor and A. Bonilla, On a universality property of certain integral oper-
ators, J. Math. Anal. Appl. 250 (2000), 162–180.
74. G.D. Birkhoff, Surface transformations and their dynamical applications, Acta
Math. 43 (1920), 1–119.
370 References

75. G.D. Birkhoff, Démonstration d’un théorème élémentaire sur les fonctions entières,
C. R. Acad. Sci. Paris 189 (1929), 473–475.
76. O. Blasco, A. Bonilla and K.-G. Grosse-Erdmann, Rate of growth of frequently
hypercyclic functions, Proc. Edinb. Math. Soc. (2) 53 (2010), 39–59.
77. B. Bollobás, Linear analysis, second edition, Cambridge University Press, Cam-
bridge, 1999.
78. J. Bonet, Hypercyclic and chaotic convolution operators, J. London Math. Soc. (2)
62 (2000), 253–262.
79. J. Bonet, Dynamics of the differentiation operator on weighted spaces of entire
functions, Math. Z. 261 (2009), 649–657.
80. J. Bonet, A problem on the structure of Fréchet spaces, Rev. R. Acad. Cienc.
Exactas Fís. Nat. Ser. A Mat. RACSAM 104 (2010), 427–434.
81. J. Bonet, L. Frerick, A. Peris and J. Wengenroth, Transitive and hypercyclic op-
erators on locally convex spaces, Bull. London Math. Soc. 37 (2005), 254–264.
82. J. Bonet, F. Martínez-Giménez and A. Peris, A Banach space which admits no
chaotic operator, Bull. London Math. Soc. 33 (2001), 196–198.
83. J. Bonet, F. Martínez-Giménez and A. Peris, Linear chaos on Fréchet spaces, In-
ternat. J. Bifur. Chaos Appl. Sci. Engrg. 13 (2003), 1649–1655.
84. J. Bonet, F. Martínez-Giménez and A. Peris, Universal and chaotic multipliers on
spaces of operators, J. Math. Anal. Appl. 297 (2004), 599–611.
85. J. Bonet and A. Peris, Hypercyclic operators on non-normable Fréchet spaces, J.
Funct. Anal. 159 (1998), 587–595.
86. A. Bonilla and K.-G. Grosse-Erdmann, On a theorem of Godefroy and Shapiro,
Integral Equations Operator Theory 56 (2006), 151–162.
87. A. Bonilla and K.-G. Grosse-Erdmann, Frequently hypercyclic operators and vec-
tors, Ergodic Theory Dynam. Systems 27 (2007), 383–404. Erratum: Ergodic Theory
Dynam. Systems 29 (2009), 1993–1994.
88. A. Bonilla and K.-G. Grosse-Erdmann, Frequently hypercyclic subspaces, preprint.
89. A. Bonilla and P. Miana, Hypercyclic and topologically mixing cosine functions on
Banach spaces, Proc. Amer. Math. Soc. 136 (2008), 519–528.
90. P.S. Bourdon, Invariant manifolds of hypercyclic vectors, Proc. Amer. Math. Soc.
118 (1993), 845–847.
91. P.S. Bourdon, The second iterate of a map with dense orbit, Proc. Amer. Math.
Soc. 124 (1996), 1577–1581.
92. P.S. Bourdon, Orbits of hyponormal operators, Michigan Math. J. 44 (1997), 345–
353.
93. P.S. Bourdon and N.S. Feldman, Somewhere dense orbits are everywhere dense,
Indiana Univ. Math. J. 52 (2003), 811–819.
94. P.S. Bourdon and J.H. Shapiro, Cyclic composition operators on H 2 , Operator
Theory: Operator Algebras and Applications, Part 2 (Proc. Summer Res. Inst.,
Durham, NH, 1988), 43–53, Amer. Math. Soc., Providence, RI, 1990.
95. P.S. Bourdon and J.H. Shapiro, Cyclic phenomena for composition operators, Mem.
Amer. Math. Soc. 125 (1997), no. 596.
96. P.S. Bourdon and J.H. Shapiro, Hypercyclic operators that commute with the
Bergman backward shift, Trans. Amer. Math. Soc. 352 (2000), 5293–5316.
97. M. Brin and G. Stuck, Introduction to dynamical systems, Cambridge University
Press, Cambridge, 2002.
98. L. Carleson and T.W. Gamelin, Complex dynamics, Springer, New York, 1993.
99. K.C. Chan, Hypercyclicity of the operator algebra for a separable Hilbert space,
J. Operator Theory 42 (1999), 231–244.
100. K.C. Chan, The density of hypercyclic operators on a Hilbert space, J. Operator
Theory 47 (2002), 131–143.
101. K.C. Chan and R. Sanders, A weakly hypercyclic operator that is not norm hy-
percyclic, J. Operator Theory 52 (2004), 39–59.
References 371

102. K.C. Chan and R. Sanders, Two criteria for a path of operators to have common
hypercyclic vectors, J. Operator Theory 61 (2009), 191–223.
103. K.C. Chan and R. Sanders, Common hypercyclic vectors for the conjugate class of
a hypercyclic operator, J. Math. Anal. Appl. 375 (2011), 139–148.
104. K.C. Chan and I. Seceleanu, Orbital limit points and hypercyclicity of operators
on analytic function spaces, Math. Proc. R. Ir. Acad. 110A (2010), 99–109.
105. K.C. Chan and I. Seceleanu, Hypercyclicity of shifts as a zero-one law of orbital
limit points, J. Operator Theory, to appear.
106. K.C. Chan and J.H. Shapiro, The cyclic behavior of translation operators on
Hilbert spaces of entire functions, Indiana Univ. Math. J. 40 (1991), 1421–1449.
107. K.C. Chan and R.D. Taylor, Jr., Hypercyclic subspaces of a Banach space, Integral
Equations Operator Theory 41 (2001), 381–388.
108. J.A. Conejero, Operadores y semigrupos de operadores en espacios de Fréchet y es-
pacios localmente convexos, Thesis, Univ. Politécnica de Valencia, Valencia, 2004.
109. J.A. Conejero and E.M. Mangino, Hypercyclic semigroups generated by Ornstein–
Uhlenbeck operators, Mediterr. J. Math. 7 (2010), 101–109.
110. J.A. Conejero, V. Müller and A. Peris, Hypercyclic behaviour of operators in a
hypercyclic C0 -semigroup, J. Funct. Anal. 244 (2007), 342–348.
111. J.A. Conejero and A. Peris, Linear transitivity criteria, Topology Appl. 153 (2005),
767–773.
112. J.A. Conejero and A. Peris, Chaotic translation semigroups, Discrete Contin. Dyn.
Syst. 2007, Dynamical Systems and Differential Equations (Proc. Conf., Poitiers,
2006), suppl., 269–276.
113. J.A. Conejero and A. Peris, Hypercyclic translation C0 -semigroups on complex
sectors, Discrete Contin. Dyn. Syst. 25 (2009), 1195–1208.
114. J.A. Conejero, A. Peris and M. Trujillo, Chaotic asymptotic behaviour of the hyper-
bolic heat transfer equation solutions, Internat. J. Bifur. Chaos Appl. Sci. Engrg.
20 (2010), 2943–2947.
115. J.B. Conway, Functions of one complex variable, second edition, Springer, New
York-Berlin, 1978.
116. J.B. Conway, A course in functional analysis, second edition, Springer, New York-
Berlin, 1990.
117. G. Costakis, On a conjecture of D. Herrero concerning hypercyclic operators, C.
R. Acad. Sci. Paris Sér. I Math. 330 (2000), 179–182.
118. G. Costakis and A. Manoussos, J-class weighted shifts on the space of bounded
sequences of complex numbers, Integral Equations Operator Theory 62 (2008), 149–
158.
119. G. Costakis and A. Manoussos, J-class operators and hypercyclicity, J. Operator
Theory, to appear.
120. G. Costakis and P. Mavroudis, Common hypercyclic entire functions for multiples
of differential operators, Colloq. Math. 111 (2008), 199–203.
121. G. Costakis and A. Peris, Hypercyclic semigroups and somewhere dense orbits, C.
R. Math. Acad. Sci. Paris 335 (2002), 895–898.
122. G. Costakis and I.Z. Ruzsa, Frequently Cesàro hypercylic operators are hypercyclic,
preprint.
123. G. Costakis and M. Sambarino, Genericity of wild holomorphic functions and com-
mon hypercyclic vectors, Adv. Math. 182 (2004), 278–306.
124. G. Costakis and M. Sambarino, Topologically mixing hypercyclic operators, Proc.
Amer. Math. Soc. 132 (2004), 385–389.
125. C.C. Cowen and B.D. MacCluer, Composition operators on spaces of analytic func-
tions, CRC Press, Boca Raton, FL, 1995.
126. M. De la Rosa and C. Read, A hypercyclic operator whose direct sum T ⊕ T is not
hypercyclic, J. Operator Th. 61 (2009), 369–380.
127. M. De la Rosa, L. Frerick, S. Grivaux and A. Peris, Frequent hypercyclicity, chaos,
and unconditional Schauder decompositions, Israel J. Math., to appear.
372 References

128. R. deLaubenfels and H. Emamirad, Chaos for functions of discrete and continuous
weighted shift operators, Ergodic Theory Dynam. Systems 21 (2001), 1411–1427.
129. W. Desch and W. Schappacher, On products of hypercyclic semigroups, Semigroup
Forum 71 (2005), 301–311.
130. W. Desch and W. Schappacher, Spectral characterization of weak topological tran-
sitivity, Rev. R. Acad. Cienc. Exactas Fís. Nat. Ser. A Mat. RACSAM, to appear.
131. W. Desch, W. Schappacher and G.F. Webb, Hypercyclic and chaotic semigroups
of linear operators, Ergodic Theory Dynam. Systems 17 (1997), 793–819.
132. L.R. Devaney, An introduction to chaotic dynamical systems, Benjamin/Cum-
mings, Menlo Park, CA, 1986; second edition, Addison-Wesley, Redwood City,
CA, 1989.
133. J. Diestel, Sequences and series in Banach spaces, Springer, New York, 1984.
134. J. Diestel, H. Jarchow and A. Tonge, Absolutely summing operators, Cambridge
University Press, Cambridge, 1995.
135. S.J. Dilworth and V.G. Troitsky, Spectrum of a weakly hypercyclic operator meets
the unit circle, Trends in Banach spaces and operator theory (Proc. Conf., Mem-
phis, TN, 2001), 67–69, Amer. Math. Soc., Providence, RI, 2003.
136. P.L. Duren, Theory of Hp spaces, Academic Press, New York-London, 1970.
137. S.M. Duyos-Ruiz, On the existence of universal functions, Soviet Math. Dokl. 27
(1983), 9–13.
138. O.P. Dzagnidze, The universal harmonic function in the space En (Russian), Soob-
shch. Akad. Nauk Gruzin. SSR 55 (1969), 41–44.
139. S. El Mourchid, On a hypercyclicity criterion for strongly continuous semigroups,
Discrete Contin. Dyn. Syst. 13 (2005), 271–275.
140. S. El Mourchid, The imaginary point spectrum and hypercyclicity, Semigroup Fo-
rum 73 (2006), 313–316.
141. H. Emamirad, G.R. Goldstein and J.A. Goldstein, Chaotic solution for the Black–
Scholes equation, preprint.
142. P. Enflo, On the invariant subspace problem for Banach spaces, Acta Math. 158
(1987), 213–313.
143. K.-J. Engel and R. Nagel, One-parameter semigroups for linear evolution equations,
Springer, New York-Berlin, 2000.
144. K.-J. Engel and R. Nagel, A short course on operator semigroups, Springer, New
York, 2006.
145. J. Esterle, Mittag-Leffler methods in the theory of Banach algebras and a new
approach to Michael’s problem, Banach algebras and several complex variables
(Proc. Conf., New Haven, CT, 1983), 107–129, American Mathematical Society,
Providence, RI, 1984.
146. A. Fathi, Existence de systèmes dynamiques minimaux sur l’espace de Hilbert
séparable, Topology 22 (1983), 165–167.
147. N.S. Feldman, Linear chaos?, <http://home.wlu.edu/~feldmann/research.html>,
2001.
148. N.S. Feldman, Perturbations of hypercyclic vectors, J. Math. Anal. Appl. 273
(2002), 67–74.
149. N.S. Feldman, Countably hypercyclic operators, J. Operator Theory 50 (2003),
107–117.
150. N.S. Feldman, Hypercyclicity and supercyclicity for invertible bilateral weighted
shifts, Proc. Amer. Math. Soc. 131 (2003), 479–485.
151. N.S. Feldman, V.G. Miller and T.L. Miller, Hypercyclic and supercyclic cohyponor-
mal operators, Acta Sci. Math. (Szeged) 68 (2002), 965–990.
152. E. Flytzanis, Mixing properties of linear operators in Hilbert spaces, Séminaire
d’Initiation à l’Analyse 34ème année (1994/1995), Exposé no. 6.
153. E. Flytzanis, Unimodular eigenvalues and linear chaos in Hilbert spaces, Geom.
Funct. Anal. 5 (1995), 1–13.
References 373

154. G.L. Forti, Various notions of chaos for discrete dynamical systems. A brief survey,
Aequationes Math. 70 (2005), 1–13.
155. L. Frerick and A. Peris, Hypercyclic operators for which every non-zero multiple
is also hypercyclic, preprint.
156. H. Furstenberg, The structure of distal flows, Amer. J. Math. 85 (1963) 477–515.
157. H. Furstenberg, Disjointness in ergodic theory, minimal sets, and a problem in
Diophantine approximation, Math. Systems Theory 1 (1967), 1–49.
158. E.A. Gallardo-Gutiérrez and A. Montes-Rodríguez, The role of the spectrum in
the cyclic behavior of composition operators, Mem. Amer. Math. Soc. 167 (2004),
no. 791.
159. E.A. Gallardo-Gutiérrez and A. Montes-Rodríguez, The Volterra operator is not
supercyclic, Integral Equations Operator Theory 50 (2004), 211–216.
160. E.A. Gallardo-Gutiérrez and J.R. Partington, Common hypercyclic vectors for fam-
ilies of operators, Proc. Amer. Math. Soc. 136 (2008), 119–126.
161. R.M. Gethner and J.H. Shapiro, Universal vectors for operators on spaces of holo-
morphic functions, Proc. Amer. Math. Soc. 100 (1987), 281–288.
162. E. Glasner, Ergodic theory via joinings, American Mathematical Society, Provi-
dence, RI, 2003.
163. E. Glasner and B. Weiss, Sensitive dependence on initial conditions, Nonlinearity
6 (1993), 1067–1075.
164. G. Godefroy, Linear dynamics, Advanced courses of mathematical analysis. II
(Proc. 2nd Int. School, Granada, 2004), 57–75, World Sci. Publ., Hackensack, NJ,
2007.
165. G. Godefroy and J.H. Shapiro, Operators with dense, invariant, cyclic vector man-
ifolds, J. Funct. Anal. 98 (1991), 229–269.
166. M.C. Gómez-Collado, F. Martínez-Giménez, A. Peris and F. Rodenas, Slow growth
for universal harmonic functions, J. Inequal. Appl. 2010, Art. ID 253690, 6 pp.
167. M. González, F. León-Saavedra and A. Montes-Rodríguez, Semi-Fredholm theory:
hypercyclic and supercyclic subspaces, Proc. London Math. Soc. (3) 81 (2000),
169–189.
168. P. Gorkin, F. León-Saavedra and R. Mortini, Bounded universal functions in one
and several complex variables, Math. Z. 258 (2008), 745–762.
169. S. Grivaux, Construction of operators with prescribed behaviour, Arch. Math.
(Basel) 81 (2003), 291–299.
170. S. Grivaux, Sums of hypercyclic operators, J. Funct. Anal. 202 (2003), 486–503.
171. S. Grivaux, Hypercyclic operators with an infinite dimensional closed subspace of
periodic points, Rev. Mat. Complut. 16 (2003), 383–390.
172. S. Grivaux, Hypercyclic operators, mixing operators, and the bounded steps prob-
lem, J. Operator Theory 54 (2005), 147–168.
173. S. Grivaux, A probabilistic version of the frequent hypercyclicity criterion, Studia
Math. 176 (2006), 279–290.
174. S. Grivaux, A new class of frequently hypercyclic operators, with applications,
Indiana Univ. Math. J., to appear.
175. S. Grivaux, A hypercyclic rank one perturbation of a unitary operator, preprint.
176. S. Grivaux and S. Shkarin, Non-mixing hypercyclic operators, unpublished (2007).
177. K.-G. Grosse-Erdmann, Holomorphe Monster und universelle Funktionen, Mitt.
Math. Sem. Giessen 176 (1987).
178. K.-G. Grosse-Erdmann, On the universal functions of G.R. MacLane, Complex
Variables Theory Appl. 15 (1990), 193–196.
179. K.-G. Grosse-Erdmann, Universal families and hypercyclic operators, Bull. Amer.
Math. Soc. (N.S.) 36 (1999), 345–381.
180. K.-G. Grosse-Erdmann, Hypercyclic and chaotic weighted shifts, Studia Math. 139
(2000), 47–68.
181. K.-G. Grosse-Erdmann, Rate of growth of hypercyclic entire functions, Indag.
Math. (N.S.) 11 (2000), 561–571.
374 References

182. K.-G. Grosse-Erdmann, Dynamics of linear operators, Topics in complex analysis

and operator theory (Proc. Winter School, Málaga, 2006), 41–84, Univ. Málaga,
Málaga, 2007.
183. K.-G. Grosse-Erdmann, F. León-Saavedra and A. Piqueras-Lerena, The iterates of
a map with dense orbit, Acta Sci. Math. (Szeged) 74 (2008), 245–257.
184. K.-G. Grosse-Erdmann and R. Mortini, Universal functions for composition oper-
ators with non-automorphic symbol, J. Anal. Math. 107 (2009), 355–376.
185. K.-G. Grosse-Erdmann and A. Peris, Frequently dense orbits, C. R. Math. Acad.
Sci. Paris 341 (2005), 123–128.
186. K.-G. Grosse-Erdmann and A. Peris, Corrigendum to the note “Frequently dense
orbits” (2008), not accepted for publication in C. R. Math. Acad. Sci. Paris.
187. K.-G. Grosse-Erdmann and A. Peris, Weakly mixing operators on topological vec-
tor spaces, Rev. R. Acad. Cienc. Exactas Fís. Nat. Ser. A Mat. RACSAM 104
(2010), 413–426.
188. D. Gulick, Encounters with chaos, McGraw-Hill, New York, 1992.
189. D.W. Hadwin, E.A. Nordgren, H. Radjavi and P. Rosenthal, Most similarity orbits
are strongly dense, Proc. Amer. Math. Soc. 76 (1979), 250–252.
190. P. Hájek and P. Vivi, On ω-limit sets of ordinary differential equations in Banach
spaces, J. Math. Anal. Appl. 371 (2010), 793–812.
191. P.R. Halmos, A Hilbert space problem book, second edition, Springer, New York-
Berlin, 1982.
192. I. Halperin, C. Kitai and P. Rosenthal, On orbits of linear operators, J. London
Math. Soc. (2) 31 (1985), 561–565.
193. G.H. Hardy, J.E. Littlewood and G. Pólya, Inequalities, second edition, Cambridge
University Press, Cambridge, 1952.
194. D.A. Herrero, Limits of hypercyclic and supercyclic operators, J. Funct. Anal. 99
(1991), 179–190.
195. D.A. Herrero, Hypercyclic operators and chaos, J. Operator Theory 28 (1992),
93–103.
196. D.A. Herrero and C. Kitai, On invertible hypercyclic operators, Proc. Amer. Math.
Soc. 116 (1992), 873–875.
197. G. Herzog, On linear operators having supercyclic vectors, Studia Math. 103 (1992),
295–298.
198. G. Herzog, On a universality of the heat equation, Math. Nachr. 188 (1997), 169–
171.
199. G. Herzog and R. Lemmert, Über Endomorphismen mit dichten Bahnen, Math. Z.
213 (1993), 473–477.
200. G. Herzog and C. Schmoeger, On operators T such that f (T ) is hypercyclic, Studia
Math. 108 (1994), 209–216.
201. G. Herzog and A. Weber, A class of hypercyclic Volterra composition operators,
Demonstratio Math. 39 (2006), 465–468.
202. H.M. Hilden and L.J. Wallen, Some cyclic and non-cyclic vectors of certain oper-
ators, Indiana Univ. Math. J. 23 (1973/74), 557–565.
203. E. Hille and R.S. Phillips, Functional analysis and semi-groups, American Mathe-
matical Society, Providence, R.I., 1957.
204. T. Hosokawa, Chaotic behavior of composition operators on the Hardy space, Acta
Sci. Math. (Szeged) 69 (2003), 801–811.
205. P.R. Hurst, Relating composition operators on different weighted Hardy spaces,
Arch. Math. (Basel) 68 (1997), 503–513.
206. V. Istrăţescu, T. Saitô and T. Yoshino, On a class of operators, Tôhoku Math. J.
(2) 18 (1966), 410–413.
207. L. Ji and A. Weber, Lp spectral theory and heat dynamics of locally symmetric
spaces, J. Funct. Anal. 258 (2010), 1121–1139.
208. L. Ji and A. Weber, Dynamics of the heat semigroup on symmetric spaces, Ergodic
Theory Dynam. Systems 30 (2010), 457–468.
References 375

209. T. Kalmes, Hypercyclic, mixing, and chaotic C0 -semigroups induced by semiflows,

Ergodic Theory Dynam. Systems 27 (2007), 1599–1631.
210. T. Kalmes, Hypercyclic C0 -semigroups and evolution families generated by first
order differential operators, Proc. Amer. Math. Soc. 137 (2009), 3833–3848.
211. T. Kalmes, Hypercyclicity and mixing for cosine operator functions generated by
second order partial differential operators, J. Math. Anal. Appl. 365 (2010), 363–
375.
212. N.J. Kalton, N.T. Peck and J.W. Roberts, An F-space sampler, Cambridge Uni-
versity Press, Cambridge, 1984.
213. P.K. Kamthan and M. Gupta, Sequence spaces and series, Marcel Dekker, New
York, 1981.
214. K.-T. Kim and S.G. Krantz, The automorphism groups of domains, Amer. Math.
Monthly 112 (2005), 585–601.
215. C. Kitai, Invariant closed sets for linear operators, Thesis, University of Toronto,
Toronto, 1982.
216. S.F. Kolyada, Li-Yorke sensitivity and other concepts of chaos, Ukrainian Math.
J. 56 (2004), 1242–1257.
217. S. Kolyada and L. Snoha, Some aspects of topological transitivity—a survey, Iter-
ation theory (Proc. Conf., Opava, 1994) 3–35, Grazer Math. Ber. 334, Graz, 1997.
218. C.S. Kubrusly, Hilbert space operators. A problem solving approach, Birkhäuser,
Boston, MA, 2003.
219. F. León-Saavedra, Notes about the hypercyclicity criterion, Math. Slovaca 53
(2003), 313–319.
220. F. León-Saavedra and A. Montes-Rodríguez, Linear structure of hypercyclic vec-
tors, J. Funct. Anal. 148 (1997), 524–545.
221. F. León-Saavedra and A. Montes-Rodríguez, Spectral theory and hypercyclic sub-
spaces, Trans. Amer. Math. Soc. 353 (2001), 247–267.
222. F. León-Saavedra and V. Müller, Rotations of hypercyclic and supercyclic opera-
tors, Integral Equations Operator Theory 50 (2004), 385–391.
223. F. León-Saavedra and V. Müller, Hypercyclic sequences of operators, Studia Math.
175 (2006), 1–18.
224. F. León-Saavedra, A. Piqueras-Lerena and J.B. Seoane-Sepúlveda, Orbits of Cesàro
type operators, Math. Nachr. 282 (2009), 764–773.
225. G.R. MacLane, Sequences of derivatives and normal families, J. Analyse Math. 2
(1952/53), 72–87.
226. M. Marano and H.N. Salas, Maps with dense orbits: Ansari’s theorem revisited
and the infinite torus, Bull. Belg. Math. Soc. Simon Stevin 16 (2009), 481–492.
227. F. Martínez-Giménez, Operadores hipercíclicos en espacios de Fréchet, Rev. Colom-
biana Mat. 33 (1999), 51–76.
228. F. Martínez-Giménez, Chaos for power series of backward shift operators, Proc.
Amer. Math. Soc. 135 (2007), 1741–1752.
229. F. Martínez-Giménez and A. Peris, Chaos for backward shift operators, Internat.
J. Bifur. Chaos Appl. Sci. Engrg. 12 (2002), 1703–1715.
230. F. Martínez-Giménez and A. Peris, Hypercyclic differential operators, Analysis and
operators (3rd Conf. EU Network, Tenerife, 2003).
231. F. Martínez-Giménez and A. Peris, Universality and chaos for tensor products of
operators, J. Approx. Theory 124 (2003), 7–24.
232. F. Martínez-Giménez and A. Peris, Hypercyclic differential operators on spaces of
entire and harmonic functions, preprint.
233. V. Matache, Notes on hypercyclic operators, Acta Sci. Math. (Szeged) 58 (1993),
401–410.
234. V. Matache, Spectral properties of operators having dense orbits, Topics in Oper-
ator Theory, Operator Algebras and Applications (Proc. Conf., Timişoara, 1994),
221–237, Inst. Math. Roman. Acad., Bucharest, 1995.
376 References

235. É. Matheron, Subsemigroups of transitive semigroups, Ergodic Theory Dynam.

Systems, to appear.
236. M. Matsui, M. Yamada and F. Takeo, Supercyclic and chaotic translation semi-
groups, Proc. Amer. Math. Soc. 131 (2003), 3535–3546. Erratum: Proc. Amer.
Math. Soc. 132 (2004), 3751–3752.
237. R. Meise and D. Vogt, Introduction to functional analysis, Oxford University Press,
New York, 1997.
238. G. Metafune and V.B. Moscatelli, Dense subspaces with continuous norm in Fréchet
spaces, Bull. Polish Acad. Sci. Math. 37 (1989), 477–479.
239. T.L. Miller and V.G. Miller, Local spectral theory and orbits of operators, Proc.
Amer. Math. Soc. 127 (1999), 1029–1037.
240. A. Montes-Rodríguez, Banach spaces of hypercyclic vectors, Michigan Math. J. 43
(1996), 419–436.
241. A. Montes-Rodríguez, A Birkhoff theorem for Riemann surfaces, Rocky Mountain
J. Math. 28 (1998), 663–693.
242. A. Montes-Rodríguez and M.C. Romero-Moreno, Supercyclicity in the operator
algebra, Studia Math. 150 (2002), 201–213.
243. A. Montes-Rodríguez and H.N. Salas, Supercyclic subspaces, Bull. London Math.
Soc. 35 (2003), 721–737.
244. T.K.S. Moothathu, Quantitative views of recurrence and proximality, Nonlinearity
21 (2008), 2981–2992.
245. T.K.S. Moothathu, Weak mixing and mixing of a single transformation of a topo-
logical (semi)group, Aequationes Math. 78 (2009), 147–155.
246. V. Müller, Local behaviour of the polynomial calculus of operators, J. Reine Angew.
Math. 430 (1992), 61–68.
247. V. Müller, Spectral theory of linear operators and spectral systems in Banach alge-
bras, second edition, Birkhäuser, Basel, 2007.
248. V. Müller, On the Salas theorem and hypercyclicity of f (T ), Integral Equations
Operator Theory 67 (2010), 439–448.
249. V. Nestoridis, Universal Taylor series, Ann. Inst. Fourier (Grenoble) 46 (1996),
1293–1306.
250. M.H.A. Newman, Elements of the topology of plane sets of points, second edition,
Cambridge University Press, Cambridge, 1951.
251. J.C. Oxtoby and S.M. Ulam, Measure-preserving homeomorphisms and metrical
transitivity, Ann. of Math. (2) 42 (1941), 874–920.
252. A. Peris, Common hypercyclic vectors for backward shifts, Operator Theory Sem-
inar, Michigan State University, 2000–2001.
253. A. Peris, Hypercyclicity criteria and the Mittag-Leffler theorem, Bull. Soc. Roy.
Sci. Liège 70 (2001), 365–371.
254. A. Peris, Multi-hypercyclic operators are hypercyclic, Math. Z. 236 (2001), 779–
786.
255. A. Peris, Chaotic polynomials on Banach spaces, J. Math. Anal. Appl. 287 (2003),
487–493.
256. A. Peris, Set-valued discrete chaos, Chaos Solitons Fractals 26 (2005), 19–23.
257. A. Peris and L. Saldivia, Syndetically hypercyclic operators, Integral Equations
Operator Theory 51 (2005), 275–281.
258. H. Petersson, Hypercyclic subspaces for Fréchet space operators, J. Math. Anal.
Appl. 319 (2006), 764–782.
259. A. Pietsch, Operator ideals, North-Holland, Amsterdam-New York, 1980.
260. G.T. Prajitura, The density of the hypercyclic operators in the strong operator
topology, Integral Equations Operator Theory 49 (2004), 559–560.
261. G. Prǎjiturǎ, Irregular vectors of Hilbert space operators, J. Math. Anal. Appl. 354
(2009), 689–697.
262. V. Protopopescu, Linear vs nonlinear and infinite vs finite: An interpretation of
chaos, Oak Ridge National Laboratory Report TM-11667, Oak Ridge, TN, 1990.
References 377

263. V. Protopopescu and Y.Y. Azmy, Topological chaos for a class of linear models,
Math. Models Methods Appl. Sci. 2 (1992), 79–90.
264. H. Radjavi and P. Rosenthal, Invariant subspaces, second edition, Dover Publica-
tions, Mineola, NY, 2003.
265. C.J. Read, A solution to the invariant subspace problem, Bull. London Math. Soc.
16 (1984), 337–401.
266. C.J. Read, The invariant subspace problem for a class of Banach spaces. II. Hy-
percyclic operators, Israel J. Math. 63 (1988), 1–40.
267. C. Robinson, Dynamical systems. Stability, symbolic dynamics, and chaos, second
edition, CRC Press, Boca Raton, FL, 1999.
268. S. Rolewicz, On orbits of elements, Studia Math. 32 (1969), 17–22.
269. S. Rolewicz, Metric linear spaces, second edition, D. Reidel Publishing Co., Dor-
drecht, 1985.
270. W. Rudin, Real and complex analysis, third edition, McGraw-Hill, New York, 1987.
271. W. Rudin, Functional analysis, second edition, McGraw-Hill, New York, 1991.
272. R. Rudnicki, Gaussian measure-preserving linear transformations, Univ. Iagel. Acta
Math. 30 (1993), 105–112.
273. H. Salas, A hypercyclic operator whose adjoint is also hypercyclic, Proc. Amer.
Math. Soc. 112 (1991), 765–770.
274. H.N. Salas, Hypercyclic weighted shifts, Trans. Amer. Math. Soc. 347 (1995), 993–
1004.
275. H.N. Salas, Supercyclicity and weighted shifts, Studia Math. 135 (1999), 55–74.
276. H.N. Salas, Banach spaces with separable duals support dual hypercyclic operators,
Glasg. Math. J. 49 (2007), 281–290.
277. H.N. Salas, Dual disjoint hypercyclic operators, J. Math. Anal. Appl. 374 (2011),
106–117.
278. W. Seidel and J.L. Walsh, On approximation by Euclidean and non-Euclidean
translations of an analytic function, Bull. Amer. Math. Soc. 47 (1941), 916–920.
279. J.H. Shapiro, Composition operators and classical function theory, Springer, New
York, 1993.
280. J.H. Shapiro, Simple connectivity and linear chaos, Rend. Circ. Mat. Palermo (2)
Suppl. (1998), no. 56, 27–48.
281. J.H. Shapiro, Notes on the dynamics of linear operators,
<http://www.mth.msu.edu/~shapiro>, 2001.
282. S. Shirali and H.L. Vasudeva, Metric spaces, Springer, London, 2006.
283. S.A. Shkarin, On the growth of D-universal functions, Moscow Univ. Math. Bull.
48 (1993), no. 6, 49–51.
284. S. Shkarin, Universal elements for non-linear operators and their applications, J.
Math. Anal. Appl. 348 (2008), 193–210.
285. S. Shkarin, The Kitai criterion and backward shifts, Proc. Amer. Math. Soc. 136
(2008), 1659–1670.
286. S. Shkarin, Existence theorems in linear chaos, arXiv:0810.1192v2 (2008).
287. S. Shkarin, On the spectrum of frequently hypercyclic operators, Proc. Amer. Math.
Soc. 137 (2009), 123–134.
288. S. Shkarin, Remarks on common hypercyclic vectors, J. Funct. Anal. 258 (2010),
132–160.
289. S. Shkarin, A hypercyclic finite rank perturbation of a unitary operator, Math.
Ann. 348 (2010), 379–393.
290. S. Shkarin, On the set of hypercyclic vectors for the differentiation operator, Israel
J. Math. 180 (2010), 271–283.
291. S. Shkarin, A short proof of existence of disjoint hypercyclic operators, J. Math.
Anal. Appl. 367 (2010), 713–715.
292. S. Shkarin, Chaotic Banach algebras, preprint.
293. S. Shkarin, Hypercyclic operators on topological vector spaces, preprint.
378 References

294. S. Silverman, On maps with dense orbits and the definition of chaos, Rocky Moun-
tain J. Math. 22 (1992), 353–375.
295. L.A. Smith, Chaos: A very short introduction, Oxford University Press, Oxford,
2007.
296. C.L. Stewart and R. Tijdeman, On infinite-difference sets, Canad. J. Math. 31
(1979), 897–910.
297. F. Takeo, Chaos and hypercyclicity for solution semigroups to some partial differ-
ential equations, Nonlinear Anal. 63 (2005), e1943–e1953.
298. M. Taniguchi, Chaotic composition operators on the classical holomorphic spaces,
Complex Var. Theory Appl. 49 (2004), 529–538.
299. P. Touhey, Yet another definition of chaos, Amer. Math. Monthly 104 (1997), 411–
414.
300. P. Walters, An introduction to ergodic theory, Springer, New York-Berlin, 1982.
301. J. Wengenroth, Hypercyclic operators on non-locally convex spaces, Proc. Amer.
Math. Soc. 131 (2003), 1759–1761.
302. P.Y. Wu, Sums and products of cyclic operators, Proc. Amer. Math. Soc. 122
(1994), 1053–1063.
303. B. Yousefi and H. Rezaei, Conditions for hypercyclicity criterion, Int. J. Contemp.
Math. Sci. 1 (2006), 99–107.
304. N. Zorboska, Cyclic composition operators on smooth weighted Hardy spaces,
Rocky Mountain J. Math. 29 (1999), 725–740.
Index

A2 , 93 Mϕ , 120
A2α , 132 Mϕ∗ , 120
B, 36, 98 N (A, B), 19, 24
Bw , 89, 99 Per((Tt )), 186
C∗ , 112 Per(T ), 12
, 359
C Sν , 133
C[a, b], 33 S ⊕ T , 36
C ∞ (R), 60 S × T , 17
C0 (R+ ), 60, 188 T, 5
Cϕ , 123 T, 41
C0,v (R+ ), 206 T ∗ , 352, 354
D, 70 Ta , 36
D, 130 U0 (X), 333
Dα , 132 V , 139
D, 36  41
X,
Dw , 218 X ⊕ Y , 36
Eτ2 , 129 X × Y , 17
E p (γ), 228 X ∗
, 352
F , 92, 100 ∞
( n=1 Xn )p , 47
F HC(T ), 237 ∞
Fw , 98, 102 ( n=1 Xn )c0 , 47
H(C), 32  · , · , 352, 353
HC(T ), 37 σ(T ), 361
HC(Ψ ), 168 σp (T ), 361
H 2 , 118 ϕ, 335
H 2 (β), 133 ϕ(D), 105
I, 4, 352 ω, 32
J(x), 26 ω(Z), 99
J mix (x), 27 c0 , 33
K, 32 c0 (v), 93
K(X), 352 dens(A), 236
K(X, Y ), 352 ∞ , 33
L(X), 352 p , 33
L(X, Y ), 352 p (Z), 33
Lp (T), 263 p (v), 353
Lp [a, b], 33, 334 orb(x, (Tn )), 23
Lpv (R+ ), 183 orb(x, (Tt )), 186

K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, 379

DOI 10.1007/978-1-4471-2170-1, © Springer-Verlag London Limited 2011
380 Index

orb(x, T ), 4 Birkhoff transitivity theorem, 10, 29, 30,

2-hypercyclic operator, 64 39, 207
4-set trick, 19, 24, 30 Birkhoff’s operators, 39, 45, 46, 65, 71,
72, 106, 109, 110, 239, 275, 316, 319
absolutely convex set, 333 birth-and-death process, 211
absolutely summing operator, 160 Black–Scholes semigroup, 211
abstract Cauchy problem, 185 bounded operator, 347
adjoint, 352, 354 bounded set, 347, 355
Hilbert space, 354 Bourdon’s theorem, 53, 64, 66, 190, 337
adjoint multiplication operator, see Bourdon–Feldman theorem, 164, 167,
adjoint multiplier 177, 256, 337
adjoint multiplier, 120, 134, 264, 302, 327,
330 C0 -group, 207
admissible comparison function, 228 hypercyclic, 207
admissible weight function, 60, 184, 206 C0 -semigroup, 181, 210
Ansari’s theorem, 162, 163, 173, 175, 177, chaotic, 186, 222
254, 337 conjugate, 189
Ansari–Bernal theorem, 220, 232 frequently hypercyclic, 256
applications of C0 -semigroups, 201 hereditarily hypercyclic, 196
approximation number, 229 hypercyclic, 186
approximation property, 299 mixing, 186, 222
Argyros–Haydon theorem, 221, 232, 300 quasiconjugate, 189
Arzelà–Ascoli theorem, 157 topologically ergodic, 208
Auslander–Yorke chaotic operator, 43, 65 topologically transitive, 186
automorphism, 110 weakly mixing, 186
autonomous discretization, 192 Casorati–Weierstrass theorem, 358
Cauchy problem, see abstract Cauchy
Bès–Peris theorem, 76, 84, 88 problem
backward orbit, 26 Cesàro operator, 84, 149, 342
backward shift operator, 36, 89 chaotic C0 -semigroup, 186, 222
extended, 219 chaotic dynamical system, 13, 14, 30
generalized, 229, 233 chaotic operator, 42, 140, 173, 221, 225,
Baire category theorem, 351 245, 265, 336
balanced set, 332 classical solution of (ACP), 185
Banach sequence space, 90, 99 closed graph, 184
Banach space, 32, 351 closed graph theorem, 355
Banach–Steinhaus theorem, 353, 355 coefficient functional, 356
Banks–Brooks–Cairns–Davis–Stacey cohyponormal operator, 158
theorem, 13, 30 common hypercyclic subspace, 321
basic sequence, 268, 302 common hypercyclic vector, 329
basis, 356 Common Hypercyclicity Criterion, 309,
unconditional, 357 329
Bergman space, 93, 133, 134 commuting operators, 352
weighted, 132 compact operator, 142, 160, 347, 352
big Picard theorem, 359 compact perturbation, 147, 160, 289
bilateral backward shift operator, 98 complemented subspace, 230
bilateral weighted backward shift completely continuous operator, 160
operator, see bilateral weighted shift complexification, 41, 51, 65, 189
operator composition operator, 110, 123, 130, 266,
bilateral weighted forward shift operator, 289, 302, 322, 328, 329
102 Conejero–Müller–Peris theorem, 168, 177,
bilateral weighted shift operator, 99, 134, 196, 259, 306
302 conformal map, 111
Birkhoff ergodic theorem, 236 conjugate C0 -semigroups, 189
Index 381

conjugate dynamical systems, 6 mixing, 17, 30

continuous family of operators, 307 quasiconjugate, 6, 338
continuous linear functional, 352 topologically ergodic, 28
continuous norm, 220, 268 topologically transitive, 8, 30
contraction, 141 totally transitive, 28
cosine operator function, 211 weakly mixing, 18, 30, 173
Costakis–Peris theorem, 166, 167, 177,
337 eigenvalue, 189
countably hypercyclic operator, 175 unimodular, 249
criterion eigenvalue criterion, 69, 345
Common Hypercyclicity, 309, 329 for C0 -semigroups, 197, 199, 211
Disjoint Hypercyclicity, 86 for frequent hypercyclicity, 249, 265
eigenvalue, 69, 345 eigenvector field
for C0 -semigroups, 197, 199, 211
spanning, 250
for frequent hypercyclicity, 249, 265
elliptic linear fractional transformation,
Frequent Hypercyclicity, 242, 262, 265
125
Gethner–Shapiro, 73
ε-hypercyclic operator, 61, 156
Godefroy–Shapiro, 69, 87, 345
ε-hypercyclic vector, 61
Hypercyclicity, 74, 87, 279, 346
equicontinuous family of operators, 355
for C0 -semigroups, 196, 210
Erdős–Sárközy theorem, 241, 265
for sequences of operators, 84, 88
ergodic map, 235
Kitai, 71, 232, 345
ergodic theorem
Recurrent Hypercyclicity, 67, 211
Birkhoff, 236
Universality, 24, 30
ergodic theory, 235, 264
cyclic operator, 37
essential spectrum, 302
cyclic vector, 37
essentially quasinilpotent operator, 160
exhaustion, 110
d-dense orbit, 61
exponential type, 104, 128, 228
De la Rosa–Read theorem, 48, 66
extended backward shift operator, 219
decomposition theorem, 176
extended complex plane, 359
dense orbit, 9
Devaney chaos, 13, 14, 30, 65
difference set, 240 F-norm, 35, 65
differential operator, 105, 129, 134, 330 F-space, 67, 333
differentiation operator, 36 finite codimension, 282
direct p -sum, 47 finite sequence, 33
direct c0 -sum, 47 finite-dimensional Fréchet space, 55, 358
Dirichlet space, 130, 133 finite-dimensional topological vector
weighted, 132 space, 333
discretization, 167, 192 finite-rank operator, 142, 156
autonomous, 192 finite-rank perturbation, 147, 159
disjoint hypercyclic operator, 86, 88, 230 finitely connected domain, 113
Disjoint Hypercyclicity Criterion, 86 first-order partial differential equation,
doubling map, 5, 6 201
dual, 352, 354 fixed point, 12
Dunford–Pettis operator, 160 flip transitive dynamical system, 28
Duyos-Ruiz’s theorem, 109, 134, 253 flip transitive operator, 49
dynamical system, 4, 30 forward shift operator, 92, 100
chaotic, 13, 14, 30 Fréchet sequence space, 90, 99
conjugate, 6 Fréchet space, 34, 333, 354
flip transitive, 28 Fredholm operator, 160
hypertransitive, 29 Frequent Hypercyclicity Criterion, 242,
linear, 37 262, 265
minimal, 26, 177 frequently hypercyclic C0 -semigroup, 256
382 Index

frequently hypercyclic operator, 237, 262, hypercyclic C0 -semigroup, 186

264 hypercyclic comparison principle, 61, 65,
frequently hypercyclic semigroup action, 338
257 hypercyclic operator, 37, 65, 334
frequently hypercyclic sequence of hypercyclic semigroup action, 168
operators, 262 hypercyclic sequence of operators, 83
frequently hypercyclic vector, 237, 256, hypercyclic subspace, 267, 301
257 common, 321
frequently recurrent orbit, 237 hypercyclic vector, 37, 65, 83, 186, 334
full support, 236 common, 305
functional calculus, 362, 364 Hypercyclicity Criterion, 74, 87, 173, 279,
Furstenberg’s theorem, 20, 30 346
Common, 309, 329
Gδ -set, 351 for C0 -semigroups, 196, 210
gauge, 333 for sequences of operators, 84, 88, 281
generalized backward shift operator, 229, Frequent, 242
233 hyperextension, 29
generalized kernel, 51 hyperspace, 29
generator, see infinitesimal generator hypertransitive dynamical system, 29
Gethner–Shapiro criterion, 73 hyponormal operator, 152
Godefroy–Shapiro criterion, 69, 87, 345
Godefroy–Shapiro theorem, 107, 134, 252 identity operator, 352
González–León–Montes theorem, 302 index, 360
Grivaux’s theorem, 233 inductive limit, 334
infinitely connected domain, 113
Hahn–Banach theorem, 355 infinitesimal generator, 184
Hardy space, 118, 133, 134 integral operator
weighted, 133 Volterra, 141
Hausdorff topology, 331 invariant map, 235
heat equation, 210 invariant set, 7
hereditarily ergodic operator, 62, 67 invariant subset problem, 65
hereditarily hypercyclic C0 -semigroup, invariant subspace, 38
196 nontrivial, 38
hereditarily hypercyclic operator, 76 invariant subspace problem, 38, 65
hereditarily hypercyclic sequence of inverse mapping theorem, 355
operators, 83 irrational rotation, 6
hereditarily syndetic operator, 67 irregular vector, 156, 159
hereditarily transitive sequence of maps, isolated point, 351
29 iterate, 4
Herrero’s problem, 48, 66, 104, 211
Herrero–Bourdon theorem, 53, 66, 190, J-class operator, 61, 156
262, 325, 337 J-set, 26
Herzog–Lemmert theorem, 66 Jensen’s formula, 359
Hilbert space, 33, 353 Julia set, 58
Hilbert–Schmidt operator, 160
hole, 113 kernel, 142
homotopic map, 360 Kitai’s criterion, 71, 232, 345
homotopically trivial map, 360 Kitai’s theorem, 140, 159
homotopy, 360 Korenblum space, 348
homotopy theory, 171
hyperbolic heat transfer equation, 202 León–Müller theorem, 167, 177, 259, 306,
hyperbolic linear fractional transforma- 339
tion, 125 left-multiplication operator, 276, 302
hypercyclic C0 -group, 207 lemma
Index 383

Riemann–Lebesgue, 198, 250 norm, 32

Riesz, 143 continuous, 220, 268
Schwarz, 359 normal operator, 150
Wengenroth, 337 normaloid operator, 158
linear chaos, 42, 58, 336 nuclear operator, 160
linear dynamical system, 37 nuclear perturbation, 303
linear fractional transformation, 115, 125,
360 open mapping theorem, 355, 358
Liouville’s theorem, 358 operator, 35, 352
Lipschitz constant, 317 2-hypercyclic, 64
Lipschitz continuous function, 317 J-class, 61, 156
Littlewood’s subordination principle, 131 ε-hypercyclic, 61, 156
locally convex space, 333 absolutely summing, 160
locally equicontinuous family, 182 adjoint multiplication, see adjoint
logistic map, 5 multiplier
lower density, 236, 256 Auslander–Yorke chaotic, 43, 65
lower triangular operator, 148 backward shift, 36, 89
loxodromic linear fractional transforma- bilateral backward shift, 98
tion, 125 bilateral weighted backward shift, see
bilateral weighted shift operator
m-homogeneous polynomial, 65 bilateral weighted forward shift, 102
Möbius transformation, 360 bilateral weighted shift, 99, 134, 302
MacLane’s operator, 40, 45, 46, 65, 71, Birkhoff, 39, 45, 46, 65, 71, 72, 106,
72, 98, 108, 245, 275, 315 109, 110, 239, 275, 316, 319
map bounded, 347
doubling, 5, 6 Cesàro, 84, 149, 342
ergodic, 235 chaotic, 42, 140, 173, 221, 225, 245,
invariant, 235 265, 336
logistic, 5 cohyponormal, 158
quadratic, 5 compact, 142, 160, 347, 352
rotation, 6 completely continuous, 160
shift, 6 composition, 110, 123, 130, 266, 289,
tent, 6 302, 322, 328, 329
maximum term, 129 countably hypercyclic, 175
Mazur’s theorem, 270 cyclic, 37
metric space, 351 differential, 105, 129, 134, 330
metrizable topological vector space, 333 differentiation, 36
minimal dynamical system, 26, 177 disjoint hypercyclic, 86, 88, 230
Minkowski functional, 333 Dunford–Pettis, 160
Mittag-Leffler theorem, 79, 88 essentially quasinilpotent, 160
mixing C0 -semigroup, 186, 222 extended backward shift, 219
mixing dynamical system, 17, 30 finite-rank, 142, 156
mixing operator, 46, 220, 225, 245, 247, flip transitive, 49
265 forward shift, 92, 100
mixing sequence of maps, 24 Fredholm, 160
mixing sequence of operators, 83 frequently hypercyclic, 237, 262, 264
Montes’ theorem, 268, 302 generalized backward shift, 229, 233
multi-hypercyclic operator, 166 hereditarily ergodic, 62, 67
multiplication operator, see multiplier hereditarily hypercyclic, 76
multiplier, 120 hereditarily syndetic, 67
multiplier of a linear fractional transfor- Hilbert–Schmidt, 160
mation, 125 hypercyclic, 37, 65, 334
hyponormal, 152
nontrivial periodic point, 94 identity, 352
384 Index

left-multiplication, 276, 302 d-dense, 61

lower triangular, 148 backward, 26
MacLane, 40, 45, 46, 65, 71, 72, 98, dense, 9
108, 245, 275, 315 frequently recurrent, 237
mixing, 46, 220, 225, 245, 247, 265 prescribed, 227
multi-hypercyclic, 166 similarity, 223
multiplication, see multiplier somewhere dense, 164, 174
normal, 150 ordinary differential equation, 204
normaloid, 158 Oxtoby–Ulam theorem, 194, 210
nuclear, 160
paranormal, 153, 160 parabolic linear fractional transformation,
positive, 152 125
power-bounded, 141 paranormal operator, 153, 160
power-compact, 147, 160 partial differential equation
power-regular, 158, 160 first-order, 201
quasihyponormal, 158 second-order, 202
quasinilpotent, 141 perfectly spanning set of eigenvectors, 265
quasinormal, 158 period, 12
Riesz, 160 periodic point, 12, 43, 186
right-multiplication, 300 nontrivial, 94
Rolewicz, 40, 43, 46, 65, 70, 72, 97, 121, repelling, 58
133, 246, 249, 252, 284, 312 perturbation
self-adjoint, 150 compact, 147, 160, 289
SOT-chaotic, 299 finite-rank, 147, 159
SOT-hypercyclic, 279, 303 nuclear, 303
strictly cosingular, 160 point spectral mapping theorem, 363
strictly singular, 160 for semigroups, 364
supercyclic, 37, 159, 232 point spectrum, 361
Toeplitz, 160 polynomial, 65
topologically ergodic, 62, 173 m-homogeneous, 65
trace class, 160 positive operator, 152
translation, 36, 134, 228 power-bounded operator, 141
unilateral backward shift, see backward power-compact operator, 147, 160
shift operator power-regular operator, 158, 160
unilateral forward shift, see forward precompact set, 85
shift operator prescribed orbit, 227
unilateral weighted shift, see weighted preserved under quasiconjugacy, 7, 189,
shift operator 338
upper triangular, 63 projective limit, 340, 341
Volterra, 139 projective spectrum, 340
Volterra composition, 157 punctured plane, 112
Volterra integral, 141
weakly compact, 160 quadratic map, 5
weakly hypercyclic, 159, 347 quasiconjugate C0 -semigroups, 189
weakly mixing, 48 quasiconjugate dynamical systems, 6, 338
weakly topologically transitive, 350 quasihyponormal operator, 158
weighted backward shift, see weighted quasinilpotent operator, 141
shift operator quasinormal operator, 158
weighted forward shift, 98, 149 quotient space, 336
weighted shift, 89, 96, 134, 149, 273,
302, 330 rational rotation, 6
operator of Schatten–von Neumann class, Read’s theorem, 65, 302
160 Recurrent Hypercyclicity Criterion, 67,
orbit, 4, 23, 83, 186, 207 211
Index 385

relatively compact set, 352 SOT-hypercyclic operator, 279, 303

repelling periodic point, 58 SOT-hypercyclic sequence of operators,
reproducing kernel, 119 281
return set, 19, 24 space
Riemann integral, 357 Banach, 32, 351
Riemann–Lebesgue lemma, 198, 250 F-, 67, 333
Riesz decomposition theorem, 364 finite-dimensional, 55, 333, 358
Riesz operator, 160 Fréchet, 34, 333, 354
Riesz representation theorem, 354 Hilbert, 33, 353
Riesz’s lemma, 143 locally convex, 333
Riesz–Dunford functional calculus, 362, metric, 351
364 metrizable topological vector, 333
right-multiplication operator, 300 topological vector, 331
Rolewicz’s operators, 40, 43, 46, 65, 70, spanning eigenvector field, 250
72, 97, 121, 133, 246, 249, 252, 284, spectral mapping theorem, 363
312 spectral radius, 362
rotation, 6, 167 spectral radius formula, 362
irrational, 6 spectrum, 361
rational, 6 strictly cosingular operator, 160
Rouché’s theorem, 359 strictly singular operator, 160
run-away sequence, 111, 134 strong operator topology, 223, 277
Runge’s theorem, 359 strongly continuous semigroup, see
C0 -semigroup
Schatten–von Neumann class, 160 subspace of finite codimension, 282
Schauder’s theorem, 160 supercyclic operator, 37, 159, 232
Schwarz lemma, 359 supercyclic vector, 37, 65
second-order partial differential equation, syndetic sequence, 22
202 syndetic set, 22, 208
self-adjoint operator, 150
semigroup tent map, 6
strongly continuous, see C0 -semigroup theorem
translation, 183 Ansari, 162, 163, 173, 175, 177, 254,
uniformly continuous, 184, 185 337
semigroup action, 168 Ansari–Bernal, 220, 232
frequently hypercyclic, 257 Argyros–Haydon, 221, 232, 300
hypercyclic, 168 Arzelà–Ascoli, 157
seminorm, 32 Bès–Peris, 76, 84, 88
sensitive dependence on initial conditions, Baire category, 351
11, 42, 61 Banach–Steinhaus, 353, 355
sensitivity constant, 11 Banks–Brooks–Cairns–Davis–Stacey,
separating sequence of seminorms, 34 13, 30
separation theorem, 176 big Picard, 359
sequence space, 90 Birkhoff ergodic, 236
set of first Baire category, 351 Birkhoff transitivity, 10, 29, 30, 39, 207
set of second Baire category, 351 Bourdon, 53, 64, 66, 190, 337
shift map, 6 Bourdon–Feldman, 164, 167, 177, 256,
shift on two symbols, 15, 16 337
σ-compact set, 307 Casorati–Weierstrass, 358
similarity orbit, 223 closed graph, 355
simply connected domain, 113 Conejero–Müller–Peris, 168, 177, 196,
solution semigroup, 185 259, 306
somewhere dense orbit, 164, 174 Costakis–Peris, 166, 167, 177, 337
SOT, 223 De la Rosa–Read, 48, 66
SOT-chaotic operator, 299 decomposition, 176
386 Index

Duyos-Ruiz, 109, 134, 253 uniformly unconditionally convergent

Erdős–Sárközy, 241, 265 series, 308
Furstenberg, 20, 30 unilateral backward shift operator, see
Godefroy–Shapiro, 107, 134, 252 backward shift operator
González–León–Montes, 302 unilateral forward shift operator, see
Grivaux, 233 forward shift operator
Hahn–Banach, 355 unilateral weighted shift operator, see
Herrero–Bourdon, 53, 66, 190, 262, 325, weighted shift operator
337 unimodular eigenvalue, 249
Herzog–Lemmert, 66 unimodular multiple, 167
inverse mapping, 355 unit sequence, 90, 99
Kitai, 140, 159 universal element, 23
León–Müller, 167, 177, 259, 306, 339 universal Taylor series, 23, 30
Liouville, 358 universal vector, 83
Mazur, 270 Universality Criterion, 24, 30
Mittag-Leffler, 79, 88 upper triangular operator, 63
Montes, 268, 302
open mapping, 355, 358 vector
Oxtoby–Ulam, 194, 210 ε-hypercyclic, 61
point spectral mapping, 363 cyclic, 37
for semigroups, 364 frequently hypercyclic, 237, 256, 257
Read, 65, 302 hypercyclic, 37, 65, 83
Riesz decomposition, 364 supercyclic, 37, 65
Riesz representation, 354 universal, 83
Rouché, 359 Vietoris topology, 29
Runge, 359 Volterra composition operator, 157
Schauder, 160 Volterra integral operator, 141
separation, 176 Volterra operator, 139
spectral mapping, 363 wave equation, 209
thickly syndetic set, 62 weak topology, 347
Toeplitz operator, 160 weak-∗-topology, 277, 303
topological sequence space, 348 weakly compact operator, 160
topological vector space, 331 weakly holomorphic function, 197
topologically ergodic C0 -semigroup, 208 weakly hypercyclic operator, 159, 347
topologically ergodic dynamical system, weakly mixing C0 -semigroup, 186
28 weakly mixing dynamical system, 18, 30,
topologically ergodic operator, 62, 173 173
topologically transitive C0 -semigroup, 186 weakly mixing operator, 48
topologically transitive dynamical system, weakly mixing sequence of maps, 24
8, 30 weakly mixing sequence of operators, 83
topologically transitive sequence of maps, weakly topologically transitive operator,
24 350
topologically transitive sequence of weight function
operators, 83 admissible, 60, 184, 206
totally transitive dynamical system, 28 weight sequence, 89, 99
trace class operator, 160 weighted backward shift operator, see
translation operator, 36, 134, 228 weighted shift operator
translation semigroup, 183, 186, 191, 206, weighted Bergman space, 132, 135
207 weighted Dirichlet space, 132, 135
translation-invariant metric, 34 weighted forward shift operator, 98, 149
unconditional basis, 357 weighted Hardy space, 133, 135
unconditionally convergent series, 356 weighted shift operator, 89, 96, 134, 149,
uniformly, 308 273, 302, 330
uniformly continuous semigroup, 184, 185 Wengenroth’s lemma, 337